Release notes

All notable, user-visible changes to the Fluxtion Audit Log Analyser. Format: Keep a Changelog · Versioning: SemVer. Add a line under [Unreleased] with every user-visible change; the release workflow stamps it.

[Unreleased]

[1.40.6] - 2026-11-06

Fixed

• Creating an analyser bundle keeps its feed-position helper executable, so saved-position run and restart commands work without a manual permission change.

• A regenerated graph raises one message however many build copies it lands in, whichever copy the scan sees first; the message names the copy outside build output. Recognised Gradle build/ and unambiguous IntelliJ out/ copies count as the same graph too; ambiguous or truncated module layouts keep graphs distinct. A build or regeneration that rewrites the same bytes raises none, including after Not now or Clear all, and Accept all still opens a log whose graph was rewritten with the same bytes. If a file disappears during an open, a settled surviving copy remains available through a fresh message.

[1.40.5] - 2026-11-05

Added

• The Compiler account screen shows Signed-in email when the service provides the authenticated user's verified address. Older services show Not provided by service; the address is omitted from Copy details.
• Right-click Java or XML code and choose Explain this to copy an explanation prompt or add it to the assistant draft. It uses the selected text, or the displayed code when nothing is selected, without sending automatically.

Changed

• Project output notifications now appear in the Canvas Messages (N) tab, with the pending count and an amber unread indicator. Viewing the tab clears the unread colour; Open and Not now still act on the individual output. Notifications no longer stack above the workspace. Accept all opens the current outputs in sequence, stopping on failure; Clear all dismisses the current messages without opening them.

Fixed

• Identical GraphML copies in Maven's target/classes and target/fluxtion-staging/resources no longer produce extra regeneration messages when the corresponding source resource exists. Different graphs and standalone build outputs remain discoverable.

[1.40.4] - 2026-11-05

Added

• A text audit record can declare quoted values itself. A record whose eventLogRecord mapping opens with textEncoding: QUOTED_SCALARS is read with the quoted-scalar grammar (format spec §3a): a value such as reason: "membership not active, balance due 30.0" is one value, "123" and "true" stay text, quoted header values decode, and \0, \xXX, \uXXXX (surrogate pairs included) are decoded alongside \\ \" \n \r \t; malformed escapes are kept verbatim. Recognised only as the first line of the mapping, record by record, so a file or Follow stream mixing old and new records reads each correctly; records without it read exactly as before. An unknown value reads as legacy and is named once as a producer finding. Fixture C35.
• Format 1, revision 1.2. The format specification names the per-record declaration as revision 1.2 (additive: an unmarked file is unchanged; an older analyser ignores the line and shows quoted values with their quote marks, entry boundaries intact). The Log format page shows a marked record and the full escape list.

Changed

• flag without a kind now saves a neutral note (labelled Note / Assessment) instead of a fault, so evidence of correct behaviour is no longer recorded as a fault; fault and confirmation work as before, an existing finding keeps its kind, and the tool description names the default and all three kinds. Session recovery is forward-only once a note finding with text exists: 1.40.3 and earlier reject the note kind and then restore no findings at all from that snapshot. A bare flag (no note or fix) is still saved in the form earlier releases restore.

Fixed

• The --mcp bridge exits when its client closes stdin, even while a call to the analyser is in flight: requests received before stdin closed are still answered as long as each completes within 10 seconds; one that takes longer is dropped, together with any after it (before, a blocked call held the bridge for up to the 60-second call timeout). It also exits when the process that launched it has gone, including when that process exits before Java starts. A bridge whose client is still running is kept, however long it sits idle, by design.
• A finding with no note no longer shows an empty label: a bare confirmation flag showed a lone "Observation:" in the records table and on the topology callout, and a fix-only confirmation an empty "Observation:" line above its assessment.
• The point-at-the-fault skill flags a found fault with "kind": "fault", and says to flag a proving record as confirmation when it shows correct behaviour or to pass kind: note for a neutral observation only when the connected tool supports it; older releases default an omitted kind to a fault.
• The add-a-node skill's audit-compare.py reads records that declare quoted scalars (textEncoding: QUOTED_SCALARS as the first line, format revision 1.2): the declaration is never compared, quoted keys, values and header fields are decoded, and a capture from before a runtime upgrade matches one from after when the logged values are the same (before, the marker and a quoted reason: "a, b" gave a false DIFFER). That match is made record by record, and only when the unmarked record shows it predates quoting, so two captures from the quoting runtime are compared exactly and one real change is reported once. Unmarked records are read as before.
• A project's authoring evidence opens without granting the project root: fluxtion-authoring.json, target/fluxtion-validation.json, target/fluxtion-reconciliation.json, target/fluxtion-run.json and target/classes/fluxtion-diagnostics.json are readable at exactly those paths below the open project, so open {diagnostics: "target/fluxtion-validation.json"} works with only src/ source roots, and the run-receipt and authoring-record checks are no longer "unavailable". The list is closed and read-only: no other file or directory becomes readable, a symbolic link (the file's or a folder's) or a second hard link is refused, a source root's own file of the same name takes precedence, and a home or credential folder project grants none of it; a refused evidence file says which of these rules applied. The project root for a settings file outside .analyser/ is that file's folder, as everywhere else.
• Design discovery no longer walks .git/, .svn/, .hg/, .gradle/, node_modules/ or .idea/ below a source root, nor target/, build/ or out/ beside a build file (pom.xml, Gradle or Ant; .iml for out/), so the build's copy of a Spring design in target/classes/ is not a second candidate and the single design still opens automatically. A file removed or unreadable during the scan (a build running alongside), or a folder removed or replaced, is skipped or named, instead of making the whole discovery unavailable. The warning counts "entries" rather than "files", since folders can be among them.

[1.40.3] - 2026-11-04

Fixed

• Starters declared as compiler sign-in builds no longer show a legacy-key status or prepend the legacy key-check command; legacy runtime modes keep their own guidance.
• The server-running skill starts with the shipped processor without an authentication gate and directs graph regeneration to compiler sign-in and the project authoring command.
• Automatic project-reopen dialogs withdraw when a newer open supersedes them, so assistant-loaded evidence is visible; stale chooser answers are ignored.
• Legacy key guidance distinguishes Maven build sign-in, and the assistant guide names the onboard setup route for Both.
• Authoring skills now distinguish saved build sign-in from legacy key status, explain recovery after a failed authoring build, and describe single-run exports and optional screenshot verification.

Added

• Automatic setup is a normal macOS Claude Code release; upgrade from Analyser 1.40.2 before using the updated compiler-sign-in starter catalogue.

• Set up with your AI assistant. A docs page for the one-question install: the Claude Code plugin or the published setup instructions install, open and connect the analyser, with optional compiler sign-in, day-2 recovery and building without a system JDK. macOS and the Claude Code terminal only for now. Publication, the shared Java 21 download and starter-export prerequisites, consent exceptions, existing plugin connections and setup privacy are explicit. Manual-mode approvals, plugin declaration, build sign-in requirements, marketplace reload and day-two recovery are documented.

[1.40.2] - 2026-11-03

Changed

• Getting started now opens with installing the analyser. The guide starts with installing JBang (macOS, Linux and Windows) and the analyser, with screenshots of the start page and of AI ▸ Compiler account…'s service list; Install & run and the home page lead with the JBang install too.

[1.40.1] - 2026-11-03

Changed

• AI ▸ Compiler account… now offers the hosted compiler services by name. The service address drop-down lists Production (https://compiler.fluxtion-playground.dev) and Staging (https://compiler-staging.fluxtion-playground.dev) on a fresh install, ahead of any address you have connected with before; it was empty until you had connected once, with only a placeholder to go on. Nothing is chosen for you, and any address can still be typed.

Fixed

• Choosing the first address in that drop-down now fills the field. The list started with its first entry already selected while the field was empty, so picking that entry did nothing.

[1.40.0] - 2026-10-27

Added

• AI ▸ Compiler account… shows compile and interpret credits separately, one row each with what remains, what was used, what is in progress, and how many requests completed or failed. Below the account rows are the plan's key limits, as the service reports them: most nodes, most inputs (event types plus exported methods, each method one input), most exported interfaces, languages, WASM and software assurance; a limit set for your account is marked. A service from before the inputs limit reports only its event-type limit, which counts event types alone: it is shown as Max event types, with nothing said about exported methods, and never beside the inputs limit. Copy details includes both balances, the plan, its limits and rates, and still never the credential. A service without plans still shows compile credit and says the plan's details and interpret credit were not reported, rather than guessing.
• Getting started ▸ Investigate a compiler charge covers interpretations (creditKind, interpret: request ids), an operator's per-account limit (CAPABILITY_SET), and the two new refusals, CAPABILITY_EXCEEDED (with the limit in the row's detail, including maxInputs and maxExportedInterfaces) and INTERPRET_ALLOWANCE_EXHAUSTED.
• Getting started ▸ Investigate a compiler charge covers getting the file (the service admin page's Download audit log zip, with the log and EntitlementProcessor.graphml, or the export tool), opening that graph explicitly, and reading an operator's credit removal (ALLOWANCE_REMOVED) and the removal answers that leave no row.
• AI ▸ Compiler account… is laid out in sections (Account, Credits, This device, Your devices) with aligned rows, the credits as a row of figures, Refresh and Copy details on the left, and Close the default; connected, the top line names just the service's host. The service address is an editable drop-down that offers the addresses this machine has connected with. Each device row shows a short credential id (the whole id in its tooltip), its kind, and when it was created to the minute in UTC, so devices with the same name can be told apart. Each device's whole name fits, whatever the system's fonts: the Name column is as wide as the longest name, and the dialog widens when a longer one is read. After a disconnect the dialog shrinks back to fit, rather than keeping an empty connected-size window.
• AI ▸ Compiler account… now shows the account: its type (Personal or Organisation), your role, the plan and whether it is active; the credits remaining, used and in progress, generations and failures, and whether the allowance is one-time or when it next resets; this device's name and credential id; and your devices, with this machine marked and Revoke selected (confirmed first) for a lost or old one. Copy details puts a plain-text summary for a support request on the clipboard, never the credential. A new sign-in names the device "Fluxtion analyser on ", so several machines can be told apart. Deleting the account is not in the app yet.
• Docs: Getting started ▸ Investigate a compiler charge explains how to open a compiler service's exported investigation of an account, find one request, read why it was charged, released or refused, and chart the balance.
• AI ▸ Compiler account…: a rate-limited or briefly unavailable service no longer ends a device sign-in (the poll waits longer and retries a bounded number of times); Revoke this device keeps the device connected when the service answers 401, since that cannot prove the credential was revoked; any 127.x.x.x address may be plain http, like localhost.
• AI ▸ Compiler account… connects the analyser to a compiler account by device sign-in: it opens your browser at the service's sign-in page, shows the code to enter, honours the service's polling interval, slow-down and expiry, and can be cancelled. The credential is kept in the macOS Keychain or the Linux Secret Service keyring (an owner-only file elsewhere), never on a command line or in a project, export, report, bundle or context. The dialog shows the plan, available credits and next reset (when there is one), and offers Disconnect and Revoke this device. The analyser only reads the account; it never changes it.

[1.39.0] - 2026-10-26

Changed

• The MCP server instructions (and /manifest's instructions) now say what to do FIRST when the person asks for evidence and the run has a fresh export, and never to silently replace a log the person has open (ask first): for "show me evidence", "what happened to this event" or "what did Fluxtion do with X", open the export with its GraphML, wait until context no longer reports inFlight, spotlight the record that answers, then explain. In a stranger authoring trial, five of five connected runs answered from the YAML and never called the analyser.
• Bundle skills: load-audit-log is now chosen for explaining a run or showing evidence, not only for loading a log, and says to point at the record in a connected analyser rather than quote the YAML; point-at-the-fault covers "which record proves a claim" and points a working-run explanation to load-audit-log; add-a-node restates that eventHandlers states a design-time expectation and does not route events. The m19-skills/2 index is re-pinned.

Fixed

• Record spotlights recheck the session loading state on the UI thread and after Java-source preparation, so a queued request cannot reveal records after a log starts loading.

• read with no anchor (recordIndex, byteOffset or at) no longer fails: it reads from the first record and its note says so, with a copyable example. The first-move instructions and the load-audit-log and point-at-the-fault skills now say to flag the proving record after spotlighting it, so the finding is kept (a trial run saved no flags).

• A spotlight on a record target (records, records:row:<n>, detail, detail:node:<id>) sent while a log is still loading is refused with "log still loading; record targets not ready", touching nothing, instead of landing on the log being replaced. Any other target still lights, and its reply carries logLoading saying what was lit is not a record.

[1.38.0] - 2026-10-21

Added

• Open, person-owned projects now offer settled GraphML and audit-log outputs with Open / Not now banners. Nothing auto-switches your log; accepting a log uses the existing graph-pairing path, even with assistant file exchange off. The starter's first-build cue uses this same ongoing watcher.
• M80: Design discovers Spring XML and Java builder definitions under authorised source roots when a project opens or roots change, without relying on generated headers. Context adds kind/files/nodes/source while keeping released keys; a single candidate opens as a read-only glance, and absent or unavailable discovery is stated. Java anchors and the multi-file chooser follow separately.
• M80 discovery keeps readable candidates when individual files fail and names up to ten failed files, with an omitted count, in the Design pane and context note. Scan-limit exhaustion remains unavailable with a distinct reason. A failed first design read no longer invents a Spring kind, and the empty Design message uses plain, shorter wording.
• A spotlight source line target may now light a RANGE, not only a single line: source:java:<fqn>:line:<from>-<to> and source:design:line:<from>-<to> (plain line:<n> is unchanged, the range n-n). The range must have its first line at or before its last, and spans at most 200 lines. The cut-out covers the whole range, first line's top to last line's bottom; a range taller than the visible area clips to it and shows the range's start, reported as partial in the spotlight reply (as a single tall Java line already was) — a design range may now show partly too, where a single design line still refuses outright unless it is fully on screen. A range whose last line is past the end of the file or design is refused whole, never silently clamped.
• M80: Java builder designs now index literal addNode/addPublicNode declarations and show honest unanchored entries for computed names and discovered constructors. source {bean} navigates either XML ids or literal Java node names; source:design:node:<id> is a permanent synonym of the existing bean spotlight family.

Fixed

• A graph that cannot be read now reports failure even when an older topology remains visible. That older view is not mistaken for a completed output open.

[1.37.0] - 2026-10-19

Added

• Claude Code now shows the step it starts, elapsed time and a quiet inactivity cue while a turn runs. Keep waiting dismisses that silence episode until a new stream event; approvals show their own cue while keeping the existing modal question, with no blocked Show control. About lists tool/turn totals, and project context exposes the same read-only activity snapshot.
• The transcript renders Markdown tables, not raw pipes. A GFM table (a header row plus a valid separator row — :---, ---:, :---: or --- per column) draws as an aligned, bordered monospace grid, with alignment from the separator row and **bold**/`code`/links working inside cells; \| stays a literal pipe. When the grid would not fit the panel's current width it stacks each row as Header: value lines instead. A pipe-started block with no valid separator row still shows as the literal text it always was. The transcript also now renders headings 4–6 (bold, base size), ~~strikethrough~~, task-list items as ☐/☑, a drawn rule for ---/***/___ alone on a line, and nested block quotes (capped at three levels). Columns are aligned with pixel-measured tab stops, not padded by character count, so a cell holding a status icon or an em dash no longer pulls that row's separators out of line with the rest of the table; the header's drawn rule now reaches the full width of its longest cell.
• Transcript tables bound cell and repeated-label expansion, preserving oversized tables as raw text. Pixel widths include displayed link destinations, image captions, visible character disclosures and navigation chips, so grid alignment and narrow-panel stacking match what is actually rendered.
• Docs: a new page, The audit log ▸ From a Mongoose server to the analyser, explains how a Mongoose-hosted processor's audit records reach a ----separated file that the analyser opens. It covers two routes: your own LogRecordListener passed at boot, and Mongoose's audit capture exported by svc-admin-web. It also covers where the audit level comes from. Steps that the Mongoose sources do not settle are marked To confirm.

Changed

• Docs: deferred M81 project build-script pre-approval after reviewing execution wrappers, shell setup and automatic backgrounding. No script-trust feature ships; existing Bash permission behavior remains unchanged.
• The offline M74 onboarding scorer now binds scores to the sealed oracle bytes, refuses incomplete captures, checks ordered person actions and their references/choices, and reads the current Java permission facts. Maven seeds reject live symlink aliases. CI checks the current Java pack parser without rewriting its expected verdicts, including the non-breaking-space URL regression. The live operator harness is still not implemented; no onboarding trial is implied.
• The offline onboarding scorer refuses nonempty answer rows without string text instead of silently ignoring their contents. Operator correlation now covers current proposal-row, restart, no-log, profile/setup, external-access and starter choices, verified against actual Java records; unknown fact-bearing operator kinds also void the run. Empty answer captures remain explicitly distinguishable from missing or malformed captures.

Fixed

• Restart recovers a Claude Code process stuck starting or stopping, preserving its conversation when it can resume. Start on a healthy session now says it is already running; externally stopped processes name signal 15 or 9, and rate-limit notes explain that they do not block Start or Restart.
• An interrupted Claude Code conversation that cannot resume in its changed folder or mode starts a new chat with your last request and a recovery reminder placed, never sent. A CLI-reported missing session falls back once; ordinary failures do not retry. The session now records a start activity clock for the live activity display.
• Fixed: accepting "this log sits inside the project X — load it?" while Claude Code is running now offers the same restart as any other project switch, instead of applying at once. If you started Claude Code before opening a project (so it began in the analyser's private scratch folder), then opened an audit log and said yes to that offer, the project became active immediately but the already-running session kept working in its original folder with no warning — so its CLAUDE.md and .claude/skills were never loaded, even though a later Start or New chat would have picked them up correctly. Now you get the usual "Claude Code is in \<old folder>; restart in \<project>?" choice first. If that log's records also run out of time order, the time-order report now waits until you have answered, instead of opening on top of the restart question.
• Clarified in the Claude Code docs page, and covered by a new regression test, what was already true of the running app otherwise: with your own project open, Claude Code's hosted session works in that project's own folder (never the analyser's private scratch folder) once you trust it — only the sample project and an evidence bundle's working copy use the scratch folder.

[1.36.0] - 2026-10-19

• Full Claude Code's documentation pre-allow now runs through the analyser's URL check rather than a hostname-only CLI allow. Non-HTTPS URLs, non-default ports and user-info URLs no longer bypass that check. Existing Full conversations require a new session so they cannot retain the previous allow rule; the person's own permission rules and selected mode still apply.

Added

• source can return the code it finds. source {…, text: {from, to}} (a range) or {…, text: {around?, context?}} (a window around the chosen line) returns those lines as data, so Claude shows you code through the analyser instead of reading files itself, with no approval to ask. It is read-only and reads only files under your authorised source roots, by their real path (a link that leads out is refused); at most 200 lines and 32 KB a call, with any cut stated; a binary or non-UTF-8 file is refused. An optional parameter for external MCP and REST clients too: without text, source and its reply are unchanged.
• Each turn of the conversation shows its time. Your turns and the answers, in Claude Code and the in-app assistant, show the time quietly at the right of their header line (14:02, muted); the first turn of a day shows the day too (Sun 18 Oct · 09:05), and hovering or a screen reader gives the full local date and time. A walk's recorded dialogue shows when it was recorded, labelled recorded. The time is the one the session recorded for the turn.
• Create a project from a starter with the assistant. Claude Code in the analyser can list the analyser's starter templates (context.starters: id, name, description, how the graph is authored, where it runs, whether its build needs a Fluxtion key — Spring first and recommended) and propose one (propose {kind: starter}, hosted-only). The review shows one row — Create a project from starter Fluxtion Spring XML (Spring XML · embedded · AOT · build needs a Fluxtion key; none found on this machine), Claude's reason, a Folder with Choose… and Downloads from fluxtion-playground.dev; nothing is built or run — starting at No, with Apply chosen disabled until a new or empty folder is chosen. Claude never names a path. Only your Apply downloads; the analyser then writes the project's skills, discovers, asks one more review (every row at No), opens the project and its Spring design (reported as shown only once it was read and displayed; declining the design's source root, or a failed read, is said as such), watches for the first build's GraphML to offer it, and offers to restart Claude Code in the project's folder with a short pointer-only draft placed — nothing is sent until you press Send. A discovery that fails ends the proposal, saying whether files were already unpacked or the download failed, and a later starter request can proceed. Switching projects before it finishes cancels what is still to come — later opens, the design, the graph offer and the draft; downloaded files remain. Your Apply is the consent, so Allow the assistant to create projects does not gate this path.
• A starter's add-a-node skill names its own regeneration. After every bundled skill matches its pin, add-a-node's regeneration placeholder is replaced by the starter's steps from the analyser's first-run record (for the Spring starters ./check-fluxtion-key.sh, ./setup.sh, ./validate.sh, ./generate.sh) and whether they need a Fluxtion API key; a starter with no such record keeps the placeholder. .claude/skills/.vendored.json lists each written file's pinned source digest and its written digest, and marks the substituted file.
• A starter's run skills name its own start command. With the same pin check and record, run-mongoose-server's start placeholder becomes ./run-server.sh (its pinned ./scripts/run-server.sh guess is gone) and run-embedded's becomes ./run.sh, each with the steps the first-run record runs before it; only those known command strings are written. The placeholders the record cannot answer (export command, log and GraphML paths, stop command, the embedded class) are kept, and .vendored.json now carries one note per placeholder — filled, or marker kept: and why — in a notes list.
• The new project's review pre-ticks the Spring design folder. The row that adds src/main/fluxtion/designer as a source root starts at Yes, and says why: it is inside the project you just created, and it is needed to show the Spring design. Every other row, including opening the project, still starts at No.
• The playground's Spring design skill and contract are bundled with the publisher's permission (owner, 2026-10-18); the playground's licence statement is pending (UP-PG-10).
• A starter project gets its skills. The canonical skills for the starter's authoring and host, and for Spring the playground's design skill and XML contract, are written from the analyser's own jar under .claude/skills/, each checked against its pinned sha-256 first; a file already there is never overwritten. Writing is not transactional: if it fails part-way, the review says installation is incomplete and may have left files.
• Set up AI… — the one place to choose how AI works with the analyser: No AI, your own AI client connects, Claude inside the analyser (Safe analysis or Full Claude Code) or both. On a first run the start page asks it inline (How do you want AI to work with the analyser? · Continue · Not now); after that one line says what is in force (AI: not set up · Claude Code found — Set up AI…). It is also in the assistant's header, first in the AI menu and at the top of Settings ▸ Assistant. For Claude inside it checks that Claude Code is installed and signed in (claude --version, claude auth status; nothing is sent, nothing personal is kept), each check with Check again. A Send, Start or Ask with nothing set up opens it, and the question you asked waits: Finish places it in the composer, so Send is the next and only step. The Ask items say (set up AI first) while nothing is set up.
• Set up AI ▸ Your own AI client, step by step (and the second half of Both): choose Claude Code, Codex or another MCP client; turn on External MCP/REST access behind a confirmation whose default is Cancel; register the analyser with Register for me… (the exact command shown and confirmed first), I'll run it myself (copies it) or Check registration; then a live check says when a client with this window's key actually reached it, and the name it gives itself, labelled as its claim. Connect an AI client stays behind More options…. Cancel after the window turned access on keeps it on and offers Turn it off; No AI offers to turn access off too (a ticked box).
• Where data goes, a small picture under the option you select in Set up AI…. It shows what talks to what on this computer and what goes out to an AI service. Claude inside the analyser is two-way: a right-click Ask reaches Claude, and Claude points back. Your own AI client is one-way: it reads from the analyser, and you copy questions to it. Full Claude Code is drawn in amber. Once the sign-in check has run, the picture says where Claude is billed. Under Claude's pictures, a link opens the new FAQ entry What does Claude Code itself send to Anthropic?. Each card keeps its sentence, and the Your own AI client sentence now says "another AI app that can connect to the analyser" instead of "another MCP client".
• "Settings changed — restart Claude Code now to apply them?" A settings or profile change that can wait — Safe analysis → Full Claude Code, a full mode that asks less, the model, the spend limit, the folder — is now offered in the assistant panel with Restart and Later, instead of stopping the session. Restart restarts it on the new settings (the conversation continues where it may, else a new chat begins, as the prompt says; mid-answer, when the answer ends); Later keeps it running on its old settings with a settings changed — restart to apply chip that asks again. Changes that take a power away — Claude Code switched off, Full switched off, a mode that asks more, the folder no longer confirmed — still stop it at once, mid-answer too, and are never offered.
• ↑ recalls your previous questions in the assistant's composer, in both modes: ↑ on the first line walks back, ↓ on the last line walks forward, and ↓ past the newest puts back the draft you were typing. Kept in memory only, per assistant — never saved, never sent. Shift and Alt arrows, and the caret in a multi-line draft, behave as before.

• Two first-class pathways on the start page, entries 2 and 3 after the demo: Build a new application with me reads the starter catalogue when you press Start building and drafts "I want to build: … Help me choose components, generate it, run sample events and show me in the evidence that it behaves as intended." naming the templates (with no AI: New project from template…, New project…, Copy prompt); Teach me by building an example opens the sample DEMO project and drafts "Teach me Fluxtion by building a small example with me: explain events, triggers and dispatch from the evidence as we go." (with no AI it plays the DEMO guided tour and offers Guided tour, Sample project, Assistant settings, Copy prompt). Both work with no AI. The list now has eight entries.

• Claude Code in the Assistant tab, off by default (Settings ▸ Assistant ▸ Allow the analyser to start Claude Code). Choose Safe analysis in Set up AI… or the assistant header's quick switch (below) and press Send: the analyser runs your own installed, logged-in claude as one long-running session that drives the analyser through its verbs, and never reads or stores your Claude Code credentials. By default it runs as Safe analysis, and your full Claude Code is a separate opt-in (next entry): the folder's CLAUDE.md, Read/Grep/Glob confined to the folder, and the analyser's verbs — those that change what is open or write a file only after you approve each call; no Bash, no edits, no web, none of your skills, hooks, settings or MCP servers. Each folder is confirmed once on this machine before Start; a bundle's working copy is refused; Start contacts Anthropic (sign-in check, bridge connection) before your first Send, with no record content. The analyser checks that the session is bound to this window and running exactly that profile, or stops it and says what differed. Send is refused while Claude is answering, Stop interrupts the turn (and ends the process if it does not stop), New chat starts a new conversation, typed slash commands pass through except those that would change the profile or the conversation (/config, /clear, /add-dir, /mcp … and others are refused or offer New chat); @file mentions work as in a terminal for files inside the folder, while lines starting with ! or # are refused, and an optional spend limit (Limit spend per session, off by default, for API-key billing) stops Claude Code when it is reached; with none set Claude Code is started without --max-budget-usd, and on a Claude plan no dollar figure is shown (see What the session costs is said once, below); with an API key About this session… says what the session used. Start is always the explicit button: opening the Assistant tab starts nothing. Claude Code does not need External MCP/REST access turned on: with it off, its private listener serves only that session, refuses every external client (even one with the REST token) and publishes no endpoint file. Switching back to the in-app assistant while Claude is answering is refused by name. All of these settings are this machine's only: never in a project profile, a shared settings file or a bundle.

• Full Claude Code, a second opt-in in Settings ▸ Assistant (…with my full Claude Code permissions, off), shown with a warning whose default is Cancel: the Assistant tab then runs your own Claude Code as in a terminal — your settings, skills, plugins, subagents, hooks and MCP servers, and your own tools (Bash, edits, web) as far as your permission rules and the mode you choose (default, acceptEdits, auto, plan) allow; the analyser does not confine it to the folder, and bypassPermissions is never offered. The folder is confirmed separately for it, naming the hooks and .mcp.json servers that will run at Start, and asked again when .claude/settings.json, .claude/settings.local.json, .mcp.json or its instruction files (CLAUDE.md, .claude/CLAUDE.md, .claude/commands, .claude/skills, .claude/agents) change; the dialog says which changed. Hooks show as status lines (event and outcome, never their output); the header names the mode, how many of your MCP servers loaded, and the folders appended to the PATH so an analyser started from the Finder or the Dock can still start your hooks and npx servers. A request that would ask you is asked of you in the analyser (see Approvals in the analyser, below), and auto on a model that does not support it is stopped with a message saying so.

• The start page asks What would you like to do? — a numbered list (Run me through a demo, Build a new application with me, Teach me by building an example, Set up a project from my code, Open an evidence bundle someone sent me, Which starter template suits what I'm building?, Explain an audit log I already have, What can this tool do?), a box for your own words and one button; the cards stay below under Or open directly (see The start page is one surface, below). Choosing an entry (a click, ↑/↓ and Enter, or Alt+1–9 — ⌘1–⌘9 on macOS) sends, opens, fetches and starts nothing; only the button does. The demo plays the DEMO journey Find the first recorded breach from the jar, with no AI, key or network, labelled Simulated conversation. A bundle that does not verify is not opened. The template catalogue is read only when you press its button. The button never starts Claude Code: it places the draft and arms Start, only when Start would be accepted. With nothing open the assistant opens in its own window, showing the draft. The page asks whichever assistant can answer (the profile you chose is remembered on this machine; Claude Code counts only when it is switched on and installed) and offers Ask with Claude Code / Ask the in-app assistant when only the other one can. With no AI configured, every entry ends in a button that runs an existing action, and the start page's Copy prompt copies your question only: no record text, glossary or token.

• Prompt packs on the start page. Drop one Markdown file on the start page to replace its numbered list for this session: a # Title, an optional description line and 1–9 entries 1. **Label** — prompt text. Its entries fill the box like a built-in prompt, and ← Return to main options (Alt+R) brings the analyser's own list back. The reader is forgiving about format — -, --, – or : for the dash, an entry with no bold label (its first words become the label), - or * bullets for numbers, extra blank lines, trailing spaces and no description are read and tidied, and a quiet note under the pack's title lists what was changed — and refuses only what is unsafe or out of bounds (over 16 KB or not UTF-8, more than 9 entries or an over-long line, HTML, images, links outside an example entry, placeholders or hidden characters, an example entry not in its exact strict shape), naming its line and one cause. A pack is never saved, shared or bundled. The user guide has the format and a sample pack.
• One list of questions for the assistant. What happened here?, Compare these rows, Did this node run, and what does the evidence establish?, Explain this point, Explain this finding and Summarise this window appear on the records' right-click menu and the Records menu, the toolbar's Ask, the record detail's Ask about this, the diff dialog's Ask Claude about this diff, the topology's node menu, a chart's point menu and a new right-click menu on the time slider — labelled Ask Claude: … or Ask the assistant: … for the assistant in force. A question is added after whatever you have written in the composer, highlighted, and sent only when you press Send. It names record numbers, the log and its load, node ids, a chart and a time — never a value, a note's words or a finding's text — and it is refused at Send if the log has changed since it was drafted. A question's draft never carries the raw text of your current selection; in the in-app mode the questions need Let the in-app assistant change the views as it answers, and say so when it is off. During a walk with a recorded or scripted conversation, an ask plays the step that shows the next turn instead of contacting any model.
• The Claude Code conversation is drawn properly, in light and dark: your turns on a band, Claude's words as safe Markdown (headings, lists, quotes, code blocks with light tinting for Java, XML, YAML and JSON), HTML and images shown as their characters and never fetched, terminal escapes and invisible or direction-reversing characters shown as marks so that what you see is what you copy, Copy on each answer and Copy code on each block (saying when the text holds such characters), links opened only after a confirmation that shows the host in plain ASCII and only for http/https, record numbers, nodes and charts as chips that only navigate (muted for an answer about an earlier generation of the log), collapsible tool lines and Alt+↑/↓ between turns. ⋯ ▸ About this session… names the folder shortened, with its full path in the tooltip.
• A visible cue when your full Claude Code is in force: ◆ on the assistant's control (◆◆ and a stronger colour for acceptEdits and auto), a coloured border on the composer and the start page's box, and ◆ Send (named Send to full Claude Code); the control's tooltip says Full Claude Code () — on: it runs at the next Start, and the Assistant tab and its window carry the ◆ marker. It follows the session: on while switched on and not running, running while it runs.
• Claude Code for the sample project or an evidence bundle's working copy works in the analyser's private scratch folder (those are the analyser's own folders, never one Claude Code is started in), and the session header says so.
• Set up a project with Claude. Set up a project with Claude… on the Claude Code tab (and the start page's Set up a project from my code toward Claude Code) asks for your code's folder, makes it the session's working folder for this run and arms Start with a set-up question. Claude reads the folder and calls a new verb, propose, offered only to Claude Code in the tab (never on /manifest, the MCP bridge or the in-app assistant): it records an offer of up to ten items — source roots, the processor, a GraphML, runbooks, a log — and changes nothing. A banner (never a dialog) offers Review…: the analyser's own check of each item beside Claude's reason, labelled unverified; every answer starts at No; an item outside the folder can never be applied; Create a project here is the analyser's own row. Apply chosen applies the Yes items one at a time through your own paths, in a fixed order, then checks the pairing. One proposal at a time; context.proposals reports outcomes without Claude's reasons.
• Switching projects while Claude Code runs asks first: Stop, and get Start ready in the new project, Keep it running here, or Cancel. Nothing starts until you press Start. An assistant's open {project} is never held.
• Example entries in prompt packs: N. **Label** — example: https://github.com/OWNER/REPO/archive/<40-character commit>.zip, optionally with sha256:<hex>. Shown with a download badge; nothing is fetched on the drop or on choosing — only the entry's Download example… button, after a confirmation that leads with the publisher (host / owner / repo @ commit, marked third-party unless it is the Telamin GitHub organisation), the exact address and the folder you choose. Only https from github.com, codeload.github.com or gitlab.com, only a pinned 40-character commit (a branch or tag is refused), no git, hooks or submodules; a digest is checked before unpacking and a mismatch refused by name; the archive is unpacked through the starter templates' contained installer with no file made executable, nothing is built or run, and you are offered to open it as a project. Downloading never trusts the folder for Claude Code.
• Recording a Claude Code conversation into a walk: the conversation editor offers your sent turns and Claude's finished answers, and the walk records its source (claude-code, the model, the Claude Code version — never a session id, path, account, cost or budget) as conversation version 2; every other walk stays version 1. Long turns are cut at 16 KB and marked; past 200 turns or 512 KB the most recent are offered and the rest counted. A step whose reveal no longer matches a turn is dropped on load, and named.
• Approvals in the analyser. When Claude Code in the Assistant tab needs your approval, the analyser asks you in one dialog that says what this call does — "Claude Code asks to delete the saved walk “x”", "…to write a screenshot of the window to “s.png”", "…to run a command on your computer" — what follows from it, and the call's input exactly as Claude Code wrote it, as plain text with invisible and direction-changing characters shown escaped. Deny is the default, at the right — Enter, Escape and closing deny; Allow once stands alone at the left, is active after half a second and has no shortcut; there is no Always allow. Unanswered for 120 s (a visible countdown) it is denied; one dialog at a time, and a second request waits for it (the dialog says how many wait) and is denied if the turn ends or you press Stop first; Stop Claude Code in the dialog denies first, then stops the turn. In Safe analysis only the verbs that change what is open or write files (open, source_root, screenshot, report, walk, import) are asked — they are now visible to Claude Code, each call asked, except playing or ending a walk and listing the deleted reports or walks, which change nothing and are allowed without a dialog; a file write while file exchange is off is refused at once with the reason, never asked; with full Claude Code in default or acceptEdits, what your Claude Code would ask in a terminal (Bash, edits, web, your MCP servers' tools) is asked here. The transcript marks each call approved by you · ✓, denied by you, denied: not answered within 120 s or denied: not allowed in Safe analysis — the words a screen reader hears too — and the status line says Claude Code is waiting for your answer. A call your own rules, hooks or the auto classifier approve is not asked in the analyser.

• About this session… shows where Claude Code's requests go when ANTHROPIC_BASE_URL is set (scheme, host and port only) and a diagnostics line when the stream held something unexpected. A conversation keeps its last 500 entries, saying how many earlier ones are not shown.

• Choose how the assistant works. Four profiles: Safe analysis (the default: Claude Code reads your evidence; it changes what is open or writes a file only after you approve each call — no shell, safe with logs you did not write), Full Claude Code (your own set-up, skills and tools, with its warning and its mode; when Claude Code asks for approval, the analyser asks you), In-app assistant (an API key and a model) and Off (no assistant in the analyser; everything else still works, and external MCP and REST clients are not affected). They are set up in Set up AI… and switched with the assistant header's mode chip (both below); either writes the same settings Settings ▸ Assistant holds, on this machine only. With nothing set up, a Start, a Send or an Ask opens Set up AI, and a question asked on the start page or from a record waits and is placed for the assistant you set up when you press Finish, nothing sent until you press Send; a change while an answer is in progress is refused until it ends. A machine already set up keeps its settings and is not asked.

Changed

• Start building offers acceptEdits when Full Claude Code is already in default. The Build preselection lived only in Set up AI, which an assistant that can answer never opens, so a set-up Full session built in default and asked about every edit. The Claude Code tab now shows one line — Building: file edits and some commands can run without asking; your permission rules still apply. Switch? — with Switch (new chat), which sets the mode to acceptEdits and restarts Claude Code (no second confirmation: Full is already confirmed), and Keep asking, which changes nothing. It never switches by itself, and bypassPermissions is never offered.
• Project ▸ New project from template… now runs the same steps as a starter Claude proposes: after the download the skills are written, one review shows what opening the project puts in force (every row at No, including adding the Spring design's folder as a source root), and the project's Spring design opens at once; the first build's graph is offered when it appears. The starter's commands are still shown to copy, never run.
• Claude Code is told where it is. Every session's note now begins: it is the assistant inside the analyser, and project, starter, log, graph and report mean the analyser's; call analyser_context first, which lists the starters. Full Claude Code is also told to begin a new application by offering a starter, never by searching the web or scaffolding by hand, and to create nothing in its scratch folder. The note is part of what a session is created with, so a conversation from an earlier version is not resumed: it begins a new chat.
• Full Claude Code may read fluxtion-playground.dev documentation without asking: one fixed rule, WebFetch(domain:fluxtion-playground.dev), on its command line, and the analyser's own approvals allow without a dialog only an https fetch of exactly that host. Any other web fetch is asked; Safe analysis has no web tool. The Full folder confirmation, About this session… and the security page say so.
• Set up AI says what each button does, and checks your own client in fewer steps. Screen 1's button reads Check Claude Code, Set up my client or Continue without AI for the choice made (the start page's first-run block too); while a question waits, the last button reads Finish, then send my question. On Your own AI client, Check connection now sits under the access switch instead of behind More options…: with access off it asks the same Turn on External MCP/REST access? question first (Cancel the default), and confirming turns access on and runs the check in one step. Setting up your own client to Bridge works now takes 4 acts instead of 7.
• The assistant panel is quiet (M78 D1): the docked panel shows the Assistant ▾ control, ✎ (New chat) and ⋯, the transcript and the composer — and little else. The Claude Code, model, posture, billing, plan and provider chips, the basis line, the folder, the posture note, the limits line and the fork warning are in ⋯ ▸ About this session…, with Pop out, Switch assistant…, Start / Resume Claude Code, Set up a project with Claude… and Settings…. The state is one or two words on the control — answering…, needs you, starting…, stopped, nothing when ready — and a refusal or an exit is one banner with Restart and Details; Details appears only when something needs attention.
• One mode chip (M78 D2): the assistant's header control is now one short chip — Safe ▾, ◆ Full ▾ (◆◆ for the modes that ask least), In-app ▾, Off ▾, or Set up AI… while nothing is set up — with the state joined to it (Safe · needs you ▾, In-app · answering… ▾), and it is the Assistant ▾ quick switch: pressing it opens the switch's menu. The in-app assistant gets the same header as Claude Code — the chip, ✎ and ⋯ — instead of its chips and status line: its provider and model are in ⋯ ▸ About this session…, its Copy prompt and Connect a CLI assistant… are in ⋯ (shown only in the in-app mode), and no key is one banner, No API key — nothing is sent. with Add key…. With nothing being answered, a quick switch that Claude Code must restart for (Safe analysis → Full Claude Code) restarts it at once with no banner; while an answer runs it is offered (Restart / Later), as a Settings change is. The start page's no-AI Connect an AI client… button is now Set up AI…, which is what it opens.
• "Claude · unverified" heads every answer; the full the model's words, not a verified result is its tooltip and accessible description. An answer's Copy shows under the pointer or with the keyboard focus.
• What the session costs is said once: a note on a session's first ready — Claude Code · Safe · uses your Claude plan — again after a restart, and in About; never a dollar figure on a plan. The Approved: … note is gone (the tool line already says approved by you); denied, not-answered and withdrawn notes stay.
• The composer says who answers: it reads Ask about this log… / Ask full Claude Code…, with the keys in its tooltip.
• Stop takes Send's place while a turn runs; a permission request shows needs you on the control and — approve? on the tool line (the dialog is unchanged, Deny still the default). A placed question says Draft — not sent; the this log · opened … chip shows only when the draft is about an earlier log. The restart offer is Settings changed — restart Claude Code now? with Restart / Later (the detail in its tooltip).
• What starting Claude Code contacts is said once, in the first-Start folder question, instead of before every first Send. On the records' right-click the Ask questions come first. When a turn ends, and on ↩ Back to the answer, the transcript shows the answer from its first line.
• Assistant ▾ is a quick switch. The assistant's header control opens a short menu — ✓ Safe analysis · ◆ Full Claude Code… · In-app assistant · Off · Set up AI… — with the profile in force ticked and said in words. A row applies at once and writes the same settings as before; ◆ Full Claude Code asks for its confirmation first (Cancel the default); a change mid-answer is refused with one line. A row that cannot work as it stands opens Set up AI at the screen that finishes it instead: Safe or Full while Claude Code was not found opens Claude inside, and the in-app assistant with no API key stored opens Other options at the key field. Each row's tooltip says what it sets up and that it is not local-only. Settings ▸ Assistant keeps every option; the in-app assistant's key is now set up behind Other options.
• One door for your own AI client. AI ▸ Connect an AI client…, the start page's Connect an assistant block (Codex, Claude, Other MCP client) and the Connect an AI client… button in Settings ▸ Assistant are gone: AI ▸ Set up AI… replaces them, and opens at Your own AI client when one is set up. The panel's Connect a CLI assistant and a no-AI fallback's Connect an AI client… open it there too. The Connect an AI client dialog is reached only from Set up AI's More options…, for a project-scope command, removing a registration, Claude Desktop's note and Check connection. Your client's MCP/REST loop — the verbs, the endpoint file and the token — is unchanged.
• The first-run question is asked once. The start page's How do you want AI to work with the analyser? block is offered once, to a new installation and to one upgraded with nothing set up alike; after that the page shows the one AI line, AI: not set up — Set up AI…, which keeps the same door.
• Set up AI ▸ Your own AI client shows where data goes: the Where data goes picture now sits with the Let programs on this computer reach this window switch, and your client's arrow is solid only while access is on (it said off whatever the switch said before).
• Documentation captures refuse a symlinked home. tools/capture-docs.py and the Java capture programs refuse a user.home that is, or runs through, a symbolic link — the analyser shows a home's real path, so a symlinked one could put an account's scratch path into a published screenshot — and take Set up AI's screens on a fresh home with --setup-ai.
• Turning on External MCP/REST access now asks first, from every door — the AI menu's checkbox, Settings ▸ Assistant (on save) and Connect an AI client, as well as Set up AI: one confirmation, Cancel the default. Turning it off is still one click. --rest still turns it on at start-up without asking (nobody is at the screen). The MCP tools, the REST verbs, the endpoint file and its token are unchanged. Connect an AI client's registration and snippet confirmations now have Cancel as the default too.
• The assistant starts in the mode of the profile in force. At launch the assistant tab now follows the profile that is in force on this machine — Safe analysis or Full Claude Code start in Claude Code, the in-app assistant (or a key stored before profiles existed) in the in-app assistant — instead of the tab remembered from last time, which could contradict it (an old Claude Code trial left an in-app user's Send refused). Groundwork for Set up AI (M77): its read-only checks (claude --version, claude auth status --json, each in an empty private folder; no e-mail or organisation is kept) and its two machine-only settings, which never travel in a profile, share or bundle.
• The Event types checklist starts hidden on a fresh install, so the Records table keeps its rows (at 1200×800 the sample's breach record had fallen below the fold). A type filter still shows while it is hidden — the toggle's 1 of 7 badge and the line under the Records title — and applying a filter, from the assistant or the action socket, does not open the list. Show it once and it stays shown.
• The Format specification says how a merged nodeLogs entry is read. Fluxtion writes one entry per run of writes by one node, so a node that logs in two consecutive callbacks (@Initialise then @AfterEvent) writes one entry with repeated keys: - pe: { cb: Initialise, n: 1, cb: AfterEvent, n: 2}. The analyser already read it correctly — every pair kept in order, the last occurrence in the record as the value, so it reads as the same writes in separate entries — and conformance fixture C34 now pins that. §8 now says that tracing marks only @OnEventHandler/@OnTrigger invocations, so did not run means no dispatch callback ran, and a traced build whose nodes log from @AfterEvent or @AfterTrigger reads as untraced (may have run).
• The docs site is organised by task. Its sections are now Start here, Diagnose an incident (your own AI client over MCP or REST), Diagnose in the analyser, Build with Claude, Get started, Demo and teach, Administer and secure and Reference. A new Home page offers four doors, and a new Start here page routes you by what you came to do. A page that moves or is removed redirects to its new home, so links from the app and saved bookmarks keep working. Pages still being written say so, and link to where their content lives today.
• Diagnose an incident is written in full: the incident path from opening the evidence to handing it on, connecting your client, a first session, troubleshooting (with the bridge's and the listener's refusals quoted exactly), and a Scripts and CI over REST page for automation (the endpoint file, the pid check, GET /manifest, and every status code the endpoint returns).
• The single Analyser assistant page is split by task: Diagnose in the analyser (an overview, the assistant panel, Ask the assistant, Claude Code in the analyser, reading the transcript, approvals), Assistant profiles and Managed policy under Administer and secure, Set up a project with Claude, Spotlights, walks and reels, the Verbs reference, and Is the log whole? on the log format page. Its old address opens the overview, which says where each section went.
• Build with Claude now opens on how the work divides: you state the intent and the acceptance criteria, the AI structures the components, Fluxtion writes the mechanics, and you check your criteria against the recorded evidence of your runs (the old build-with-AI loop page redirects to it). Check your criteria against the evidence says what you can use for that today (computed answers, flags, reports that re-run against the next log, walks, bundles and replay compare, each with what it does not show), and marks the acceptance checks being designed as planned. Producing an audit log now says the analyser opens binary (FLXA) audit logs as well as text ones, and the Spring pages say which assistant can build, and that keeping an earlier run's numbers means exporting its report's PDF while that run's log is open.
• Get started and Demo and teach are written: Guided tour and sample project (the demo, the four-stop tour and what the sample project holds), Starter templates and examples (choosing a template, what a download will and will not do, and why an example entry is pinned to a commit), Prompt packs (using one, the forgiving format, what is refused, and the sample pack), Open your own log, step by step (the old Getting started, whose Settings tour is now the Settings reference under Administer and secure), and Running a demo (tour, pack, walk, reel, bundle). The start page's page describes its Build and Teach pathways, and the evidence-bundle demo is now a section of the bundles page, which redirects from its old address.
• Administer and secure is written: a one-page Security model and limits (marked as a draft for owner review) with the trust boundaries, what each assistant profile can and cannot do, including which folders source_root may add (never the home folder or a credential folder, and for the in-app assistant only inside the open project) and that open can open any file, approvals and their stated limits, why no profile is local-only, and the limit of what the analyser can see; External access (MCP / REST); What is stored where, what leaves; and Projects, profiles and sharing, which absorbs Portable context and Sharing setups (both redirect there) and has one table of what is kept where. The FAQ's security answer moved to the security page, and the FAQ links there.
• Reference is complete: a Command line page (every option, what each prints and its exit codes), a Glossary, and, on Verbs, every verb's arguments in tables generated from the analyser's own verb schemas, which a test keeps in step with GET /manifest.
• Help ▸ User guide is now a page of links into the docs site, ordered like the site, each with its address shown, plus a few offline essentials; a test checks that every link resolves. The release-notes page has one title.
• A deleted chart can be restored, and so can a chart definition a later call replaced — for every caller: you, the in-app assistant, Claude Code in the tab and external MCP/REST clients. Deleting a chart (Graph tab → More → Delete chart, graph {delete}), or a call that replaces an existing chart's stored definition (re-sent notes, series, guides or bands, clearNotes, an import replacing a chart by name), keeps the previous definition in a chart bin on this machine (the last 20, per project, never in a profile or a shared settings file), as deleted reports and walks already are. More → Restore deleted chart… brings one back; the delete confirmation now says the chart can be restored. External clients gain graph {restore}: {restore: true} lists the bin (each entry's id, name, whether it was deleted or replaced, and when) and {restore: "name or id"} restores one under its name, never over a chart that has the name now — one added optional parameter, the graph schema otherwise unchanged except delete's description, which no longer says "NOT recoverable".
• Claude Code in the tab asks before it deletes or replaces a chart. graph is judged call by call: creating a chart or showing one is still approved without asking, while deleting, renaming or restoring a chart, or replacing an existing chart's definition, opens the approval dialog — "Delete the chart “x” (it can be restored)", "Replace the chart “x” (its old definition can be restored)" — where it used to be refused (a delete) or allowed without a question (a replacement). Under a Full Claude Code mode that never asks (auto, plan), those calls are refused. The in-app assistant has no approval dialog; for it, the chart bin is what makes a delete or a replacement recoverable.

• The Claude Code operator lines are behind Details — the CLI version, where requests go, the PATH line, diagnostics and the last error line — reachable by keyboard. Details is shown only while the session's own decision says something needs you (its tooltip says what): a session that did not match its set-up, a start that failed or was refused, or a warning.

• Claude Code in the Assistant tab (review fixes, before release): only the analyser verbs on its explicit list are served to it, so a verb added later stays refused until listed; it has its own rate limit, separate from external clients; the folder question's answer and each of its tool calls are decided in the session's own order, so a call is refused once the session has been stopped for a profile mismatch; and it never starts in a folder other than the one confirmed, even when a project is opened during Start. The Turn on the REST transport button and the warm start setting are gone; a stored warm-start setting is ignored.

• Claude Code in the Assistant tab (second review fixes): pressing Stop while Claude Code is still connecting now starts nothing, where before the start could carry on and launch it briefly, leaving the conversation reported as open in another window; with External MCP/REST access off, the analyser's listener answers only the running Claude Code session, not any request that claims to be it.
• Full Claude Code, from its review: the folder dialog names the permissions a folder grants itself (allow rules, additional directories, default mode) instead of saying nothing runs, and flags a server named like the analyser's own bridge; .claude/rules, .claude/output-styles and .claude/hooks are part of the folder's trust, so a changed hook script asks again; a running session whose trusted folder files change is stopped at once and Start asks about the folder again, and a changed spend limit is offered as a restart that applies it (see Settings changed — restart Claude Code now, above); a .claude that links out of the folder is named and not read; the folder scan is bounded in folders as well as files; IDE, hook and plugin variables of a host Claude Code session and relative PATH entries no longer reach Claude Code.
• The analyser no longer prints the full REST token to its console when the REST transport starts; the token hint is shown, the endpoint file ~/.fluxtion-analyser/rest-endpoint (mode 600) carries the token, and Copy prompt carries it only when you tick Include my local REST token (see Security, below). Automation that read the token from the analyser's console must read ~/.fluxtion-analyser/rest-endpoint instead.
• The start page's hero title is always bold and larger, in every theme (its style was fixed from the theme in force when the page was built), and its drop line mentions prompt packs. The Claude Code header no longer shows the CLI's firstParty provider name.
• An unknown llmProvider is refused, and the in-app assistant says "the configured provider '…' is not one the analyser knows — choose one in Settings ▸ Assistant"; it used to be sent to Anthropic with the stored key.

• Explain selected with LLM and the toolbar's and record detail's Explain are replaced by the assistant's questions above; the toolbar's Ask is disabled, saying Select a record to ask about it, until a record is selected. The finding dialog is titled Write a finding for record N.

• One assistant panel. The in-app assistant and Claude Code are now one panel with one conversation, one composer and one header (the mode chip, below); the profile in force chooses who answers, and how it is paid for (your Claude plan or an API key) is said once. The in-app conversation is drawn as Claude Code's is — safe Markdown, code blocks, links that ask first, analyser references as chips — and its actions are the same collapsible tool lines with the same marks (✓, refused, denied). ⋯ ▸ About this session… says what the mode in force cannot do. Each keeps its own rules: in-app actions and the round limit; Claude Code's profile, permissions and approvals.

• The start page is one surface. The question What would you like to do? leads; the large cards under it became Or open directly — four small titled blocks (Explore DEMO, Open, Create, Settings), each heading above its buttons, in a grid that wraps as the window narrows; each button an icon and a short label, at least 32 px tall, with a focus ring (Guided tour, Sample project, Project, Evidence bundle, Audit log, GraphML, Project profile, From template, Settings, Fluxtion key), with the old card's sentence as its tooltip. Recent projects and evidence bundles sit side by side under Recent. Every action does what its card did.
• Settings ▸ Assistant is grouped by profile — Safe analysis, Full Claude Code, In-app assistant, and what applies to every assistant — and a control is disabled while the switch it depends on is off: the full switch, Claude Code's model and the spend limit need Allow the analyser to start Claude Code.
• Claude Code's profile is named as Set up AI and the quick switch name it: Safe analysis and Full Claude Code (mode) — no internal word ("posture", "hosted", a log "generation", a verdict code such as POSTURE_CHANGED:) reaches the screen; a log reloaded since a question was drafted is "load 3 → 4".
• The consent dialogs share one rule: the safe choice (Deny, Don't trust) is the default, at the right; the risky one (Allow once, Trust…) alone at the left. Safe analysis's folder question has verb buttons (Trust this folder / Don't trust, no longer Yes/No with Yes the default) and says so in words when the folder is the analyser's own private working folder.
• The proposal review speaks plainly: Source folder, Processor, Graph file, Runbook, Log to open, and the analyser's check in words (found: a folder with 12 Java files); Claude's reason — unverified shows for the row you point at or focus, before you answer.
• The start page's main action is its strongest control: the accented button beside the box, as tall as the box; Enter sends and Shift+Enter starts a new line, as in the assistant's composer, said under the box; the box has a placeholder and reads as a field in dark too. Choosing a prompt entry twice adds its words once; the no-AI label shows only with its buttons. On macOS the page's shortcuts are ⌘1 … ⌘9 and ⌘R (⌥ and a digit types a character there), and each tooltip names the shortcut that works. Accessible names contain the visible label (Other MCP client — set up, Fluxtion key — not found); the start page's chip remover is a 24-pixel target.
• One name for external access, External MCP/REST access, in Set up AI, the Settings checkbox (formerly Allow the assistant to drive the UI over localhost (REST)), the AI menu item (formerly Local MCP / REST enabled), the assistant's limits line, the status line (External MCP/REST access is on: listening on …) and the Connect an AI client screen (Turn on External MCP/REST access…, formerly Enable local transport…, and its status words); a question disabled in the in-app mode names the real switch, Let the in-app assistant change the views as it answers. The user guide, FAQ, install and connection pages use the same names, list all 18 tools (with import; propose is offered only to Claude Code in the tab), name all six destructive verbs, quote the bridge's not running message as it reads, and say how a script gets the token (~/.fluxtion-analyser/rest-endpoint, check its pid) and why an external client is refused while Claude Code runs in the tab with external access off or closed. Settings ▸ Assistant says where the REST URL and token are (~/.fluxtion-analyser/rest-endpoint, readable only by you; the console shows the token's first 8 characters), names the exchange folders Exchange folder (this machine) and Exchange folder (this project), and that every switch is per user with no administrator lock yet (a managed policy is planned).
• Spotlight callouts are never cut silently: a caption wraps at about 40 characters, shows at most three lines ending "…", and expands in place to the whole caption on hover or keyboard focus; its accessible name is always the whole caption. The spotlight guidance every assistant reads asks for captions under about 80 characters, with the detail in the answer.
• Colours: secondary text, control borders, the focus ring and placeholders meet WCAG AA in light and dark (secondary text was the disabled grey); full Claude Code's colour is its own — a denied call is a neutral ink with its words, a status note its own colour. Transcript expanders are 24-pixel targets.
• The words of two refusals an external client sees changed; their status codes did not. A hosted-only listener (Claude Code in the Assistant tab with external access off) answers an external MCP/REST call 403 with "External MCP/REST access is off in this analyser (Settings ▸ Assistant); this listener serves only the analyser's own Claude Code session", and the fluxtion-analyser MCP bridge's unreachable error (-32001) reads "analyser not running, or External MCP/REST access is off — turn it on in Settings ▸ Assistant, …" (it said REST transport disabled). A client should key on the status or error code, never on the text.
• Settings ▸ Assistant names the header control as it reads (the mode chip, Safe ▾ for example), not Assistant ▾.
• A drafted question no longer says (log generation N). A small muted chip by the composer says which log the drafted questions are about — this log · opened 14:05 (or an earlier log · Send will refuse it after a reload) — and the analyser tells the model the binding with the message it sends. Send still checks every drafted question against the open log.
• The shared canvas's posture is now Work mode wherever you read it: AI ▸ Work mode ▸ Investigate / Build / Guess from what is open (was AI ▸ Posture ▸ Research / support / Authoring / deploy / Derived), and the Project panel's row reads Work mode: Investigate — guessed from what is open; change it in AI ▸ Work mode. Nothing external changed: open {posture: research|authoring|derived}, context.handoff.posture and every MCP/REST schema are byte-identical, and the spotlight target menu:AI:Posture still lights the item (an alias of menu:AI:Work mode).

Fixed

• A Java source spotlight asked right after goto or topology {select, source: true} no longer bounces "superseded; retry against the current view". It now waits for the source read that verb started, and the scroll that read queued, before it captures the view. A newer spotlight or a real navigation still supersedes it.

• The status bar's line after opening a log and a graph no longer depends on which arrived first. With a graph open beside the log, the bar shows the graph line (its verdict about this log, with the log's own warnings) whether the graph was opened before the log, after it, or in the same open {log, graphml} call. Before, a graph opened first or in the same call could leave the log's line instead, and which one stayed was timing. Follow still shows the Following line.

• Full Claude Code writes a new report in your project without asking. Under Full Claude Code, a report call that creates a new report in the open project, adds a section to a report that same session wrote, or writes a new PDF or CSV into the project's exchange folder no longer needs an approval; the tool line says allowed: writes a new report inside this project. It never writes over an existing file or a report you made (the analyser re-checks just before writing and refuses a report that appeared or changed meanwhile). Overwrites, deletes, renames, restores, evidence bundles, anything outside the project, Safe analysis and external MCP and REST clients are unchanged. The permission belongs to the project it was given in, that exact output and that complete report revision: a project switch, a change of destination or your own edit of the report ends it.

• The transcript keeps every word of an answer. Inline code that names a node of the graph (such as `OrderUpdateEvent`), a record or a chart becomes a chip, and the chip's characters were not in the transcript's text: selecting, copying or reading the answer aloud gave "tracing an order from through to …". The chip's label is now the transcript's own text under it (a quoted chart name keeps its quotes); chips still only navigate.
• An answer's Copy no longer flickers. Moving the pointer onto the Copy button, a chip or the unverified mark hid that answer's Copy, because the transcript took the crossing for the pointer leaving it. Copy now stays shown while the pointer is anywhere over its answer, the button included, and hides only when the pointer leaves the transcript; its place was and is reserved, so nothing moves.
• A proposal from Claude Code can always be seen. Its Review… banner was only in the assistant panel, so with the assistant behind another tab (a spotlight on Topology or Reports selects one) or in a separate window behind the main one, nothing on screen said a proposal was waiting. A strip above the start page and the workspace now says so while one waits, and its Review… brings the assistant into sight and opens the review. A starter proposal's banner now says it creates a new project in a folder you choose, not that it sets up the session's own folder.
• Java source spotlights light beside the graph. With the Topology tab showing, every source:java: spotlight was refused as "no longer visible in its bound document": the embedded source pane opened in Split with the half it needed at height 0, or, in a narrow side column, with no width at all. The spotlight's half now gets room and the pane always keeps width, so in the sample project the node, its Java line and the records light together (found in the owner's hands-on test, 2026-10-18).
• The sample project's profile always reaches its sources. A sample profile whose source roots do not find the shipped processor is healed on the next open (the demo folder, written project-relative); a healthy profile is not rewritten.
• run-mongoose-server names the start script where the starters ship it, ./run-server.sh at the project root, not ./scripts/run-server.sh.
• context {sections: […]} says where a key lives. Asking for walks (a key — context.walks — not a section) was refused with only the list of sections; the refusal now adds that walks is carried by the project section, and likewise for any other key named as a section.
• Full Claude Code opens the sample project's own log and GraphML without asking. The sample's session works in a scratch folder, so its own files were never "inside the project folder" and every open of them needed an approval. While the sample is the active project its folder counts too, by the same real-path checks; Safe analysis, a project switch, Follow and any file outside are still asked, and the tool line says allowed: it only opens a log or GraphML inside this project's folder or the analyser's own sample. A link inside the sample or the analyser's folder cannot widen the exception; ordinary home-folder aliases still work.
• An approval no longer times out unseen. When Claude Code's approval dialog opens, the window it belongs to — the popped-out assistant's too — is restored if minimised, brought to the front and given focus, and on macOS the Dock icon bounces. If it still goes unanswered for 120 s, Claude is now told not to retry the same request but to ask you in its reply (it used to retry, each a fresh 120 s wait).
• Two misleading replies to the assistant corrected. context.processors[].from no longer says discovered under a root for the selected processor when no source root is configured (a selected or configured class is the project's or your own settings'); and source {bean} with no design open now says no design is open and, for a node of the open graph, names its class and the source {fqn} call that opens it, instead of a file path is required.
• The sample project finds its own event processor. Its profile named a placeholder class that is not in the demo set, so the processor's source never opened and context.processors said not found. The sample now selects com.acme.demo.generated.DemoQuoteProcessor (and lists the traced log's processor), an older sample profile is healed on the next open (only its processor fields change), and a new configuration names no processor until you choose one.
• Start page, no log open: a cancelled chooser no longer leaves a dead end. Pressing Open a log… on the "No log is open — the answer will be general." line and then Cancel used to leave Open the sample and Open a log… disabled until the question was asked again. The line now stays with all three actions usable, and Open the sample still opens it and goes on with the question.
• Full Claude Code opens your project's own log without asking. An open of only a log or GraphML file whose real path (links resolved) is inside the session's project folder no longer needs an approval under Full Claude Code. Switching or closing a project, Follow, any other field and any path outside the folder still ask, and Safe analysis asks for every open. External MCP and REST clients are unchanged.
• Full Claude Code is told to stay inside the project. The analyser adds one fixed instruction to a Full session (work only inside the project folder; never search for the Fluxtion API key or read key files; build with ./mvnw and the project's scripts), so it stops asking you about probes such as find / or env | grep. The folder-trust dialog and the FAQ say the analyser adds it, and that it is a request to the model, not a sandbox. Your project's CLAUDE.md and your Claude Code settings are not touched. A conversation started before this change starts a new chat, because its system note changed.
• Building with Claude asks far less. When Set up AI opens for a Build question it now preselects Full Claude Code in Claude Code's own acceptEdits mode, with the line "File edits and some commands can run without asking; your permission rules still apply." — you can still pick default, and the Full confirmation is still asked. A command your permission rules ask about, and the analyser's own changing verbs, are asked in the analyser's dialog, Deny the default; Full does not promise a dialog for every command. Also fixed: the full mode chosen in Set up AI was never written, so Full applied there always ran default.
• After a downloaded example is unpacked, the start page says which of its files were made runnable — "Runnable after unpacking: mvnw, regen.sh" (the first few and "and N more" past five, every file in the tooltip); nothing is said when none were. The archive's modes are read only while unpacking, so the line comes after it, not in the download confirmation.
• Start building with no AI set up opens Set up AI with your Build question waiting, as a first Send does, instead of a dead end. The wizard says plainly that Safe analysis cannot write code and preselects Claude inside with Full Claude Code (its confirmation is still asked); after Finish the build pathway runs as before — the starter catalogue is read, then the draft is placed, never sent. The start page no longer says "No AI assistant can answer here yet" twice, and with Build chosen its box reads "Describe the application you want to build" instead of "Ask about a log…".
• A question sent while Claude Code asks you to trust the folder is no longer lost: after Trust Claude Code starts and the question is sent once; after Don't trust it stays in the box. Before, the box was emptied and the panel said Ready with nothing sent.
• Full Claude Code with no project open starts in the analyser's own empty scratch folder. Every Start and Send was refused "The folder's settings, hooks, MCP servers or instruction files changed after you confirmed it…" although no folder had ever been confirmed, and Start Claude Code then did nothing: the start-time re-check now uses the same rule as the decision (the scratch folder counts as confirmed while it holds none of the files Start would load). A folder that was never trusted is asked about, never reported as changed.
• A downloaded example's mvnw and *.sh scripts keep the executable bit their archive records, so ./mvnw and the example's own scripts run as its README says (they failed with permission denied). No other file is made executable, a script the archive marks 0644 or as a link stays plain, and the analyser itself still never runs an example.
• The template dialog no longer shows an internal milestone id in a template's description ("The M19 onboarding bundle…" now reads "The onboarding bundle…"); a check keeps such ids out of the shown descriptions and the start page's words.
• Sources ▸ Find GraphML in source roots… now also searches the resources folder beside each …/java source root, where a generated processor's .graphml is committed (src/main/resources); before, the playground bundle's graph was reported as "No .graphml found".
• Opening the audit log of the project that is already open no longer asks "Load this project?" when the project was reached through a link (macOS's /tmp is a link to /private/tmp, and a file chooser hands back the resolved path): the two are compared by real path.
• First run, polished (a robot re-walk of the newcomer's journey, 2026-10-18). When the assistant is docked and a spotlight it lights needs another canvas tab, the spotlight still shows what it points at, and the canvas says "Spotlight on Topology — the answer is in Analyser assistant" with ↩ Back to the answer, one press away (popped out, the answer never moves, so nothing is offered). A start-page question about evidence asked with no log open is no longer answered in general without a word: the page says "No log is open — the answer will be general." with Open the sample, Open a log… and Ask anyway, and nothing is placed or sent until you choose (after an open, the question goes on about that log). The in-app assistant with no API key says one true thing — the banner No API key — nothing is sent. with Add key… — instead of "Ready" over "Press Send to ask Claude". The restart you asked for is noted as "Restarted with the new settings — the conversation continues", not "Stopped: …". The Safe analysis status says "Send starts Claude Code" once.
• A filter that hides records now says so, wherever you look. With the Event types checklist hidden, a type filter used to leave no sign but showing 3 of 21 in the Time range header. Now the toggle reads ▸ Event types (▾ while the list shows) with an accent badge 1 of 7 whenever a type filter is in force, the Records title says Records: 3 of 21 shown · 18 hidden, and a line under it names each filter that hides records with its own undo: 18 hidden by event types Show all · 9 hidden by search "hedging" × (and the time range and Flagged only likewise).
• A search set by an assistant shows in the Search box. A text filter set through the action socket or MCP (filter {text: …}), a walk step or the topology's node menu now appears in the Search box, so the person sees what the assistant filtered by, and the hidden line names it.
• The Event types checklist counts each type under the other filters: Tick (3 / 7) when the search, the time range or Flagged only keep 3 of its 7 records, and a muted row when they keep none; the rows stay in order of their totals, and a type that first appears under Follow gets a row.
• An empty Records table says why, and offers the undo. The fixed-position grey message (below contrast, under the scroll bar at narrow widths, invisible to a screen reader) is replaced by a real line in the theme's muted ink: No records: event types exclude all 21 Select all · search "hedging" excludes 9 Clear. Record detail says the filter hides every record instead of asking you to select one.
• The Event types checklist asks for its rows' height up to a maximum instead of a fixed 200px.
• The assistant is told what the filter hides. context carries a new, additive hidden field — {count, byFilter: [{filter: "eventTypes", hides, excludedTypes}, {filter: "text" | "timeRange" | "flaggedOnly", hides}]} — and the in-app assistant's About: line says it too (filtered: 1 event type hidden (RiskBreachEvent), 9 of 10 shown), so "why are records missing?" is answered with the filter that does it.
• The checklist keeps three rows in a 1200×680 window. Select all / Select none sit beside the list instead of above it, the first group needs no Event types heading under the Event types toggle, and in a narrow Facts column the Search row puts its buttons on a second line instead of laying them over the search box.
• The Event types checklist and the Records table always show at least three rows each (all of them when there are fewer). The checklist's divider stopped at 60px, which held only Select all / Select none, and dragging the Records/Record detail divider up could squeeze either list to nothing. Each floor is now measured from the list's own rows, so it follows the theme and font size; both dividers stop at it, a divider saved smaller is moved out when the checklist is shown, and when the Facts column is short the Record detail pane gives way first, down to its buttons and two lines. Only in a window too short even for that does the checklist give way, never the records table. The one-touch collapse arrows on these two dividers are gone: the Event types toggle hides the checklist.
• First run with the assistant (a robot walkthrough of a newcomer's journey, 2026-10-18). A question asked from a record, a chart or the topology before any assistant is chosen now waits and is placed for the assistant you set up, instead of being drafted for the in-app assistant and then refused by Claude Code four times; asking the same question again replaces it rather than appending a refused copy, and a question drafted for the other assistant offers Ask again for Claude Code. The assistant panel no longer shows "press Start, then Send" beside "Press Send", a "Start — ask your question" button, or a disabled "Running…" beside "Ready" (see The assistant panel is quiet, above). Add key… opens Settings on the provider's page, not Source roots. The Full Claude Code confirmation shows its whole warning, including the sentence about the home folder and the REST token that was cut off.
• Claude Code and a change in Settings ▸ Assistant (owner's hands-on test, 2026-10-18). Switching Allow the analyser to start Claude Code back on after it stopped a session no longer leaves the panel saying it is switched off; putting a setting back before an answer ends no longer stops the session anyway; changing the Claude Code model or the spend limit while a session runs no longer silently waits for the next start — it is offered as a restart; and after Keep it running here on a project switch, the next Settings save no longer stops the kept session.
• analyser --help names the switch as Settings does — External MCP/REST access, not "the REST transport" — and says what --replay-compare excepts: every record is exact except its endTime and thread lines, not endTime alone.
• The start page's Evidence bundles guide button says where it goes: its tooltip read Help ▸ User guide ▸ Evidence bundles, a section the in-app help does not have; the button opens the docs site's Opening a bundle you were sent page in your browser, and the tooltip now says so.

• The REST endpoint file ~/.fluxtion-analyser/rest-endpoint is removed when the analyser's window closes (as before, now pinned by a test) and also when the analyser quits without closing its window — the application menu's Quit on macOS, or a SIGTERM. It can remain after a crash or a kill -9; every client checks the pid it names before trusting it (the analyser's MCP bridge reports analyser not running), so a stale file is harmless, and the next analyser to start replaces it.

• An external MCP/REST client refused while Close external access while this session runs is in force is told so — "external access is closed while an in-app Claude Code session runs" — not that the REST transport is off, which Settings contradicted.

• A screen reader hears who denied a call: every denied tool line was spoken "denied by the posture", including your own Deny and the 120 s deadline.
• Claude Code's question tool no longer opens an approval dialog. In full Claude Code, when Claude tries to ask you something through its own question tool (AskUserQuestion), the analyser denies it without a dialog and the conversation says Claude asked a question — answer it in the conversation; Claude is told to ask in its reply instead.
• The permission dialog no longer says a built-in tool was "named by the server". For Claude Code's own tools (Bash, Edit, …) the secondary line reads built-in tool; a tool of an MCP server still shows its server and the name the server gave it.
• The start page no longer counts an ollama provider with no key as an assistant that can answer — the in-app assistant refuses a provider it does not know, so the page promised an answer it would not give.
• A transcript tool line shows the input as Claude Code sent it, as the permission dialog does: an integer reads 3, not 3.0, collapsed and expanded.
• A request Claude Code withdraws while it waits behind an open approval is never shown — it used to be shown, and could be approved, after the first was answered; nothing is written to Claude Code for a request it withdrew.
• The approval dialog describes the call that runs. It chose its sentence with its own reading of the input, so report {name, delete: false, sections} read as delete … can be restored while the analyser built and replaced the report; the dialog, the analyser and the file-exchange check now share one choice of operation for report and open (and walk's own).
• A spotlight callout never splits an emoji at a line break: a caption wrapped, or cut before its "…", inside an emoji or another character outside the Basic Multilingual Plane, which then drew as two boxes; it now breaks between characters as they are drawn.
• An external MCP/REST request nested more than 512 levels deep is refused with 400 "malformed JSON". It used to overflow the parser's stack: the connection closed with no reply and the error escaped into the HTTP server.
• A spotlight callout no longer flickers open and shut at the frame's edge: the expanded card always covers the collapsed one; and when a callout that has the keyboard focus goes out, the focus returns to where it was.
• A held Enter on the start page no longer sends the question it placed: the composer ignores an Enter pressed before it had the focus, until that key is released. With an entry chosen on the start page, the hint under the box says what Enter does (Enter: Open log… · Shift+Enter starts a new line). The start page's ⌘/Alt+1…9 and ⌘/Alt+R no longer act while another text field has the focus.
• The set-up proposal's review shows every Yes / No at its default width: the answer is now the column after the number, the analyser's check wraps in its own column and a long path is shortened — at 760 px the answers used to be the last column, scrolled out of view.

Security

• A source root can no longer expose your home folder or your credentials. For every caller — you, external MCP/REST clients, the in-app assistant and Claude Code in the tab — source_root {add}, Settings ▸ Sources, a project profile, a shared settings file, a set-up proposal and a bundle's remembered roots refuse a folder that is your home folder or contains it, or that is, holds or is inside a credential folder (~/.ssh, ~/.aws, ~/.config, ~/.gnupg; for ~/.m2 only the folder itself, which holds settings.xml — ~/.m2/repository and below stay allowed), with symbolic links resolved and, on macOS, without regard to case; the refusal names the reason in plain words, and a root that reached the settings some other way is never read. An assistant is also held to the open project: the in-app assistant is refused a folder outside it ("add this folder yourself in Settings ▸ Sources"), and Claude Code's approval dialog for source_root says "outside this project", the call still yours to allow. External clients keep adding any other folder as before. Claude Code's working folder is judged by the same list, which now includes ~/.m2.
• A shared settings file can no longer turn on External MCP/REST access. Project ▸ Export settings…'s Assistant category, ticked by default, carried External MCP/REST access and Let the in-app assistant change the views as it answers, and Import settings… applied them on the receiving machine. Both are now set on each machine only, as file exchange, project creation and the Claude Code switches already were: never exported, never applied on import and never in a project profile. The category is now Assistant limits and carries only the in-app assistant's two action limits; an import of a file that carries a switch ignores it, and its summary says so — External MCP/REST access is set on this machine only. A team that turned external access on through a shared settings file must now tick it once on each machine (Settings ▸ Assistant ▸ External MCP/REST access, or AI ▸ External MCP/REST access).
• Invisible "smuggled" characters are refused in prompt packs and dropped from questions. Variation selectors (U+FE00–FE0F, U+180B–180F, U+E0100–E01EF) and invisible fillers (such as U+3164, U+115F, U+1160, U+FFA0 and U+034F) are not format characters, so a pack entry reading Summarise this log could carry a hidden payload, one byte per selector, that Claude received on Send; a log or chart name could carry one into a drafted question. A prompt pack holding one is now refused (hidden characters, naming its line), except a single U+FE0F directly after an emoji (the emoji form of ⚠, for example); a name in a drafted question drops them; and Claude's words in the transcript and the proposal review show them as marks.

• No assistant profile is local-only, Safe analysis included. Set up AI, the quick switch's tooltips, the assistant's About this session…, Settings ▸ Assistant, the in-app help, the FAQ and the assistant guide now say it plainly: what you ask, and the evidence Claude reads through the analyser, is sent to the model provider (Anthropic, on your Claude plan or API key), as in any Claude Code session; the in-app assistant sends them to the model provider you chose, on your API key. Safe describes what Claude Code may do on this machine, not where your words go.

• Full Claude Code's folder trust is checked again when the process starts, by its whole digest (settings, .mcp.json and instruction files), so a hook added between the confirmation and the start is refused by name; the confirmation now says plainly what it does not re-check — files a CLAUDE.md imports with @path, CLAUDE.md files in the folders above, and the contents of a hook's scripts.

• With external MCP / REST access on, full Claude Code's own tools can drive the analyser as an external client (through ~/.fluxtion-analyser/rest-endpoint). The assistant panel now says so while such a session runs and offers Close external access while this session runs; external access opens again when the session ends.

• Turning on Full Claude Code asks for the same confirmation everywhere. Set up AI's Full Claude Code and the quick switch's ◆ Full Claude Code… open the warning as a confirmation — Cancel is the default, and Enter, Escape and closing it cancel — and nothing changes until you press Turn on full Claude Code in it. The Settings switch uses the same dialog, now under the shared consent-dialog rule (the risky choice alone at the left). Re-applying Full while it is already in force asks nothing.

• A request waiting for your approval always shows its own input, and a request whose input is not shown cannot be approved. The analyser kept the words of the last 64 Claude Code permission requests of any kind, so a burst of other requests while one dialog was open could drop the words of a request waiting behind it: its dialog then showed no input while Allow once still allowed exactly that input. A request's words are now kept until it is answered or withdrawn, and dropped then; and if they are ever missing, the dialog says so and keeps Allow once disabled, and an allow that still arrives is written to Claude Code as a deny.
• An approval in the analyser covers exactly one call. The analyser itself refuses a call of one of its verbs that change what is open or write files from Claude Code in the tab unless your Allow once for exactly that call — the same verb, input, tool use and turn — is waiting; it is used once, and expires after 60 s, at the end of the turn, on Stop and when Settings change. So a call approved by a permission rule or hook of your own, without you seeing it, is still refused for those verbs; every other rule (the exchange directory, Allow the assistant to create projects, graph {delete}) still applies.
• The session's own record holds no words of Claude's and no absolute paths. A set-up proposal's summary, reasons, paths and class names, and the folders you choose for set-up or an example, are kept by reference and shown as …/name; the analyser's own reasons have paths shortened the same way. A permission request is recorded by its class, verb, turn and decision — never its input.
• Copy prompt no longer puts your local REST token on the clipboard by default. It used to append the REST action manifest — the loopback endpoint and the token that lets anyone holding it drive this analyser — whenever External MCP/REST access was on, and its dialog said only that nothing had been sent. The token, and the manifest with it (every action call needs the token, so the manifest is useless without it), is now copied only when you tick Include my local REST token (lets the pasted assistant drive this analyser), which is off each time the dialog opens and shown only while REST is on. The dialog names every part it copied — "the analyser's standing instructions, the glossary, the raw text of 2 records and your question" — says when the token was left out, and when a very large selection was cut at 60,000 characters it names only what was copied and marks the part the cut went through.
• Text the analyser did not write is shown as its characters, never as HTML. Swing renders any label, button, menu item, tooltip, list or table cell, tab title, window title or dialog line that starts with <html> as HTML, and fetches an <img src=…> in it the moment it is built. A recent project's path or folder name, a bundle's file name and sender's note, a report or walk title, a chart, series or focus name, a node id from a log or GraphML file, an event type, a profile's runbooks and analyses, a template catalogue entry, a CSV header and an error message that echoed any of them could all reach one, and a node id in the topology tooltip and an analysis rationale were spliced into the analyser's own markup unescaped. The recent-file menus (Project, Audit log and Sources ▸ Open recent) and the verified-bundle notice are covered too. Every such surface now shows the text as typed, and nothing is fetched. A tooltip needed its own fix: html.disable on a component does not reach its tooltip.
• Under full Claude Code's auto and plan modes the analyser refuses its asked verbs outright, the calls that change nothing stored (playing or ending a walk, listing a bin) included — those modes hide the verbs, and the analyser's own admission now agrees.

[1.35.1] - 2026-10-16

Fixed

• A spotlight set of several Project-panel sections (project:log, project:graph, project:processors, project:roots) no longer draws its boxes over the wrong rows. Measuring a section used to scroll the panel to it, so each later member moved the rows under the boxes already measured for earlier ones while the verb reported ok. Measuring no longer scrolls; revealing does. A section that scrolled out of view to show another is refused by name, as every other such pair is, and one still partly in view is boxed where it shows.
• The Project panel's Event processors list shows a processor the project declares once, not twice: the declaration now rides in that processor's provenance column instead of adding a second row with the same name, with each declared name the row does not already show (project declaration "Quote pricing EU", "Quote pricing US" · declared). A declaration that names no class (a runtime processor) keeps its own row.
• Source roots inside the project are named by their path in the project, and roots that would read alike once shortened keep more of their path until they differ. Six roots used to read ~/…/main/java. Copy, Reveal and Remove still act on the full path.
• A spotlight's caption keeps clear of what it is being compared with, wherever there is room: the topology's other visible nodes for a node, and the source lines either side for a lit line. Placement among callouts and lit targets is unchanged; this only breaks ties in favour of a spot that hides neither.

[1.35.0] - 2026-10-16

Added

• An agent can create a project through the same discovery as Project ▸ New project… instead of hand-writing a profile: open {project: "<absolute dir>", create: true} previews the offer — source roots, skills, GraphML to open afterwards, whether a CLAUDE.md can be offered, and an offerDigest — and writes nothing; open {project, create: {sourceRoots, skills, referenceGuide, offerDigest}} creates the project with only the items named from that offer, makes it active and closes the open log and graph. It never overwrites an existing project, refuses a stale offerDigest, refuses unknown keys, and goes with project alone. The verb count stays 18.
• Settings ▸ Assistant ▸ Allow the assistant to create projects, off by default. An assistant — over MCP/REST or the in-app assistant — may create a project only while it is on, and is also refused inside an evidence bundle's working copy, the exchange directory or an existing project, and through a .analyser that is a symbolic link. It is this machine's setting only: never in a project profile, a shared settings file or a bundle, so a cloned repository cannot turn it on. Project ▸ New project… is never gated by it. The FAQ's security answer lists every file a create writes.

Changed

• Project ▸ New project… no longer creates a missing folder. Choosing a folder that does not exist is refused at once, before the offer is shown, with "New project… does not create one: make the folder, then choose it"; it used to create the folder as a side effect of writing the profile.

Fixed

• Project ▸ New project… no longer reports success when the profile could not be written. Before, a failed write left the settings cleared, the selection was adopted into them anyway, and the next save wrote that over the previous project's profile — or over your own settings when no project was open. Now the create is refused with a warning and the settings you had stay exactly as they were.
• Project ▸ New project… can no longer write over a project. A profile that appears while the offer dialog is open, or a dangling link at the profile's path, is refused rather than written through; a directory that already has a project asks "Open it instead?" straight away, before the offer is shown.
• A created profile lists its source roots in the order they were offered, so creating the same project twice writes the same file; the order used to change from run to run.
• Project ▸ New project… reads GraphML only under the folder you chose: a source root or a src/main/resources reached through a symbolic link to somewhere else, and a linked .graphml file, are no longer walked or parsed, and are not offered. The same holds for an assistant's preview.
• A failed project create no longer leaves a half-written .analyser/project.fluxtion-settings behind to be refused as an existing project next time: a profile the create made is removed again, with an .analyser/ folder it made.
• After Project ▸ New project from template…, the dialog that lists the commands to run now lists each template's own first run, in its guide's order, one way to do each thing. The Audit analyser bundle gets ./mvnw package then ./run-and-export.sh (the one command its CLAUDE.md names) instead of ./run-server.sh, ./export-audit.sh and ./stop-server.sh separately. The Spring starters get ./check-fluxtion-key.sh, ./setup.sh, ./validate.sh, ./generate.sh, then ./run.sh or ./run-server.sh: they ship at most a placeholder processor, so their default build calls the remote generator and needs the key their README checks first. The other Fluxtion starters get ./check-fluxtion-key.sh then ./run.sh, where the dialog offered only ./mvnw package; the Mongoose starters get ./run-server.sh, after the key check where one ships. On Windows the dialog lists no commands when a step of the first run has no .cmd form, rather than half a sequence, and says so: "No Windows form of this template's first run. See its README." No catalogue template ships a .cmd run script today, so this currently covers every one of them. The dialog no longer says such a template "declares no recognised lifecycle scripts". TemplateExecutablesTest checks the list against a record of what each playground template's guide says.
• The add-a-node skill says that any bean a ReferenceBinding names must also be listed in nodeBeans. A parent reached only through a constructor-arg ref is not a node, and validation refuses the binding with SPRING_REFERENCE_UNKNOWN_BEAN. The skills index re-pins it.

[1.34.0] - 2026-10-13

Added

• The add-a-node skill adds a node XML first: declare the bean, its ReferenceBindings (DATA for a parent that must not trigger it, parentUpdateCallback on a TRIGGER parent for an @OnParentUpdate callback) and its sink (SinkBinding under sinks) in the designer XML, validate and regenerate — regeneration writes the class shell with its annotations, so the author fills only the method bodies. It says not to hand-write the class or wire a sink outside the XML: a sink wired by hand is not declared in the design, and a hand-written class that disagrees with the XML is refused (SPRING_RECONCILE_CONFLICT) (cycle-5 findings 5.7, 5.8).
• The load-audit-log skill has a section, Diagnose in a fresh context. When checking a design change takes more than one analyser read (a chart, several records, a before/after comparison, several processors), the author hands the diagnosis to a subagent instead of continuing in the long design conversation. A single check stays where it is. The subagent runs in the foreground, and the author waits for its reply without using the analyser meanwhile. Its prompt is a brief of at most 2,000 characters: the project, each processor's export and GraphML, the change, the prediction and the question. The brief carries no evidence. For each log in turn, the subagent reads the file-identity verdict, the pairing, the time order, the producer findings and the in-flight state from the analyser itself while that log is open, and reports them with its answer, each fact cited by recordIndex. On the authoring ladder's baseline, diagnosis was 54.8% of the cost; whether this lowers it has not been measured yet. CanonicalSkillsTest checks the section, and the skills index re-pins it.
• The Spring authoring guide has a Model choice section (step 2), for authoring with the Spring XML starter and the Audit analyser bundle: Claude Haiku 5.5 is the default and meets the acceptance bar on ladder rungs 3–5, with rung 2 (the chart step) still open; Claude Sonnet 5 is not recommended (on the first ladder, no rung where it passed and the default did not, at about 18 times the cost per run); Claude Haiku 4.5 is not for the chart step; other models are unmeasured. It covers your own LLM client, not the in-app assistant. Connecting an LLM and Playground to analyser point to it. Every row is checked against the committed run evidence on each build (ModelChoiceEvidenceTest, with a digest of the cited score files), so a row's status cannot disagree with its runs' pass or chart results.

Fixed

• Project ▸ New project from template… installs every script a template ships at its root executable. The Audit analyser bundle's run-and-export.sh — the one command its CLAUDE.md names — arrived without the execute bit, so ./run-and-export.sh failed with "permission denied"; so did the Fluxtion and Spring starters' run.sh, and setup-java.sh / runWasm.sh, which some generator options emit. The installer still never takes execute permission from the downloaded archive: it grants it to a fixed list of root script names, now checked on every build against a record of what each playground template ships (TemplateExecutablesTest), including every ./ command a template's own guide tells you to run.

[1.33.0] - 2026-10-13

Added

• The add-a-node skill shows how to prove a change by comparing audit logs (#126): capture the same input before and after, compare the business records — the keys nodes chose to log — separately from the framework records invocation tracing adds (thread, method, annotation, forkedExecution, asyncMethod), and conclude "the logged business entries changed / did not change, for this input" from the business comparison alone — an output no node logs is not compared. It ships audit-compare.py (Python 3, standard library only), which prints that split, exits 1 when business entries differ, 3 when a match depends on reading an ambiguous annotation as trace (--annotations-are-trace asserts it), and 2 when a log holds nothing comparable or is in a shape it refuses to guess at, and it says to verify behaviour from the audit log rather than from generated code. The load-audit-log, run-mongoose-server and add-a-node skills now say a bundle hosting several processors exports one file per processor, logs/audit-<processor>.yaml. The skills index re-pins them, and ships the script as a pinned supporting file.

Changed

• The add-a-node skill no longer tells the author to read the regenerated processor, its GraphML text or the builder to confirm a change. It reads the Spring XML, the compiler's own messages (validation counts, authoring receipt, diagnostics) and the audit log, which shows dispatch order directly. In a G14 authoring run those reads were about a fifth of all tool output and answered nothing the compiler messages and audit log did not. CanonicalSkillsTest fails on any skill sentence that tells the author to read generated code.
• The trace key annotation, which Fluxtion runtime 1.1.1 writes after method to name the annotation that dispatched a callback, is read as framework data like thread and method: the logical log view mutes it, and read still reports a node whose entry carries only trace keys as traceLikeOnly. Which keys are trace is decided by position: only the leading run tracing writes (thread, method or the forked forkedExecution/asyncMethod, then an annotation naming an annotation) counts, so a node's own key spelled annotation or thread, or logged after that run, stays the node's own, in read and in the logical log view. A forked node's trace entry now counts as traced, and a traced binary log's trace keys are muted by the same positional rule. Conformance fixture C33 pins it.

[1.32.0] - 2026-10-06

Added

• A saved spotlight walk can be recorded as a reel: one self-contained HTML page — a title page naming the walk, the log, its record count, its time span and the event processor; a frame per step with its own words and callouts; and a finish page (#82). Reports ▸ Spotlight walks ▸ More ▸ Record reel…, or walk {name, reel: "<file>.html"}. The frames are the analyser's own window with its spotlights composited in, photographed only once the session says the step has settled.
• No reel without its bundle named on the finish page. A reel recorded from a session that came from a verified evidence bundle ends with that bundle's sha256: identity, its file name as a link, the limits the bundle already states, and "open this bundle in the Fluxtion Audit Log Analyser to replay these steps yourself against the real log". A reel recorded from an ordinary session says plainly that there is no bundle to open and no identity to check — a reel is a claim, and an unanswered question reads as one. A step the analyser could not fully show carries its state and reason onto the page rather than being hidden.
• A bundle that does not carry the frames' log is not offered as their evidence. A bundle's provenance is about the project, not whichever log is open, so a reel could be recorded with a verified bundle in force while the frames showed an unrelated log. The finish page named that bundle and invited the recipient to replay the steps against it — a false claim their own verification would have confirmed, because the bundle is genuine. The page now states that the bundle is not the evidence for those frames and why, and makes no invitation it cannot honour.
• A reel that cannot be written leaves nothing behind. The page was written straight to its destination, so an interrupted write left a truncated HTML file that a browser renders as a reel which simply stops — before the finish page, the one that names the evidence. It is now staged beside the target and moved into place.
• A reel's text carries no machine path (#113). A reel is meant to be sent, but its captions, callouts, reasons, dialogue and title- and finish-page fields went onto the page as written, so a step pointing at a local file sent your home directory with it. Every word on the page now goes through the evidence bundle's own path redaction — the same rule, Unicode paths included — and the walk answer lists what was removed. A path whose ending that rule will not guess refuses the reel, as it refuses a bundle. The frames are screenshots and are not redacted; the title page says so, so a sender knows before sending.

Fixed

• The log-findings display suite now accounts for its deliberately injected SummaryPanel failure and fails on unexpected EDT exceptions. Its failed-load recovery assertion and mutation controls remain in place.
• Native mouse regressions confirm delivery before testing a gesture and retry only an undelivered setup press. Unavailable input is an error, so it cannot be scored as a caught cancellation mutant. Cancellation assertions and controls remain unchanged.
• The bundle processor regression waits for the bundled log's source inference before checking a recipient's later choice, and exercises the real processor-selection action instead of assigning its config field directly.
• Faster development checks: focused draft mutation feedback, cancellation of superseded PR runs, and eight full-gate workers with named witness baselines; complete mutation evidence is still required for merge.
• Borrowing from an evidence bundle reads and unpacks off the event thread (#83). The guarded settings update runs on the event thread and is refused if cancelled or if the project changed during the read, including switching away and back.
• Evidence-bundle opens still create a pristine working copy (#85). Opening another bundle or choosing Private settings → Clear unused copies removes only marked, same-host copies whose ownership locks are free. Active readers in another analyser and pending opens are kept; older unmarked copies and uncertain ownership are kept too. Closing a project releases its copy for a later cleanup. The bundle files themselves are untouched.
• Saving an LLM provider key in Settings clears the assistant's "no provider configured" banner and enables Send at once, and a regression test and mutation control now hold that (#78). It already worked on main, but only as a side effect of an unrelated settings event; nothing guarded it. The assistant guide now explains valid OpenAI API model IDs, and that HTTP 429 can mean a rate, credit or usage limit.

[1.31.0] - 2026-10-05

Changed

• The analyser now runs on fluxtion runtime 1.1.0, the released stack (compiler 1.0.76, mongoose 1.0.32). Its own session processor and the DEMO and replay test fixtures were regenerated with builder 1.0.76; the existing test fixtures' audit logs and replay records are unchanged, byte for byte. The DEMO set the start page installs is unchanged.

Fixed

• Bundle provenance is bound to the accepted operation and the verified profile content, not a reused pathname (#93). Borrowed charts reach the live view before saving (#94), and unrelated edits preserve temporarily unavailable bundle source anchors while deliberate deletion still works (#95).
• Java walk steps stay Preparing until their targets are actually lit; source reads run off the UI thread, fail or expire explicitly, and native keyboard navigation supersedes pending work (#96–#98). Java captions and target reasons now disclose that the saved source revision was not compared; CURRENT remains name/line lookup, not source-content verification (#102).

• Project reopen offers use operation-scoped origin and permission in the session processor, record disabled offers as skipped, and label project and machine-fallback candidates independently (#99, #101). The Project-panel architecture gate now covers nested helpers and inherited Navigator actions (#100).

• Bundle profile redaction recognises Unicode path segments, including decomposed accents (refs #79). Quote a complete path with double, single, backtick or curly quotes, or Japanese corner brackets, to redact it whole. Quoting preserves the existing ratio, URL and path-shape exemptions. Quoted spans containing sentence punctuation, another quote type, a second path start or a parenthesised message fall back to ordinary path handling, preserving the surrounding prose. A straight or curly single closing quote needs a separator before a following letter; double quotes and corner brackets can sit beside prose. An unquoted final segment containing Han, Hiragana, Katakana, Hangul or Thai still refuses export with a request to quote the path. The existing limits for other scripts and ambiguous path spellings remain documented.

• A record broken by a line break inside one of its values no longer becomes evidence. An unquoted value — measured with an admin command's operator-typed arguments in eventToString, on mongoose 1.0.32 with fluxtion 1.1.0 — could end its line early, and the text after it was read as the record's own fields: the analyser reported a node that does not exist, blamed the graph for not declaring it, and showed an event type that was never dispatched. The analyser now reads a record's structure by one rule: a field only at the fields' own indentation, outside a quoted value and outside the node logs. A record whose structure breaks — a line less or more indented than its fields, a line at their indentation that is not a field, a second record key, a field written twice — is read only up to the break, before the first copy of any repeated field, and no further than its event text, node logs or an unknown field. Both repeated copies are withheld even when the first came from an unquoted grouping ID or thread name; a genuine first copy can also be withheld. None of its node logs are read, and a new producer finding names it by record and line. A value quoted across lines is never read as fields, a new node or the end of the node logs. An exported service call's two-line description, which the generator writes, is read whole, as before.
• Node logs the analyser declined to read are no longer reported as absent. Coverage, the report's coverage table, the graph pairing, the records table and the CSV export now say a broken record's node logs were withheld, where they said a node never wrote audit output, that no node output was recorded, or that the record held 0 node logs.
• An unknown field written after a record's node logs ends them and is ignored, as the format says. It used to be read into the last node's value.
• Opening a graph with, or after, a log no longer wipes the log's warnings from the status bar. The pairing sentence replaced the whole line, so a broken record, run-together records or a time-order problem stayed in context and the tooltip while the bar read clean.

Checked

• Mongoose 1.0.32 admin commands on a processor generated for fluxtion 1.1.0 audit as their own AdminCommandEvent record: the analyser frames, tables, covers and steps through them correctly, and an event a command raises follows as its own record.

[1.30.1] - 2026-09-30

Fixed

• A walk step can point at Java source (#72). It was refused with the design, toolbar and menu targets on the grounds that source has no basis a walk can check — but a class's identity is a NAME resolved against the source roots in force, which is a firmer basis than a record index bound to one log. The most natural thing to point at in "how is this price calculated" is the line that calculates it.

[1.30.0] - 2026-09-30

Fixed

• The event-type checklist and the records table now share the Facts column on a divider you can drag, remembered between sessions. A fixed band could not serve both a log with three event types and one with forty.
• Reopening a project offers its topology as well as its log. "Inside this project" meant "under the project directory", so a project whose code lives in sibling checkouts — its own declared source roots — had its topologies excluded and only ever offered logs. Each list now also falls back on its own.
• The recent-projects list no longer fills with bundle working copies. Opening a bundle unpacks a fresh copy each time, and each one was recorded as a project — ten opens pushed every real project off a list of ten. Copies already recorded are forgotten on the next start; the bundles themselves are remembered in their own list.
• Open project… can reach a profile again: the chooser hides dotted directories by default, and a profile lives in <project>/.analyser/, so the dialog could not show the file it was asking for. It also opens in the current project's directory.
• The window titles a bundle by its own file name — recorded-run.fexp — rather than by the temp folder it was unpacked into (bundle-<hex>-<random>), which nobody can follow. The identity is still beside it.
• Deleting a bundle's source root now sticks. The deletion was never recorded — an empty set of roots was refused as "not an answer" — so closing and reopening the bundle brought the root back.
• The event-type checklist can no longer starve the records table: it takes at most a third of the Facts column and scrolls beyond that.
• Clearing a project's default focus now sticks. defaultFocus was not an owned key family, so the previous value was copied back over the blank one: the setting could be changed but never cleared.
• Starting the analyser no longer rewrites a restored project's committed profile. Applying its settings entered the save half of the funnel, so a launch alone produced a diff nobody asked for.
• Borrowing from an evidence bundle no longer leaves a full working copy behind while reporting "nothing was changed". It reads one file, from a scratch copy it removes.
• context.project.bundle now names the event processor as the sender's claim and says it is not paired against the log, rather than adopting it silently.
• context.bundles says whether the exchange directory in force was declared by the project or is this machine's own, instead of calling both "this project's".
• A project restored when the analyser starts now has its settings in force, not merely loaded. Its source roots reached the config but nothing applied them, so every event processor read "source not found" and the Source tab was empty while the Project panel listed all the roots.
• Opening an audit log that lives inside the project you already have open no longer offers to "load the project" and move you onto a different profile of it. The offer compared the profile file, so a named profile (project.<name>.fluxtion-settings) looked like a different project from its own canonical one.
• Open on an event-processor row now opens that processor's source. It reached "the Source tab", which shows whichever processor is selected — so a list of six had one button repeated six times. The active processor is unchanged; making one active is its own action.
• Listing a project's evidence no longer fails while a capture is in flight: the walk went into the capture's own working folder as it was created and deleted, and the resulting error escaped a context call.
• The start page no longer leaves "Verified …, the audit log is loading" behind once the bundle is open; the Project panel carries its identity, working copy and limits for the whole session.
• A chart explanation or note mentioning something like ~1/price is no longer mistaken for a home-directory path and replaced with "path removed" inside an evidence bundle. Digit-leading usernames such as ~7dev/logs/x.yaml are still redacted.
• Opening an evidence bundle whose project fails to apply no longer leaves the next ordinary project labelled as that bundle's evidence.
• Restarting the analyser no longer comes up inside a bundle's working copy. The copy is the project while the bundle is open and was remembered like any other, so the next launch showed an experiment with no graph, no log and no source: all three arrive by opening the bundle, never by loading a profile. It now starts with no project, and the bundle is one click away in Recent evidence bundles.
• Reopening a bundle you had already anchored in the same session puts its source tree back. It was remembered but only ever restored once.
• Clearing your own source roots while a bundle is open no longer erases that bundle's anchor.
• Two evidence bundles sharing a file name are now tellable apart on the start page: each row carries just enough of its folder to differ. The same bundle copied into two projects used to give two identical rows, separated only by a tooltip.
• Opening an evidence bundle no longer arms crash recovery for the working copy.
• The start page no longer keeps a "session captured" banner at the top after the capture is done.
• Add source root now opens in the current project or experiment's directory rather than your home directory.

Added

• The three columns are named: Context (what the system is), Facts (what the run did), Canvas (what you make of it), each saying so on hover. Opening a project offers the audit logs and topologies it knows about — either, both or neither — instead of leaving you with a project and nothing on screen; a profile still carries settings and never what was open. Project ▸ Open recent log and topology… asks for the same offer at any time.
• Private settings, under Context: what this machine remembers about the open project — the focus you were last on, the bundles opened here, the working copies unpacked. Never written to a profile, never shared.
• A project can name a default focus — its own "start here", carried in the profile and travelling to a colleague — while the focus you were last on is remembered per project on this machine only. Yours wins when both exist; a fresh checkout gets the project's. Set it from the focus picker.
• Event-processor rows carry their own actions: open, set active, remove, add source, settings.
• The event-type checklist moved into the Facts column, with its own button beside Search, above the records it filters.
• A source root can be removed from the Project panel row that shows it, and Settings… on that row arrives with the root selected rather than merely on the right page.
• An evidence bundle now names the event processor its log came from, so a recipient opens it with that processor selected instead of none. It is a class name, not a path: the graph names the nodes and never the processor, so there was nothing to guess from.
• The window title and the Project panel now say when the session came from an evidence bundle, and which one, for as long as it is open — not only in the dialog at the moment it opens.
• Evidence bundles you have opened are remembered: a "Recent evidence bundles" list on the start page, showing what the sender said each one is, and the same list in context.bundles.
• The assistant can drive evidence bundles: open {bundle} opens one, and import takes named parts of a bundle into the project you are already in, so a bundle can be investigated without reaching for the desktop UI.
• Recent evidence bundles and recent projects each carry a … menu: remove from the list, reveal in the file manager, or rename.
• An evidence bundle can be pointed at your own copy of the source it describes. A bundle carries no source paths, so the code behind its records was unreachable; the tree you choose is remembered against that bundle and restored next time you open it.

[1.29.0] - 2026-09-29

• Assistant turns now attribute an open to the action that requested it through the reader's completion. A person's competing open stops the old turn, while the assistant's own open waits until its log and view are applied before running the next action. Changing the investigation filter during a turn stops its pending reply before it can chart the new view; changing it after a turn completes keeps the conversation usable, and the next Send reads the current filter.
• Assistant retries keep their tool instructions after a cancelled first request; late Java-source preparations cannot light after the turn ends. Provider failure bodies and unknown action names stay out of session diagnostics.

• Journey dialogue limits include the author and serialized text; oversized profile counts and overflowing schema versions are refused without unbounded turn allocation.

• The first conversation journey: Find the first recorded breach (OA-5). A four-step DEMO journey over the quote-series log, with a scripted conversation — finding the record where the application first logged a breach (record 16, its own RiskBreachEvent), the chart that marks every one, and what supports the answer. It is an ordinary .fexp evidence bundle with no replay inputs, on its own docs page: Conversation journeys ▸ Find the first recorded breach, which states its size, sha256 and identity. tools/capture-journey.py builds it from a real run, asking the analyser for every number the dialogue states, and JourneyCatalogueTest re-derives those numbers from the bundled log on every build. An analyser from before conversation journeys (checked with the released 1.27.0 and 1.28.0) opens the bundle and plays all four steps without the dialogue, and keeps the dialogue if it saves the project.

• Conversation journeys play beside the real evidence (OA-4). Playing a walk that carries dialogue shows it in the assistant, labelled Simulated, Recorded or Edited recorded conversation in words, beside what the analyser actually shows for that step: Preparing evidence… until the step settles, then each target's real state. Next, Back and playing from a step reveal exactly that step's part of the conversation — nothing is re-run or duplicated. A step whose evidence is not shown says so and keeps the conversation at the last step that was shown, never presenting that step's words as if its evidence were. Nothing is sent to a provider and nothing is written while a journey plays; there is no Send in a demonstration. Ask about this evidence ends it and opens a fresh live conversation with none of the journey's words; Connect your CLI assistant is beside it. A dialogue journey waits for a live turn in progress to finish or be cancelled, and a live question ends a journey. The strip names the journey's mode, and the conversation can be hidden without stopping the walk.

• Walks can carry dialogue (OA-3). A spotlight walk may hold a conversation — written as a script, captured from completed turns of the live chat, or captured and edited — each labelled for what it is (Simulated, Recorded or Edited recorded conversation). Each step names the last turn visible there, bound by a stable step id that survives rename, replace and reorder. Open it from the Reports tab's walk list (More ▾ ▸ Conversation…) or the spotlight save menu (Include conversation…): nothing is preselected, only a finished turn's question and answer are offered (never the hidden context, manifest or system prompt), and Preview what leaves this machine shows the exact words that travel. The walk verb saves dialogue with steps or attaches it to a saved walk, and context.walks names a walk's dialogue without repeating its words. Dialogue travels with the walk through projects, sharing, the bin and evidence bundles; a machine path in a turn is removed on export and named, and malformed or oversized dialogue is refused by name before anything is installed. A walk without dialogue is stored exactly as before; a newer dialogue version is kept intact and the walk plays without it.

• Pop out the assistant (OA-2). The assistant can move to its own resizable window and back (Pop out / Dock, or close the window to dock it). It is the same conversation, draft and request in either place — moving it never sends again or cancels — and it stays open while the analyser changes tabs. The docked tab says where it went, with Show and Dock. The window's place is remembered on this machine only and restored within today's screens; it follows the theme. Scrolling or copying in the docked assistant no longer ends a showing walk.

• The onboard assistant's turns are decided by the session (OA-1). Send, Cancel and New chat are session facts; a reply, an error or an action result that arrives after Cancel, New chat or a change of project, log or graph can no longer append, run an action or clear a newer busy state. Every action result — a render success too — goes back to the model, and each action shown is the dispatcher's actual result, labelled OK, REFUSED or NOT RUN. With no log open, context, topology and source verbs work; record verbs refuse by name. A turn states when it is requesting, running an action, complete, cancelled, superseded, failed, or at a limit; a new per-turn action budget (Settings ▸ Assistant) sits beside the per-reply cap and round limit. With no key, nothing is sent: the panel offers Configure provider, Copy prompt and Connect a CLI assistant. A provider reply with no text is a bounded failure rather than a raw body.

[1.28.0] - 2026-09-29

• Opening a topology while an evidence bundle is still being prepared now keeps the newer graph; the late bundle cannot switch projects or load its log.

Changed

• The start page groups choices by outcome. Explore DEMO and Create a project form the left column, Open your work and Assistant and settings the right; the four groups stack on a narrow window. Evidence opening is distinct from project creation, file-drop guidance is visible, recent entries identify their profiles, and assistant connection precedes optional regeneration setup.
• Start and narrow panes remain usable throughout an investigation. Start keeps its opening choices and a return link when a workspace is active, offers an explicit existing-project action, and shows bundle progress and drop refusals in the page. Native drops on explanatory text work; extra dropped filenames are named. A late bundle verification cannot replace a newer project choice. Walk, series and producer-finding controls reflow in narrow panes.
• The Swing analyser gives charts and reports more room. Open graphs use one selector instead of a wrapping tab strip; report, walk and chart actions use compact menus; saved walks offer a selectable step list. The Edit series panel moves below the plot in a narrow pane, and reports have a distinct reading surface and human-readable titles in their list. Chart grids use round value and UTC time ticks, and the value range includes zero when doing so does not flatten a narrow-range series.
• The start page uses the full workspace. It offers evidence-bundle loading, incident investigation, project authoring, direct file opens, a bundled sample, a saved spotlight tour, and a clickable recent-project list. Opening a recent project goes to its workspace, and each start choice has a theme-aware icon. Producer findings now separate severity, explanation, fix, source declaration and location actions into readable cards. The template picker puts choices beside their explanation, and file drops route evidence bundles, GraphML, audit logs and Spring XML designs.

Added

• An evidence bundle can carry the run's replay records. Name them with report {bundle: {path, replay: <file>}}. They must pair with the open log: each replay record is one of the log's records, at its eventTime, in order. A replay from another run is refused, naming its first record that does not match. It is also refused with a time window, or while the log is still growing, because a replay needs the whole run. The records are packed as the replay/ member of a format 2 bundle, and a bundle without them stays format 1, unchanged. The capture says how many recorded inputs it carries, and how many exported-service calls the log holds that replay records cannot carry. --verify prints the same, with the bundle's replay limit. Nothing in the analyser replays them: you replay them into your own build, then compare.
• --replay-compare <bundle.fexp> <replayed-audit.yaml>: does a replay give the same audit log? It verifies the bundle, then compares its log with a replayed one record by record. Every line must be exact except endTime, the live clock reading at the end of a cycle, which a replay cannot know. It prints AGREES, N of N records and exits 0. Otherwise it prints DIVERGES at record k (Event): path: 'bundled' ≠ 'replayed', the first difference with both values. A line one side has and the other lacks, such as a node that logged nothing on replay, is named as missing. The first record one side has and the other does not is named too. Either way it exits 1. A bundle with no replay records, or with an excerpt, is refused. Besides endTime, the thread a cycle ran on is excepted, because a replay runs on its own thread.
• The replay runner: tools/replay/ReplayBundle.java. Run it with JBang against your own build: jbang tools/replay/ReplayBundle.java --bundle run.fexp --processor <class> --cp <your build> --out replayed.yaml. It checks that your build is the bundle's processor, comparing the nodes and edges of your build's GraphML with the bundle's, and refuses by name when they differ. It loads only the event types your processor handles, and replays each record at its recorded instant on a data-driven clock. It writes the audit log for --replay-compare.
• The DEMO records a replay. examples/fixture-generator gains a replay writer compiled into the processor. It records only the inputs the caller names, stamped with the instant each cycle ran at, for exactly the event types the processor handles. It also writes a recorded run's log, replay records and graph, and writes them only after replaying them reproduces the log.

[1.27.0] - 2026-09-28

Added

• Evidence bundles: an investigation as one verifiable file. Ask the running analyser with report {bundle: {path, notes?, from?, to?}} and it writes a .fexp into the exchange directory: the log, its graph, the saved charts and named focuses, reports and walks, hidden columns, and your notes as notes/NOTES.md. It refuses, by name, when the capture would not be coherent: no log open, a load pending, a log file that changed or whose identity is not established, a log that is not one plain file, or a bundle already being written. A log still growing under Follow is captured as the records read so far, and says so. It takes the settings in force now, not the file on disk, so an edit a read-only profile could not save is still in the bundle. It pauses Follow while it copies, and if another log is opened or the log is closed meanwhile, it refuses and deletes what it wrote. In an exchange directory shared by several analysers, it never touches another capture's working folder; a folder left by a capture that was killed is cleared once its owner is provably gone. context.capture says when it is done, its identity, and what was left out, redacted or excerpted.
• An excerpt instead of the whole log: from and to pack only the records in that time window, each record's exact text, re-read and matched against the source before the bundle stands. Walks and reports are re-based onto the excerpt, so their steps are current on the other side. One that points outside the window is left out and named. The manifest records the cut.
• Nothing that describes your machine leaves: no source roots, runbooks, environments, processors, keys or assistant settings. A chart with an external CSV series or markers is left out and named. A setting whose whole value is a machine path refuses the capture, naming it. A path written inside prose (a report narrative, a caption) is replaced by ‹path removed› and listed. Relative paths, URLs, times and ratios are left alone.
• --verify <bundle.fexp> and --unpack <bundle.fexp> [--into <dir>] for the recipient. They check every member against the manifest in constant memory whatever the log's size. They refuse a changed, missing, unlisted, duplicated, oversized or escaping member, naming it, and say when the log is an excerpt. --unpack extracts to a fresh working copy only after the whole bundle verifies, and never modifies the received file. A bundle is unsigned (it shows that nothing changed, not who sent it) and is not a replay, and every output says so.
• An Evidence bundles section on the docs site: what a bundle is, sending one, opening one, the operation, commands and format, and a demo. tools/evidence-bundle-demo.py runs the round trip on the DEMO log with a sender and a cold recipient, whole log and excerpt, and checks every step.
• context.project.unsavedEdits: whether a project edit is still waiting for its (debounced) write to the profile file.

Changed

• A spotlight walk states the "not re-checked" caveat once, on its first step that rests on records or charts, instead of on every such step. With Follow off it was always present, and it filled the strip. Stepping a walk from the socket with walk {play, step} now continues the walk already showing, so the caveat is not stated again on each step.
• context.log.generation: the session's log generation, so a tool copying the open log can tell whether another log was opened while it copied.

[1.26.1] - 2026-09-28

• Cancel unfinished record-table and time-slider drags when an owned modal takes focus, so opening a modal during a drag cannot leave selection scrolling or time-window panning running after release. Non-modal focus changes preserve an ongoing drag and defer its final selection notification until release. Switching to another application does not cancel the drag; the original reported incident remains unconfirmed.

Added

• A mouse trace, for one open defect. -Danalyser.mouseTrace=<path> (or =stderr) records every mouse press, release and click with the component AND window that received it, the buttons it believes are down, and the records table's selection state — and flags the moment the table believes a drag is in progress while no button is down. It exists to catch a table that has been seen scrolling and extending its selection on its own until the app is restarted. Off unless asked; it consumes no events and changes no behaviour. See docs/handoff/runaway-table-mouse-trace.md.

[1.26.0] - 2026-09-28

Added

• Spotlight walks — an explanation you can replay. A walk is a named sequence of spotlight steps, saved in the project like a report, that you step through on the analyser itself with ◀ ▶ (or the ← → keys) on a strip on the spotlight — later, by someone else, with no assistant connected. Right-click a spotlight to Save as new walk…, Add to walk, or during a walk Replace this step; the right-click keeps the spotlight on. Or ask the assistant: the new walk verb (the seventeenth) saves, plays, ends, renames, deletes and restores walks, one operation per call. The Reports tab has a Spotlight walks list with Play, Play from step…, Rename, Delete and Restore deleted…; the Project panel counts them; context.walks reports each walk and the showing step's targets. Each step restores its tab, filter, record, open chart and saved focus without editing any of them, and every record, chart and graph node it points at is bound to what it was saved against: played against another run, a changed record is shown as not available instead of being lit somewhere else, and a callout is marked (historical) or (unresolved) when it no longer describes what is under it. A click elsewhere, or a change of view you make, ends a walk; Play from step… resumes it. Settings export now shares "Investigation reports and spotlight walks". A walk never claims more than it can show:
• when the log's file changes after it was read, the walk's record targets are unresolved and go out;
• a step whose record its own filter hides is not shown;
• a step that names a record outside the log is refused whole, and the previous step stays on screen;
• a step without a filter resets to the defaults, not to whatever you had selected;
• each callout carries the number the strip states;
• a saved record whose raw-text representation is unknown or differs is not lit, even when its text digest matches;
• a step that points at the topology canvas is not shown when the record it names is not the one selected — the canvas follows the selection, so lighting it would point at a different record;
• importing settings over a playing walk ends it with the reason, exactly as saving over it does, and switching project away from it ends it too;
• on a log whose file has never been re-checked (no Follow, or a reader that cannot say), a step resting on records or charts says so, because current there means unchanged since the step was saved, not unchanged on disk.

Changing a walk while it plays ends it with the reason, and renaming it keeps it playing. Adding to a walk whose chart steps were saved against another run is refused rather than silently moving those charts to this run.

Fixed

• A cleared topology no longer leaves its file's digest behind. The topology panel's graph digest survived Close graph and a graph supplied by the log's source, so session recovery could record the previous graph's digest beside a different graph. The digest now belongs to the graph on screen, and is absent when none has an established identity.

[1.25.0] - 2026-09-27

Added

• graph {name, delete: true} — remove a chart's definition over the socket, the way Graph ▸ Delete chart does on the desktop. An investigation's probe charts used to persist, reopen with the project and sit in the tab strip looking like findings, and an assistant had no way to clear them up. Unlike a report delete it is not recoverable. Deleting the last chart no longer saves or reports a blank placeholder as a chart that remains.
• graph {name, close: true} — take a chart off the screen and keep its definition, the way Close graph has always worked on the desktop. Over the socket, delete used to be the only way to get a chart out of the way, and it is not recoverable; the safe half of removal is now available to an assistant putting away a chart worth keeping. It is refused while chart definitions are withheld as ambiguous, because nothing is saved then and a close would discard the chart.

Fixed

• Updating a report no longer empties it. report {name} without sections meant empty, so retitling a report, adding a note or adding a path to render its PDF silently destroyed every section. That was not recoverable, because a replace never entered the recently-deleted list. Now a call changes only what it names. The sections are kept, together with the log and view they were written against, and the reply says so (sectionsKept, writtenAgainst). sections: [] still empties a report deliberately.
• Re-sent chart notes replace, instead of doubling. notes on graph appended, so re-sending a definition to adjust one thing doubled every pinned note, and the plot collapsed the duplicates into one badge that hid it. Notes now replace like every other collection parameter. A re-send that drops notes names them in the reply.
• Pinning a note no longer erases the chart's time axis. The x-axis labels were drawn at the component's bottom edge, inside the explanation footer, whose background covered them.
• A step chart's closing value holds to the edge of the window. The last step stopped dead at its own point, so a window pinned past the last record looked half empty, which reads the same as no data. A line chart still stops at its last point, because it interpolates only between points it has.
• Spotlight callouts on neighbouring lines no longer cover each other. Callouts pointing at consecutive source lines could run out of places and draw one box across another's text. They may now slide along their side; a single spotlight lands exactly where it always did.
• An explicit null no longer wipes a report. report {name, sections: null} — what a client sends for an optional argument it has no value for — read as "replace with nothing". Absent and null now both mean unchanged; sections: [] is still how a report is emptied deliberately.
• A replaced report is dated when it was replaced. Rebuilding a report's sections kept the original created date, so an exported PDF showed CREATED and WRITTEN AGAINST describing two different reports. The date now travels with the sections, as the log fingerprint already did.

[1.24.0] - 2026-09-27

Added

• The window title names the profile, not just the project. With several profiles in one project, the title now says which one is active (project — profile), so you can see which profile your edits are saving into.
• A project can say where the assistant's exchange directory is. A project-relative assistant.exchangeDir places exported reports and screenshots beside the evidence they are about. It can only choose WHERE — turning the exchange on stays a setting on your machine — and it must be a directory inside the project: an absolute path, .., or a link that leads out of the project is refused, and the machine setting is used instead.
• A workspace anchor can be declared from Settings ▸ Source roots. The Project panel told you to declare one and there was no control to do it. Each choice says where it lands and how many of your roots it makes portable.
• Reports can be deleted, renamed and restored. Delete… and Rename… sit in the Reports tab, and the assistant can do the same. A delete is recoverable: Restore deleted… brings the report back into its project. The last 20 deletions are kept on this machine, never in the project profile.
• Report actions remain visible in a narrow sidebar. Delete and Restore no longer wrap out of sight at the default window size. Restore accepts any report name, including true, and refuses a request that mixes restoration with another operation. Boolean true lists deleted reports.
• Exchange paths cannot escape through nested links. Reads and exports check the actual destination inside the exchange directory, including linked subdirectories, rather than only the configured directory itself. Absolute paths through an alias of that same directory remain usable; unresolved symbolic links are refused with an explanation.
• aggregate says where its counted records begin and end. With the total it now returns firstRecordIndex and lastRecordIndex, so "when did the application first log X?" is one filtered count. In virgin runs a smaller model counted breaches from a window it had read, and named the first value over the limit as the first breach, not the application's own RiskBreachEvent. The tool descriptions now say that records read are only a sample, and that breach_count counts the application's breach flags, not values over a limit. Earlier wording that defined a "first" occurrence in prose was dropped: in two 10-vs-10 trials it lowered correct first-breach answers. The aggregate description shows the filtered call itself ({metric: count, filter: {dimensions: ["<EventName>"]}}), because models that made that call received the first record and answered correctly.
• Spotlight guidance points menu answers at the item. The spotlight description asks for the item (menu:<Menu>:<item>), not the menu alone, the status bar or an unrelated toolbar button. Its example targets no longer suggest the status bar or the Flag button, the two a model lit for menu answers.

Fixed

• Follow reports a time-order violation it appends. A record that arrives out of order while a log is followed now reaches context.timeOrder, the status line and the time-order caveat on time-anchored answers. Until now the order was checked once, when the log opened.
• The Follow status line says what the load line says. It keeps the log's provenance and the time-order warning instead of dropping both on the first tick, and a failed Follow read is added to the line rather than replacing it.
• The status line never mixes two revisions of a followed log. Its record count and its findings always describe the same content; between a poll and the rescan it asks for, the line keeps the previous revision's whole line.

[1.23.0] - 2026-09-26

Added

• A report says what the log itself shows. An empty file, a damaged one, or one holding a document with no record key is now stated on the exported PDF, on the Reports tab and in the report reply, under Log findings — the same findings the status bar and context give, damage first. A report made over a clean log is unchanged.
• A report's coverage table carries level changes. A node the log set to a quiet audit level is annotated on its row (levelChange) and in the notes under the table, exactly as the coverage verb returns it. It stays uncovered and in the ratio: the annotation explains, it never excuses.
• Six conformance fixtures for the producer findings (C25–C30): four empty-file shapes, a complete file that still holds a corrupt document, and a per-node level change. Both reading paths must now agree on the producer findings for every fixture, not only on the records.

Changed

• The empty-file message reads the same however the file was opened, and fits what was read. A file that may still be written says "No records in this file yet."; a file whose own stream-end marker says the writer finished with nothing says so instead of suggesting a buffering writer — also when that marker arrives while the file is followed, and when the file is opened as a rolled set of one; an empty rolled set of several files is named as a set. A file being followed and the same file opened cold say the same thing.
• A rolled set's completeness note is drawn as a note on the Reports tab, muted as on the PDF, rather than in the warning banner a fault gets.

Fixed

• Turning Follow on no longer hides the log's warning. The status line Follow starts with dropped it, so an empty file being followed read "watching for new records…" and nothing else until its first record arrived.

• Charts and the record detail pane now say when the file behind the log has changed. After a log was rewritten in place, the table warned that its rows were the log as it was indexed, while the charts and the detail pane went on showing values with nothing to say they might no longer match the file. Both now carry the same warning, with the same reason, above every open chart — including one opened afterwards — and above the record. Reopening the log clears all three.

[1.22.1] - 2026-09-26

• The Design view shows its XML at the default window size. The file path and note at the top of the Design view wrapped without limit in the narrow side panel and took the whole height, and the bean list kept a fixed 210 px, so on a fresh 1200×800 window the XML itself had no room. Every source:design:bean: and source:design:line: spotlight was refused, while context listed the bean. The note now keeps three rows and scrolls, and the bean list takes at most 30% of the pane (210 px when there is room; a dragged divider is kept, only capped, including a first drag after a theme switch).
• A refused design spotlight says why. It said "session design is unavailable, or the anchor is missing, ambiguous or outside the document" for every cause. It now names the one that applies: no design open, no such bean, a bean id declared more than once, a line outside the file, or a line that is not on screen.
• Neighbouring spotlights no longer draw through each other's line. Two spotlights on adjacent lines each outlined 6 px beyond their line, so each outline crossed the other line's text. Their outlines now meet in one separator halfway between the lines.
• Opening a project's file by path points you to the project. A design outside the source roots was refused with a suggestion to authorise its parent directory; assistants widened that to the whole project directory. When the file sits in a project that is not open, the refusal now names open {project: <dir>} first, which applies the project's own roots; the project already open is never suggested, even when it was opened through a directory alias. A relative path says what it was resolved against (and that a graphml path resolves against the analyser's working directory, not a project) and asks for an absolute path.

[1.22.0] - 2026-09-26

Added

• An empty log now says it is empty. A file with no records reads as exactly that, in all six shapes it can take, instead of opening silently with nothing in it and leaving you to guess whether the run produced nothing or the reader found nothing. A file with no index supplied is untouched — that means no index, not no records.
• A document that is not a log is named as one. A YAML file that never opens a record is reported rather than read as a log with no content. The test is the file's framing — its first non-blank, non-comment line — not a search for the key somewhere in the text, so a document that merely mentions eventLogRecord is still not a log. The warning quotes the line it found and says what a stream-end marker would and would not establish, rather than asserting a verdict.
• An uncovered node whose audit level was changed now says so. A node set to WARN still runs, but its info lines are suppressed, so it carried no entries and coverage listed it as uncovered with no explanation — while the record stating the change sat in the same file. That change is now shown beside the node, naming the records that open and close it. It is an annotation, never an excuse: the node stays in the uncovered list and in the ratio, and the annotation is read even when a filter hides the record it came from. A change applies exactly as the runtime applies it: a change naming no node sets every node, a change addressed to another processor grouping does not apply, and a later change — per-node or global — ends it. A node name is compared exactly as written, so a name with extra spaces or punctuation addresses no node. Records are matched to a processor by the grouping each declares, so one processor's change never explains or ends another's, and a record that declares no grouping is qualified rather than assumed. For records after a stream-end marker the explanation is conditional: the log does not say whether the level survived into the later run. Every explanation states each thing the log leaves open — whether the change named no node or a node literally called "null", whether it applied, whether it survived a marker — in the one condition its conclusion rests on, and the change that ends a window is described with the same care as the one that began it. A change the log does not show to have applied is described as recorded, never as having set the level; one described as setting the level applied whenever the change before it did, within the same run. A control record the analyser cannot read ends the explanation there and says so, rather than being passed over as if it changed nothing, and a change after a stream-end marker is named without claiming the level lasted until it. Every conclusion is bounded — the definite one and the conditional one about a later run alike: it speaks for the records after the change (or after the marker) and before the window ends — the next change, an unreadable control record or a stream-end marker — in the change's own grouping, never for the whole log. An annotation stops at the second stream-end marker after its change — counting every marker, so two written back to back with an empty run between them are two — and records past it get no level explanation and stay uncovered. When the records in view include some the annotation does not concern, it says it speaks for those it does. A condition in force — that the change named no node, that it applied — is carried into every sentence resting on it. A record a log reader could not read at all, but whose header names the control event, is described as that, with the time its header states; every header field the format permits is read, and nothing from the payload.
• An empty or blank file now opens by its extension rather than being refused as unreadable.

Fixed

• Follow no longer calls a replaced, unreadable file unchanged. A failed UTF-8 read retires the earlier verification and opening digest immediately, including when the file size is unchanged. The session receives that failure on the same poll, and the diagnostic identifies the displayed records as a retained snapshot.
• A byte-order mark no longer changes a verdict. A record behind a BOM lost its thread, level and logger, and because the finest level is read from there, auditLevelFinest fell from DEBUG to INFO and coverage went on to say debug calls might be missing. It was not only a first-line problem: a file made by concatenating two runs carries a mark in the middle, and every record behind it was affected the same way.
• A BOM before a file's first --- stopped it separating, so the head of a healthy file ran together; a file containing only byte-order marks framed as one record instead of reading as empty. A BOM counts only at the very start of a file: an event value containing a BOM-prefixed --- line and marker lines is not a record boundary and cannot make a log read as complete. Two BOM'd files concatenated therefore no longer separate at the join, and the missing-separator warning says so.
• Following a growing log no longer fails when a poll lands inside a multi-byte character; the rest of the character is awaited, and until it arrives the log does not claim to be complete. Bytes that can never form a character fail loudly rather than being waited for, and the log then says its completeness is unknown until it is reopened, rather than keeping the verdict it had before them. That failure is reported as damage to the source, first among the log's findings, and reaches the assistant and the status tooltip even while the file keeps growing. If the file is later replaced by a longer readable one, it is reloaded rather than read as though the new content had been appended; a replacement of the same length is not detected, and the log keeps saying its completeness is unknown until reopened.
• A healthy record read through the binary reader is no longer reported as missing its record key.
• An event whose name merely resembles the framework's own control event is no longer counted as one.

Changed

• Record lines are trimmed of ASCII whitespace only — space, tab, CR, LF — matching the format specification and the rest of the reader, where one path previously trimmed every Unicode space. A line indented with an ideographic or em space is no longer trimmed to its content, and the fields on it are lost. YAML permits only the space character for indentation and no known producer emits one.

[1.21.0] - 2026-09-26

• Source panes show the file as it is now. After a regeneration, the Topology tab's source pane kept showing the old processor, even after the log was reopened; and after renaming a node's class, opening that node could still land on the removed class. Every navigation now re-reads the file, reopening a log or graph rechecks both source panes, and node navigation uses the processor as it is on disk. Reads run in the background with a visible "not yet rechecked" state, so a slow disk no longer freezes the window.
• Source panes no longer let an older read win, and a hung disk cannot pile up threads. With both source panes reading the processor, an older read that arrived last could put back the pre-rename model; the later request now always wins. context and the design view no longer read the processor on the window's thread (they say "not yet read" instead), Ctrl-click checks existence in the background, a timed-out read says so in the pane body, a node that could not open says why, and at most two threads wait on a hung disk. Superseded reads and Ctrl-click checks are removed from the queue rather than left waiting, and a Ctrl-click check has the same time limit as a read, so a stale click can no longer open a class long after it was made.
• A Ctrl-click that cannot be checked now says why. If the source reads are backed up or the lookup fails, the node pane explains it instead of silently doing nothing, and Back, opening a file, a spotlight or opening a node all cancel a pending Ctrl-click so it can no longer open a class over what you navigated to. In Processor-only mode, a failed check opens Split view so its explanation is visible alongside the processor.
• A rolled set now stops serving a member file that changed in place, and says which one. A single log whose file was rewritten in place is refused until it is reopened, but the same file inside a rolled set was still read, through an index that no longer matched it. A rolled set now reports its members' changes, naming the member, and refuses record reads while any member's reads are suspended. A log whose reader cannot check for changes (a plugin reader) now says identity: not assessed in context instead of saying nothing, which read like a check that passed.
• An exported coverage table now says when coverage is refused. The report scored coverage by its own rules, which knew only one of the four reasons coverage can be meaningless. So where the coverage verb refused (a kept graph that does not describe the log, a graph built from what ran, a graph with no audit logger), the PDF still printed a ratio. The table now follows the same decision: it prints the refusal instead of a number. A qualified figure, such as one from a log captured below TRACE, carries its qualification as the first note.
• A coverage answer can no longer be credited to a log or graph opened while it ran. Coverage took the log, the session's identity for it and the graph at three different moments, and read the view filter while it scanned. A log opened in between gave the old log's comparison the new log's identity, so an id from the old file was shown as a finding about the new one. Everything coverage scores is now captured at one moment, and the scan uses that copy. If the log or graph changes while it runs, the reply says superseded, and the comparison qualifies nothing.
• A legal one-record log is no longer reported as several records run together. The check for missing --- separators counted the words eventLogRecord: anywhere in a record, including inside a quoted value. So a record that mentioned them in its own text was reported as "2 records run together". The check now looks only at lines that start a record, outside quoted values. It reports a real collapse as a suspicion, naming the lines it read and the lines that look like new records. A record too long to check in full now says it was not checked, instead of passing as clean — and so does the record still being written under Follow. Under Follow, a log that never writes a separator is noticed before its first record is complete. A file that starts with a byte-order mark is checked the same way as one without (it used to hide a collapse) (M68.3).
• More assistant requests are now honoured whole or refused whole, and a refusal no longer changes the view. This covers the calls below. Some calls still keep what worked and name what did not, on purpose: a report keeps its good sections beside a rejected one, graph applies its series, markers and bands one by one, source_root adds and removes each path on its own, and a saved analysis stops at the first failing step. Unknown keys are named, not refused — at the top level and inside items too, by path (notes[1].txt, sections[0].focuss); a section's free-form call is left to the verb it names. The full list is in the evidence-integrity spec (D-E3).
• open with a rolled set and a graph opened only the logs, and the graph was dropped without a word. It now opens both, and the graph stays. The same applies to a log with an explicit format together with a graph.
• topology checks every field before applying any. A bad scope or unknown node used to be refused after the selection had already changed. An orientation it did not know silently became top-down, and is now refused.
• topology {recordIndex} with nothing selected did nothing and replied with record 0. It now selects that record, or refuses saying why (no log, not in the log, hidden by the filter).
• flag refuses a record that is not in the log instead of attaching the finding to the last record.
• goto on an empty log is refused, where it used to answer "record 0".
• topology with saveFocusAs and nothing to save is refused before anything is applied — whether there was no focus, or the same call removes it (pop, or showAll). It used to select the node it was given, or clear the focus, and then refuse. The check runs the call's own steps on a detached copy of the view first. focus: true with saveFocusAs in one call still saves the focus that call applies.
• A refused call leaves the spotlight lit. Dropped anchors, a failed source-root removal, a rename's extra fields and misspelled parameters are now named (M68.4).
• A chart can no longer be given a name that the assistant cannot point at. graph accepted a name with a colon in it, and spotlight then could not address that chart. A name containing : or ", or exactly note or series, is now refused when it is given, and nothing is created. Charts already saved under such a name still open and still answer to the graph verb, and are never renamed for you. Spotlight reaches a saved name with a colon or a reserved word quoted, graph:"a:b":note:2, and context.graphAddresses lists that address. A saved name that contains " cannot be written as an address at all: context.graphAddresses gives it none (null), and context.graphAddressUnavailable says why. Renaming the chart gives it one. Repair names… applies the same rule to the names it gives (M68.6).
• Charts in exported reports are drawn at page size, and a section that cannot be drawn says so. A report's chart was a capture of its tab at the tab's current size, controls included. A tab that was not showing made the picture a sliver, and the chart inside it then claimed "No data under the current filter" over a series that had data. Charts are now drawn off-screen at the page's size. A plot with no room to draw says that, and the no-data sentence is kept for charts that really have no data. A requested topology section, or a chart that produced no picture, used to leave nothing on the page. It now prints NOT RENDERED with the reason. A topology section for a saved focus is now drawn in the PDF, captioned with its node count. So is a series section, drawn from exactly its stored call — the same expression, resolution (STRICT unless the call says LOCF) and filter the series verb uses for that call, so the picture shows the points the verb counts. The view filter on screen does not apply to it, and the caption says which scope and resolution were drawn. A call a single drawn series cannot carry (crossings, buckets, limit, both key and expr, an unknown key) says NOT RENDERED and why, as does a focus that no longer resolves (M68.2).
• series now refuses a resolve it does not know. Anything other than STRICT or LOCF used to become STRICT without a word; it is now an error naming the value.
• Report series calls using key preserve the literal field name, including punctuation and spaces; a key such as node.value+1 is not evaluated as a formula. Formula evaluation remains explicit through expr.
• A project pointer that cannot be followed now says which project root it tried. A runbook or glossary pointer whose file was missing said only "NOT found under the project root", without saying which root. One that could not be resolved at all showed no warning: either no project was open, or its path left the project folder. The Project panel and context now state the failure. A missing file names the root and the path it resolved to. A path outside the root is refused, naming the root. With no project open, the pointer says there is no root to resolve it against. An environment's log directory and a report destination directory are checked the same way; a remote destination (s3://, https://) says it is not checked rather than passing as fine (M68.5).
• The assistant's record verbs now check whether a log you are not following has changed on disk, before they answer. Before, read, aggregate, series, coverage and the others kept serving rows after the file changed. For a large log read directly from the file, an in-place rewrite could make those rows describe bytes that were no longer there. Such a rewrite now suspends those verbs until the log is reopened, and says why. A file replaced at its path is labelled superseded, because what is shown is still the file that was opened. context.log.identity reports it, and so does the status line when you return to the window. The log table now shows the same warning above its rows, saying they are the log as it was indexed. A log opened through a plugin reader is not checked (it says not assessed) (M68.5).
• Follow now notices when the file it is following is replaced, not only when it shrinks. A file rewritten at the same length used to count as "no growth" and was ignored. A rewrite in the middle combined with an append was indexed as an append, over records that had changed. Follow now compares every byte already read. If they changed, or a different file now has the path, the change is announced and the log is reopened; context.log.identity says why. A file that cannot be verified says so rather than passing as unchanged (M68.5).
• Opening a log and a graph together no longer loses the graph. open {log, graphml} replied ok, then closed the graph a moment later if it did not fit the log. The analyser treated a graph opened while the log was still loading as left over from the previous investigation. A graph somebody opens after asking for a log is now kept for that log, and the mismatch is announced instead. A graph that was already open before the request, or one a log's reader supplied, is still closed when it does not fit (M68.4).
• A coverage run now carries the log and graph it was made against. Coverage scans the whole log in the background, then records what it found against the graph's pairing verdict. If another log or graph opened while the scan ran, the result was recorded against the new pair, although it described the old one. The comparison now carries the pair it names, and the analyser refuses to apply it to any other (M44.4c). That was only sound once the log itself was captured at the same moment as that identity — see "A coverage answer can no longer be credited to a log or graph opened while it ran", above.
• Following a live log no longer pushes the analyser's own session history out of its audit record. Each Follow update is kept in a small ring of its own. Opening, closing and switching projects therefore stay on the record however long a log is followed. The export interleaves both rings and says how many of each it dropped. The coverage verb also reads the session's verdict as of the last completed step, and no longer waits on the window to refresh it (M44.4b).
• Coverage now refuses a graph that a log's reader inferred, as it always said it would. The session tracked a graph by its file, and a graph supplied by a log's reader has none. So the analyser's session believed no graph was open while one was on screen. The refusal to score coverage against a graph built from what ran therefore never came from the session. Internally, graphs and log closes now reach the session as facts at the place they happen, not through a menu-refresh observation that skipped any change made mid-operation (M44.4a).
• Spring authoring scripts are installed runnable. A Spring template installed from the analyser left setup.sh, validate.sh and generate.sh without the execute bit, so ./setup.sh failed with "permission denied". They are now made executable with the other bundle scripts. Archive file modes are still never trusted.
• A recreated project is not offered the old project's session. Recovery was keyed by the profile's path, so deleting a project and recreating it in the same place showed the previous project's "Restore last session" offer. A project profile now carries a random profileNonce, written once when the analyser creates the profile and kept by every save; each saved session records the nonce of the profile that captured it. A different profile at the same path gets no offer (capturedBy: "different profile at this path"), and a session or profile without a nonce is withheld as capturedBy: "unknown" rather than guessed. An existing profile gains a nonce the first time the analyser saves a change to it, so that profile's committed file gets one new line. A withheld offer still names when it was captured and what it would have opened, and the offer message now names its capture time.
• An assistant can ask context for only the part it needs. context {sections: ["pairing"]} returns just those sections — log, project, pairing, processors, source, topology, view, charts, menus, design, handoff — and skips the file reads and lookups behind the rest. Each section is exactly what the full context would say, and any warning that qualifies it (a load still in flight, producer faults, time disorder, a partial dispatch order, and a rolled set's file list with the view's file-local offsets) comes with it; a scope block says what was selected and carried. An unknown or empty list is refused. Calling context with no sections is unchanged.
• A downloaded template's source folders must lie inside its own project. Installing a template now refuses it if any of its project settings files — the main one, a named one, or a nested module's — names a source folder outside the project, the whole project, a home, absolute or Windows-style path, or a workspace anchor. Nothing is installed when that happens. This covers source folders only: a template's saved charts can still name external CSV files, which are not checked. Projects you set up yourself can still point at folders outside the project, such as a neighbouring module.
• After a failed log open, the Topology tab again says the graph was not compared.
• A graph opened with no log says it was not compared. The Topology tab now states that nothing is shown as matched or executed, instead of saying nothing, so a design-first look at a project (design, topology and Java, no log) cannot be mistaken for a checked one.

[1.20.1] - 2026-09-24

• An assistant can find its way around the new menus. Pointing at a menu item that is not where it was asked for now says where it is — "'Follow (tail)' is in the Audit log menu" — and a renamed item names its new name: File > Reset is now Project > Close log and topology. Asking for the old File menu says what replaced it. The assistant's context also lists every menu and its items, so an assistant can read where an action lives instead of guessing.

[1.20.0] - 2026-09-24

• The README names both Export settings and Import settings under Project for sharing setups.

• Project, Sources and Audit log now have separate menus. Project holds profiles, saved analyses and settings; Sources holds source configuration, topology, design and producer diagnostics; Audit log holds acquisition, Follow and record export. Source settings shortcuts open their named page. Records, Theme, AI and Help remain separate. The toolbar and assistant verbs keep their existing behavior. Reset is renamed Close log and topology and moved to its own group in Project. Recent logs and topologies sit beside their open actions. Saved menu:File… spotlight steps are now refused; update them to the new visible menu names. Guides and screenshots follow the layout.

[1.19.3] - 2026-09-24

• An action that cannot happen now says so. Open on a saved chart that cannot be shown used to bring the Graph tab forward and then do nothing at all; it now explains why — no log loaded, duplicate names withholding the definitions, or no such chart — naming the chart. Nothing is said when it succeeds.

• Zoom and pin now say which one keeps its window. Both set the visible time range and sit side by side, but a zoom is a view and is forgotten, while a pin is saved with the chart and comes back on reload. The zoom controls had no tooltip at all and the pin's did not mention that it persists.

• Duplicate chart names can now be repaired in the app. A project holding two charts under one name still withholds both rather than guessing, but the Graph panel now offers Repair names…, which names each contested chart, says what it contains, and offers rename or delete for each. Nothing is chosen by default and a partial answer is refused, so no definition is removed without being asked for.

• An ambiguous project no longer blocks unrelated chart work. Only Delete is withheld while duplicate names are unresolved; New graph, Rename and Close work as usual, both in the app and through the assistant's graph action, and a chart made while names are unresolved is kept when they are repaired. Nothing is written to the project until then. Duplicates were creatable by earlier releases, so this affected people who had done nothing wrong.

• An assistant that sends a chart series in the wrong shape is now told so. graph {series} takes "instanceId.key" strings. An object such as {expr, label} used to be turned into a key that could never match, saved with the chart, and answered as a success, leaving an empty chart that looked finished. A bare string instead of a list was silently ignored. Both are now refused with the right shape named (exprs for a labelled or computed series), and nothing is changed.

• An assistant can now read back a chart's plot style. The graph reply and the context list of saved charts report each chart's style (step, line or points), and saved charts list their series as instanceId.key rather than in the internal stored form.

[1.19.2] - 2026-09-24

• Legacy global settings with duplicate chart names no longer interrupt log loading. The Graph panel explains why the whole chart set is withheld, chart actions refuse with that explanation, and all definitions are retained for manual correction. A valid project still opens normally; no chart is silently renamed or chosen over another.

• Chart imports now apply incoming definitions before autosave snapshots the old tabs. Named actions reopen saved charts with their metadata; explicit new-tab names cannot overwrite another chart. Duplicate chart names in a project/import are refused before applying it instead of choosing one. Delete confirmations cannot delete a replacement tab loaded while the question was open. Rename collisions on the action socket return a refusal without opening a blocking dialog.

• Choosing a plot style from the dropdown is now saved. Setting a chart to Line or Points from the style control kept the change on screen but never asked to be persisted, so it reverted to stairs on the next load. Only the assistant's graph {style} path saved correctly. Fixes the user-facing half of the style persistence added earlier in this release.

• Sharing or importing settings no longer resets a chart's style or reopens a closed chart. Rewriting an external series or marker path rebuilt the chart and silently dropped both.

• Closing a chart no longer deletes it. Close now puts a chart away and keeps its definition — series, formulas, right axis, explanation and pinned notes — so it stays listed in the Project panel and reopens from there, and stays closed across a reload rather than reappearing. Previously the project's saved-chart list mirrored the open tabs, so closing a tab silently and unrecoverably destroyed the chart and its annotations.

• New: Delete chart, beside Close on the Graph toolbar, for removing a chart's definition on purpose. It names the chart, says what is lost, and asks first. Existing projects are unaffected: a chart that has never been closed carries no new setting and opens exactly as before.

• Open on a Project-panel row now opens that row's thing. Open on a saved report reveals that report instead of whichever one was already selected, and saved charts gain an Open they never had — those rows previously offered no action at all. A chart that is saved but not currently a tab is opened from the profile and selected; one already open is selected rather than rebuilt, so nothing you changed since is discarded.

• A chart's plot style is saved with it. Stairs, line and points are part of a saved chart and survive a reload. Previously the choice was never written to the profile, so a chart deliberately set to line or points silently came back as stairs — the reading of the chart changed without anyone touching it. A chart saved before this release carries no stored style and opens as stairs, exactly as it did before; the first save after upgrading then records every chart's current style, so an older project file does gain a style line per chart once you save it. An unrecognised style in a hand-edited profile is dropped rather than applied.

• Coverage no longer tells you a node is missing from a graph that declares it. The analyser now reads the graph's own declaration of which nodes are framework plumbing (fluxtion.framework) before falling back to guessing from class names, so a framework class you used as a node and named is counted as yours; and it checks whether a logged node is in the graph against every declared node rather than only the authored ones, so a framework node that writes audit output no longer raises a warning. On the recovery-packet graph this changes coverage from 2 declared and 2 covered to the honest 3 and 3. coverage and the graph-open echo say how authorship was decided (authorshipBasis: declared, inferred or mixed), and framework nodes are listed separately under frameworkNodesNotScored rather than vanishing.

• Mismatch messages state the fact and stop. The three messages that used to conclude, from a node-name mismatch, that a graph and its log came from separate builds — the coverage warning, the topology match line and the pairing's own reason — now name the ids that disagree and leave the judgement to you, because matching node names do not establish which build either file came from.
• A graph with no node eligible to score now reports no ratio instead of a vacuous 100%. coverage carries ratioAvailable and a ratioNote, and membership is reported separately under membership — it can be fully established even when there is nothing to score.
• A graph kept against a log is no longer described as fitting it unless it does. The Topology panel says kept, not confirmed when the log recorded no node output and kept on a partial match when some logged ids are not in the graph; context.graphPairing and the graph-open echo add membershipEstablished, everyObservedIdDeclared, pairingScope, pairingSampled and appliesMeans, because applies is a keep-it-open policy, not proof of fit. Coverage's policy no longer returns its fullest verdict for either case, and when the pairing was judged on a sample it says which records it looked at.
• An exported report PDF now prints the notes under each table that the Reports tab already shows on screen. A coverage table used to export as "declared 3 · covered 3" with no sign that the log wrote a node id the graph does not declare, while the same report on screen said so; the page now carries that warning, the exclusions note and the audit-level caveat.
• The graph pairing now says what it looked at however you open the files. Opening a log and a graph in one call, or the graph first, used to publish a 500-record sample as a claim about the whole log — "declares all 3 node(s) this log writes" about a log that writes 4 — while opening the log first labelled it correctly. context.graphPairing, the Topology panel and GraphML discovery now all state the same scope (pairingScope, pairingSampled), and discovery's candidates carry the same membershipEstablished / everyObservedIdDeclared / appliesMeans facts as context.
• When coverage compares the whole log, it now qualifies the pairing shown on open and says so. If it finds a node id the sample could not see, context.graphPairing.qualifiedBy and the Topology panel say the sampled pairing is superseded and name the id, and the coverage reply carries qualifiedPublishedPairing. The qualification is dropped when a different log or graph is opened. When the same log grows under Follow, it is kept but marked stale — qualifiedBy.stale, with logRecordsAtComparison and logRecordsNow — and it restates exactly what it compared ("first 600 of 601 records") instead of speaking for the whole log.
• Under Follow, the graph pairing now counts appended records. A followed log that grew from 600 to 601 records used to go on publishing "first 500 of 600 records" in context, the Topology panel and the coverage claim; the pairing is now re-judged against the current log on every append, so all three say "first 500 of 601".
• Narrowing the view no longer erases a whole-log finding. Running coverage on a filter after running it on the whole log used to replace the whole-log result, so the Topology panel went back to leading with the sampled verdict the whole-log run had already disproved. The whole-log result now stays in the lead, the filtered result is stated beside it (qualifiedBy.narrower), and the filtered coverage reply says which whole-log finding is still in force. A later whole-log run replaces an earlier filtered one, because it covers everything the filter could find.
• Two coverage runs on different filters no longer erase each other's findings. Each undeclared node id a filtered run finds is kept, with the filter it was found under, in qualifiedBy.undeclaredFromEarlierFilters, until a whole-log run replaces them all; the next filtered reply says which comparison it replaced and what that had found; and with no whole-log run, the Topology panel note leads with the undeclared ids filters have found.
• A filtered comparison now remembers its filter. After you change the filter it is reported as "an earlier filter (…)" with filterStale: true, instead of going on calling itself the current filter.
• A stale comparison's fields now say what its words say: qualifiedBy.scope states what was compared ("first 600 of 601 records") and supersedesSample is false, where they used to still read "whole log" and true beside stale: true.
• Following a live log no longer fills the session's own audit record. Each append used to write one record to a 2,000-record ring, evicting the session's real history in about half an hour at one append a second; the session's view of the log is now refreshed when coverage reads it. The session's record of a log arriving also states the sample it was judged on.
• More messages state a mismatch without concluding which build a file came from: the warning on a single-record finding export, the step-through status ("N not declared in this topology"), both named-focus recall messages, the in-app help and three passages of the published docs. The help page used to tell you to treat a mismatch as a versioning fault; it now says the two files disagree about which nodes exist.
• The Topology panel's status line now leads with the pairing verdict, which it used to put fifth, where it was cut off at every window size, and the verdict leads with what qualifies it: the sample it was judged on ("first 500 of 600 records: …"), or, once coverage has compared the whole log, what that found. Hovering the line shows all of it.
• A coverage claim qualified by the pairing no longer hides the capture-level caveat: when the log was not captured at TRACE, the claim note says so as well.

[1.19.1] - 2026-09-24

• Named project profiles (.analyser/project.<name>.fluxtion-settings) now resolve relative paths from the project root, matching the canonical profile. Source roots, runbooks and other project-relative pointers no longer resolve one directory too deep.

• Docs: describe starter 1.0.74 standalone and hosted Spring authoring, generated callback audit facts, matching-version upgrades, and the version-scoped protection for dependency classes. Older-release evidence and remaining limitations stay explicit.

[1.19.0] - 2026-09-23

• Java spotlight echoes omit partial when no line band is measurable, rather than describing an invisible line as partly visible. Line numbers follow the editor, including its final empty line after a trailing newline.

• Spotlight explicit Java documents and logical lines beside topology. Ask an assistant for a Java spotlight to use this feature; opening a log does not activate it automatically. The source pane label and assistant echo disclose what is highlighted. Source reads prepare off the UI thread, expire after ten seconds, and disclose origin, text revision, first-match lookup and unverified relationship to the run. Wrapped Java lines light their visible portion with a partial qualification.

• Source viewport changes now remeasure or extinguish Java and design spotlights. Design lines keep the released wholly-visible refusal rule; Java captions go out when their bound source revision changes.

[1.18.0] - 2026-09-23

Added

• A log can say whether it is whole — audit format 1.1 §1a adds an optional stream-end marker, and context reports log.streamEnd as complete, missing_records, more_than_declared, unverified, unterminated_marker or unknown. A file that makes no claim reads as unknown, never as complete, so "this node never ran" stays a conclusion you have earned rather than one the file's shape implied. The status bar says complete when a file claims it. The marker is never shown as a record. A rolled set is never reported as complete, however many of its files say they are: a marker vouches for the file that carries it, and nothing records how many files a set should hold, so a set with a whole file missing looks exactly like one with nothing missing. A member that lost records still makes the set say so, and names the file. See Analyser assistant ▸ Is the log whole? and format specification §1a.
• A record is never removed from a log because of its own contents. An event whose toString happens to contain a line shaped like the stream-end marker is an ordinary record, and is indexed, counted and shown like any other.
• The marker must be followed by its --- separator to count. Until it is, a writer may be halfway through writing it, and the same bytes mean two different things — so an unterminated final record is an ordinary record, never a completeness claim. This is what keeps a live tail and a fresh open of the same file in agreement.
• Every number is reported in the scope it belongs to: a run inside a file, a file inside a rolled set, the whole log at the top. A verdict about one run of several names that run, and a set's verdict names the file it came from. Every run that fails its count is reported, not only the first.
• A log read through a reader plugin, or followed as it grows, shows its completeness verdict to the person at the screen as well as to an assistant.
• A log file saved with a byte-order mark reads its stream-end marker correctly.
• A file that ends with a stream-end marker its writer never terminated is reported as exactly that, and the unfinished marker is not shown as a record. Producers must write the separator after their marker; until they do, completeness is unknown rather than guessed at.
• Inside a rolled set, a verdict about one run reports the rows read and goto accept — the whole set's numbering — with that file's own numbering kept beside it.

[1.17.0] - 2026-09-21

• Context keeps loaded log sizes separate from on-disk metadata, flags changed graph/log/result inputs, and labels producer hash comparisons as of intake. The Project panel and exported reports carry the qualification.

• Design spotlights settle scrolling before measuring and refuse lines outside their text viewport; adding a distant target reports departed highlights.

• New marker conditions use same-record values by default. Explicit LOCF and older saved definitions retain carried state, disclosed in the legend, echo and PDF notes.

• Saved chart pins now state their scope and explain windows outside the new log’s series; context, graph echoes and report captions carry filters and extraction status.

• Hidden Project spotlight sections reveal before measurement; marker-only targets give an explicit unsupported-target refusal. Design-root refusals name the exact opt-in call, and PDF flag glyphs use readable text.

• Formulas accept true/false literals and typed text equality/inequality, shared by charts, series queries, markers and report row highlights.

• Topology and coverage disclose unknown dispatch hierarchy; missing supertype edges no longer shade unlogged branches as off-path. Complete invocation traces retain their stronger claim.

• Assistants can start/stop the existing Follow control with standalone open {follow: true|false}; unsupported or still-loading readers refuse.

Fixed

• Topology Show all exits every focus through both the toolbar and MCP, including calls that change scaffolding visibility. Chart windows keep left/right scales separate through pin/filter/restore.
• Chart legends reserve space beside the plot; explanations and note text sit below it. Nearby note pins combine into numbered ranges above the plot, and exported charts include the legend.
• Ordinary YAML opens include the final EOF record, consistent across heap and mapped stores. Explicit Follow reopens an EOF snapshot as a live read and withholds its tail until a complete separator arrives; a quiet interval never completes it. Missing separators do not prove damage.
• Windowed rolling-series queries retain earlier history, so delta changes at the lower bound match whole-log results and an existing threshold state is not reported as a new crossing.
• Confirmation findings use Observation / Assessment across the shared views and PDF exports. Flag categories and text survive explicit session recovery against the verified same log.
• Graph discovery reports disagreeing copies with fingerprints, node counts and missing logged ids. Opening a graph also checks configured roots and its directory, announcing conflicts without refusing it.
• Graph/log pairing includes every declared node, including framework nodes that log, consistently across discovery, the topology view and session state. Hiding scaffolding no longer changes the verdict.

Documentation

• Refresh demo, chart, topology, conversation and Spring guide screenshots; update the guides and recorded conversation echoes for pending tails, unknown hierarchy and the chart layout.
• Canonical Mongoose guidance documents snapshot audit export, the observed 1.0.43 tail/listing limits, and the live-reader delivery boundary; no polling command or completed starter update is implied.
• point-at-the-fault now opens with what to do when fixing a fault — keep the broken run's export, rebuild rather than regenerate for a body change, never edit shipped data, and no running analyser is needed — and add-a-node asks authors to log the state behind a decision. Skills index re-pinned.
• The point-at-the-fault skill now says how to write up a fault: the symptom from the cited record, the cause from the fix diff plus before/after copies of the changed sources, so a reader without the session can check both. Skills index m19-skills/2 re-pinned; the playground must re-vendor.
• Qualify C++ as a preview target and limit the reported Java/C++ equivalence to the tested graph and event stream; the vendor experiment does not establish general supplier compatibility.
• Add a vendor-integration worked example and supplier checklist, qualify composition and topology claims, and document the open dependency-shadowing and new-node audit-scaffolding gaps.
• Add a guided Spring authoring path with real project/design/validation screenshots, a client prompt, runbook and prerequisite checkpoints.
• Update the Spring guide for the verified 1.0.73 provisioning fix, with upgrade instructions and the separate hosted-template tooling limitation.

[1.16.0] - 2026-09-20

Fixed

• Session recovery returns to an actionable offer when a newer open supersedes it, and stale start-page buttons cannot answer another project's offer. Recovery rechecks are decided by the session graph; changed log sets are withheld before publication as one unit.
• Built-in YAML and rolled-log opening compute full content identities from the indexing read rather than making two extra file traversals. Full verification remains required at every file size.

Added

• Cold-start corpus check — tools/check_coldstart_corpus.py replays the reviewed scorer defects against preserved public evidence, without starting model sessions or using a key. It also fingerprints finished projects with --project and --baseline, without a journal; the corpus and observer-clock regressions run in static CI on every commit.
• Whole starter catalogue — File ▸ New project from template and the start page's authoring action show every entry, marking recommended starting points and disclosing declared build/regeneration key needs and agent entry files. Missing and explicitly empty declarations remain distinct.
• Explicit session recovery — project reopen and app relaunch offer the previous local session. Accept or dismiss on the start page or with open {restore}. Recovery verifies file bytes, restores unchanged log sets, topology, design and diagnostics, and reports missing inputs and withheld views. Remembered global log and topology paths no longer open implicitly at launch.
• Project landing and processor intent — a project opened without a log shows its own roots, runbooks, saved charts and declarations, with explicit evidence-open actions. Profiles can distinguish a declared generated type, runtime processing and an unspecified type without inventing an FQCN.
• Saved chart definitions before a log is opened — the Project panel and context.savedGraphs list saved charts separately from live graph tabs, with an explicit waiting-for-input state. An open tab does not claim that its series are bound to valid data.
• Local Spring rehearsal launcher — tools/start-spring-demo.sh provisions the existing acceptance sample, its project MCP configuration and an isolated analyser, plus playground/docs previews. tools/stop-spring-demo.sh stops its services while retaining project edits and evidence.
• Spring design on the canvas (M66) — open a session XML design, browse its bean index in Source ▸ Design, and use source for a fresh, read-only file/bean/line/class glance. Design Follow works without a log; bean spotlights follow identity across edits and qualify old captions. File ▸ Open producer diagnostics shows validation, reconciliation and compiler findings with explicit location and input-freshness limits. Bean/node/record links are navigation by name; their relationship to a loaded run stays unverified. Configured source roots now also authorise XML and producer JSON reads; opening a project does not grant access by itself. Unsupported or unreadable producer results clear the previous findings.
• Spring authoring guide — working with an LLM as a design partner, the local XML/source workflow, and the analyser as the shared canvas for evidence, charts and investigation reports. Flow diagrams and illustrative design/"what if?" conversations show how a question becomes a repeatable scenario. The guide distinguishes the published local starter workflow from analyser features and names the minimum analyser version for design and producer findings.

Fixed

• Correct the Mermaid sequence-diagram syntax on the Spring authoring conversations page.
• Keep both Spring authoring guides tied to the compiler release that includes the starter tool; the downloaded project holds the exact version pin.

Changed

• Upgrade the analyser's Fluxtion runtime to released 1.0.16 and its regeneration builder to 1.0.71. This is an owner-requested dependency update alongside M66; XML rendering itself needs no newer API.

[1.15.0] - 2026-09-18

Added

• A spotlight target can name its chart. graph:<name>:note:<n> and graph:<name>:series:<label> light a note or series on the chart called <name>, selecting that chart first; the bare forms still mean the selected chart. When the selected chart does not have the target, the refusal now names the selected chart and the charts that do, in the form to send next.
• A spotlight target can be a menu item. menu:File lights a top-level menu's open popup and menu:File:New project from template… lights one item in it: the menu opens, the item is cut out inside the window, a screenshot paints the open menu into the shot, a click on the lit item chooses it, and the spotlight goes out when the menu closes (Escape, a click elsewhere, {clear: true}). A submenu's items and dialogs are not reachable. An assistant asked "where is that setting?" can now point at it.

Changed

• The spotlight guidance every client is handed says that a call replaces what is lit unless it sends add: true, so a step-by-step walk lights its steps in one call or adds — a context-free client narrating a cycle lost step 1 when it lit step 2.

Fixed

• graph given the same external label twice in one call now draws one series, not two — and keeps it in its first slot, so with x, y, x the legend's swatches still name their own lines (the first cut of the fix appended the replacement, and the legend and plot disagreed on the colours).
• The REST /manifest now carries instructions — the same text the MCP bridge hands every client — so a REST client reads the one guidance rather than a copy. external is replace-by-label, but one call carrying a label twice drew two identical legend rows — which a spotlight then rightly refused to point at. The later entry applies and the echo's warnings says so.

[1.14.1] - 2026-09-17

Changed

• screenshot {scope: "menu:File"} says where each menu item is. With a menu open, the reply now lists the menu's items — text, whether enabled, and bounds relative to the window it reports — so an assistant, or the docs build, can point at New project from template… instead of describing where it is.
• The tutorial's first step shows where to click. The File-menu and template-picker pictures on From playground to analyser are now marked: a ring on New project from template…, then ① the bundle and ② Use this template. The marks are drawn by the documentation build from positions the app reports, and are styled unlike the analyser's own spotlight on purpose — that covers the main window, not menus or dialogs.
• The two getting-started pages now show what they describe. From playground to analyser adds the in-app route to the same bundle — File ▸ New project from template…, with the File menu and the template picker pictured — and says what the bundle had all along but the page never mentioned: it declares guided-start as one of its runbooks, so with an AI client connected "give me the guided tour of this project" is the whole prompt. Guided start gains its first screenshots (the Start page it opens on, and a spotlight mid-tour) and a closing pointer to the tutorial. Its prompt still uses the demo set, on purpose: nothing to build, no key.

Fixed

• The left column no longer starts expanded and empty. With Event types and Project both toggled off, the analyser opened with the left column as wide as it was last dragged (517 px in the report) and nothing in it — it looked like a panel that had failed to draw. Toggling the panels off while running collapsed the column to its rail; starting that way did not, because the rule was written twice and the startup copy ignored what was showing. It is one rule now, and a panel that IS showing still reopens at the width you chose.
• context no longer tells an assistant to call a verb that does not exist. While the session's posture is only a guess, context.handoff.posture.note said "set it with handoff {posture}" — but that verb was folded into open before 1.14.0 shipped, so an assistant that followed the note was refused. It now says open {posture}, and a test holds every shipped text to never showing a call to the removed verb.

[1.14.0] - 2026-09-17

Added

• An AI client can now point. It could already open, filter, select, draw and screenshot; what it could not do was say "this, here". The new spotlight {target, caption} action dims the window, cuts one named thing out, and draws a short caption with an arrow to it — so when an assistant explains a result it can light the record, the node's lines in the detail, the node on the graph, a numbered note on a chart or a series in its legend, rather than describing where to look. Targets are a small fixed vocabulary named as you would say them (tab:topology, records:row:12, detail:node:<id>, topology:node:<id>, topology:verdict, graph:note:2, graph:series:<label>, project:log, toolbar:flag, status). A target that is off screen is brought on screen first — a filtered-out record is revealed the way goto reveals one — and one that does not exist is refused with the reason, never lit on nothing. A chart series is named by its exact label: a prefix of one (graph:series:quote) is refused rather than bound to the first series that starts with it, and a label that two series on the chart share is refused too, saying there are two — lighting the first would be a guess. The caption is tagged assistant, because it is the client's words and not something the analyser established. A spotlight goes out on any click, Escape, {clear: true} or any action that changes the view, and nothing about one is ever saved. The guided-start skill now spotlights before each beat speaks. MCP clients see 15 tools — analyser_spotlight is the new one.
• A runbook that finds a fault now ends in "here it is". A new skill ships with the analyser, point-at-the-fault (AI ▸ Find skills…, and in every generated starter once the playground re-vendors): the shape of a runbook that checks a log for ONE class of fault — state it as something checkable, let the analyser scan for it, confirm it on the record, then light the record and the node that reported it, with a caption quoted from the match. Its two rules are the point: the pointer follows the evidence, never the script; and finding nothing is a result — it says what it checked and lights nothing. Its worked example runs on the demo data, and the release checklist replays it call for call.
• Several things can be lit at once, each with its own numbered callout — and assistants are now told when to point. A finding is usually a relation: this node, that record, the crossing on the chart. spotlight {targets: [{target, caption}, …]} lights up to six together; they are numbered on screen and in the reply, so the assistant's sentence ("1 feeds 2, and 2 never logged") finds its place on your screen. {add: true} keeps what is lit, {clear: true, target} puts out just one, and a set is all-or-nothing — one target that is not there, or two that cannot be on screen together (different tabs), refuses the whole call with the reason, and a call that is simply wrong (a misspelt target, a seventh, a record the log does not have) is refused before anything moves: your filter and selection are exactly as they were. A number is never reused while the set stands, even after the highest one is put out, because your chat has already said "2". Callouts are placed to avoid covering one another or another lit target where there is room; in a crowded window the least-covering side wins, and a callout can still sit over content that is not lit. The guidance point before you explain now reaches an assistant however it connects — the built-in assistant, a copied prompt, or an MCP client — instead of living only in one skill; so you can simply say "show me" or "highlight it". The user guide's new Ask it to show you section has what to say, in light and dark.
• A screenshot shows the spotlight. The screenshot action paints the window's content, which a glass-pane overlay is not part of, so a client checking what it had lit would have got an undimmed image. It now composites a live spotlight exactly where it is on screen.
• A shared canvas for you and an AI client: the session's posture, and the authoring handoff. Two things both of you can now see and either of you can set. Posture — whether this session is research/support or authoring/deploy. The analyser guesses it from what is open and says when it is only guessing; AI ▸ Posture, or the assistant's open {posture}, sets it, because intent changes before any file does. The authoring mode selector's record — the modes in force, the figures the catalogue resolved and what is left to author — placed with open {record} or AI ▸ Place mode-selector record…. Both appear on the Project panel and in context.handoff, each attributed to whoever set it. The analyser never runs the selector and does not check the record; a malformed one is refused whole, with the reason — including one whose branch is not a string, which used to be stringified into a record that looked valid. It is session state: a project switch clears it, open {close: "handoff"} takes it off again (the same idiom as open {close: "project"}), and nothing is written to the project profile. It lives on open — the action that already means put this in force — so it adds no tool; a canvas write goes alone, and combined with anything else open does it is refused rather than half applied.
• The in-app help explains posture and the mode-selector record. AI ▸ Posture and AI ▸ Place mode-selector record… were described only by their tooltips and on the docs site; the Help menu never said what posture is for, what the authoring mode selector is, or that the analyser carries its record without running or checking it. The help's LLM assistant section and the docs' Working with AI page now do.
• The built-in assistant is told about every action and every parameter. Its list of actions is the one written by hand, and it had drifted: it never mentioned open, topology, coverage, report, screenshot, context or source_root, nor seven parameters of the ones it did (a flag's fix, a chart's notes and explanation, a rolled set's file, …) — while an MCP or REST client, whose list is generated, was offered all of them. The missing ones are now derived from the same schemas, and a test holds every published action AND parameter to having a line, so the next one cannot be forgotten.

Changed

• Closing the log while one is still loading now cancels that load. File ▸ Close log, Reset and the assistant's open {close: "log" | "all"} used to close the previous log and let the pending one land anyway — you asked for nothing to be open and a log arrived two seconds later. A close is now a request the session processor hears, and one that covers the log supersedes a pending open exactly as a newer open or a project switch does: the late result is refused, the busy indicator clears at once, and the assistant's echo names what was cancelled in supersededPendingOpen. Closing only the graph leaves a loading log alone. Close log and Reset are available while a log is still arriving — including the very first load on a fresh analyser, when nothing is open yet and there was otherwise no menu command to ask with.
• The open {graphml} echo names the graph's size and its authored count separately. One key, nodes, held the AUTHORED count — 10 for the demo graph whose status bar says 20 — beside a pairing verdict, where it read as the graph's size. It is now graphNodes and authoredNodes, and open {discover} lists authoredNodes. This removes a reply key, deliberately and with no alias: there is no nodes key any more, because a key that misleads is not made safe by putting a better one beside it. A client that read nodes gets nothing where it used to get a number — read authoredNodes for the same value, or graphNodes for what it was usually taken to mean.
• Follow now refreshes open graphs (M65). With Follow on, a chart that was already open kept its old points — through zoom, Fit, a narrower time range and even an identical re-send of its definition — because the follow poll told the table, slider and status bar about new records but never the graphs. Every open chart now re-extracts as records arrive. Where the view lands is a rule, not a reset — the view moves only to reveal a point that would otherwise be hidden: a chart that already shows the new point does not move; one showing the whole log grows with it; one pressed to the live edge slides with it; one zoomed into the middle holds exactly; a pinned chart keeps its window; changing the chart's definition still resets the view as before.
• graph gains refresh: true — re-extract now, for an agent driving an analyser whose Follow is off. The echo reports refreshed: "scheduled" | false; a re-send that changes nothing (same series, same markers, same bands) now re-extracts nothing, where markers/bands used to re-extract on every presence.

Fixed

• Recalling a saved analysis over the action socket works again. Since 1.13.0 every agent open {analysis, bind} ran its steps and then failed with "SessionDriver is confined to the thread that created it": the recall deliberately runs off the UI thread, but the decision it records afterwards was submitted to the session processor from that same thread. The decision is now made on the processor's thread. Found by re-running the sample-conversations harness, which had not been run since the asynchronous open landed; tools/verify-m46-agent-api.py now checks it on the built jar.
• A log reopened at startup is no longer attributed to "you". context.log.openedBy had two values, so a fresh --rest instance that restored the previous session's log told an agent it had opened a log it had never asked for — a sibling run's, when two runs shared a home. It now says "the previous session — restored at startup, not opened in this one", and a path given on the command line says so too. The Project panel shows the same words.
• topology no longer says "no records" while records are open. Until a record is selected the step cursor is empty, and its label claimed the log was. The position now reads "no record selected — 10 open; goto {recordIndex} to step through one" and the echo carries recordsOpen.
• The slider's follow echo no longer resets an unpinned zoom on every growing tick. Extending the slider's range re-sent the same time window, and a chart re-windowed on it each second — invisible until now only because the chart never refreshed at all.
• The log store is safe to read while Follow appends to it. Series extraction and the series verb walk a bounded view (size and spans captured under the index lock, then the text); the store publishes the text before the rows that point into it. Before, a walk overlapping an append could throw on its thread, and the graph's best-effort error path swallowed it — the chart just failed to update.

[1.13.2] - 2026-09-17

Changed

• Opening a log is now a decision of the session processor, not something it is told about afterwards (M44.3). Every open — File menu, drop, Recent, socket open {log}, S3, a rolled set — is a request with an operation id; the load is the effect the processor asks for, answered Pending at once and LogOpened when it lands. A load that lands after a later open superseded it is refused and discarded instead of shown over the newer log. context reports inFlight: "opening …" while a load is outstanding, so a hung load no longer looks idle. The session driver is confined to the Swing thread and says so loudly if crossed.
• A graph is judged only when a log really arrives (M44.3a). The refresh a menu close performs used to re-judge the unchanged log and could close the graph that had REPLACED the one judged (1.13.1 review R2-F3); the judgement now fires on the arrival only, and the close names the graph it judged. A consequence a person will notice: closing a mismatching graph from the File menu no longer pops the "graph closed — does not fit" warning, which was that spurious re-judgement.
• Two more formula golden fixtures, and a stricter fixture parser. min(max(x, lo), hi) is pinned as an elementwise clamp and min(4, 2) as the number 2 — the M28 guarantee that kept a min(x, N) window overload out of the language (rollingMin/rollingMax are the windowed forms). A fixture with a doubled metadata line (expr: twice) is now rejected instead of silently taking the last (hardening N1).
• Canonical skills reworded after the 1.13.1 review (docs/skills, index m19-skills/2 re-pinned): add-a-node says constructor mapping covers eligible instance fields (static fields never; @ConstructorArg / @AssignToField opt in); run-mongoose-server says the capture ROLLS daily and nothing deletes old files, prefers a fresh capture location over deletion, and scopes the input-tailing advice to the starter's file source.
• Install page: how to pick up a newer release under JBang. The examples now include jbang --fresh … and jbang cache clear, and the page says plainly that JBang runs its cached jar until you do one of them (a user on 1.13.0 kept getting 1.13.0 after 1.13.1 shipped).

Fixed

• Three lifecycle defects in the asynchronous open (M44.3), found by its independent review. The completing operation's audience is now established before its effects run, so a person re-opening a graph from Recent while a socket load is in flight no longer makes that load's arrival modal (B1). A project switch during a pending load retires the load's "opening …" description at once, and the busy indicator, the pending pairing and a graph opened next follow the gate rather than the discarded reader — immediately, whether the switch then succeeds or fails; a no-op re-open of the active project and a bad project path still leave the load alone (B2, and its second-pass remainder). A superseded load's failure is recorded as a stale result but never shown: no status overwrite, no dialog (B3). Each is pinned on a real frame with a latch-controlled reader, with a positive control beside it.
• The session processor's exported audit snapshot reopens as one record per dispatch. The export carried no --- framing, so the analyser read fifteen records as one (pre-existing; review F4). Framed now, and the test opens the file through the real store and counts.
• Note rules stay on their points when zoomed in. The dashed rule a note pins to a moment was placed from the view bounds truncated to whole milliseconds, while the series use the exact bounds; zoomed to a window of a few milliseconds every rule drifted up to a millisecond's width (tens of pixels) right of its point. Notes now use the series' own pixel mapping.

[1.13.1] - 2026-09-16

Added

• read … fields names the nodes that ran but logged nothing. A record's projection carries tracedOnly: [instanceId…] when a wire trace marker said the node ran and it logged no value in the record, and for legacy text logs traceLikeOnly when every entry the node wrote is the tracing regime's thread/method and a method entry is present. Both are decided per node over ALL its contributions to the record, so a trace-only first contribution beside a value-bearing second one is not "logged nothing", and a lone business key spelled thread is not trace evidence (independent review B2). An empty projection read exactly like "did not appear"; the difference is now visible where the values are.

Changed

• Builds against fluxtion-builder 1.0.68 (was 1.0.67) for the fixture-regeneration profile; the default build still resolves only fluxtion-runtime 1.0.15.
• The canonical skills learned from an agent session on the template bundle (docs/skills, index m19-skills/2 re-pinned): add-a-node also covers making an existing node log its values (a body-only change, no regeneration) and states the constructor-mapping rule — derived state transient or @FluxtionIgnore — before the regenerate step; run-mongoose-server says the export is cumulative across restarts, that a stop must be confirmed clean (the registry entry has been seen left behind), and that the input file is edited only between a stop and a start; load-audit-log says the open echo cannot judge the pairing and context.graphPairing is the authority. Bundles pick the new bytes up when the playground next vendors them.

Fixed

• The Source panel no longer crashes the event thread with a StackOverflowError on a generated processor. The Java highlighter matched string and char literals with a regex whose alternation sat inside a repetition, which the regex engine matches by recursing once per character; an UNPAIRED quote — the apostrophe in a generated javadoc's "the auditor's HashMaps" — made it scan the 4 KB of file after it and overflow the stack (reported on 1.13.0 from a JBang install). Literals are now scanned by hand: a literal ends at a line end (LF or CR, even after a backslash), an unterminated one colours nothing, and a text block ("""…""") is coloured whole across its lines, an escaped triple quote inside it being content rather than its end (reviews R2-F4, R3-F2).
• A graph opened while a log is loading is not judged against the previous log — anywhere. The log loads in the background, so the graph was judged against whatever was loaded when the call ran: "no log is open" on a first open, the old log's node counts on a re-open, while context was right a moment later. The log echo now says loading: true; a graph opened while a load is in flight (same call or the next one) echoes pairing: pending; and — independent review B1 — the verdict in force retires with the log it was about, so context.graphPairing, the topology note and the Project panel say pending during the load instead of attaching the previous pair's verdict to the new graph. A load that fails restores the still-true verdict. The explicit format path (open {log, format: "yaml"}) now starts the same lifecycle: it used to bypass the load-start bookkeeping and judge a graph against the previous log (second-pass review R2-B1).
• A fresh window judges the pair when the log lands, and a socket-driven arrival never waits on a dialog. With no project opened first, nothing had built the session driver, so the promised verdict was null and context showed log and graph with no applies at all (pre-existing on 1.13.0; independent review F3). A log arriving now builds it — and the warnings that arrival raises (a mismatched graph closed) are rendered for THAT request's audience: a socket caller gets the text in the status bar, never a modal it cannot dismiss (second-pass review R2-B2; the flag used to be set only by project transitions). The audience belongs to each OPERATION, not to the session: the socket verbs close, openGraphml and selectProcessor declare it on entry, and the File-menu close/reset actions, File ▸ Open GraphML, a file drop and Open recent GraphML declare it at their entrance, so a socket close after a person opened the log no longer inherits the person's audience (reviews R3-B1, R4-F1/F2).
• A configuration refresh no longer moves the reader. Re-applying unchanged roots and processor used to scroll the processor pane back to its type declaration; an unchanged hit now keeps its viewport and caret (second-pass review R2-F5, pre-existing).
• The Source panel re-reads both panes when the source roots change, and its "No source to show" placeholder says where the roots came from. Switching project (or adding a root) with the same processor selected left the previous placeholder on screen — naming the previous project's root — while the new roots already resolved the file, because an unchanged class name was never re-read. Both panes now re-read a file whose content changed with the roots, a pane showing a miss is always re-rendered so its placeholder names the roots NOW searched (independent review F4: the node pane kept the old root when the file was missing under both), and a miss is retried on every navigation. The placeholder names the project the listed roots belong to and, when the open log sits inside a project that is not in force (a socket-driven open never shows the "Load this project?" dialog, so the offer was only a status-line note), names that project with both ways to load it: File ▸ Open project… and open {project: …}.

[1.13.0] - 2026-09-16

Added

• The binary format has a specification and a conformance corpus, and the analyser passes it. The runtime now publishes FLXA — the binary audit log format with twenty-six fixtures shipped in its jar (f01–f26); the analyser's FlxaConformanceTest reads every one through its reader, parser and tokenizer, and the format page links the two. "Reads FLXA" now means passing that suite.
• Open a binary audit log in the analyser. A FLXA binary log now opens like any other, recognised by its magic bytes rather than a file extension, so the faster record format is no longer command-line only. Requires fluxtion-runtime 1.0.15 or later. Truncated logs open too — a half-written trailing record is the normal end state of a crashed process, and the reader reports the unusable bytes instead of refusing the file. What the binary format does not carry (groupingId, thread) is left out rather than invented.
• Compare two audit logs on business outcomes, headlessly. ScoreCommand reads both logs through the shipped reader and parser and reports whether every published figure agrees after every scored event. It reads Fluxtion's natural form (book: { mid: 17.1} → book.mid) and a tagged stage/value convention — the dialect is declared by the caller, never inferred from key names, because the record format reserves none. It refuses to report a score it cannot stand behind: an event-sequence mismatch, a differing event type, an empty expectation, an expectation carrying no figures at all, a figure never published, a figure outside the contract, and any non-finite value are each a distinct verdict rather than a silently favourable number. Ten guards, one per comparison defect found in this project's history — five of them found by independent reviewers executing against the first implementation. Exit codes: 0 pass, 1 differences, 2 untrustworthy, 3 usage or I/O.
• The analyser now reads the Fluxtion compiler's graph metadata when a .graphml carries it, and says when it does not. A processor built with a recent compiler can declare which nodes are able to write audit output at all, what order nodes dispatch in, and whether an update crosses each edge — facts the analyser previously had to work out by reading source, and often could not. Coverage is the visible difference: a node the graph says cannot log is no longer counted as one that stayed silent, and that now works without the source in hand, which is the normal case for someone else's log. Graphs without the metadata — every graph produced before it existed — are read exactly as before, and every answer says which of the two it came from.
• Guided start: install and tour the analyser with an AI assistant. A new docs page carries a copy-paste prompt that installs the analyser via JBang, opens it, has you connect your assistant, and then walks you through three things by driving the UI. A matching guided-start skill runs the tour inside an already-connected session. The assistant is instructed to point at the screen rather than state figures you cannot see, to use your own log if one is open, and never to open anything over your work without asking. No key or account is needed.
• Create a playground starter without leaving the analyser. File ▸ New project from template… reads the live versioned catalogue, lets you choose a starter and identity, safely downloads and atomically extracts it, then opens its project profile. The HTTPS origin is pinned; traversal, expansion, overwrite and executable-bit attacks are refused. Downloaded code is never executed—the final dialog only shows and copies fixed lifecycle commands.
• New project now offers the setup already present in the directory. After choosing a directory, one confirmation lists detected Java source roots, skill-shaped runbooks and GraphML. Every choice starts off: discovery never silently adopts project content. Confirmed source roots and skill pointers are saved to the new profile, and at most one confirmed topology opens. An empty directory is an ordinary empty offer and can still become an empty project.
• A new project can be given a CLAUDE.md that points at the canonical Fluxtion authoring resources. New project… offers it as one unchecked box beside the source roots, skills and GraphML it found — including for an empty directory, where it helps most. Only entries marked agreed are ever written; a Spring-only link is written only for a Spring-authored project; an existing CLAUDE.md is never overwritten and you are told rather than left wondering; and the file carries links with a reason each rather than restating any rule, so improving those pages improves the project too.
• A downloaded template can be given a CLAUDE.md too. New project from template… now offers it as one unchecked box on the destination dialog. Only one of the catalogue's templates ships agent instructions of its own, so the rest arrive with nothing for an AI assistant to read; a template that does ship one keeps it untouched, and the status line says which happened.
• A generated project's skills are now selected by the template it came from. The canonical library gains an m19-skills/2 index: common is always shipped and a template names the specialisations it wants (mongoose, embedded, spring). A new spring/add-a-node skill covers adding a node to a Spring-XML graph, including the two ways that fail silently. m19-skills/1 is unchanged.
• A versioned canonical skills index for generated bundles. Build/release tooling can retrieve the analyser-owned m19-skills/1 Mongoose snapshot directly from the repository's public raw HTTPS root; tests pin the selected tiers, source revision and exact skill bytes. Generated projects remain offline snapshots and never fetch this index at runtime.
• Local Fluxtion build-key management without a first-run gate. The Start page and AI ▸ Fluxtion API key… now open one masked dialog for the established ~/.fluxtion/fluxtion.apiKeyFile, including named local profiles. The Project panel states only whether that canonical file has a configured key and documents the builder rule: a -Dfluxtion.apiKey passed to a future build overrides it; FLUXTION_API_KEY is not read. The analyser never validates the key, redisplays it, stores it in app settings, or puts it in a project, share export, action response, status message or console output.
• A loopback-only playground origin override for local experiments. -Dfluxtion.analyser.playgroundOrigin=http://127.0.0.1:PORT points New project from template… at a playground served locally. Plain http is accepted only for 127.0.0.1, [::1] and localhost; every other origin keeps the HTTPS rule, and the origin must still be bare — no path, query or credentials. The template dialog states which origin is in force whenever the override is set. It is a JVM property by design and is not reachable from Settings or storable in a project profile, because a persisted origin would outlive the experiment and travel with a shared project.
• Collect a PGO profile until the native build lands, then keep the profile. tools/bench/land-native.py. A GraalVM image lands at either ~1.6 ns/event or ~5.5, and the profile decides which: hold it fixed and four rebuilds reproduce it (1.60/1.66/1.68/1.67), while a profile that misses reproduces that too (5.71/5.63/5.61). The compiler is deterministic in the decision that matters; what varies is profile collection, which is a measurement — two collections of one workload minutes apart differ in over a thousand call-count contexts. So the harness collects, builds and measures until one lands, then keeps both the binary and the profile pair that produced it, and prints the --profile command that rebuilds it with nothing left to chance. It refuses rather than reports: arms that disagree on output are discarded unmeasured, a figure at or below the elimination floor is a deleted loop and not a result, a run that produced no RESULT line is a failure and never a zero, and every attempt — including every discarded one — is printed, so exhausting the attempts cannot read as coverage.

Changed

• The analyser builds against the released Fluxtion runtime 1.0.15 and compiler 1.0.67, and its own committed processor carries a build fingerprint for the first time. Nothing about the application behaves differently; the fingerprint is what lets a future build prove its generated source still matches the graph it was generated from. The analyser's own processor graph therefore ships with the compiler's full metadata vocabulary (the default since builder 1.0.66): which nodes are framework plumbing, which can write audit output, and the dispatch order — the same facts it reads from anyone else's graph.
• The audited event path is faster, and level with native. Measured on the release candidate (RECORDED-BASELINES M61, GraalVM 25.3.4, six-node quote engine with sparse logging): unaudited dispatch 7.5 ns JIT / 13.2 native; the audited binary record under LOW_LATENCY_AUDIT 23.8 JIT / 23.1 native, with Temurin 21 C2 at 23.2. Separately, an earlier JFR profile had put 56% of the audited path in code that resolves names which never change: an IdentityHashMap lookup per event to id the event type, while the identity table built for exactly that sat unused; per-logger key caches spread across three cache lines per entry; and a resolved-once decision re-checked on every entry. The 1.5–1.9× native deficit recorded through this work was never a property of the toolchain — it was the JIT speculating through pointer-chasing that closed-world compilation has to execute. Interim figures along the way (M40–M60) stay in the baselines file as history; these notes carry only what the release candidate measured.
• An ordinal audit-key API, added during this cycle, was dropped before release. It let a code model replace auditLog.info("v", v) with an indexed call, and measured a real gain — against the data-structure fault above. With that fixed it was slower than the plain String path on both toolchains. No API that shipped in 1.0.14 is removed or changed in 1.0.15; node code stays auditLog.info("v", v).
• The analyser's own session audit log now shows where each transaction closed and its effects ran. Effects the session processor decides on — closing a log, applying a project, showing a warning — used to be carried out entirely outside the audit record; only the decision appeared. They now run at the processor's transaction boundary, so opening the analyser's own log shows the decision, the boundary and each outcome in one sequence. No change to what the application does.
• Project transitions are now decided by an auditable Fluxtion processor, and the analyser can open its own audit log to see why. The rule that a project switch closes the log and graph — and the exceptions that it must not, when the project is being adopted because a log was just opened, when the load failed, and when the project is already active — lived in three places in the UI and could only be checked by running the app. It is now one decision graph with a replayable record, which distinguishes what was decided from what actually happened: asked to close and closed are separate entries. Behaviour is unchanged, with one improvement: open {close: "project"} now reports what it really closed rather than what it predicted before closing it.
• coverage now refuses to print a number in two more cases where it would have been misleading, and qualifies it in a third. It already declined over a graph inferred from what ran. It now also declines over a graph whose processor was built without audit logging — every declared node would read as never logged, blaming the nodes for the build — and over a graph you deliberately opened against a log it does not describe, where the denominator belongs to a different system. Keeping that graph on screen is still right; scoring against it was not. And where the log was captured below TRACE, the number is still given but now says what it hides: a node may have run, logged, and had its output discarded.
• When one open call names several things at once, the reply lists what it did not honour in the order they would have been honoured — largest act first — rather than in an arbitrary order. The same parameters are reported; only the ordering changed, and it now tells you which act won.
• The runnable jar is about 1.2 MB larger (2.43 MB → 3.64 MB), which is the Fluxtion runtime and its one transitive dependency. Building the analyser still needs no Fluxtion API key and no compiler.
• The playground-to-analyser tutorial now describes the released bundle honestly. It uses the real catalogue entry and paths, opens project/GraphML/log as separate session-safe actions, distinguishes the fixed Chronicle export from a followable file, and adds generated anonymous screenshots for records, source navigation, graphing and AI-client setup. It also names the remaining starter gap: the current audit records contain no numeric business value to graph.
• The canonical Mongoose skills now describe the real Chronicle export beat. They discover the running server through its registry entry, run the generated project's own YAML export command, then open that concrete export with GraphML. They no longer imply that starting Mongoose directly writes an analyser-readable YAML file or that a deployment descriptor reveals one.

Fixed

• A click in the record detail selects the entry you clicked, never a different property's series. The exact-key click used to extract an identifier from the displayed line, so the reader's @unkeyed: 42 marker resolved to a property named unkeyed, a click inside the quoted key "price: adjusted" or the dotted key desk.price resolved to price, and a click inside a string value containing price: 42 did the same. The parser now records each key's complete source position with its parsed entry and the panel resolves a click only through those positions; a click on a value, on a marker, or on a multiline item without an exact position falls back to the node's named-key menu. The Logical view's right-click resolves its node through its own layout rather than the raw text's offsets. Logical view, step status and report evidence share one formatter that keeps keyless values (shown as @unkeyed) distinct from business keys named null or "@unkeyed", while legacy text values are shown exactly as written.
• A report table read binary evidence under the wrong grammar. The table assembly re-parsed each record's text as legacy YAML instead of taking the store's parsed record, so a binary log's quoted business key "@invoked": 123 became the trace marker's true, a keyless value became a named field @unkeyed, and a string value kept its quotes. The table now reads the same parsed record the read verb does, and a test holds the two to the same answer.
• A logged String or char can no longer pose as a figure, a null, or another record's identity. The binary reader wrote string values bare into the record text it constructs, so "ok, price: 42.0" read as a second entry carrying a number the producer never published, a logged ' character deleted the entry after it, and a value with a line break could rewrite eventType. Values that would be syntax are now written in a quoted form the format specification defines (§3a, fixture C16), which the reader declares through the plugin SPI (AuditLogReader.textEncoding()) and the tokenizer then decodes losslessly: a quoted scalar is a string whatever it spells. Nothing in the text selects a grammar, so a text log — every log the text runtime writes — is read exactly as before, including a multiline value whose middle line is spelled like a field (fixture C17).
• The record diff compares values by kind, and numbers exactly. A number 42.0 and the string "42.0" were SAME; they are now CHANGED, and the kind is shown when it is what differs. Two different logged longs above 2^53 are CHANGED, not narrowed to one double.
• A keyless value is unnamed evidence, never a figure. A value logged under a null key reached the scorer as the business name null and could hide a change under a real key of that name; it reached the topology's graph menu and threw. It is kept beside the node and offered to nothing that addresses a name.
• Trace provenance is node metadata, and completeness is never inferred for a binary log. The binary reader's trace marker is a reserved key only it can emit bare (@invoked), carried in the model as traced beside the node's entries — set only by a wire TRACE entry with key 0, never by a business property spelled invoked: true, and never sharing a value slot with a business key of the same spelling (a review had shown the diff and an agent's field read returning true where the log said 99). Nothing infers completeness from it, because the binary format carries no such declaration; the text format's method-on-every-node heuristic applies only to records the text reader produced.
• Scoring a damaged binary log is untrustworthy, not a PASS. The score command read a cut file through a path that discarded the reader's damage report and printed a normal PASS on the readable prefix. It now carries the report: the comparison is printed as readable-prefix only, stderr names the damage on whichever side, and the exit code is 2 — for a cut tail and for names the file never defined alike.
• A YAML export of a binary-derived log is refused. It would re-open as legacy text and change every quoted String value; the .flxa file is the lossless artefact. CSV export is unaffected.
• "Copy selected as YAML" consults the same eligibility as the file export, and declines for a binary-derived log with the same explanation.
• A binary log no longer claims to follow or to anchor by byte. Its reader declared both and the store can do neither, so an agent reading by byteOffset was addressing nothing. It declares random access by row only, and the index refuses anchoring.
• A damaged binary log says so. A cut tail, or references to names the file never defined (including String values), used to open silently as a whole log. The reader now reports them through the plugin SPI, the store keeps them, and they appear as a source damage finding in the status bar and the context echo — beside the records, never as one.
• A binary log whose time unit is wrong, undefined, or unstated is refused before a record is delivered, not after. The analyser no longer assumes a header that states no unit means milliseconds — a pre-release runtime could write nanoseconds under it. Declare the unit into a copy with the runtime's AuditLogTool --declare-unit millis|nanos --out <copy>, then open the copy.
• The neighbours of the round-6 fixes, pre-empted. The score command reads a binary log through the binary reader under its grammar instead of hard-coding the text reader; a binary file in a roll set is refused by name rather than probed as an untimed text file; a binary record whose producer did not record endTime reads as absent rather than as an instant in 1970; a dictionary id redefined in a file is reported as source damage; the scorer's verdict says within tolerance rather than identical, which is what it always was. The FLXA conformance corpus grew to the twenty-six fixtures the runtime ships, and the analyser passes all of them.
• The binary record's clock readings are the framework's, and endTime is a choice the profile makes. BinaryLogRecord took both logTime and endTime from a method chosen by a -Dclock= system property — a benchmark switch that had reached production code — and its default re-read the wall clock instead of using the reading Clock.eventReceived had already taken. logTime is now that reading, endTime a live second reading when it is recorded, and the property selects nothing. What ships: the default clock strategy is unchanged, System::currentTimeMillis (12.9 ns a call, millisecond resolution); ClockStrategy.fastEpochMillisClock() (8.0 ns) and nanoEpochClock() are opt-in projections of nanoTime that never re-anchor to a wall-clock correction, which is why neither is the default; and endTime stays on by default — 1.0.14 and every release before it emitted it on every text record — and is off under LOW_LATENCY_AUDIT and addLowLatencyEventLog, restorable with recordEndTime(true). The unit of getWallClockTime() does not change; nanosecond timestamps are the next release. Measured on the release candidate (tools/bench/latency-kit/RECORDED-BASELINES.md M61, GraalVM 25.3.4): the audited binary record under the profile is 23.8 ns on JIT and 23.1 native, against 7.5 / 13.2 unaudited; recording endTime costs a further 13 ns. Figures this project published before M61 were measured under development defaults later reverted by review; M61 supersedes them.
• Method tracing works in a binary audit log. addTrace wrote into a byte buffer that the record's length() does not describe, so a trace produced no visible bytes — and because nothing marked the record as having content, a trace-only record was never published at all. AUDITED + BINARY therefore lost every trace, silently. Traces are now ordinary two-slot entries carrying a new TAG_TRACE(8): a node id, no key, no value. Every entry stays exactly two slots, which is the property that lets a reader skip one without decoding it. A second defect surfaced with the first: the reader counted a trace's absent key as an unresolved id, the diagnostic that distinguishes a rolled file from a corrupt one, so every traced log would have looked corrupt. Tag 8 was unallocated and no log in the wild contains a trace — LOW_LATENCY_AUDIT disables tracing — so nothing that exists is broken by the change, and readers built against the earlier format report tag 8 as unknown rather than mis-decoding it. The normative format specification is updated.
• A near-miss path on the local assistant socket no longer reaches the handler. The JDK's HTTP server dispatches by longest path prefix, so POST /action/not-a-route with a valid token executed the action, and /manifest/anything served the manifest. Both routes now require their exact path, enforce their method, and refuse any request carrying Origin before anything else.
• A wrong path on the local assistant socket now says where the right one is. Requesting anything other than /action or /manifest returned a bare "no context found", so a client holding a valid URL and token had no way to learn the protocol — and /manifest, which lists every verb and schema, was itself undiscoverable. The 404 now names both routes, the request envelope and the auth header.
• The Project panel now says WHY a processor's source is missing. "Source not found" had two causes with opposite remedies — no source roots configured, or roots configured and the class simply absent — and one message for both sent half of readers to add a root that could not help.
• The canonical load-log skill now respects project switching as a session boundary. When the generated project is not already active, it opens the project first and the log plus GraphML in a second call; combining them causes the analyser to ignore the log/graph parameters deliberately.
• The REST endpoint file is now published atomically. MCP clients and startup automation no longer catch the file between creation and JSON write and mistake an empty token file for a broken analyser.
• New-project discovery now finds committed AOT GraphML under src/main/resources. This is the standard Maven location used by generated bundles; previously only Java source roots were scanned.
• Rejected Fluxtion key-profile names now wipe the submitted credential buffer. Validation failures receive the same caller-buffer hygiene as successful saves and later write failures.

[1.12.0] - 2026-08-28

Fixed

• The MCP status light no longer says "MCP starting" for ever after another analyser window closes. Two windows: the second takes the endpoint (the first correctly reads MCP elsewhere); the second closes and takes the endpoint file with it — and the first window, whose server had been listening all along, read MCP starting indefinitely because nothing ever wrote its endpoint again. The 5-second watch now re-publishes this window's endpoint when the file is missing or names a dead process, and never displaces a live owner. Found by the owner's first eyeball run of tools/verify-m43.py, which now checks the reclaim too. The light's ATTENTION states are also amber, not red, as D-AI9 promised — the previous colour was the ⚠ brick shared with the Project panel, which reads as red in a one-word status label.

• A refused table highlight rule no longer prints its legend on screen. When a rowWhen is refused because the table's rows have no records (aggregate buckets, coverage rows, series buckets/stats), the Reports tab no longer prints "… — rows where …" under the un-highlighted table; the screen now matches the PDF, which was already correctly silent.

• A secret on the analyser's own command line can no longer reach an AI client's config file. The MCP bridge command reconstructed from a java -jar launch copied every JVM option before -jar, and that command is written into the client's own configuration (claude mcp add, Codex) and offered for copying — so starting the analyser with something like -Dhttp.proxyPassword=… put it in a file a user may sync or commit. Credential-shaped options are now refused by the same gate that already guards report destinations, using one shared definition rather than a second copy.

• Report table calls now retain nested parameters after restart. A report can store and reissue structured filter, crossings, and similar call values without flattening them into Java display text; existing scalar-only report settings remain compatible.

Changed

• Docs: the AI menu is now the documented route to connecting an LLM. Connecting an LLM leads with AI ▸ Connect an AI client… (reachable mid-session, unlike the Start page), shows the menu, and adds a table for reading the status light — including MCP elsewhere, the state that quietly wastes time when two analysers are open. The Project panel page notes the runbook description and where pointers are edited.
• The reports guide and recorded AI conversation now cover every table source. The guide explains aggregate, series, coverage, structured call persistence, scalar context, and record-only highlighting; the regenerated “Chart it and write it up” conversation proves an aggregate table and CSV export from a real isolated run.

Added

• Find skills… — the runbooks your project already has. AI ▸ Runbooks… can now scan the project for skill-shaped files (SKILL.md with name/description frontmatter) and offer them, instead of making you go and find them with a file chooser. Ones already declared are shown greyed rather than hidden. Finding is not adding: picking one opens the same confirm step a hand-typed pointer goes through. The scan never leaves the project, skips build output and vendored trees, and names its own cap when it stops early.
• An AI menu — everything an AI client needs, where you can find it. MCP setup used to live in Settings and on the Start page, which only appears via Help ▸ Start page — so it was invisible mid-session, which is exactly when you think to connect an LLM. The menu gathers Connect an AI client…, the local MCP/REST toggle, the report exchange directory and the docs, and adds the one thing that had no UI at all: Runbooks… and Domain glossary… for the pointers a project declares. An item whose precondition is missing is disabled with the reason in its tooltip rather than opening a dialog that explains itself after the click.
• An AI status light in the status bar. It answers one question — would an AI client asking right now reach this window? MCP ready (green) means the analyser is reachable, deliberately not "connected": whether a client is actually talking to it is a separate fact. MCP elsewhere (amber) is the one worth building it for — another analyser window owns the endpoint, so your AI client is answering questions about a different log. MCP off is grey, not red: it is a choice, not a fault.
• Runbooks can say what they are for. A runbook pointer now carries an optional one-line description, served in context.runbooks[] and shown on the Project panel row, so an AI client can choose the relevant runbook without opening every file. When you point Add runbook… at a skill-shaped file, its name/description frontmatter prefills the fields — but what is stored is what you leave in them: editing that file later never silently changes what the analyser reports.
• Report tables now assemble aggregate and coverage calls. Aggregate buckets carry their metric, population, scope and truncation; coverage exports the complete graph-ordered ledger of covered, uncovered and excluded nodes with the same denominator and caveats as the MCP action. Empty tables state the verb's reason rather than presenting a blank grid.
• Report tables now assemble all three series shapes. Bucket, crossing, and whole-window statistic tables use the series verb's own result data. Crossing rows link back to their source records and retain rowWhen; ambiguous bucket-plus-crossing calls refuse explicitly.

[1.11.0] - 2026-08-27

Fixed

• A direct-JAR MCP registration now refuses debug and instrumentation JVM flags. Setup will not copy a JDWP or Java-agent option into the client bridge command, where it could bind an occupied debug port or suspend before the client can talk to it.
• Generic MCP setup shows its complete identity disclosure immediately. The JSON field is more compact and the dialog opens tall enough on ordinary screens to show the JBang, registration, and bridge names without clipping the final explanation.
• Client-registration confirmations no longer squeeze the command. Codex and Claude Code now use a wider confirmation window and command field, so the exact Java/JBang bridge vector is readable before either client CLI is allowed to run.

Changed

• Installation and MCP names are now explained consistently. JBang's installed command is analyser; the separately named MCP client registration is fluxtion-analyser. The guides now state that this is intentional, correct a stale catalogue slug in the specification, and show where each name appears.
• Working-with-AI MCP guidance now follows the in-app setup. It explains the resolved JBang or Java bridge command, confirmed Codex/Claude Code registration, and the no-token generic fallback instead of implying that every client needs a hand-written jar command.
• The connection guide now shows the setup before it changes anything. Anonymised screenshots cover Generic MCP's copy/save JSON and the explicit Claude Code confirmation, including what Cancel and a successful client command do—and do not—prove.

Added

• Connect an AI client, without a first-run pop-up. The Start page now offers Codex, Claude and generic MCP setup, with the same entry under Settings ▸ Assistant. It explains which current analyser window an agent can operate, asks before it persistently enables local transport, and can check the exact local bridge command without exposing the endpoint token. Its status distinguishes this app, the bridge, and a third-party client registration instead of calling a local check “connected”.
• Packaged launches can check their own bridge. When the documented JBang launcher is not present, setup now safely uses a directly launched single jar from the current process. Ambiguous classpaths remain unavailable rather than becoming a guessed command.
• Bridge checks retain the app’s local-home context. A packaged app launched with a custom Java user.home now passes that same setting to its bridge child, so the child reads this window’s endpoint instead of a different home directory’s stale or absent one.
• Long bridge commands are now readable. MCP setup renders the resolved command in a selectable, word-wrapping multi-line field instead of clipping an absolute Java or jar path at the window edge.
• Codex can now receive a confirmed local MCP registration. Setup detects the local CLI without launching it; an explicit check, add, replace, or remove names only fluxtion-analyser, shows the exact no-token command, and uses the Codex CLI rather than editing its configuration file. A command success is kept distinct from a connected client session or approved tool call.
• Claude Code can now receive a confirmed user-scoped MCP registration. Setup uses its current CLI with an explicit --scope user; a project-scope command is copy-only, so the analyser never creates or modifies a repository’s .mcp.json. A check makes clear that Claude Code may start its configured bridge, and a client scope can override another one.
• Claude Desktop now has an explicit, honest path. Setup distinguishes it from Claude Code and explains why no portable .mcpb package is shipped for a per-machine JBang/Java bridge. It directs to Generic MCP setup instead of inventing a second launcher or writing Claude Desktop configuration.
• Generic MCP setup now gives you the complete configuration. It renders the exact local stdio command as selectable JSON, lets you copy it, or saves it only to a file you choose (with overwrite confirmation). It never puts the per-run endpoint or token into that configuration.

[1.10.0] - 2026-08-27

Fixed

• A report's table section kept its numeric call parameters. recordIndex: 0 in a table's call arrived as a JSON double and was re-issued to read as "0.0", which is no anchor — so the table failed with read needs a recordIndex although the author had given one. Integral numbers are now rendered whole. Found while recording the sample conversations: the harness refuses to publish a failed call, which is what it is for.
• Coverage no longer counts event classes as nodes that never ran. The denominator included things that can never write audit output — event classes (data entering the graph, not code that runs) and exported service interfaces — so a gap that is a category error was reported as a low score. On the shipped demo that alone read 50% where the honest figure is 100% of what can log; it now reads 83% with the one genuinely silent node remaining. Anything left out is named, with its reason, in excludedFromDenominator and summarised in excludedNote — a denominator that quietly shrinks is the same dishonesty as one that quietly includes, and harder to notice because the number improves. The excluded list and the scored list now come back in graph order, as documented, so a reader can line them up against the Topology tab — previously both were scrambled by an unordered copy.

Changed

• The Project panel and the Event types panel follow a theme switch (Theme ▸ …) — both painted their rows, section borders and group headers from the theme at build time and kept the old colours; they now re-render on the switch, and the panel's warning colour is theme-aware.
• Project panel buttons say what they do: Show → Show file (the file manager), Go → Open (the thing in its tab). A runbook or glossary row's Open now reads the file in a read-only viewer inside the app — as written, nothing run, nothing handed to an agent — with Show file and Copy path beside it.

Added

• Docs: Working with AI ▸ Sample conversations — seven LLM → MCP → analyser conversations on the demo set, among them chart it and write it up (a report with a derived table) and deploy the fix to UAT and show me it worked (a project environment stamps the UAT log's provenance; the report header says it was matched, not declared). The asks and answers are authored; every tool call and echo is recorded from a real run by tools/capture-conversations.py (the transcript counterpart of the screenshot harness), so the page regenerates with the verbs instead of going stale. The tool calls sit in collapsed blocks — open one to see the wire — so the page reads as a conversation first. The shipped demo project now declares a uat environment over logs/uat. The compiler-generated no-audit graph is now a checked-in fixture and a real-negative test for M40.1.
• Coverage no longer counts nodes that cannot log, and says which. A class with no way to reach an audit logger is silent by construction, so its absence from a log is not evidence of anything. With source configured, coverage proves this from the class itself — and the note is careful not to sound like a clean bill of health: such a node cannot be observed in any audit log, so the ratio is silent about whether it ran rather than vouching for it. On the shipped demo this is the last remaining gap, taking it from 83% to 100% of what can log with the unobservable node named. Proof is required to exclude: no source, or a supertype the analyser cannot recognise, leaves the node counted.
• A coverage gap can be the audit LEVEL, not a silence — and the answer now says so. When nodes are uncovered, the echo carries the levels the log was actually captured at and names what that would have discarded: "every record in this log was written at INFO; if any node logs at debug or trace, those calls wrote nothing here." It states the two facts and renders no verdict — the log genuinely cannot distinguish "the threshold excluded them" from "nothing called debug()", so it does not pretend to.
• Docs: Working with AI ▸ Runbooks, glossary and saved analyses with an AI — how the in-app assistant or an LLM over MCP finds them in context, uses them (read the file yourself; the analyser never runs one), and creates them (write the file and the pointer into the repository, show the diff — there is no verb, by design). Runbook files should be written in the skill shape (frontmatter name/description, then steps) so one file serves the team, Claude Code and the analyser.

Fixed

• An older analyser no longer strips a newer one's settings on save. Both the project profile and the own-settings file used to be rebuilt from scratch, so a build that did not know a key dropped it the next time it saved — for a team on mixed versions, a committed profile silently losing facts. Writers now carry over every key family they do not own and rewrite only the ones they do, so lists still shrink and nothing unknown is lost.

Added

• Path anchors — a sibling checkout travels. A project may declare workspaceRoot=.. (or ../..): a source root or Maven repo under that anchor is written relative to the project with .. steps instead of ~/…, so ../shared-lib/src/main/java resolves on a colleague's checkout. One automatic rule, no per-path toggle. Runbook and glossary pointers keep refusing ... The Project panel shows each root's stored form and warns when a root under a project is absolute or ~ — before the profile is shared, not after it fails.
• Environments — which system a log came from, declared once per project. environment.N.name, .provenance and optional .logDir (plus environment.default) in the profile. A log opened without a declared provenance takes the environment whose log directory contains it, else the default — never a guess — and context.provenanceSource (and the Project panel) say who supplied it: declared by the opener always wins. A report written against a matched provenance says so in its header — matched by directory, not declared — so the qualification survives the session. Travels by default under an Environments share checkbox whose label names the cargo.
• Vocabulary — the project's glossary reaches the assistant. vocabulary=docs/glossary.md in the profile points at a markdown file in the repository (the same pointer rules as a runbook). Its text is served in context.vocabulary.text and placed first in every Explain prompt, so live is read the way this system means it — for an LLM on an unseen processor and for a support engineer in their first week. Travels by default under a new Domain glossary LOCATION share checkbox; shown in the Project panel.
• Report destinations — a place, never a credential. A project can record where its reports are published (destination.N.name/location: an S3 bucket, a directory, a ticket system's base URL). The analyser states it — context.reportDestinations, the Project panel's Reports section — and never publishes; the publisher acts with its own credentials. Anything shaped like a credential (user info, query tokens, AKIA…, password=) is refused with the reason at every entrance, and so are the known webhook hosts — a webhook URL is a credential in path form. The Report destinations share checkbox is off by default for that reason; the M38 share categories are now complete and documented.
• Repeatable analyses. A project can save the analysis we run every time — a named sequence of analyser verbs with its rationale and declared parameters (analysis.N.* in the profile; {log}-style placeholders bound at run time). Recall it from File ▸ Run analysis or open {analysis, bind}; context.analyses lists the offer. Steps run through the action socket's own dispatcher and stop at the first failure. A step can only be an analyser verb, never a project switch, and every path a step names must be inside the project (or a {parameter} bound at run time) — so a shared analysis can drive this viewer, read and write inside the project and the exchange directory, and nothing else; travels by default under a Saved analyses share checkbox.
• Runbook pointers — the first slice of portable context. A project can now record where its runbooks live — runbook.0.name=deploy / runbook.0.path=ops/deploy.md in the committed profile — a pointer into the repository, never the commands, and the analyser never executes it. Absolute paths, .., URLs and anything shaped like a command are refused with a reason at every entrance — the profile loader (the status bar says what was dropped and why), share import and share export. Deliberately no action-socket verb: the reviewed file is the writer. Pointers appear in the Project panel and in context.runbooks, and travel only under a new Runbook LOCATIONS share checkbox that is off by default. Docs: User guide ▸ Portable context — the three-tier rule the rest of the milestone follows.
• The Project panel — what is in force, in one place. A new Project toggle on the left rail, under Event types, states the active project (name, directory, settings file), the open audit log as you named it, the graph and its pairing verdict, every configured event processor with whether its source was found, and each source root with the tier that supplied it. Every empty section is a sentence saying what would fill it. Every button reveals or navigates — copy a path, show it in the file manager, go to the tab — nothing on it changes state; a processor whose source was not found offers Add source (Settings ▸ Source roots) instead of a Go with nowhere to go. It is drawn from the same context payload the assistant reads, so a person and an agent see one set of facts. Paths are abbreviated on screen; Copy gives the full value. The left column is now draggable and its width persists. A sixth section, Reports, shows where files leave — the assistant's exchange directory, or File exchange off and where to turn it on (the state that silently refuses an agent's screenshot) — and the project's saved reports by title.
• context reports more, and earlier. The graph is reported whether or not a log is open; log.openedBy; graphPairing.graphPath; processors (class, selected, source found, from); source.rootTiers; and source now appears on a fresh start instead of only after the first log has loaded. Also exports (enabled, dir) and reports (name, title, sections, from).
• The analyser tells you a processor will produce no audit log — before you run it. Open a .graphml and it reads whether audit logging was installed at build time: the compiler puts an EventLogManager node on the graph when addEventAudit() was called, and leaves it out when it was not. Without it a processor writes nothing — not a sparse log, none — however carefully its nodes call auditLog.info(…). Measured, not assumed: the same program with and without that one call produced two records and an empty file. The Project panel's Graph section says so outright, and the verdict rides context.graphPairing.auditLogging, so a person and an agent are both told before opening an empty log and concluding the system was quiet.

Fixed

• A project's saved graphs are no longer wiped when you open a log. With a project active, the first log opened under it replaced the profile's graphs with a single empty "Graph 1" — silently, on the next auto-save, and on every release since 1.1. Binding a log opened a placeholder tab and that counted as an edit, so the profile was rewritten from the one placeholder a moment before the saved graphs were restored from it. A placeholder is structure, not an edit; the saved graphs are also snapshotted before the log binds. Reported as "opening a project with no log destroys the graphs" — that path was checked and is fine; the loss happens on the log open that follows.
• The Topology tab's split-view EventProcessor dropdown is now populated. It is a separate SourcePanel instance from the Source tab and was never handed the processor list, so it stayed empty even when the Source tab's dropdown was full — the topology's selected processor still navigated (it shares the SourceService), but you could not switch processors from the split view. The choices are now mirrored into it whenever the Source tab's are refreshed.

Added

• A guide for answering questions about a running system — the support path, written round the questions rather than the features: what to check in the first two minutes, how to tell a quiet system from an incomplete log, the four questions that actually get asked (including "the check never fired — is that true?", which is the one a log alone cannot answer), and how to hand an answer back so someone who was not watching your screen can review it. It is marked as a draft to be corrected by the people doing the work.
• The docs say the analyser answers to people who did not build the system. Both top-level pages assumed you wrote it — the home page opened "Your processor…", and the loop reads Author → Guarantee → Deploy → Prove → Assure, which is the builder's path and easy to mistake for a prerequisite. Support and on-call arrive cold at Prove, and the record answers them anyway, because it was emitted by the processor rather than written by whoever built it. Named on the home page, in Why it matters, and at the point in the loop where you can join it.
• A "Working with AI" section on the docs site, replacing the lone The build-with-AI loop page: the loop, connecting an assistant (moved out of Getting started, so the AI story is in one place), and the support guide above. Getting started is now just install and open a log. The support guide is also linked from the top of the User guide, because most of it needs no assistant at all.
• The analyser now says what the PRODUCER got wrong, instead of opening a broken log without comment. Three mistakes, all of which used to fail in silence:
  • missing --- separators — the whole file is read as one record, so the count is wrong and every record after the first is invisible. Nothing errored; the log just looked short.
  • no nodeLogs in any record — usually a graph built without addEventAudit(), so there is nothing to read, filter or plot.
  • nothing but EventLogControlEvent — the sink was attached before the audit level was set, so the log holds the logging configuration and none of the run. Each names the cause and the fix. They appear in the status bar, in full on its tooltip, and as producer in the context echo — reported, never repaired, and never as a dialog: an open arrives from an agent as often as from a person.

Changed

• "Producing an audit log" can now be followed to a working log. The page explained the ideas and then handed the one thing you need — how to actually turn it on — to another project's documentation. It now carries a complete, runnable example, and names the three things that must all be true: the node extends EventLogNode, the graph calls addEventAudit(), and something is attached as the sink. It also documents the trap found by following it: record.toString() does not write the --- document separator, and a file without it opens fine and is read as one record however many it holds — silently. Setting the audit level dispatches a control event through the graph, so the page says to set the level before attaching the sink.
• Getting started is two minutes and needs no configuration. The Quick start told you the analyser opens Settings on a first run — it hasn't since the start page shipped — and sent you off to download a sample log, which the jar now contains. It is now: run it, click Open the demo log. Source roots, the event processor and the topology moved below, where they belong: things you do when you bring your own log, none of which block you from reading records. The start page is documented and has a generated screenshot (M36.5).

[1.9.0] - 2026-08-25

Added

• "All routes" from a node everything feeds is bounded, and says so. Focusing scope=routes on a sink — a publisher, a P&L aggregate — used to select most of the graph (198 of 309 in the case that raised it), because every route into a sink is the graph. On a large graph where routes would cover more than half, the scope now stops at 3 hops each way and the status line says exactly that — how many nodes "all routes" would have been, and that the ≤3 hops checkbox in the Topology toolbar lifts the bound. Small graphs and mid-graph nodes are unchanged. An agent gets the same switch: topology {routeBound: false} returns every route, and the echo carries routeBound, scopeBounded and a note naming the parameter — being told the answer was narrowed is no use without a way to widen it.
• A finding PDF's "where it sits" picture is readable on a large processor. Above 60 nodes the whole-estate view rendered as a grey band — checked at 309 nodes: 8% zoom, the lit path reduced to specks. The second picture now shows the cycle's nodes and their immediate neighbours, and its caption counts what was left out ("+249 nodes not shown"), so the reader sees the unlit nodes next to the path — where "the check never fired" is actually visible. Small processors are unchanged.
• Report pictures fit their labels. In a finding PDF, node boxes now widen to the longest label instead of eliding to "Category…" — on screen hover reveals the rest; on a page nothing does. The on-screen view is unchanged.

[1.8.0] - 2026-08-25

Added

• analyser --rest — start it for an agent (M19.7). Opens the app with the localhost REST transport on, so a process can drive it on a machine that has never run the analyser; on a first run it prints where the REST endpoint file is, rather than expecting anyone to read a dialog. The setting persists (Settings ▸ Assistant shows it on) and the console says so. The MCP bridge's "analyser not running" error now names this command. Human launches are unchanged.
• The analyser opens on something you can use (M36). With no log loaded, the records pane now shows a start page instead of an empty table: what the tool does, three questions a log alone will not answer, where it sits in the cycle (build → deploy → ANALYSE → commit), and three ways in phrased as the sentence you would say — I am building a processor / Something is wrong in production / I want the numbers out of this. Every action on it runs against a demo log that ships inside the jar, so nothing on the page needs a server, a source root or an API key first. It is a state, not a screen: opening a log replaces it, closing one (File ▸ Close log) brings it back, and there is nothing to dismiss. Help ▸ Start page recalls it at any time without closing the log you are working on — the page offers its own way back.
• The audit record format has a published specification (M34.3) — The audit log ▸ Format specification on the docs site: Format 1, normative, with MUST/SHOULD rules for framing, every field, nodeLogs values and attribution, what order means under a TOTAL or PARTIAL reader, what absence means, and what a reader declares. It is honest about the two things the design specifies that the record does not yet carry. Alongside it, a conformance fixture set (src/test/resources/ conformance/) pins thirteen semantics, and every fixture is run through both the built-in reader and the plugin SPI — the two must agree record for record, which is the guarantee to anyone writing an emitter or an adapter: emit these records and you get exactly what a native log gets.
• A graph built from what ran no longer implies nothing was missed (M34.2). On a topology a source inferred from its own run, every node logged by construction — so "did not run" and "not on this path" can never appear, and their absence reads as a clean bill of health. The Topology status now says so outright, and the topology echo carries the same caveat for agents. coverage already refuses such a graph; this is the shading half of the same argument.
• When a log's source offers a graph you already have one open, the analyser says so. The graph you opened still wins — a graph someone named outranks one that merely arrived — but the offer is no longer silent, so you know a second, possibly better-matched topology came with this log.
• A source that cannot promise an order no longer gets one drawn for it (M34.2). The topology's ordinal badge — the small number saying "this ran third" — is the most confident thing the canvas draws, and on a concurrent engine that number is arrival order, not causality. When a reader declares ordering: PARTIAL the badge is not painted (the node still shows that it ran, which is true, and stops claiming when), step-through says "logged 3 / 5" rather than "step 3 / 5", the Topology status carries a standing warning, and the topology echo carries orderMeaningful with its caveat so an agent reading the data rather than the picture is told too.
• An agent can open a project — and close one (M35.8): open {project: "<dir or its .analyser/project.fluxtion-settings>"} and open {close: "project"}. Until now context could report "this log sits inside a project" and nothing at the socket could accept it — an offer with no accept button — and an agent could open and close a log and a graph but not the thing that owns both. The verb applies; it does not ask: a dialog cannot be answered over the socket, so the safety lives in the echo, which names every category the switch replaced with before/after counts (source roots, Maven repos, event processors, named graphs, focuses, reports, hidden columns), what it closed and where those were (the log and graph — a project is a session boundary), the project that was active before, and the one call that puts it back. Re-opening the already-active project changes nothing and says so. context now names the project in force.
• A log source can now supply its own graph (M34.1). A reader plugin that knows its engine's structure — a workflow registry, a compiler's output — can hand it over with the log, so the topology is there without hunting for a .graphml. A graph you opened is never displaced by one that merely arrived: the source's graph fills an empty slot, and open {graphml} replaces anything. The view always names which it is, because a graph is evidence and evidence has a source. And when a graph was inferred from what ran, coverage refuses to answer rather than printing 100% — it subtracts what ran from what was declared, and there the two are the same set.
• A log can say which system it came from — open {log, provenance: "risk-engine · localhost:8081 · ~/dev/risk"}. A file name is not a system: export three servers' logs and you get three artefacts nobody can tell apart, and a report headed "written against export-1.yaml" tells a reader nothing. Provenance rides the status bar, context, report headers and PDFs, and lets the mismatch banner say "same content but a different system" — the case two servers running the same build produce, where the file name never could. Never inferred: omit it and the analyser says nothing rather than guessing. Shared reports carry it, so the Reports sharing row now names it.
• An agent-driven open can no longer strand the app on a dialog (M35.7). Opening a log that sits inside a project popped a modal "load this project's settings?" question from inside the load path — fine for a human, fatal over the action socket, where nobody can answer it and every later verb blocks behind it. Worse, the new log was already live behind the invisible dialog, so it was answerable against the previous log's graph. Opens that come from the socket now report the offer instead: it appears in context as projectOffer, and in the status bar. Opening by hand is unchanged — you still get the dialog. Loading a project is still never automatic, because it replaces your source roots, event processors, graphs and hidden columns.
• Closing a log no longer leaves a stale record count on screen. The time-range header went on reading "showing 582 of 582" after the log had gone.
• The Topology tab now says whether its graph fits the open log (M35.6) — permanently, beside the thing it qualifies: "⚠ DOES NOT FIT THIS LOG — the graph declares only 0 of the 211 node(s) this log writes". The main status bar said it at load time, but 32 other things write that line, so the warning vanished on the next filter change — while the shading, step order and coverage figures it qualifies stayed on screen.
• Switching or closing a project now closes the log and graph with it (M35.5). A project owns your source roots, event processors, named graphs, focuses and reports — swap it and all of those change underneath the open log, so keeping it meant viewing one project's log through another's settings, with focuses pointing at nodes from a graph that was no longer the right one. The one exception: accepting the "this log sits inside a project" offer adopts the settings and keeps the log, because there the project is being adopted precisely because that log was opened.
• Find the GraphML for the log you have open (M35.4). File ▸ Find GraphML in source roots… lists every .graphml under your configured roots, ranked by how well each fits the open log — node count, and how many of the log's nodes it declares. Agents get the same list from open {discover: "graphml"}. Nothing is opened until you pick one: an auto-selected graph is a graph nobody chose, and the moment it is wrong the analyser is confidently describing a system it is not looking at. Unreadable files are listed with the reason rather than silently dropped, and a scan that hit its bound says so instead of presenting a partial list as the whole answer.
• Switch processor without reopening the log (M35.3). A multi-processor server emits one GraphML per processor; opening a second one now judges it against the log you already have and says so immediately — open {graphml} returns the node count, how many of the log's nodes the graph declares, and the verdict. A graph you opened deliberately is kept even when it does not fit, because comparing one build's graph with another build's log is a real thing to want; only a stale graph, found sitting there when a new log arrives, is closed.
• A graph loaded for one log no longer survives into another (M35.2). Opening a log now re-checks the loaded topology against it: kept when it still declares the nodes the log writes, otherwise closed, with the reason and its counts — "the graph declares only 0 of the 17 node(s) this log writes, so it describes a different system or build. Reopen it deliberately if you meant to compare them." Previously the first log's graph stayed silently, and coverage, "did not run" shading and step-through all answered from it. The verdict is in context as graphPairing, so an agent can see the pairing before deriving anything from it.
• Close a log, close a graph, or reset (M35.1) — File ▸ Close log / Close graph / Reset, and open {close: "log"|"graph"|"all"} for agents. Until now the app could only ever load: there was a "Close project" but no "Close log", and opening a second log left the first log's topology on screen, so coverage, "did not run" shading and step-through described a graph that had nothing to do with the records. Closing clears everything derived from the log — records, filter, search, time slider, event checklist, summary, detail, flags, execution shading, the step cursor, follow — while profile state survives: named graphs, focuses, source roots and saved reports are yours, not the log's. Anything that can no longer resolve says so instead of vanishing, so a saved report with no log reads "written against risk.yaml · 726 records; no log is loaded" with every anchor named.
• A log now says whether its order means anything (M34.1, first slice). Fluxtion's nodeLogs order is derived by the AOT compiler, so step-through and the topology's order badges are reading back real causality. A source whose components run concurrently has no such order to report — and the M34.0 spike found the on-screen presentation identical either way. Readers now declare ordering: TOTAL | PARTIAL, the claim reaches the index, context reports it before anything is derived from position, and Settings ▸ Plugins marks a partial-order reader. Nothing changes for an audit log: text containers are totally ordered and say so. Readers published against 1.5.0 keep compiling — the capability is additive and defaults to TOTAL.

Changed

• First run no longer opens a dialog before you can use the app. Launching without a saved configuration used to announce "No configuration was found" and force Settings open — reporting the ordinary condition of a first launch as a fault, and demanding source roots and an API key from someone who had not yet seen a single record. The start page does that job now, and Settings is where it always was (File ▸ Settings, and one click from the start page).

Fixed

• The loop bench cleans up after itself when a run fails. If the analyser died mid-run the bench printed a traceback instead of a named failure and left the analyser and the stub server running.
• Reports no longer run off the right-hand edge. A report's widest line — usually the "written against … · N record(s) · " provenance header — was drawn on a single unwrappable line, and because it set the preferred width of the whole detail pane, every section under it was laid out to match and the tab grew a horizontal scrollbar. Report text now wraps to the pane at any width, and callout boxes and tables are as tall as their content instead of stretching to fill the pane.
• A rotating log no longer starts interrupting an agent. With Follow on, a log an assistant had opened would re-open as though a person had asked for it when the file rotated — so every dialog the assistant path suppresses came back, on the one workflow where nobody is watching the screen. A reload now keeps both facts about the open that started it: what it declared, and who asked.
• Two more dialogs no longer hang an agent-driven open (M35.9). open {logs: [...]} — the rolled-set verb — never carried the "this came from the socket" fact, so a set with out-of-order records put the time-order report on screen as a modal; and open {log} on one member of a rolled set stopped at the "open the whole set?" question that only a person can answer. Both are now data: the report reaches context as before, and the set offer appears as context.rolledSetOffer with the member files, so the agent decides. Under the hood the who-asked-and-what-they-declared facts now travel with each load (OpenRequest) instead of sitting in fields that one step could consume before another read them — the shape behind four earlier defects in this area, including a suppression that had never worked.
• Relative paths in a committed project profile are relative to the project, not to .analyser/. A profile that said sourceRoot.0=src/main/java was resolved against the folder the file sits in, landing at <project>/.analyser/src/main/java — a directory that does not exist — so the playground bundle contract's "open the log and everything is configured" could not have worked. The canonical .analyser/project.fluxtion-settings now anchors at the project root; a loose .fluxtion-settings file imported from elsewhere still anchors at its own directory.
• Opening a project no longer rewrites its committed profile. Every flush wrote the profile back with absolute paths and a fresh share.exportedAt, so a project you merely opened produced a diff on every teammate's machine. A project profile is now written the way it is read — paths under the project are project-relative (before the ~ rule, so a project inside your home still travels), no timestamp, no date comment — and a write that would change nothing does not happen. A diff in .analyser/project.fluxtion-settings now means a setting changed. One-off share exports are unchanged and keep their timestamp.
• The time-order report no longer interrupts an agent. Opening a log with out-of-order records from an assistant or MCP client put a modal dialog on screen that nobody at the keyboard had asked for — and which a later passer-by would find describing a log that might already be closed. The suppression added for this was reading a flag another step in the same load had already consumed, so it never took effect. The report still reaches the status bar, context, and the caveat on every time-anchored verb; opening by hand is unchanged.
• context no longer fails on a fresh start. Before any log had ever been loaded, the first context call an agent makes against a just-launched analyser threw instead of answering "nothing is open" — found driving M35.8 against an isolated, never-used config. It now answers with what it knows: which project is in force, the source roots, and an empty filter.

[1.7.0] - 2026-08-20

Added

• Investigation reports — the account of an investigation, not just its evidence. A report is an ordered list of references with connective prose, never a free-form document: findings, records, charts, focuses, derived tables and visibly-labelled narrative, built with report {name, sections}, rendered in the new Reports tab (every evidence section clicks through to what it references), exported to PDF, persisted with your profile and shared under their own category — a shared report carries prose written about your data, which deserves its own consent checkbox. Two rules are enforced, not encouraged: a finding section renders what flag wrote byte-identically (the verb cannot author or override a diagnosis), and narrative always wears its standing label so an assertion can never pass as a record. Evidence re-renders live — a stored report is a re-runnable claim — and the report captures the log fingerprint and filter it was authored under, announcing before anything renders when either differs (a renamed copy with identical content gets a softer "same content — a different file" notice). Clicking an evidence link whose record is hidden by the filter offers to widen it, never fails silently. Table rows are derived from a stored query with declared presentation; a row highlight is a rule that is printed with the table and re-evaluates strictly against each row's own record; tables export to CSV.

Fixed

• The chart's empty state no longer lies about why. "No numeric series selected" painted equally for a chart with no series and for one whose configured series matched no records under the current filter — the second case now says so and names the filter as the reason.
• A changed external-marker CSV is re-read. The per-definition cache now checks the file's modification time, so a file that changed on disk cannot leave the chart showing evidence that no longer exists; previously a stale read survived until the definition changed or a restart.
• A row-highlight rule that needs history is now refused and named, never quietly applied to a one-sample window. rowWhen evaluates against each row's own record, so a rolling window holds a single sample: mean, sum, rollingMin and rollingMax collapsed to their bare argument and highlighted rows as though the window had been computed, while the label printed under the table still claimed the window — a report stating a rule the analyser never applied. lag, delta and rate failed the other way and could never fire. All seven are now rejected with the reason, in the table, in the verb echo and in the parameter schema; compute the window with series and compare on a plain value.
• The "different log" banner names the log you are actually on. It described the loaded log using the report's own stored file name, so re-opening a report against another log announced the mismatch while naming the file you were not looking at. The verdict still turns on content (record count and time range), so a renamed or moved file is still not a mismatch.

[1.6.0] - 2026-08-18

Added

• Marker series now appear in the chart legend — each with its own glyph, in its own colour, and the number of events it holds. Previously a chart could draw triangles and crosses while the key named only the value series, so nothing on screen said what a glyph meant. A marker series that resolved to nothing still gets a row, reading (0) with the reason on hover — an event type that never fired is a finding, not something to hide. Right-click a row to remove that marker series.
• The topology index is sorted, and reads numbers as numbers. Each of Nodes, Events and Services is now alphabetical, with digit runs compared by value — so CHILL-2 comes before CHILL-10 instead of after it. On a generated estate (chillers, tills, zones) the previous graph-emission order was arbitrary, and plain alphabetical order would have been actively misleading.

[1.5.0] - 2026-08-18

Added

• External series — plot what the outside world did. An agent (or you) adapts a FIX log, GC log or venue export into a (timestamp, value) CSV and the analyser plots it beside the audit-derived series: File ▸ Add series from CSV…, or graph {external: [{path, label, time, timeFormat, zone, value, offsetMillis}]}. The clock domain is declared, never guessed (a wrong guess reverses causality invisibly; a wrong declaration is at least visible); the chart is stamped with each external series' clock and offset so a foreign line can never pass as audit evidence in a PNG or PDF; loads report rows skipped, reordered and the resolved time range, with diagnostics that never echo file contents. Verb reads are confined to the exchange directory — the renamed Allow assistant file exchange opt-in now covers writes and these reads, one switch, one directory — or to files you picked in a chooser yourself. External definitions persist with the graph and travel in shared setups (the data never does — a recipient without the file sees the rest of the graph and a note naming what did not resolve).

• Rolled log sets open as one log. Opening any member of a set (maker.log.1, date-stamped siblings) offers the whole set; the load order comes from each file's content — its first timed record — never the name, so logrotate's newest-first .1 and an incrementing writer's oldest-first .1 both load correctly with zero configuration. Record numbering is global and gap-free; byte offsets stay real per-file offsets (verbs take file alongside byteOffset on a set). Memory scales with the set total, not per member. Agents: open {logs: [...]}.

• Time order is now validated on every load — within each file and across roll boundaries — and violations are reported with record anchors, never silently repaired: a backwards timestamp is a finding, and re-sorting would destroy the evidence. While violations exist, time-anchored answers (at, windows, buckets) carry a caveat; the full report is in context and the load dialog.

• Log-source plugins. The analyser can read audit records from other containers — parquet, Chronicle, a database — through reader plugins: jars you explicitly install in ~/.fluxtion-analyser/plugins/, listed with their declared time base and capabilities in Settings ▸ Plugins. The trust boundary is stated plainly there and in the FAQ: installing a jar is arbitrary code execution, nothing is ever bundled or downloaded, and a plugin can only be a reader — it cannot add verbs. Every reader declares its clock domain up front, capabilities degrade loudly (a container without byte anchors is anchored by record index and says so), and open {format} forces a specific reader. Without plugins, nothing changes.

• Point-snapped chart tooltips. Hovering a plot now snaps to the nearest actual sample within a small radius and reads series · time · value — instead of reporting the cursor's raw coordinates whether or not data was there. A series dense enough to be decimated on screen answers its cursor column's min/max range rather than pretending one sample is the truth; with no sample in range, the coordinate readout remains as before.

• Marker series — events on a value chart. Plot buys/sells (or any event) as glyphs on a price line, each point carrying a payload (a client order id) shown on hover — and clicking a marker selects its record, because a marker is a signpost to the evidence, never a substitute for it. Three pieces: graph {markers: [{label, glyph, when, y, payload}]} — a bare key fires wherever it was logged, a formula fires where truthy; y can ride a plotted series or axis for a rug lane under the plot; dense columns render one glyph with a count badge instead of soup or silence. Payloads are display cargo only — they never enter formulas or filters. Marker definitions persist and share with the graph (never their extracted values), and a marker pinned to a series that isn't on the graph says so instead of vanishing. A marker that rides a series rides its scale too: pinned to a right-axis series, the glyphs draw against the right scale — and follow if the series is later moved between axes. Markers draw from their own palette, distinct from every series colour, so a glyph riding a line never vanishes into it.

Fixed

• The series verb could answer from superseded data, disagreeing with the chart for the same formula. A record carrying no logTime cannot be plotted, but it has still observed values — and under locf the carry is "last known value". The verb skipped such records outright, so a formula whose inputs moved on an untimed record was evaluated afterwards from values already known to be stale: on one log the chart plotted 10 where series reported 22. Both looked plausible, which is the dangerous kind of wrong for a forensic tool. The verb now updates the carry from every record and only declines to emit a point where there is no time, matching the chart exactly.

[1.4.0] - 2026-08-17

Added

• Conditional formulas — graph and series expressions can now judge their inputs: comparisons (> < >= <= == !=, Unicode forms accepted) and if(condition, then[, else]) with and/or/not. The two-argument if plots only while the condition holds — if(ask − bid > 0.004, ask − bid) draws the spread only in breach, gaps elsewhere — and an unknowable condition plots nothing rather than guessing a branch. Existing formulas are untouched (min(4, 2) still means the smaller of 4 and 2).
• Threshold guide lines — graph {guides: [{value: 0.004, label: "4bp limit"}]} draws a labelled horizontal rule on either scale, so the threshold an investigation turns on is visible on the plot instead of interpolated by eye. Persisted with the graph, shared and exported with it.
• Condition bands — graph {bands: [{expr: "ask − bid > 0.004", label: "in breach"}]} shades the time intervals where the condition held. The condition is what persists; its intervals recompute with the data by the same extraction pass as the series, so a band can never disagree with a plotted series about when the condition was true.
• Rolling-window formulas — lag(x, N), delta(x), and mean/sum/rollingMin/rollingMax (x, N) over the last N samples, in graphs and the series verb alike. Windows fill before they speak, a non-numeric sample leaves a window unchanged (so a rolling mean survives a no-quote gap instead of erasing its history — and, stated plainly, a full count window holds its value indefinitely after the last contributing sample; time windows go empty instead), and conditionals compose: mean(if(c, x), 10) is the mean of the last 10 samples where c held; if(c, mean(x, 10)) shows the all-samples mean only while c holds. Time-windowed forms take a duration instead of a count — mean(x, "5m"), rate(x, "1m") (change per minute, scaled from however much of the minute the samples actually cover, so a filling window reads the true rate rather than understating it) — and age samples out against each record's own clock, which is the right tool when record arrival rate varies.

[1.3.0] - 2026-08-17

Changed

• Topology focus is now a filter, not a toggle — and this changes two familiar gestures. Applying Focus (button or F) filters the view to the selection's scope, and that context becomes the whole graph: clicking nodes explores within it, and focusing again drills deeper. Esc steps back out one level; Show all returns to the full graph; a clickable breadcrumb (All (62) ▸ hedge path (12) ▸ …) shows where you are. Clicking empty canvas now clears only the selection and dimming — it no longer exits the focus. If a shown cycle ran through nodes the current context can't show, the status line says so rather than cropping the propagation silently.

Added

• series verb — stats, threshold crossings and minute/hour buckets over any key or formula, computed inside the analyser (assistant, REST and MCP alike). "Where does the spread exceed 0.004?" is now one call returning the exact crossing records — each with a recordIndex/byteOffset anchor for a follow-up read — instead of an agent paging raw text to do arithmetic. Crossing lists are capped with an explicit truncated flag, and filter.text is refused loudly rather than running an index-speed verb at scan speed.
• Time anchors on read and goto — pass at (epoch millis) and the analyser resolves it to the record at-or-before that moment, so "show me 09:14:03" is one call instead of estimating record indexes from record rates. Clamping to the first timed record (when at predates the log) is declared in the reply, never silent.
• Field projection on read — pass fields: ["instanceId.key", "instanceId.*"] and each record comes back as compact {recordIndex, logTime, event, values{}} rows instead of ~2 KB of raw text — a 10–50× token saving when an agent needs two numbers, not the whole record. Last occurrence per record (identical to graphing, so a projected value always matches the plotted one); requested fields that matched nothing are named in the reply. Raw text stays the default.
• Verb echoes now name what they ignored — a parameter no verb schema declares (usually a typo) is listed as ignoredParams in the reply instead of vanishing; and a graph call whose rightAxis or note names a series that isn't on the graph gets a warnings entry naming it (previously a silent no-op, discoverable only by looking at the plot). The aggregate verb's limit parameter is now declared in its schema (it always worked; the manifest just didn't say so).
• Named focuses — save the current topology context by name with a rationale (Focuses ▾ on the toolbar, or topology {saveFocusAs, rationale} from an agent), recall it by name (picker or topology {focus: "name"}), delete from the picker. Saved with the project (never the API key), shared like saved graphs (replace-by-name), and honest across builds: recalling a focus whose nodes aren't all in the loaded topology says how many resolved instead of silently dropping the rest. Agents stepping out of contexts use topology {pop: true | "all"}.

Fixed

• The share dialog now discloses that named focuses travel with the Graphs category — the checkbox reads "Graphs and named focuses" and the sharing guide's category table says what rides along (node sets and the rationale; nothing secret can — the API key is excluded by construction). Found in review: the import preview said it, the export consent didn't.
• Graphs made while a project was active were never saved to the project — and could be lost entirely. Graph edits only reached disk at exit, and the project file was always written with a stale (empty) graph list; quitting with no log open could even wipe previously saved graphs. Graph changes (from the UI or the graph verb alike) now persist as they happen, to the active project's file; every profile write captures the live tabs first; and the project is flushed on quit.

[1.2.0] - 2026-08-17

Added

• An assistant can show you where a control is. analyser_screenshot accepts scope: "menu:File", which opens a top-level menu and captures it — so "it's under File" can be a picture instead of a sentence. The window is raised first, because a screen capture would otherwise include whatever happens to be sitting on top of it.
• Project profiles. Settings now have two tiers: machine things (API key, theme, recent files) stay in the global config, while a project's source roots, Maven repos, event processors, saved graphs and hidden columns can live in a .analyser/project.fluxtion-settings file beside the project. Switching projects replaces those settings instead of piling them on top of the last project's. A profile can never contain your API key, so a team can commit one. File ▸ Open project / New project / Save project as / Close project, with a recent-projects list and the active project in the window title. Edits save to the project as you make them (debounced, so a committed profile keeps legible diffs) — including edits made by an assistant over the socket. File ▸ Import settings now asks whether you mean Merge (share a setup, additive) or Open as project (replace). Documented in Working across projects. Opening a log that sits inside a project offers to load that project — so a downloaded bundle configures itself. It asks once per log, never for a project already open, and takes "no" for an answer for the rest of the session.

Fixed

• The action manifest advertised six verbs while thirteen shipped. GET /manifest hardcoded its verbs list, so it contradicted its own schemas field and an external agent never learned that topology, open, source_root, screenshot, report, context and coverage existed. The copy-prompt handed to agents named five. Both are now derived from the schema set, with a test that fails if either is written out again.
• analyser_coverage shipped undocumented, and the assistant guide still described file exports as able to overwrite anything — which stopped being true when exports became opt-in and confined.
• Documentation screenshots can be regenerated again. tools/capture-docs.py now uses a throwaway export directory and unique capture names, so it works with the export guard rather than around it.

[1.1.0] - 2026-08-16

The topology release. The analyser now draws the processor's graph, walks events across it step by step, and is honest about what the log can and cannot prove. The whole app is scriptable by AI agents over MCP, findings export as evidence-grade PDFs — and anything that writes a file is locked behind an opt-in.

Added

• Topology tab — open the processor's build-time .graphml (File ▸ Open GraphML…, or just drag it onto the window — drop a log + graphml pair together and both open) and see the node graph laid out by dispatch layers: pan/zoom/fit, top-down or left-right, node colours for events / handlers / nodes / exported services, spacing and text-size sliders, a collapsible name index with javadoc tooltips, and a status line that counts what the filters are hiding. The topology you had open reopens next start; Open recent is split into audit log and GraphML.
  • Explore: click a node repeatedly to widen its scope (node → neighbours → routes → whole graph), F to focus, Cmd/Ctrl-click to multi-select, Show all (or click empty canvas) to reset; Fluxtion's scaffolding nodes hide behind a checkbox.
  • Act on a node: right-click to open its source, plot one of its logged values, filter every view to records mentioning it, or copy its instance id; double-click jumps to source (nested holder classes resolve).
  • Source opens beside the graph (Processor · Node · Split — the dispatch call site above, the method below) with a Sync toggle: follows your stepping, or stays where you put it.
  • Build mismatch is surfaced, never hidden: the status line names instance ids that appear in the log but not the graph — treat that as a version mismatch, not a curiosity.
• Step through an event on the graph — select a record and its cycle lights up in dispatch order; one cursor walks record → each nodeLogs row → next record with ↓/↑ (plus whole-record skip and autoplay): the entry point is marked, the path trails behind the current node, the table selection and the detail viewer's highlighted line track every step. The readout says whether you are walking logged rows or every invocation; a node that logs twice gets two steps; only edges whose both ends ran are lit.
• Execution honesty — what makes the picture trustworthy:
  • A silent node is no longer drawn as "didn't run". The graph distinguishes logged / ran but logged nothing (it was the only route into something that ran) / may have run / not on this path — with an on-canvas legend and the claim in words on hover.
  • A fully-traced log (processor built with an audit level) is detected, and the tab stops hedging: absence becomes did not run, and the legend says so.
  • Exported services are entry points, drawn alongside events — an operator action reads like an event, not an unexplained cycle. Re-dispatched events (processReentrantEvent) are in the test fixtures because they are the case most likely to be misread.
  • The assistant knows what silence means too: its prompt carries the two audit regimes, how to settle them from the processor source it is given, and dirty/@OnTrigger(dirty=false) propagation — so "absent" can be read as "the branch not taken" instead of guessed at.
• Diagnose one cycle, not just a trend — write a finding on a record (Records ▸ Write a finding…, or flag {note, fix} from an agent). One write site, three readers: the table's note column, a callout painted on the topology, and Export finding to PDF — the explanation and suggested fix, the full event record and node log, and two graph views (the cycle alone, and the whole processor with that cycle lit, so you see what the event didn't reach). Rendered for the page, not screenshotted; dependency-free.
• Charts explain themselves — a multi-line explanation block, notes pinned to moments in time (right-click the plot where you are reading), and a second right axis for series of different magnitude. Drawn on the plot, so an exported PNG carries them; saved with the graph; scriptable via analyser_graph.
• Node coverage — "which of this processor's nodes never logged in this run?" (analyser_coverage): compares the GraphML against the log, keeping covered / never logged / silent-by-design apart, with a separate build-mismatch signal. Born from a 309-node test estate where 54 nodes were unreachable — with every test green and nothing to say so.
• MCP bridge — java -jar analyser.jar --mcp runs the analyser as an MCP server over stdio: an MCP-native client (Claude Code, Claude Desktop, Codex) discovers one tool per assistant verb — thirteen of them — automatically, with no prompting and no copied token. The app publishes its live REST endpoint to ~/.fluxtion-analyser/rest-endpoint (mode 600) while the transport runs, so a client configured once keeps working across analyser restarts; both the legacy MCP handshake and the current handshake-free revision (2026-07-28) are spoken.
• Fully scriptable — new verbs topology, open, source_root, screenshot, report, coverage join the existing set, published as MCP tools automatically; tools/drive-analyser.sh scripts them over the localhost REST transport. Scripted plots bring the Graph tab forward so you see what the agent built.
• Node log, two ways — Logical (each node a block, values on their own lines, dispatch order) and Text (the raw audit YAML). Traced-only keys are muted rather than hidden, so a traced record is recognisable at a glance.
• Collapsible event-type panel behind a vertical nav rail (state persists); column checkboxes on the rail and the table right-click. Better code typography (best installed monospaced family).

Changed

• The launcher rejects unknown --options with usage and exit 2, and adds --help. Upgrading note: JBang serves its cached jar, so run jbang --fresh analyser@… to get this version — on the cached 1.0.0 jar, analyser --mcp shows usage instead of starting the bridge.
• JBang launches no longer print JVM native-access warnings (the catalog passes --enable-native-access=ALL-UNNAMED).
• Right-click the records table for record actions (flag, copy as YAML, diff, export) and the column chooser; Columns is no longer a top-level menu. Open from S3… is now Open log from S3….
• One shared content surface (source, record detail, topology) with hairline edges, derived from the active theme; switching side tabs no longer moves the main divider.

Security

• Documentation screenshots were replaced. The images shipped with 1.0.0 were captured against a real audit log and contained live venue, vendor and project names; they are regenerated from the anonymised demo fixture (five that could not be regenerated were withdrawn). Screenshots are now produced by a capture harness that only ever loads the demo fixture.
• File-writing verbs are opt-in and confined. screenshot and report require Allow file exports (Settings ▸ Assistant, off by default), write only inside the export directory you choose, and never overwrite an existing file — so the action socket's out-of-box promise stays "nothing outside the loaded log". The FAQ's security answer documents every mutating verb, and a test now fails the build if that ever stops being true.

Fixed

• The app could become impossible to close — a failing step in the quit sequence escaped the window-closing handler before the exit call. Quitting is now step-isolated and the exit always runs (most likely to bite when the jar is rebuilt underneath a running app).
• Topology: a node that ran was drawn faded when execution shading and selection shading compounded — evidence now wins; clicking a node no longer resets the zoom; a focused view no longer empties on click; the embedded source pane re-themes on theme change and no longer clips the graph or steals ↓/↑ from stepping; adding a source root re-runs EventProcessor inference.
• A release now refreshes the docs site's release-notes page automatically.

Docs

• User guide: Topology & step-through — reading the graph, opening a .graphml, stepping a cycle, node actions, the build-mismatch warning, and why the offline case is the one this tab is for.
• Assistant guide: "Connect an MCP client" — copy-paste config for Claude Code, Claude Desktop and Codex, a worked session transcript, troubleshooting, and exactly what an agent can and cannot reach (your client does not launch the analyser — it starts only the bridge).
• Install guide: the JBang first-run trust prompt, and that a new release needs jbang --fresh.

[1.0.0] - 2026-08-14

Added

• Assistant: read verb — fetch the raw text of N records around a record/byte anchor over the localhost socket, so an agent can seek the log without filesystem access (rate-limited).
• Assistant: per-verb JSON schemas in GET /manifest — every verb's params are now self-describing.
• Assistant: goto {reveal:true} un-hides a filtered-out record; otherwise the echo names which filter hides it.
• Graphs: agent-built graphs can carry a rationale caption (provenance) shown under the plot.
• Records ▸ Copy selected as YAML; the record detail pane gained a Copy button.
• Diff viewer: export as CSV, JSON or PDF.
• Source viewer: a Wrap toggle (off by default).
• Share your analysis setup: File ▸ Export settings… / Import settings… — save a versioned .fluxtion-settings file (or copy it to the clipboard / email it) carrying source roots, Maven repos, event processors, named graphs (formulas and pins included), hidden columns and assistant preferences. Import merges safely (lists add, graphs replace by name) behind a summary you confirm. API keys, AWS details, recent files and search history are never included.
• Settings: Maven repos tab — local repositories (default ~/.m2/repository) are searched for *-sources.jar when a class isn't under any source root; "don't search" opt-out.
• Settings: Event processor tab is now an add/edit/remove list of FQNs with one marked active (double-click to edit).
• Settings: History tab Clear all button.
• First run with no saved configuration opens the Settings dialog automatically.
• Graphs: formula list per graph with Edit and Remove; the f(x) field autocompletes from discovered keys and existing formula labels; formulas can reference other formulas by label.
• Detail viewer: right-click an attribute to add it as a series to the current, a named, or a new graph.

Changed

• Toolbar buttons now pair each label with a small icon (hand-drawn, theme-aware).
• Event types panel: split into Event types and Callbacks sections with Select all / Select none and a right-click Only this / Add / Remove (the group-by radio is gone).
• Detail viewer: right-click anywhere on a node line to add any of its values to a graph (no longer only when you click exactly on a key); the event/eventToString line navigates to the handler source.
• Summary rows: a left-click no longer changes the filter — right-click a row to filter by it.
• The LLM tab is now Analyser assistant; the search field grows to fill the width with a Clear history button; chart gridlines have more contrast.
• Graphs: the f(x) formula field now shows a dropdown of matching keys/labels (↓/↑ to move, Enter or Tab to accept, Esc to dismiss) instead of inline ghost-text completion.
• Record detail word-wrap is now off by default (toggle still in the detail toolbar).
• Subtle panel backgrounds: the graph plot in light mode and the source viewer now read apart from the surrounding panels.
• Settings dialog no longer opens over-wide; Assistant tab fields are no longer clipped.
• Main window layout rebalanced — the records table leaves more room for the detail panel and tabs.

Fixed

• Help ▸ About now shows the analyser's own version instead of a bundled dependency's version.
• Help ▸ User guide now renders as a readable light document in the dark theme (previously dark text on a dark background), and its links now open in the system browser.