flutter_agent_harness 0.1.366

Cross-platform AI agent harness for Dart and Flutter: streaming provider adapters, agent loop with tools, session persistence, context compaction.

Changelog

0.1.358

• fix(hub): boot online from the persisted DAP credential — with no DAP_MASTER_SECRET/DAP_CLIENT_SECRET in the environment, the CLI now seeds the hub kill switch from ~/.dap/config.json clientSecret (the explicit prior opt-in from /dap start), restoring the documented "the next boot is online by itself" behavior: hub mail delivery and hub peers in agent_directory (the browser extension, embedded hosts) work on a fresh boot. The fabric gate also reads the mutable environment overlay the plugin actually uses, not the read-only process environment.

• fix(195): images follow-up from the #190 review. F2: a stale [Image N] citation in history no longer silently rebinds to the WRONG image after compaction renumbering — once a renumbering boundary (compaction summary, structured marker, local-trim note) exists in the window, authored history citations degrade to the (image no longer available) note while generated content-keyed refs, carrier labels and the current message stay intact. F3: the current message's in-place images now carry their [Image N] label so the model can cite what it sees. F4: the app logs per-request cap drops through AppLog (logs/app.log) instead of silently dropping. F5: estimateContextTokens charges repeated images at the wire-replacement cost (first occurrence full, repeats ~32 chars), fixing the transcript-vs-wire estimation asymmetry; per-message estimateTokens is unchanged. F1: the images: config section gained the previously claimed tests — CLI parse/round-trip/strictness (bad bools, unknown keys, non-positive ints, malformed sections) and the config-service validator dispatch pin.

• feat(169): declarative-only secured theme API — theme packs (theme.json + optional wallpaper image in a .zip; strict schema validation with unknown-key rejection, path-traversal/symlink/size screens, WCAG contrast warnings), a Settings section to import, switch and remove packs (a colors-only apply keeps the current wallpaper; removing the active pack reverts atomically), wallpaper layers in the chat, apps grid and launcher screens, and the jsr.fa.theme bridge — list/current/apply behind the theme permission with a consent dialog per apply. Apps can never install or delete packs (no such API exists; pinned by a byte-scan test).

• chore(flutter): raise the Flutter floor to 3.47 (>=3.47.0) across flutter_app, packages/fa_ui, packages/fa_llm_flutter and yoclip; drop the meta: 1.18.0 dependency overrides (the 3.47 SDK declares meta ^1.18.3, so 1.19.0 resolves for dart_tui without an override); regenerate the flutter_app goldens for the 3.47.4 rendering (text/geometry drift only); clean up new-SDK lints in two config tests; ios Podfile now forces the iOS 16 deployment floor on pod TARGET-level build configs too (podspec-declared 15.0 targets broke the build against 16.0-only Promises under the newer toolchain).

• chore(deps): js_widget_runtime ^0.4.121 — typed fallback hardening for legacy manifest keys, widget manifest i18n compatibility, and the JSR video onError bridge.

0.1.357

• feat(155): backend agent mode — fah as a server-side agent engine for a Go supervisor. --output events[=full] switches stdout to a HEP v1 JSONL stream (header + agent_start/turn lifecycle frames, message and tool deltas, usage passthrough in turn_done; events=full also caps tool args at 4000 chars instead of the 100-char summary); prose mode is untouched. --attach <path> (repeatable) reads image files, sniffs the type from magic bytes, and rides them as base64 data-URI image blocks on the first user message. SIGTERM (and SIGINT) in headless mode now abort the run gracefully — partial transcript persists (persistAbortedPartials), a cancelled frame closes the HEP stream, stdout is flushed, exit code 130; a second SIGTERM forces exit 143. An unknown --session key on a clean root starts a fresh session instead of failing. --version --output json prints {"version":…,"hep":"v1"}. (Per-request [Image N] image dedup was also prototyped here and is superseded by the session image registry from #190, which this change adopts instead.)

0.1.356

• ci(177): PR-level CI speed restructure — a pull request now pays only for what it touched. Path-aware gating (changes job), the sequential "Quality gates" monolith split into parallel jobs (static, test-core ×3 duration-balanced shards with merged coverage, coverage-gate, guards, flutter-tests on ubuntu), pub/flutter caching everywhere, concurrency cancel-in-progress on all PR workflows, and one aggregate Quality gate required check. PRs ratchet coverage on CHANGED lines (scripts/diff_coverage.py); main/tags/nightly keep the full 80% ratchet. The pre-commit hook is now a thin wrapper over the SAME scripts/ci_fast_gate.sh CI runs, with the same path filters.
• ci(177): nightly full-matrix workflow (nightly.yml) — the original monolith gates + the PTY/CLI integration suites (never gated on PRs before; live-provider tests self-skip without keys) + the terminal-visual screenshot suite (never ran in CI at all) — with a deduped nightly-red auto-issue. packages/fa_ui non-golden tests join CI for the first time.
• ci(177): automatically growing terminal-test coverage — weekly coverage-gardener.yml re-measures PTY-suite coverage of lib/src/cli/** and commits the higher baseline; scripts/check_cli_coverage.py enforces it nightly (ratchet only up).
• ci(177): duration-balanced test sharding — scripts/test_shards.json + scripts/rebalance_shards.py (junit timing or file-count fallback), rebalanced weekly by shard-rebalance.yml.

0.1.357

• feat(171): session image registry — unique images ride a provider request exactly once per window; every other occurrence becomes a stable [Image N] text ref. rewriteHistoryImages (lib/src/agent/image_registry.dart) rewrites the OUTBOUND payload only (session JSONL byte-identical): carriers [{text:"[Image N]"},{image}] anchor before the first referencing user message or after the tool-result run (never inside call/result pairs); the current user message rides images in place; the per-request cap drops current-first-then-newest with a drop notice (never silent); dangling refs (compaction, cap) resolve to (image no longer available). Stateless per-request rebuild — determinism, eviction and resume come free. Config: images.registry kill switch (false = byte-for-byte legacy shape) + images.maxPerRequest (default 20), strict parse, CLI and app honor it. Fixes #171.

0.1.354

• feat(148): structured compaction engine — context hiding that is structured, addressable, and expandable. New branch records hidden_range / compact_checkpoint store hide/cover state keyed by stable record ids (append-only; classic compaction's lossy summary is untouched); the context projection replaces hidden records with one-line markers (≤12 tokens each) at their original positions, hides tool pairs atomically so the wire stays valid for both provider shapes, and keeps a stable numeric alias per record (1-based JSONL line number — the zero-tool fallback resolves to the same bytes). The two-pass engine (lib/src/compaction/structured/) hides first via a cheap judge over the context ledger and checkpoints only when hiding is insufficient (nested checkpoints flatten at depth cap). The compact_expand tool restores any id/range on demand under a per-turn token budget with paging; the agent loop resets the budget per user turn. Engine choice: compaction.engine: classic|structured in config (strict parse), resolved session < project < global, default classic; CLI host flag and settings entry included, the Flutter app honors the same chain through loadAppCompactionEngine (config parity, no new settings screen). Trajectory ledger folds both record kinds as collapsible compacted rows. Fixes #148.

0.1.344

• fix(97): the TUI secret sheet accepts multiline and over-wide pastes safely: a pasted NAME drops CR/LF (UPPER_SNAKE can never carry line breaks), a pasted VALUE normalizes CRLF to LF and grants PEM-style multiline credentials verbatim, and the masked/revealed value renders one frame row per line instead of embedding a raw newline inside a row (which physically tore the sheet frame in the terminal). The exact-width row off-by-one seen in the same frames is issue #109 and is fixed there. A shortening (CRLF) paste now advances the value cursor by the normalized length - pinned to the raw paste length it overran the buffer and crashed the next edit keystroke (review follow-up).

0.1.336

• fix(97): TUI secret sheet opens on the VALUE field with the suggested name as a dim placeholder (typing replaces it wholesale, Tab toggles fields, ▸ marks focus), the value renders masked from the first keystroke (Ctrl+R reveals), and a blocked Enter shows the reason inline instead of dying silently. SECURITY ADVISORY: with the sheet as shipped before this fix, a credential typed on name focus rendered UNMASKED in the Name row (visible in scrollback, screen recordings, transcripts) — treat secrets entered through the CLI secret sheet before 0.1.336 as potentially exposed and rotate them.

0.1.335

• feat(dap): password-protected local hub — LocalHub(masterSecret:) requires a credential on every WS upgrade (Authorization: Bearer for native clients, the dap_token query param for browser clients that cannot set headers), rejects strangers with 401 + Connection: close (keep-alive pools never replay a dead socket), and gates {"t":"enroll"} to master-authenticated connections: enrolling issues a per-client secret persisted (agentId → secret, mode 0600) in ~/.dap/hub.json alongside the password, so hub restarts keep both. fa hub serve resolves the password as --secret > DAP_HUB_SECRET > the state file, and the first interactive start (bare fa hub serve or the CLI's one-button /dap start) offers to set a password once (empty = stay open, remembered). The browser extension joins a protected hub via a new Hub password field in the panel (faDap.secret, sent as dap_token); an empty field keeps the stored password, and hub.bind rebinds preserve it. Docs: docs/dap.md §8.1/§8.3.
• fix(59): scheduled-message ownership release on dispose — tearing a session down now clears the owner tag on its pending self-addressed records, so the next session's queue adopts them (the #65 recreate contract) while #88's anti-theft tagging keeps protecting LIVE foreign owners. Fixes the pre-existing flutter_app regression from the #88 merge; also repairs the branch's golden gaps (DapHubMark golden + guard registration, resume-refresher exemption, regenerated DAP page snapshots) and the golden-test fake missing the DapHubService binding members.
• test(dap): protected-hub e2e — the Playwright dap spec now also runs against a PASSWORD-PROTECTED hub (FakeHub(masterSecret:) + DAP_E2E_HUB_SECRET on the e2e hub server): a stranger boot is 401'd before any hello (client stays reconnecting), the password boot connects, and a CLI enrolled with the same password exchanges DMs with the extension end to end. The shared scripted mock gains resetScript() so every test's CLI gets the dap_dm tool call as call #1.
• feat(dap): the app-mode Agent network editor gains the Hub password field (write-only — an empty field keeps the stored one), wired through DapHubService.saveConnection: the extension variant rides hub.save, the desktop variant persists clientSecret into ~/.dap/config.json.
• fix(dap): /dap start no longer 401-loops against its own protected hub — the session's client now dials with the HUB's password (prompted or read back from the state file), persisted as clientSecret so stale secrets from an older hub can never win the resolution precedence.
• fix(extension): a session created from ANY surface now shows up in every other surface's sessions drawer. The SW broadcast the attach trio after session_new/session_open only to the requesting port, so a panel/app-tab that missed it kept a stale live-session id and its drawer filter hid the freshly archived row ("added a session, still see one"). The trio now broadcasts to all connected ports, and the drawer filter no longer drops archived rows by a possibly-stale id.
• fix(app): clicking a session no longer teleports it to the top of the sidebar. The pending-click jump-to-top reordered the list while the row's stamps were still old, parking a "1:08 PM" row above an "8:42 PM" row. The click now highlights the row in place; the list reorders only when a session's real activity time changes.
• fix(app): ONE session-selection source for every surface. The wide sidebar and the narrow drawer had divergent active-row logic (manager slot vs SW live id), so a switch lit two rows in the wide sidebar (pending + stale slot) and none moved in the drawer. FlutterSessionManager.hostedLiveId is now the single truth (set on every attach broadcast), pending clicks REPLACE it until it lands, and both surfaces derive their dot from the same effective id.
• fix(app): the wide sidebar's active dot follows session switches. The live session is always the freshest row, so the dot sat pinned to the first row and a switch had no immediate feedback. The clicked session now highlights and jumps to the top instantly (pendingSessionId), clearing when the manager's slot catches the attach broadcast.
• feat(app): the Agent network page is now connections-first: a list (active connection on top, always present even when never bookmarked)
  • Add connection; tapping a row opens that connection's details in the SAME page — the active one shows url/status/identity/incoming routing/channels/edit, a bookmark shows its coordinates with Make active and Remove (the switch stays live).
• feat(app/extension): multiple hub connections in Agent network. The extension keeps a bookmark list (faDap.savedConnections): Add connection saves a new hub and goes live on it, tapping a bookmark switches the live agent (hub.switch — the entry's own secret applies, an open bookmark clears the stored password), re-saving the active connection preserves the list (hub.connections.set manages it). SW logs: [dap-hub] connection saved / switched / list set.
• fix(extension): session switches now follow the UI everywhere. The manager's active slot kept its boot-time id/stamps after any session_new/session_open, so the active dot stayed on a row whose label never changed, and the wide-layout sidebar (app tab) listed only one session — it read the page-local repo while the session files live in the service worker. The relay now re-keys the manager slot on an adopted broadcast, the sidebar/drawer take fresh stamps from the SW poll, and listPersistedSessions delegates to the relay's sessions_query on hosted pages. Logs: session switches, dispatches, busy refusals, and mail-driven bound-session switches are printed on both sides ([fah][relay] / [dap-host]).
• fix(dap): the browser client stops the fast retry loop when the hub keeps rejecting the credential (wrong Hub password): after three fast closes with a credential configured it holds at a 30 s re-check and surfaces the new unauthorized phase instead of hammering 401s.

0.1.331

• fix(86): compaction checkpoint honors the owner review on #82 — the hardcoded imperative-verb ask detector (_requestMarkerPattern) is gone; every user-role message (initial + steering, any phrasing or language) now reaches the summarizer as a USER REQUEST CANDIDATES line and the LLM itself judges which are open asks (only structural non-user content — <system-notice> envelopes, agent mail, projected branch summaries — is skipped; detectUserRequestCandidates → userRequestCandidateLines), so marker-less asks ("можешь глянуть почему тест падает?") can no longer be evicted before the summarizer sees them. The compaction prompts (system, first-time, update, turn-prefix, branch) gain an explicit tool-results assessment: important outputs (test verdicts, command results, error traces, fetched data) are checkpointed with what produced them; trivial banners may be omitted, with the LLM assessing importance. The owner's ≤500-char instruction budget (issue #81 AC7, restated by #86 AC4) holds at net delta 423: the summary_update.md section hints were tightened to make room while every operational rule (PRESERVE/ADD/UPDATE/ASSESS and the open-ask close-out rules) stays intact. Tests: marker-less-ask regression (unit + end-to-end into ## Open User Requests), non-lossy candidate pin, tool-results wording pin across the five prompts, and a checkpoint behavior test (failing-test output preserved, success banner not required).

0.1.330

• feat(27): A2A boundary gateway (issue #27 phase 3) — cross-machine agent_message rides A2A: name@machine addresses resolve the machine against the a2a: config section and deliver as message/send with a faMail metadata envelope (A2aMailGateway, lib/src/a2a/ a2a_mail_gateway.dart); failed remote tasks error instead of dead- dropping, unconfigured machines fail with the exact config hint. Inbound: fa serve --a2a now accepts envelope-carrying sends and deposits them into the project's file inboxes (by mailbox id or session display name) instead of running an agent turn. The hub remains the intra-machine transport; hosts still carry no transport code.

0.1.324

• fix(43): typing during minutes-long thinking streams no longer degrades — the 32KB tail hard-split in _appendOutput used to land on TranscriptMarkdown's commit boundary (fresh substring identities defeat both resumability sentinels) and force a full O(transcript) markdown+wrap rebuild, so frame builds grew with the session (real-PTY e2e at 200 reasoning deltas/s: frame-build p99 24.8ms / max 81ms after 60s, keystroke echo p99 59.5ms). The grown-tail rollback now also recognizes the split shape — the boundary line survives as a prefix of the concatenated chunks (_boundarySurvivesAsChunks, one O(boundary) string compare per suspected split) — rolls back one source line and re-walks the chunks; replaced content still takes the documented rebuild path. Rebuilds in the flood regime: 0 (bench scripts/tui_typing_bench.dart --grow-tail; frame-build p99 5.0ms / max 7.4ms flat). Contracts: split-resume trio in test/cli/transcript_markdown_perf_test.dart; perf log: docs/performance-cli-tui.md.

• feat(skills): the fa-self-config skill (issue #29, phase 1) — fa configures itself by editing the config files the CLI settings commands write, with a test-enforced parity guard: every settings-affecting CLI command (/provider, /models, /model, /model-edit, /memory, /tools, /cube, /mcp, /redact, /skills, /approval, /allow, /mode family, /settings) must stay documented in the skill with its config-file equivalent, and a new settings command fails CI until it is classified and documented. The skill encodes precedence (project memory:/cube:/tools: win over the user file), strict vs tolerant parse behavior, live vs next-boot application, and the never-inline-API-keys rule; an accuracy test pins every documented key against the real parsers (CliConfig/MemoryConfig/ToolsConfig/ McpConfig/roles) so no phantom keys or stale names can ship.

0.1.322

• feat(flutter_app): CodeMie sign-in works inside the browser extension — no webview, no localhost callback, no API key. The app page detects the extension host (chrome.runtime.id), opens the login page in a NORMAL browser tab (IdPs forbid framing; MV3 has no webview), and polls the models endpoint with credentials: 'include' until the shared cookie jar holds a session. The redirect-interception dance (localhost callback server) stays a desktop/mobile-only concern; the saved provider keeps an EMPTY key — the service worker's streaming fetch carries the jar. Poll/parse core is pure and tested (codemie_extension_signin.dart); extOpenTab/extFetchString bindings live beside the existing chrome.runtime interop.
• feat(browser_ext): CodeMie cookie sign-in without any API key or SSO dance — the panel now talks to a small ext_request op surface on the service worker (cookies.get_all, SW-relayed fetch with credentials: 'include' — MV3 + host permissions mean no CORS and the browser jar rides along — and tabs.create for the login page). codeMieLogin probes llm_models: 200 → cookies alive, model ids prefill, the key field stays empty; 401/403 → the login tab opens and the probe repeats until the jar holds a session. A CodeMie base URL (isCookieAuthUrl) streams DIRECT from the SW — never through the bridge relay, which would strip the cookies and 401. New ext_request/ext_result protocol kinds; every op validates its params and answers a structured result, unknown ops never crash.
• fix(browser_ext): the exfil gate's ask under yolo now stays silent — the user's contract is that yolo asks for NOTHING (the extension has no bash to carry critical patterns), so only the interactive modes (ask/write) surface the outbound dialog; unattended keeps allowing without asking.

0.1.321

• fix(browser_ext): cross-origin tabs_open/downloads_start no longer hard-fail without asking — the exfil gate threw approval_required straight at the model, so even yolo could not open a never-visited site (the agent looped on the tool error, "reads the page but can't open apple.com"). Flagged outbound actions now route through the host's approval prompt: one dialog per new ORIGIN (an allow seeds the visited set), deny keeps the tool error, and unattended hosts allow without asking so autonomous runs never stall on the 120s backstop. The ask is an injectable ExfilApprovalAsk on the tool surface; unwired surfaces keep the conservative hard error.
• fix(browser_ext): the approval loop could not be answered or rescued — the extension panel's primary chat surface (the narrow-layout session sheet) never installed an approval handler, so every gated tool call sat out the SW's 120s backstop and was denied with the user never asked: the turn looped "busy", produced no output, and the model kept retrying gated tools. Fixes, three ways:
  • FaChatSurfaceHandlers (fa_ui): a reusable binder installing the approval dialog + ask sheet + secret-request sheet on the active service for as long as a chat surface shows; the session sheet now binds it (moves on session switch, clears exactly its own handlers on dispose, never clobbers a foreign handler).
  • ApprovalFlow (browser_ext/dart, pure + VM-tested): the SW host's pending-approval core extracted from agent_host (same wire events); resolveAll completes every pending prompt at once.
  • live approval-mode changes: reconfigure no longer drops the mode behind the busy guard — the mode applies mid-run, and flipping to yolo/unattended resolves pending prompts as allowed (the user's rescue gesture); only provider/mailbox/hub/tool changes still require an idle host.

0.1.313

• feat(js-ext): JS extension core (issue #32) — JsrRuntime engine seam + engine-agnostic bootstrap (jsr.ext.*: registerTool tiers, six hook events, slash commands, provider flows, session/fs/exec/keys/io/has bridges) over three transports (qjs stdio, flutter_js send-message, web worker); JsExtensionHost wires commits into the agent with deny-precedence hook composition (JS can only tighten), append-only afterToolCall redacted twice (before JS sees it, before persist), read-only prepareNextTurn, E14 follow-up collapse, and 30/10/120 s load/hook/tool budgets that degrade to log lines, never crash the loop; manifest/trust/install machinery — strict accumulating manifest parse (E12), TOFU trust with capability-diff re-prompt + hash-only silent re-grant, local/zip/gh/catalog/bundled planners with sha256-verified hostile-zip rules, and the tolerant v2 catalog client (unions widgets

  • extensions).

• feat(cli-ext): fa ext list|install|remove|update|audit with --pin <sha256>/--trust/--strict/--bundled/--json (enable/ disable stay REPL-only), the quickjs-ng engine process (qjs --std, binary via explicit override → FA_QJS_BIN → PATH, missing binary degrades to an engine unavailable skip), the /ext REPL family (list/enable/disable/audit/remove/update/reload), and idempotent .fah/bootstrap.yaml applies at every start — project then user, E16 project-wins shadowing, E15 named soft-fail lines unless FA_EXT_BOOTSTRAP_STRICT=1.

• feat(app-ext): flutter_app/lib/services/ext/ — AppExtensionService owns the on-device store roots + host; flutter_js (native) and web worker (web) runtimes behind one factory. v1 scope: trusted-only load (no trust prompt — untrusted extensions tombstone-skip); the app trust UI lands in a later wave.

• feat(js-ext): bundled crap-guard reference extension (post-edit CRAP + 2800-line guard via crap4dart, 2 s edit debounce, one aggregated {append} per burst, E17 single missing-tool note) compiled in as const strings, byte-mirrored into js-ext-registry/crap-guard/ for publishing (sync test enforces equality) with a README covering the fa_widgets catalog.json entry, zip + shasum -a 256 recipe, and gh-repo layout; authoring guide in docs/js-extensions.md.

• feat(browser-ext): browser extension v2.1 (issue #30) — fa web moves into the extension and grows browser superpowers. The panel is now an app-hosting bootstrap: scripts/build_browser_ext.sh --with-app bundles the Flutter web build as browser_ext/app/ and the side panel/full tab loads it (v1 minimal chat stays as fallback). The service worker gains the UI protocol server — FaUiProtocol over a chrome.runtime port (17 message kinds: hello/attach/prompt/steer/ cancel/stream/approval/sessions/settings, prompt-id dedup + attach replay), WorkerRelayTransport vs LocalStreamTransport with capability detection — and a 34-tool browser surface (browser_api_tools.dart) over the typed ChromeApi facade (23 chrome.* groups): tabs/windows/groups/sessions/history/bookmarks/ downloads/cookies CRUD, first-class inject_js (ISOLATED/MAIN worlds, MAIN always prompts) + inject_css, cdp_eval, page_screenshot, app_screenshot, nav_wait; restricted pages refuse scripting. Permission⇄tool matrix (permission_matrix.dart): 28 core rows / 9 second-tier optional / 11 excluded (incl. the impossible-by- construction passwords row), unpacked vs store manifest profiles (store strips debugger+cookies); manifest 0.2.0 with 26 core permissions, second tier in optional_permissions, ask-fa command (Ctrl+Shift+1) and omnibox keyword fa. Background residency: badge state machine (idle/busy/mail, resync after SW kill, denied- notification fallback), exactly-once scheduled tasks over chrome.alarms, lifetime-capped offscreen documents, EntryPointHub steering omnibox/commands/contextMenus into one agent. Security layer: quarantine + instruction-hierarchy classifier (page bytes are always data), injection validator (login-form/keylogger refusals), exfil gate (page-derived/cross-origin/data-exit approval), and a new layer_credential.dart redaction layer masking credential-shaped form values pre-persist. Docs rewritten: browser_ext/README.md, docs/browser-extension.md, new docs/browser-extension-permissions.md. Pending: flutter_app worker-relay chat integration, Playwright e2e, store publication.

0.1.297

• fix(tui): a long thinking burst froze the whole TUI (typing dead, spinner stuck — user report: "пиздец как тормозит" on a glm-5.3 session). Root cause: the CLI wrapped every thinking delta in inline markdown, and the escape-dense result drove AnsiMarkdown's inline link regex into quadratic backtracking — measured 10.9 s of synchronous work per coalesced 185 KB flush. Two fixes: thinking deltas now dim verbatim (spans cannot pair across deltas anyway), and inline span substitution is skipped for lines over 4 KB (AnsiMarkdown.inlineFormatMaxChars) — degenerate spans render verbatim, fences/rules/tables unchanged. Regression-tested by the new tool/thinking_lag_probe.dart PTY probe (burst streams: keypress echo p95 stays under budget) and a markdown unit test.
• docs(readme): reflect the shipped surface — status, design contract, what's inside (agent core, providers, sessions, tools, skills, approval gate, trajectory, messaging), CLI slash commands.

0.1.296

• feat(bash): automatic transient-failure retry — timeout-class failures (the model's per-call cap, or an outer future cap on a hung transport, e.g. a stalled gh API call) are retried up to 2 times (3 attempts total) with a 1s backoff before the error surfaces. Retries are visible as [bash attempt N/3 ... — retrying] notices in the tool output, and the final error carries them too. Aborts and real command failures (non-zero exit, exec errors, shell unavailable) are never retried.
• fix(app): the macOS/iOS app could boot to a black screen — the native Firebase SDK auto-configures the [DEFAULT] app from GoogleService-Info.plist at plugin registration, so the Dart Firebase.initializeApp threw an unhandled [core/duplicate-app] that killed main() before the first frame. The duplicate-app case now reuses the natively configured app.

0.1.296

• feat(bash): automatic transient-failure retry — timeout-class failures (the model's per-call cap, or an outer future cap on a hung transport, e.g. a stalled gh API call) are retried up to 2 times (3 attempts total) with a 1s backoff before the error surfaces. Retries are visible as [bash attempt N/3 ... — retrying] notices in the tool output, and the final error carries them too. Aborts and real command failures (non-zero exit, exec errors, shell unavailable) are never retried.

0.1.295

• docs: fa1.dev is now referenced from the README (website + live web demo + iOS beta + installer) and from homepage in pubspec.yaml, so the link renders on pub.dev next to the package name.

0.1.294

• feat(redact): layered secret redaction pipeline (issue #24) — ten pure layers (registered/path/vendor/prefix/pem/asn1/connection/context/ entropy/pii) merged priority-first into RedactionPipeline.scan/redact with idempotent [REDACTED:<kind>] markers, allowlist + data-URL pass-through, live RedactionConfig, and per-layer/per-tool stats. Agent hooks mask tool results BEFORE session persist, mask the outgoing context, deny credential-file read/bash in blockMode, and mask user prompts; write-side tools (write/edit/checkpoint/MCP write-ish) are never filtered. redact: config section, /redact command, app wiring, docs/redaction.md.
• test: tui_prototype_snapshot marked integration — the real-PTY + dart run cold start intermittently exceeded the gate timeouts on CI, which had been silently blocking every tag publish since ~0.1.214.

0.1.289

• fix(roles): transient transport failures no longer kill the turn. A dropped connection ("Connection closed while receiving data", resets, refusals, DNS/TLS handshake failures, 502/503/504) during a provider call is now retried in place with the retry policy's backoff budget (retry.retriesPerEntry, default 2), then falls over to the next chain entry — the run survives instead of ending with an error event. Key rotation is deliberately skipped for this class (the endpoint dropped, not the credential), the observable-output guard still holds (a stream that already emitted content is never silently replayed), and every retry is announced via FallbackNoticeKind.transportRetry (the CLI prints [roles] connection lost on … — retrying in Ns).
• fix(app): the onboarding flow follows the dark theme — provider cards were hardcoded Colors.white with near-white title text (white on white), the wide ghost column's "Ask when needed" used the brand blue on a near-black background, and step dots/progress tracks/dividers were light-themed; all now route through the _onb* dark-aware helpers with a readable _onbPrimary accent in dark mode.
• fix(cli): steering robustness — file-path-prefixed input sent while a run is busy is steered with its attachment instead of dying on the busy gate, steered ~//./ paths resolve like interactive input, leftover steering after an aborted/interrupted run runs as a fresh turn or is dropped loudly (texts printed), and queue drops always show what fell out instead of vanishing silently.
• fix(cli): a !command executed locally now also steers a compact <system-notice> (command + exit code + capped output tail) into the conversation — the agent learns what ran without being woken.
• deps: the DAP hub client moved to the published fa_hub_client 0.2.8, whose reconnect backoff no longer overflows after ~64 attempts (Dart int shift semantics: 1 << 64 == 0) into a zero-delay tight reconnect loop against a dead hub — the root cause of the multi-hour CPU storms.
• fix(tools): media slot overrides no longer fall back to the main provider API key — a slot override is its own provider configuration (CLI generate_image + the app's MediaModelsStore.resolve).
• fix(tools): generate_image surfaces MiniMax base_resp errors that hide inside HTTP 200 responses.
• fix(cli): per-folder model memory — /model and /provider switches are scoped to the launch folder; the global config keeps its seed triple, so a switch in one workspace no longer leaks into the others across restarts.
• fix(cli): model/provider persistence callbacks are awaited — /model x
  • /exit can no longer lose the switch.
• fix(cli): the over-window context guard renders a calm yellow note instead of a red error; explicit guidance when compaction cannot free the window.

0.1.282

• feat(cli): a DAP / Hub entry in the /settings hub — view the live hub snapshot (resolved url, agent name, connection state), set the hub url and the agent name (persisted through the hub client's ~/.dap/config.json read-modify-write, so channels and invites survive), test the connection, and write the hub: false plugin opt-out into .fah/packages.yaml (existing sections preserved). The flow reads the hub state through an injectable snapshot seam the executable wires to the hub plugin — no sockets, fully fake-drivable in tests.

• fix(cli): .fah/packages.yaml entries actually load now — package:yaml reifies its map entries as dynamic-keyed, so the loader's String-keyed whereType filter dropped every entry and the whole file (plugin enabling and configuration, hub: url/name, hub: false opt-outs) was inert. The loader now deep-converts the parsed yaml tree to plain Dart values before handing sections to plugins and the DAP / Hub settings flow. The loader now lives in lib/src/plugins/packages_config.dart (reading through the ExecutionEnv), so tests can pin it without importing the executable.

• feat(app): DAP/Hub settings section in the Flutter app — a settings row opening the hub page (resolved URL, connection probe, agent name/agentId, channels) with add/edit of the machine-shared ~/.dap/config.json connection via fah_hub_client; web degrades to an honest not-supported note (the hub client is IO-bound).

• fix(dap): a dap_dm to a known-but-offline peer no longer reads like a typo — the no-match error lists the known online peers and points at dap_invite for offline peers and dap_peers for typos.

• feat(plugins): FahPlugin.dispose — the CLI calls it once per plugin at shutdown (errors swallowed per plugin, so one bad plugin cannot block exit); plugins override it to release sockets, processes, and timers. The hub plugin host disposes the vendored hub client.

• chore(cli): the fah_hub_client import in bin/fah.dart is now marked as the ONLY core-CLI import of the hub client — downstream forks wanting a different or no hub client patch that import plus the 'hub' case in _builtInPlugin.

• fix(messaging): agent_directory no longer drowns in graveyard mailboxes — it lists LIVE mailboxes (recent activity) plus anything holding pending mail and the agent's own address; stale mailboxes from long-finished sessions appear only with the new all: true parameter. Liveness is source-defined MailboxEntry.lastActivity (MailboxEntry.isLive, 15-minute default window): the file repository scans the .heartbeat marker plus inbox/read content mtimes (the .id identity marker and _scheduled/dot directories are excluded), and running hosts keep the heartbeat fresh from their existing inbox-watch timers via the new MessagingRepository.touch (CLI every ~4s, app every ~6s; best-effort). Unknown-activity entries (custom repositories) are never hidden.

• refactor(cli): the fah executable's startup is decomposed — the pure phases (serve --a2a argument interception, provider/model restoration, the secure-store preload set, the startup API-key decision, roles-secret collection, secret-redactor and web-search assembly) moved to lib/src/cli/startup.dart with unit tests in test/cli/startup_test.dart; bin/fah.dart keeps only process glue and _runApp reads as a phase sequence. The CRAP ratchet is green again (worst 12.00): the plugin-resolution test importing bin/fah.dart had dragged the whole executable into the coverage trace at 0% hits, scoring _runApp CRAP 3906.

• fix(dap): vendored hub client warns loudly when the identity file's chmod 600 cannot be applied (no chmod on Windows, or the call fails) — the Ed25519/X25519 private seeds would otherwise persist with default permissive ACLs, readable by other local users.

• fix(dap): vendored hub client's HKDF expand enforces the RFC 5869 255-block cap (throws StateError) — a zero-length MAC from a misbehaving crypto dependency can no longer hang the hub connection.

• fix(dap): .fah/packages.yaml plugin opt-out is value-aware — hub: false (or an empty value) now really disables the plugin instead of the key-only de-dup keeping it on; a failing connect to the zero-config default hub prints one quiet hint line instead of a raw error, so plain CLI starts stay clean.

• fix(dap): the hub plugin's fire-and-forget connect carries a defensive .catchError, so a future escape can never kill startup silently.

• feat(memory): the long-term-memory LLM slot is wired — a new HarnessLlmProvider adapts fa_llm's LlmProvider onto the harness streaming contract and is resolved per call (memory role → smol → main) in BOTH the CLI and the app. Memory consolidation and semantic search now actually run instead of being silently skipped.

0.1.280

• fix(providers): correct the Copilot token guidance (0.1.278 had it backwards). Per the official GitHub Copilot CLI docs (2026-09), the supported credential types are fine-grained PATs (github_pat_…, v2) WITH the "Copilot Requests" permission and OAuth tokens — while classic PATs (ghp_…) are NOT supported by Copilot at all. 0.1.278 rejected all github_pat_ tokens up front and recommended ghp_ — blocking working tokens and pointing at dead ones. Now: a pasted fine-grained PAT is accepted with a hint that the "Copilot Requests" permission is required (its absence is what makes the exchange 404), a pasted classic ghp_ token is warned about and re-asked, and the exchange 404 message names the missing permission.

0.1.278

• fix(providers): GitHub Copilot rejects fine-grained PATs clearly, at connect time. The Copilot token exchange answers github_pat_… tokens with HTTP 404 (GitHub's Copilot credential API only accepts classic PATs and OAuth tokens), which used to surface as a bare "token exchange failed (HTTP 404)" on the first message and, during connect, as a silent fall-back to manual model entry (the /user lookup accepts fine-grained PATs, masking the problem). Now: the paste-token step rejects a fine-grained PAT immediately with the fix (use the GitHub device flow or a classic ghp_… token), the exchange short-circuits the known-dead token type without a round-trip, the 404 names both likely causes (token type / no Copilot plan), and a failing exchange is reported during connect instead of an unexplained empty model list.
• chore(cli): split the banner/key-status block out of agent_cli.dart and disentangle _runPrompt (error-stop and empty-continue branches into named helpers) — the file-size guard and the CRAP ratchet are green again.

0.1.277

• fix(cli): endpoint-reported context windows now apply in roles mode. The Copilot /models limits (capabilities.limits. max_context_window_tokens / max_output_tokens) were parsed and applied only when no roles: resolver was configured; with one, a chain entry riding the catalog default silently kept the provider's default window — the copilot default is 1M, so a 256k model (live- verified: the endpoint reports kimi-k2.7-code = 256000/32000) ran with compaction thresholds sized for 1M and the over-window guard never fired in time. Roles entries with EXPLICIT contextWindow/ maxTokens still win over the endpoint (per-limit).
• fix(cli): the line REPL warms the model cache like the TUI, so /model <id> switches see endpoint limits there too.

0.1.276

• fix(cli): the over-window guard no longer strands the agent mid-task. When the guard stops a run (outgoing context past the model window), the post-run auto-compaction frees the window and the CLI CONTINUES the interrupted turn on its own — once per user prompt, delivering a <system-notice> that names what happened and tells the model to avoid re-reading the outputs that filled the window. Ending the run there left live sessions idle until a manual "continue" (seen on a glm-5.3-flash session: guard at 200676/200k, local trim to 20k, then silence). _maybeAutoCompact now reports whether it actually shrank the transcript; the continuation fires only when the window was really freed.
• fix(cli): compaction settings scale with the model window (CompactionSettings.forWindow) unless compactionSettings: is pinned in the config — matching the Flutter app's existing rule. The pi-fixed defaults (keep 20000) structurally prevented compaction on small-window models: with keep 20000 on an 8k window the kept region always covers the whole transcript, so nothing could ever be summarized away and the over-window guard could never be satisfied.
• app: the same guard + compact + continue flow in AgentService (once per user text, reset on the next real input).
• agent_loop: exported contextWindowExhaustedMarker and isContextWindowExhaustedError so hosts recognize the guard without parsing numbers out of the error text.

0.1.274

• fix(compaction): the summarizer can no longer hold the turn hostage. A dead/dribbling summarizer endpoint used to keep the "Compacting context…" row up for up to ~20 minutes (10-minute per-attempt budget from 0.1.240, smol + main): the per-attempt budget is now 90 s and a new whole-run totalBudget (4 min) skips any further attempts once burned, falling to the mechanical local trim. New AutoCompactorHooks.onAttemptStart surfaces each attempt in the busy row ("smol=provider/model, attempt 1, 90s cap") so the wait reads as bounded, and the CLI resets its delta tail per attempt.
• fix(agent-loop): mid-turn over-window guard — before each provider request the outgoing context is estimated; past the model's window the run stops with a clear "Context window exhausted" error instead of silently sending a 287k-token context to a 200k model (seen live: compaction only runs at turn boundaries, and a single verification turn with full-suite logs ballooned far past it). Gross overflow only — between the compaction trigger and the window the post-run compaction flow still owns the decision.

0.1.273

• fix(agent): Agent.abort() disarms the run idle watchdog — it guards a WEDGED stream going silent, not a host that cancelled on purpose; an 8-minute timer no longer outlives an aborted run. The app's AgentService.dispose now aborts an in-flight run (its error callback skips notifyListeners once disposed), which un-wedged the widget-test floor: 8 pre-existing flutter_app failures from the 0.1.271 watchdog (chat screen, session binding, work bar, session chat sheet) pass again.
• refactor(tui): split _handleBusyMsg's transition diagnostic/copy out of the dispatcher (CRAP 15.15 → under the pinned 12.0 ratchet).

0.1.272

• fix(providers): no provider carries a default model anymore — every connect/switch flow picks the model explicitly. /provider kimi and /provider copilot now fetch the endpoint's /models (the Copilot dialect runs the GitHub→Copilot token exchange first) and ask the user to pick, falling back to a manual id entry when the fetch is empty; the silent gpt-4.1 seed is gone from the Copilot connect flow.
• fix(providers): the Copilot /models dialect lists only picker-eligible chat models — pi-mono parity (model_picker_enabled, policy.state, supports.tool_calls) plus a supported_endpoints check: responses-only models (e.g. gpt-5.6-sol) 400 with "not accessible via the /chat/completions endpoint" on our chat transport and are no longer offered.
• feat(cli): /model and /provider switches print a role-models note naming the models the smol/subagent/memory roles still run, so a main-model switch never silently strands a mismatched combination ("/settings → Agent models to adjust").

0.1.271

• fix(tui): the busy row now names its owner and dies on its own. Every arm/release carries a provenance tag rendered in the row (Working… 91s · run) and logged to fa.log; a stretch silent for 3+ minutes shows a quiet Nm hint instead of pretending steady progress; and a model-level watchdog force-releases any row with zero activity for 10 minutes — the last-resort finally for the immortal "Working…" class, with the last armer named in the diagnostic log.
• test(tui): provenance render, transition forensics, quiet hint, and the watchdog release are all pinned in fa_tui_test.

0.1.270

• fix(memory): flutter_agent_memory 0.2.1 — the deletion ledger is now strictly append-only (a delete never rewrites existing content; tolerant parser infers a missing type: from the id prefix and recomputes missing fingerprints), closing the clobber class that erased 144 tombstones in production. ExecutionEnvKbStorage implements the new KbAppendCapable through the existing ExecutionEnv.appendFile — racing deletes append instead of read-modify-write last-writer-wins.
• test(memory): regression pins the incident: a legacy 144-entry ledger with the count: header survives a delete with every tombstone intact and the new entry appended; the adapter's native append is asserted directly.

0.1.268

• fix(tui): the "Working… Ns" immortal-spinner wedge (found live: a session burned 100% CPU for 8 hours after the run had cleanly ended — Working… 28301s). Root cause in the TUI busy state machine: a raw BusyMsg(true) landing on an IDLE model — a post-run straggler like a compaction finally-branch calling setBusyPhase('') after the busy bracket released — re-armed the spinner with a fresh elapsed window and scheduled a NEW 100ms tick chain; nothing ever sent the matching BusyMsg(false), and every chain re-rendered the full transcript (268k tokens) ten times a second — the 99.8% CPU storm. Two guards now close the class: the model drops phase-relabels and duplicate busy starts while idle/already-busy (no new window, no extra chain — one chain per busy stretch), and the controller's setBusyPhase no-ops at busy-depth zero. Regression tests: post-run straggler cannot resurrect the row, duplicate starts keep the single chain.

0.1.267

• feat(memory): flutter_agent_memory 0.2.0 — merge-friendly note ids (n_0447_a1b2 = sequential index + 4-hex md5 of the normalized text; parallel branches union-merge, legacy n_0001 ids valid forever), and MemoryRepoInit.ensureGitSupport() wired into the project store: the memory dir self-maintains .gitignore (derived artifacts never committed) and .gitattributes (DELETIONS.md merge=union).
• feat(memory): the memory_add tool description IS the memory repo's policy (MemoryPolicy.memoryAddPolicy — docs/memory/memory_add_policy.md): durable facts only, solved problems are superseded via delete+add instead of rotting, project scope is PUBLIC (committed to git — no secrets or personal data), user scope stays machine-local. One source of truth for every future agent.
• fix(build): CRAP ratchet green again — MemoryController delegates path resolution to MemoryConfig (one source, covered through the controller tests), and AgentCli.run sheds the retry-notice closure into _wireTransientRetryNotice (the 12.61 breach).
• docs(agents): the git-backed memory layout is documented in AGENTS.md (paths resolution, derived artifacts, id scheme, the commit convention: memory/ changes ride with the task's commit).

0.1.266

• feat(memory): project-level .fah/config.yaml — the memory: section now also resolves from the project's own config file and WINS over the user-level one, so the git-backed memory pointer travels with the repo: memory: {projectPath: ./memory} in .fah/config.yaml makes the repo's memory/ directory the project memory for anyone who clones it (CLI and app, web keeps the default).
• feat(repo): flutter_agent's own memory moved into git — the full revision landed first (446 notes audited one by one; 142 deleted through tombstone memory_delete: 84 duplicates, 44 stale, 14 superseded), the surviving 304 notes now live in the committable memory/ directory with derived artifacts (GRAPH.md, MEMORY.revision, indexes) gitignored and DELETIONS.md under merge=union. .gitignore flipped from ignoring all of .fah/ to a whitelist: config.yaml, rules.yaml, lsp.json, mcp.json, agents/, skills/, packages.yaml are committable; logs/sessions/bash_jobs stay local.

0.1.265

• feat(memory): configurable long-term memory storage paths — the new memory: section of ~/.fah/config.yaml (projectPath, userPath, strict schema like the other sections) feeds MemoryController's new storage-path overrides: a relative projectPath resolves against the project root, ~/ in userPath expands against the user home, and null keeps the historical .fah/memory layout. Point projectPath inside the repository (e.g. ./memory) and the project memory becomes committable — anyone cloning the repo gets its memory. Wired in the CLI (AgentCliConfig.memoryConfig) and the app (AgentService reads the same section through a conditional IO/stub loader — web keeps the default). Step P0 of the git-backed memory program.

0.1.264

• fix(cli): macOS Cmd+Left/Right no longer types "aaaa" in the composer — those keys arrive as ^A/^E control bytes, and the dart_tui decoder puts the BASE LETTER into the event text, so every unhandled ctrl combo leaked its letter through the catch-all insert. Three insert paths (composer, picker type-to-filter, prompt re-shaper) now drop command-modified keystrokes (ctrl/alt/meta/hyper/super — shift stays, kitty-protocol shifted letters carry real text), and ctrl+a/ctrl+e do what the user meant: readline start/end of line.
• feat(cli): the request_secret sheet gets readline Ctrl+U — one keystroke clears the suggested name on name focus (the 16-backspace "erase SUDO_PASSWORD" nuisance) and kills the value back to the cursor on value focus; both hints name the key.
• test(cli): PTY visual coverage with screenshots — a real terminal run proves the composer edits (Xhello worldYZ after ^A/^E plus silent ctrl+x/g/z) and the full secret-sheet lifecycle end-to-end through a canned LLM that emits request_secret: suggested name, Ctrl+U clear, dot masking, Ctrl+R reveal, value kill, save — with the grant never echoing the secret into the transcript (screens 102–108).

0.1.263

• feat(cli): the request_secret TUI sheet can reveal the typed value — Ctrl+R toggles between the masked dots (default) and clear text, with the header line naming the current state (value hidden (Ctrl+R reveals) / value visible (Ctrl+R hides)). Long sudo passwords and tokens pasted into a blind field were unverifiable; now one keystroke shows what was actually typed before Enter.

0.1.262

• fix(providers): GitHub Copilot enterprise sign-in actually connects — the token exchange sent the legacy Authorization: token <githubToken> scheme, which api.github.com/copilot_internal/v2/token rejects with a bare 401 for enterprise-managed (EMU) accounts; pi's actively maintained oauth/github-copilot.ts uses Bearer for the exchange, and so do we now (OAuth device tokens are gho_/ghu_ Bearer credentials). The account-type picker hardcoding also bit: an enterprise tenant's real API host lives INSIDE the exchanged token (proxy-ep=proxy.<tenant>.github copilot.com), so both the chat path and the /models listing now derive the tenant host from the token (pi getGitHubCopilotBaseUrl parity) and the picked tier host is only the fallback. The app gets the fix for free through the shared core exchange.

0.1.261

• feat(providers): transient network retry on every provider call — a Wi-Fi/VPN switch mid-turn killed the stream with "Connection reset by peer" and the turn ended in a hard error. providerStreamFunction (the one chokepoint for chat turns, roles chains, compaction summaries, and memory extraction) now wraps every adapter with transientRetryStreamFunction: a socket-level failure (reset / refused / unreachable / timed out / broken pipe / TLS handshake cut — certificates excluded, a bad cert never heals in 5s) sleeps 5s and replays the call, up to 3 attempts. omp's observable-output guard is kept: a stream that already emitted content is never replayed, so a retried generation can't duplicate text. The CLI surfaces each retry as a dim [net] connection lost — retrying in 5s (attempt 2/3) line plus an fa.log entry instead of a mysterious pause; transientRetryNotice/transientRetrySleeper are the host/test seams.

0.1.260

• fix(app): Copilot parity with the CLI — the app's model pickers now fetch through the copilot dialect (GitHub→Copilot token exchange, capabilities/limits) instead of 401ing the raw GitHub token against <host>/models; a picked Copilot model connects as the copilot wire dialect, and the entry-scoped FA_KEY_COPILOT_<NAME> key resolves as the fallback. Deleting a Copilot account now removes its entry-scoped token from the Keychain and the saved-keys store (before: only the shared FA_KEY_<HOST> slot went away and the token leaked); the duplicated name→key algorithm is gone — both surfaces use the one core CustomProviderRegistry.copilotEntryKeyName.

0.1.259

• feat(cli): fa.log now opens with fa boot sid=… version=… — the shared diagnostics log had no way to attribute a wedged "Working…" row to the BUILD that held it (today's 1949s spinner post-mortem stalled on exactly that: the wedged process predated the 0.1.255 busy-bracket fix, but only circumstantial evidence proved it). Every process now names its version next to its session id before any lifecycle line.

0.1.258

• fix(cli): the Copilot preset row now actually starts the connect flow — the "Add provider" picker shipped the row WITHOUT its handler (null?.call() closed the picker and nothing happened; the typed /provider copilot always worked). The handler map gained the copilot entry, and the picker test now asserts rows == handlers in BOTH directions so a dead row can never ship again, plus a line-mode routing test and a PTY visual test (screenshots 28/29/34) drive the real TUI from /provider to the "Copilot sign-in" step.
• fix(cli): fa --session <name> no longer opens another project's session — same-named sessions resolve with the LAUNCH folder first (a single local match wins); an ambiguous remainder (several in one folder, or none in it) auto-resolves to the most recently updated with a printed note naming every candidate id, the TUI offers the scoped "which one?" picker right after boot, and mid-session /session <name> asks with a numbered/wizard list. The switch itself now runs DETACHED under the flow gate (an awaited interactive pick deadlocked the sequential line REPL), redelivers lines typed during the switch instead of dropping them as flow junk, and redraws the idle prompt so the zeroed status meter shows.

0.1.257

• fix(cli): GitHub Copilot appears in the TUI "Add provider" picker — the preset list is hand-maintained and shipped without a Copilot row even though /provider copilot (device flow) worked; the row routes to the same connect command, and a picker test now pins every preset key to a handler so a future provider can't silently miss the menu.

0.1.256

• fix(cli): /model <id> on a saved CodeMie/custom entry no longer fails with 'no usable chain entry: set OPENAI_API_KEY' — the roles-mode pin carried the entry's key NAME, but the resolver's secrets never held the key material (the CodeMie JWT/SSO switch paths bypass the resolver), so the chain failed to resolve and the switch silently did nothing (the status line kept the old model). _switchModel now seeds the resolver with the session's live key under the pinned name (env-ring fallback) before pinning, and cookie-header auth (CodeMie SSO) keeps the direct model set — a cookie can never ride a Bearer chain.

• chore(sync): resolve the orphaned stash-pop conflicts to origin/main

• fix(cli): central busy bracket in _startRun — no more spinner-after-settle

• feat(fa_ui): always-visible custom answer in the ask sheet

• feat(apps): JsMediaHost wiring — video/audio nodes play for real

• fix(agent): abort wedged runs — SSE event-level idle watchdog + Agent.runIdleTimeout + fa.log run forensics

• feat(cli): env preconfig — required fields, base64 twins, all-roles default

• feat(app): chat text-size setting + dense markdown + unclipped copy

• feat(apps): adaptive layout host support (js_widget_runtime ^0.4.87)

• test(integration): fix latent env-key crash in real-model test

• feat(providers): Z.AI first-class + env-activated providers + FA_PROVIDER_* env preconfig

0.1.255

• fix(cli): the "Working… forever with an idle agent" wedge — the TUI busy bracket lived only in the submit handler, so runs triggered elsewhere (the idle inbox wake, a shell-job settle, a scheduled message) streamed with phase labels but released nothing: the spinner stayed on after the run settled. The busy bracket is now CENTRAL in _startRun (paired with the settle future) and the counter is the extracted, unit-tested BusyDepth — nested brackets (submit around run around wake) collapse to one edge pair. Forensics: fa.log showed run end with the spinner still up in two live sessions.

0.1.254

• Wedged "Working…"/"Compacting…" turns now abort instead of hanging forever. Two layered fixes after three live sessions (working, compacting, post-compact) pinned the busy row for 30-60 minutes with no recovery:
  • The SSE idle watchdog moved from the raw byte stream to a per-moveNext timer on decoded events: gateways keep dead generations alive with : comment heartbeat bytes, and the byte-level timer reset on every heartbeat — event-level silence is the honest signal. (Stream.timeout after the async* SseDecoder never fires at all — a Dart quirk this sidesteps entirely.)
  • New Agent.runIdleTimeout backstop (default 8 min, Duration.zero disables, onRunIdleTimeout callback): a run with no events outside tool execution is cancelled with a TimeoutException reason and ends as aborted. Streaming deltas re-arm, tool phases disarm — a long legitimate test gate never trips it.
• Run-lifecycle forensics in ~/.fah/logs/fa.log: one line per phase transition (run/turn/tool start|end, auto-compact start, watchdog fires), each tagged with the short session id — a wedged busy row can now be attributed to the exact phase (provider turn vs named tool vs compaction) that never finished, across parallel fa processes.

0.1.253

• fix(cli): the busy row stops lying about compaction hangs
• feat(catalog): category filter chips above the widget list
• refactor(cli): split for the 2800-line size gate
• feat(providers): multi-account ChatGPT — per-entry keys, account picker, app parity
• fix(providers): review fixes — string-typed oauth json, pre-send cookie baseline, SSE incomplete/failed as stream events
• fix: drop duplicate models_for_endpoint import left by the main merge
• fix(codex): pin chatgpt catalog visibility; note review fixes in changelog
• fix(codex): terminate failed/incomplete SSE turns as error events
• fix(codex): serialize OAuth expiry as a string; decode stays tolerant of int blobs
• docs(goal): codex gpt auth — tick checklist, fill implementation log
• feat(providers): unhide chatgpt provider — ship docs, changelog, live smoke
• test(providers): chatgpt oauth token-leak guard — persisted registry + transcript carry no tokens
• feat(providers): codex models — live GET /models with bundled-catalog fallback
• feat(providers): codex http sse transport — header parity, cloudflare cookie replay, full event coverage
• feat(providers): chatgpt oauth expiry tracking — expiresAt + needsRefresh
• feat(providers): codex transport helpers — header parity, cloudflare cookie jar, rate-limit parsing

0.1.252

• fix(cli): the busy row stops lying about compaction hangs — the pre-flight compaction's 'Compacting context…' label no longer stays up for the whole run (a long tool call read as a compaction hang); the busy row now names the executing tool ('Running bash…') and drops back to 'Working…' between calls; resumed sessions drop generation-time usage anchors, so a compacted branch no longer phantom-reports its pre-compaction size (no more no-op compaction on every resume + an honest context gauge).

0.1.251

• feat(catalog): stacked equal-width action buttons on tile rows
• feat(launcher): Open menu item + icons, icon-square-only hover
• feat(app): interactive live tiles via widget.interactive opt-in
• feat(app): drop bundled demos — the catalog is the source of apps
• fix(app): actually pass sessionInfoNames to the sidebar (torn-write casualty)
• feat(app): the sidebar shows the CLI-written session_info names
• fix(app): reset/remove/install refresh the grid + sane dialog width

0.1.250

• feat(providers): ChatGPT goes multi-account — each account saves as its own named entry with its own name-scoped secure-store slot (FA_KEY_CHATGPT_COM_<NAME>, the CodeMie per-entry key pattern), so a refresh-token rotation never overwrites a sibling account's blob and the refresh callback writes to the ACTIVE entry's slot.
• feat(provider): /provider chatgpt oauth offers the saved ChatGPT accounts first (switch to one, or add another) before running OAuth; a NEW account gets a guided model pick from the live Codex /models (bundled default when the fetch answers nothing), while a re-login to the SAME entry keeps its last-used model.
• feat(app): the ChatGPT OAuth flow names the entry from the OAuth account's email (id_token claim) and matches re-auths by name + baseUrl — a second ChatGPT account lands in its own entry and its own entry-scoped Keychain slot, mirroring the CLI.
• feat(chatgpt): the ChatGPT provider (Codex backend) is now visible across the pickers and /provider — it ships with the real Codex HTTP SSE transport: Codex header parity (originator, session/thread ids, account id), cloudflare cookie replay, and a single challenge retry.
• feat(chatgpt): OAuth access tokens refresh proactively (expiry is tracked on the credentials blob) and the rotated blob is re-persisted.
• feat(chatgpt): reasoning deltas surface as thinking blocks; response.incomplete / response.failed end the stream as a terminal error event, preserving any partial text already streamed.
• fix(chatgpt): OAuth credentials toJson returns Map<String, String> again (expiry serialized as an epoch-millisecond string); decoding still accepts blobs with a raw int expires_at.
• fix(chatgpt): the cookie-replay baseline is the header the request actually carried (snapshotted before the send), so the single challenge retry fires only when the jar truly learned a new cookie.
• feat(chatgpt): /models probes the live Codex /models endpoint and falls back to the bundled catalog on 401 / challenge / malformed bodies.
• feat(memory): the long-term-memory LLM slot is wired — a new HarnessLlmProvider adapts fa_llm's LlmProvider onto the harness streaming contract and is resolved per call (memory role → smol → main) in BOTH the CLI and the app. Memory consolidation and semantic search now actually run instead of being silently skipped.

0.1.249

• fix(app): self-heal stale-catalog sha mismatches + card-ify the catalog list
• test(app): hold-release on a classic tile now expects the menu
• fix(app): installed widgets really swap Preview for Remove + tests
• fix(app): re-add the _localApps field declaration (clobbered again)
• fix(app): restore the _localApps field + refresh (clobbered mid-commit)
• feat(app): catalog sheet — Remove for installed, Created by me, real avatars
• fix(app): tile menu opens for plain icon widgets + right-click + soft hover
• fix(site): openPreview really uses the RUNNER url — the torn write had resurrected the /widgets/preview/ path
• fix(site): repair torn index.html (duplicate tail after ) + restore the app-only marks
• feat(site): platform widgets are marked 'runs in the Fa app' instead of a broken preview
• fix(site): preview iframe points at the jsr repo's own Pages runner
• feat(site): widget preview runs the real Flutter/jsr runner — DOM shim removed

0.1.248

• fix(cli): a run counts as busy from the moment it is STARTED, not from the first streamed byte — the inbox watcher / shell-job settle path no longer starts a parallel run during the pre-flight compaction window (live session showed Bad state: Agent is already processing a prompt right after [auto-compacted]).

• feat(app,widgets): catalog entries parse the optional platforms manifest field; the catalog sheet shows iOS/macOS platform chips.

• feat(site): widget gallery cards show platform tags and no longer wrap the size/jsr ≥ text or the Preview/Download button labels.

• feat(provider): GitHub Copilot as a first-class provider — catalog entry plus the /provider copilot CLI flow (GitHub device flow with user_code + verification_uri, or paste an existing PAT; works headless) and the app's fa_ui Copilot connect sheet.

• feat(providers): Copilot protocol core in lib/src/providers/ — copilot.dart (streamCopilot: token exchange, mandatory Copilot headers, errors-as-events), copilot_oauth.dart (short-lived token

  • header builder), copilot_device_flow.dart (grant + poll).

• feat(keys): entry-scoped FA_KEY_COPILOT_<NAME> secure-store keys with an env-first _2… ring — CI supplies keys without a store.

• feat(models): live copilot /models dialect — the GitHub token is exchanged for the Copilot token, capability/limit fields parsed.

• feat(provider): multi-account isolation — each GitHub account saves as its own named entry (copilot-<login>); re-auth updates only its own entry.

• fix(providers): copilot kind-dispatch audit — every provider-kind dispatch in lib/src now handles copilot: inspect_image streams through streamCopilot (GitHub token exchanged for the short-lived API token, never sent as the Bearer header) and the roles fallback chain accepts copilot entries (providerStreamFunction → _CatalogStreamFunction → streamCopilot). A single dispatch-chain test drives role takeover + inspect_image + /models in one run.

• fix(models): copilot /models limits override — the catalog's static copilot contextWindow: 1000000 / maxTokens: 32768 are wrong for several real models (400s from the backend); the endpoint-reported capabilities.limits.max_context_window_tokens / max_output_tokens now replace them per model when present, and the catalog defaults stay when the payload has no opinion.

0.1.247

• feat(tools): Hailuo 2.3 video dialect + headless pre-flight compaction
• fix(tui): key parser never swallows a trailing control byte into a text run
• fix(cli): restored sessions label the model with the pinned-key account

0.1.246

• feat(tools): HailuoVideoDialect — Hailuo 2.3 video generation on the MiniMax V1 contract (POST /v1/video_generation → poll task_id → file_id → /v1/files/retrieve → download_url), with size mapping (1080P/768P) and full MockClient flow tests. Dialect registry reordered most-specific-first: both MiniMax dialects share the minimax baseUrl marker, so the H3 dialect previously swallowed Hailuo endpoints into the wrong (V2) contract.
• fix(cli): headless runs compact BEFORE the first request — a resumed session already over the threshold went out over-window and the endpoint rejected the turn (the same pre-flight guard the REPL had).
• fix(tui): key parser never swallows a trailing control byte into a text run — a burst PTY read delivering text\r in one chunk ate the Enter into the text and the composer never submitted.
• chore: drop the build artifacts (fa-local/bundle/…) from git — install_local.sh rebuilds them and the directory was already gitignored (they were force-added once and churned 13 MB per install).

0.1.245

• fix(cli): restored sessions label the model with the RIGHT account too — 0.1.244 covered live switches via _activeCustomName, but a restarted session has none, so two same-endpoint accounts (kimi-ira1/kimi_me) still showed the first registry entry. The status label now disambiguates by the apiKeyName the persisted roles chain pins for the endpoint (_endpointEntryLabel + _chainKeyNameFor); the first-match scan remains the pin-less fallback.

• fix(cli): status bar labels model with the ACTIVE saved provider

0.1.244

• fix(cli): the status bar labels the model with the ACTIVE saved provider entry, not the first baseUrl match — two accounts on one endpoint (e.g. kimi-ira1 + kimi_me both on api.kimi.com/coding/v1) used to show the first registry entry's name no matter which one was picked (/model kimi_me → status read kimi-ira1/<model>). _statusProviderLabel now prefers _activeCustomName and only falls back to the endpoint scan when no custom entry is active; the key resolution was never affected.

• fix(cli): media slot flow propagates custom provider keyName

0.1.243

• fix: gen_prompts trims description trailing newline
• fix: skip subagent integration tests when MiniMax key missing
• feat: v0.1.242 — MiniMax media picker fix + generate_video tool
• fix(provider): minimax /model picker shows the full catalog, not the saved modelId

0.1.242

• feat(catalog,ui): remote models catalog + thinking markdown + table CRAP fix (0.1.241)

0.1.241

• feat(catalog): remote provider-models catalog at fa1.dev/models-catalog.json preloaded once per process — fills contextWindow defaults the MiniMax-style /v1/models endpoint doesn't publish (M3=1M, M2.x=204800) and ships per-slot media-model lists (imageGeneration/videoGeneration/speech/transcription) the chat endpoint never returns. Pickers stay endpoint-driven: the catalog NEVER seeds the chat model id list (the provider's own /v1/models is the source of truth) — it only enriches metadata and the media slots. Endpoint-reported values always win; failures are silent (10s budget, never throws); pickers keep their manual-entry fallback. RemoteCatalogEnrichment (lib/src/providers/) is the host seam — preloaded from bin/fah.dart at boot.

• fix(tui): thinking-streaming deltas get inline markdown (bold, italic, inline code) inside their dim wrapper — reasoning snippets that quote function names or emphasise alternatives render properly instead of leaving stray **…** in the dim background. Renderer is a small single-pass scanner split into _matchInlineSpan + per-marker helpers (lib/src/cli/ansi_markdown.dart); block constructs (fences, headings, lists) still belong to AnsiMarkdown for full answers.

• refactor: AnsiMarkdown._renderTableRows split into _tableFragments / _tableSeparator / _renderTableRow / _renderTableLine — CC dropped, CRAP ratchet back to 12.00 (pre-existing method was 16.27 at 69% coverage; now 12.00 at 100%).

• fix(compaction): 10-minute attempt budget — a hung summarizer can no longer wedge the turn (0.1.240)

0.1.240

• feat(catalog): remote provider-models catalog at fa1.dev/models-catalog.json preloaded once per process — fills contextWindows the endpoint didn't publish, suggests a defaultModelId, and ships per-slot media-model lists (imageGeneration, videoGeneration, speech, transcription) for providers like MiniMax whose chat endpoint doesn't return their media models. Endpoint-reported values always win; failures are silent (10s budget, never throws); pickers keep their manual-entry fallback. The catalog lives in RemoteCatalogEnrichment (lib/src/providers/) — preloaded from bin/fah.dart at boot.

• fix(tui): thinking-streaming deltas get inline markdown (bold, italic, inline code) inside their dim wrapper — reasoning snippets that quote function names or emphasise alternatives render properly instead of leaving stray **…** in the dim background.

• fix(compaction): a summarizer endpoint that accepts the request and never answers can no longer wedge the turn on the compaction spinner forever — every summarization attempt now runs under a 10-minute wall-clock budget (AutoCompactor.attemptBudget); the timeout fails the attempt (not transient — no retry spin), the pass falls through to the next summarizer or the honest local trim, and the turn goes on.

• fix(compaction): the emergency local trim zeroes the kept generations' usage anchors — a stale generation-time anchor kept the post-trim estimate over the window and retriggered the compactor every turn.

• fix(app,test): the whole quality gate goes green — CRAP ratchet, app suite, l10n and goldens (0.1.239)

• fix(replay): compact system-notice rows in restored transcripts (0.1.238)

• fix(provider): entry-name status label + persist catalog model picks (0.1.237)

• docs(goal): codex gpt auth — cross-platform notes (Rust portability does not transfer)

• docs(goal): codex gpt auth — point the reference at the open-source GitHub repo

• docs(goal): codex gpt auth — ship the ChatGPT/Codex-backend provider

• perf+fix: growing-tail throttle, codemie scoped key, status bar (0.1.236)

• docs(goal): copilot — translate to English, add tickable implementation checklist

• docs(goal): copilot — multi-account contract, mandatory TDD plan, Keychain-only tokens

• docs(goal): copilot provider — multi-account support as first-class scope

• docs(goal): add copilot provider goal — migrate copilot-proxy-go protocol into fa_llm

• deps: flutter_agent_memory ^0.1.1 (0.1.235)

• fix(memory): tombstoned delete via KBMemoryStore; hosted dep (0.1.234)

• feat: schedule_message + bash stdin param; crash hardening (0.1.233)

• perf(tui): O(1) in-place rollback of grown-tail arrays (0.1.232)

• perf(tui): flush streamed output every 16ms (0.1.231)

• perf(tui): cache formatted sticky echo rows (0.1.230)

• feat(memory): memory_delete tool; quote-style service blocks (0.1.229)

• fix(tui): add the missing system_notice_render import in fa_tui (0.1.228 fixup)

• feat(tui): render system notices as dim blockquotes with a gear marker (0.1.228)

• perf(tui): ASCII fast path + bounded line-width memo in tuiTextWidth (0.1.227)

• fix(tui): hoist terminal reset before SIGINT use; wire compaction delta tail (0.1.226 follow-up)

• fix(tui): restore terminal modes on Ctrl+C exit (0.1.226)

• fix(tui): keep the input caret visible while a run streams (0.1.225)

• fix(messaging): cross-project agent_message delivers to the recipient root (0.1.224)

• feat: instant key echo + Ctrl+C exits like /exit (0.1.223)

• fix(cli): visible hint after mid-run Ctrl+C abort (0.1.222)

• chore(tool): land the frame-build bisect probe (gitignored path)

• fix(tui): no-newline stream deltas stay incremental — typing lag gone (0.1.221)

• chore(tool): land the key-latency probe (path is gitignored)

• feat(tools): warn when a stale fa build writes old-format job logs here (0.1.220)

• fix(cli,compaction): paste-a-path sends, honest compact status, bounded extraction, per-delta ctx memo (0.1.219)

• fix(cli): FaTuiController stub carries setBusyPhase — web build compiles again

• fix(app,site): widgets entry points visible and verifiable

• docs(widgets): mark C2/M1 app-catalog milestone shipped with live checks

• feat(app): widgets catalog — install from fa_widgets releases, slim the bundle

• test(tools): assert shell job ids by shape after unique-id change

• chore(messaging): skip messages-registry write when unchanged

• fix(tools): collision-proof background job ids and image filenames

• fix(session): serialize JSONL writers per file and quarantine torn lines on open

• chore(local): rebuild fa-local bundle binary at 0.1.218

• fix(tui,fork): serialize tracer writes through a flush queue

• fix(tui,fork): FA_TUI_TRACE file sink flushes per event (crash-safe)

• perf(tui,fork): render-aware cursor dedupe — zero-byte idle frames

• perf(tui,fork): lazy frames, output dedupe, keypress-paint tracer

• fix(tui): replay restored messages in full — no per-message head caps

• fix(tui): wrap overflowing table cells — keep the box grid readable

• perf(tui,fork): drop-frame fps throttle — stop sleeping the event loop

• chore(vendor): inline dart_tui 2.0.0 under vendor/ for perf work

• perf(tui): amortized history-cap trim — no full re-parse per streaming flush

• feat(tui): truthful busy-row phase labels

• feat(cli): pre-flight context guard — compact BEFORE an over-window request

0.1.239

• fix(app): the macOS app builds again against the current core — the harness barrel now exports ScheduledMessageQueue + scheduleMessageTool, the service declares its _scheduledMessages field and arms the queue, the fabric repo's root: is a plain String, _AutoCompactorFlutterHooks implements onDelta, and the app's flutter_agent_memory constraint moved to ^0.1.1 (the stale lock had resolved harness 0.1.218).
• refactor(cli): the perf-sprint methods fit the CRAP ratchet again — pure extractions (busy-edge sendBusy, _rollingTail, _maybeSwitchToSavedEntry, fold-XOR fence scan, _needsRewrite/_nextOpenTag/_taskOpenerEnd, _resumable/ _tailThrottled, _replayUserTui/_chromeMarkerLine); no behavior change. Max CRAP back to 12.00.
• test(app): the flutter_app suite is green under the hook again — the compaction-failure test matches the honest local-trim contract, the drawer tests drain the real-async lazy open, the map goldens skip when the demo moved to the widgets catalog, the ctrl+s visual test cancels the full sh-<n>-<uniq> job id, the /agents visual test seeds mail under the app-group sessions root, the corrupt-session test matches the quarantine-on-open heal, and WidgetsCatalogSheet passes the l10n + golden-coverage guards (new arb keys en/ru + two goldens).

0.1.238

• fix(replay): a restored <system-notice> message (background-shell job settle, inter-agent mail, task result) replays as ONE dim chrome line instead of the raw block — the full multi-line command dump, log paths and the closing tag no longer wall up the transcript after a resume. Mixed content still replays verbatim.

0.1.237

• fix(status): the provider shown next to the model is now the saved provider ENTRY name (z.ai, codemie-personal, …) matched by endpoint — the model's provider field carries the catalog protocol kind ("openai"), which read as "it switched to OpenAI" although the pick was z.ai.
• fix(provider): /model switches on catalog providers (no active custom entry) now persist the picked model — the host's onModelChanged fired only for saved-entry switches, so a restart restored the last provider switch's model (gemini from a codemie test) instead of the one the user chose.

0.1.236

• fix(roles): model switches on NON-catalog endpoints (CodeMie SSO, DIAL, self-hosted) resolve the endpoint-scoped store key (FA_KEY_
• perf(tui): growing-tail throttle — a streamed paragraph whose last line keeps growing (a long thinking burst with no newline) re-formats + re-wraps that line on every 16ms flush, O(line length) per tick, which saturated the event loop and froze the screen (neither the thinking nor typed input rendered). Expensive tails (>8k chars) now re-render at ~10 Hz and complete instantly on the final newline; short lines and tests keep the byte-exact immediate path (debugTailThrottled counter).
• fix(codemie): the SSO wait now shows a live "still waiting…" status and bails after 5 minutes instead of hanging forever — the page usually signs in by itself (existing browser session), so the callback just needs patient waiting, not user action.
• fix(status): hide the $0.0000 price when the model has no cost data, humanize big token counts (12.3k / 4.6M), and show the provider name with the model in the status bar. MemoryRevisionService (+ MemoryRevision, ConcurrentRevisionException) are now public exports, so pendingDeletions/markConsolidated and the consolidate(expectedRevisionHash:) contract are reachable directly.

0.1.234

• fix(memory): memory_delete now goes through KBMemoryStore.deleteRecord (flutter_agent_memory 0.1.0) — tombstones + revision bump mean consolidation can no longer resurrect deleted entries; graph rebuilds on delete. deleteEntityById stays as a raw-storage escape hatch only.
• deps: flutter_agent_memory flipped from git tag to hosted ^0.1.0 (published on pub.dev) — unblocks fa's own future publish.

0.1.233

• fix(tui): a throwing update()/picker command can never close the app — dart_tui wraps model.update in try/catch (log + keep running), matching the existing runCmd guard. Cross-provider model switches surface errors as notices instead of dying (the CodeMie crash).
• fix(cli): stdout.terminalColumns/terminalLines fall back to 80/24 when stdout is not a TTY (session-switch replay crashed twice in crash.log).
• fix(tui): sendBusy is reference-counted — a second submit during a running turn (slash menus work mid-stream) no longer resets the elapsed timer + sticky echo, and its finally no longer kills the spinner of the live stream.
• feat(tools): bash gains a stdin param — text written to the command's stdin right after start (ask the USER for a passphrase via the ask tool, then feed it; ssh-add/sudo no longer hang on a raw prompt).
• feat(messaging): schedule_message — persisted delayed notes to an agent mailbox (self by default); pending records survive restarts and delivery rides the inbox idle-wake. Wired into the CLI and the app.
• deps: flutter_agent_memory 0.1.0 (git tag v0.1.0) — tombstoned deletes, timeout-guarded search, public graph-overview API.

0.1.232

• perf(tui): grown-tail rollback truncates the transcript arrays in place instead of sublist-copying them per streaming chunk — the per-chunk cost no longer grows with session length (the last time-coupled term; everything else is bounded or O(delta)).

0.1.231

• perf(tui): streaming flush interval 50ms → 16ms (~60 fps). A traced real huge session showed p50 frame build of 37µs, so flushing thrice as often is nearly free and streamed text (thinking included) appears smooth instead of in 50ms chunks.

0.1.230

• perf(tui): cache the formatted sticky echo. During a stream every keystroke forces a paint, and each paint re-ran markdown formatting over the entire echoed prompt — typing cost was O(echo lines) per key, painful with long prompts in big sessions. Rows are now formatted once per echo change (content+width keyed cache carried across model copies); view frames just write the cached bytes.

0.1.229

• feat(memory): memory_delete tool — remove stale entries by id (scan project then user scope, or restrict with scope); hosts get the prompt-section refresh hook like memory_add.
• feat(tui): <task-result> blocks and [auto-compacted] / [context trimmed] / [memory maintained] service receipts render as dim blockquotes with a ⚙ marker, same as <system-notice>.

0.1.228

• feat(tui): <system-notice> blocks (background-shell settlements, inter-agent mail) render as dim blockquotes with a ⚙ marker instead of raw tags — session notes look like notes. Model-visible records keep the raw tags; only the TUI view restyles them.

0.1.227

• perf(tui): width measurement fast path — printable-ASCII runs skip grapheme iteration entirely and a bounded whole-line memo (8192 entries, insertion-order eviction) catches re-measured rows. Profiled with crap4dart on the transcript suite: wrapAnsiLine 30.7s → 0.62s total (×49); tuiTextWidth/tuiGraphemeWidth dropped out of the top-10 hot methods; formatLine total 37.0s → 32.8s.

0.1.226

• fix(tui): Ctrl+C exit now restores terminal modes before exit(130) (mouse tracking off, alt-screen exit, cursor show) — the shell prompt no longer inherits mouse reporting, so wheel scrolling after fa quits stops printing escape garbage. The reset block is shared by the natural end and the SIGINT path.

0.1.225

• fix(tui): the input caret stays visible while a run streams. The old hide-while-busy guarded a pre-force-home artifact (the cursor jumping inside streamed text); the renderer now re-homes after every painting frame, so hiding only stranded invisible typing. Pickers and prompt dialogs still hide it.

0.1.224

• fix(messaging): agent_message to a peer in ANOTHER project now lands in the recipient's messages root. Delivery used to write into the sender's cwd-slug root, so the other fa (draining only its own root) never saw the mail — cross-project chats silently vanished. FileMessagingRepository resolves the mailbox's real root (messages-registry.json slug map, then a sibling-slug scan by .id marker); unknown ids stay local. Diagnosed live: a message fa<->crap4dart agent was found sitting in the wrong project's inbox.

0.1.223

• feat(tui): input-driven frames bypass the fps throttle — a keystroke or wheel event paints immediately instead of waiting out the ~18ms frame window (kitty's input_delay/repaint_delay split). Measured echo latency during a 500-delta/s stream over a 20k-line transcript: p50 12-17ms → 0.1-0.2ms, p95 ≤0.4ms.
• feat(cli): Ctrl+C now exits exactly like /exit — aborts any in-flight run (bounded 5s wait), persists the partial transcript, prints the fa --session '...' resume hint to the real stdout, exits 130. Esc stays the abort-without-exit key inside the TUI. Adds AgentCli.waitForIdle (bounded settle) and sessionResumeHint.

0.1.222

• fix(cli): Ctrl+C while a run streams now prints "run aborted — press Ctrl+C again to exit". The first press always aborted the run but gave no feedback, so the suddenly-silent screen read as a hung process (PTY-verified: the SIGINT abort path works; only the hint was missing). A second press exits 130 as before.

0.1.221

• fix(tui): streaming without trailing newlines no longer force-rebuilds the whole transcript render — a delta that grows the current line (the coalescer's exact no-newline flush shape) used to break the incremental boundary's identity sentinel and trigger a full format+wrap pass (~220ms on a 150k-token transcript) on EVERY flush, saturating the UI loop so typing during a stream rubber-banded or froze. A prefix-extended tail now rolls the durable caches back one source line and resumes: measured 4327ms → 15ms over 20 growing appends; appends with newlines stay O(delta) as before.

0.1.220

• feat(tools): one-per-session warning when a freshly written old-format background-job log (sh-<n>.log, the pre-unique-id scheme) appears in this directory — proof a stale fa build shares the cwd and its job output can interleave with stale logs; surfaced as a loud CLI hint to restart that instance on the current binary.
• chore(tool): tool/key_latency_probe.dart — headless key-echo latency probe for the TUI (stream + burst + status-line scenarios), used to verify that typing during an active stream stays sub-frame on current builds.

0.1.219

• fix(cli): a pasted absolute path that EXISTS is sent as a message with the file attached instead of being refused with a "filesystem path, not a command" hint; a nonexistent path keeps the hint.
• fix(cli): auto-compaction reports honestly — a failed pass no longer prints the success-looking "[auto-compacted] N tokens summarized" line, and a no-op pass stays quiet.
• fix(compaction): emergency local trim — when both summarizers are down and the transcript is over the window, the most recent keepRecentTokens stay live behind a user-role marker (in-memory only; the session file keeps the full history) so the agent can keep working instead of being stuck over-window until the endpoint recovers.
• fix(cli): compaction-time memory extraction is bounded — the extraction stream is cancelled after 90s and force-skipped after 120s (the phase label no longer hangs for the whole role-chain retry ladder), and the "Compacting context…" phase is restored afterwards.
• perf(cli): the status line's context estimate memoizes the settled transcript by list identity + length only — the in-flight stream message is estimated per render and never invalidates the memo. Keying the memo on the stream's growing length used to force a full O(context) re-scan on EVERY streamed delta (dozens per second — the "typing lag" while a run streams over a large transcript).

0.1.218

• fix(tui,fork): the FA_TUI_TRACE file sink flushes every row — a traced session is the one you post-mortem, so a hard kill can no longer eat the trace tail. Opt-in only; zero cost when tracing is off.

0.1.217

• perf(tui,fork): cursor home dedupe is render-aware — a frame that painted rows force-homes the physical cursor exactly once; fully idle frames now emit ZERO bytes (the identical-content skip no longer pays the CUP).
• perf(cli): drop the nonce-SGR suffix from the idle cursor line — the renderer's render-aware re-home replaces it, so idle appends stop repainting the status row every frame.

0.1.216

• perf(tui,fork): dropped frames no longer build the view — Program now builds the screen lazily inside the frame budget check, so frames the fps throttle discards stop paying a full assembly for nothing.
• perf(tui,fork): renderer output hygiene — the window title OSC sequence and an unmoved cursor's CUP are emitted once (deduped, invalidated on clearScreen/alt-screen/scroll/insert) instead of on every frame.
• feat(tui,fork): keypress→paint tracer — attach withTracer(...) or set FA_TUI_TRACE=<path> to get an ordered JSONL timeline of stdin arrivals, drained batches and painted/dropped frames (build_us/render_us), for offline latency joins.
• perf(cli): the frame builder stops allocating a List

0.1.215

• fix(cli): restored sessions replay messages IN FULL — the per-message head caps (TUI: 20 rows, line mode: 2 rows) hid long answers behind a trailing ellipsis after a restart. The global row budget still bounds a marathon replay, but it now drops OLDER WHOLE messages instead of decapitating every entry; tool-call-only runs keep collapsing.

0.1.214

• fix(cli): wide GFM tables keep their box grid — cells wider than the terminal wrap onto aligned continuation rows instead of collapsing the whole table back to raw markdown; only a degenerate budget (tiny width, many columns) still prints raw rows.
• perf(tui): dart_tui vendored under vendor/ and its fps throttle fixed to drop-frame semantics — an early frame no longer sleeps the event loop (~16 ms stalls per frame made typing/scrolling rubber-band during long streaming answers); drops repaint only the latest view via an internal RenderTickMsg, invisible to models.

0.1.213

• style(env): brace single-statement if in CwdOverrideEnv.backgroundJobsSupported

0.1.212

• fix(bench): drop stale fa_tui show-import from tui_stream_bench
• fix(tui): open-table boundary invariant — streamed tables never lose rows
• feat(site): widgets gallery page + machine index; widgets GOAL
• perf(tui): incremental transcript markdown+wrap — streaming flush x729 faster
• fix(cli,app): attach delivery, pasted-path attachments; split cli inbox part file

0.1.211

• fix(cli): a pasted filesystem path is not a slash command

0.1.210

• ci(mobile): skip the iOS artifact download when the IPA build was skipped

0.1.208

• feat(attach): live CLI sessions in the app — presence, 1:1 view, input handover

0.1.207

• fix(app): explain empty responses that follow an image-bearing prompt

0.1.206

• fix(app): skill discovery scans user-level roots (~/.claude, ~/.copilot, ...)
• ci(macos): tolerate an existing keychain when packaging the TestFlight PKG

0.1.205

• fix(cli): TUI renders rows in terminal cells (grapheme clusters) - markdown no longer slides after emoji/CJK; status row padding and menu truncation are cell-width aware; output history cap raised 200 -> 2000 (fence-safe replay).
• feat(cli): unattended approval mode - auto-approve everything including critical bash patterns, for runs without a user present; yolo regains its critical-pattern safety net.
• feat(cli): approval prompts accept a typed note and deliver it to the agent as steering feedback alongside the decision.
• fix(cli): empty assistant responses retry once with a 'continue' nudge (loop-level maxEmptyRetries); errors/aborts never trigger it.
• feat(session): incremental message persistence (crash-safe appends); sessions with persisted records survive empty-session cleanup.
• fix(cli): status meter tracks streaming thinking/tool-call sizes; session switches reset tok/cost/turn; failed runs keep the last real ctx anchor.
• feat(cli): skills access defaults to granted (opt-out); bare /skills access opens an interactive picker; disabled-skills hints on listings; /skills dispatch decomposed for the CRAP ratchet.
• fix(providers): OpenRouter OAuth keys persist endpoint-scoped (FA_KEY_OPENROUTER_AI) and the auth picker offers the stored key first.
• feat(tui): mouse capture defaults ON (two-finger scroll); FA_TUI_MOUSE=0 opts out.
• test(cli): waitForIt poll budget 2s -> 25s (coverage runs on loaded boxes).

0.1.204

• fix(providers): do not close shared HTTP client during OpenRouter OAuth exchange

0.1.203

• fix(cli): provider picker, CodeMie auth refresh, skills access
• feat(session): unify CLI and macOS app session storage
• fix(providers): restore Kimi endpoint (api.kimi.com/coding/v1) and default model k3
• fix(cli): auto-refresh expired CodeMie SSO cookie on startup

0.1.202

• fix(install): remove broken Dart fallback, respect FA_INSTALL_DIR, sign macOS CLI in CI

0.1.200

• fix(ios): correct force_load path — pod products live in a per-pod subdir
• fix(ios): force-load cupertino_http pod binary + CI gate on its FFI symbols
• fix(flutter_app): project mount sets agent cwd to /project so sessions are folder-scoped

0.1.199

• fix(cli): show auth-method picker when adding openrouter/codemie from TUI

0.1.195

• fix(install): macOS CLI bundle + quarantine/sign handling

0.1.194

• fix(ci): merge Release.entitlements into the signed macOS build, don't strip them

0.1.191

• feat(cli): add auth-method picker for CodeMie SSO/JWT and OpenRouter OAuth/key
• feat(cli): auto-restart CodeMie SSO when the saved cookie expired
• fix(cli): catch uncaught errors and harden provider switch against crashes
• fix(ui): make ChatComposer transparent and regenerate goldens
• fix(network): use platform HTTP client for sandbox env, allow local HTTP, log bookmark failures
• fix(macos): add app-scope bookmark entitlement and surface folder picker errors

0.1.188

• feat(app): js_widget_runtime back on hosted pub (^0.4.79, the git pin dropped) — JS apps gain the Material 3 catalog: appBar/navigationBar/ navigationRail/tabBar/fab/segmentedButton/radio/searchBar/tooltip/ popupMenu/banner/bottomAppBar/carousel/drawer, modal overlay nodes (bottomSheet/dialog/snackBar/datePicker/timePicker), flChart (line/bar/pie/radar/scatter via fl_chart), new layout/display/input nodes (wrap/align/flexible/spacer/scroll/clipRRect/svg/markdown/chip/badge/ progress indicators/switch/checkbox/slider/dropdown, textButton/ outlinedButton/iconButton), M3 motion tokens — and the seeded js-apps skill documents the whole catalog
• fix(providers): one shared keep-alive HTTP client for all provider streams instead of a fresh client per call — per-turn TCP churn piled up TIME_WAIT sockets until connect() stalled into the watchdog (TimeoutException: Future not completed); kimi-cli/pi reuse one client for exactly this reason. Aborts still close just their own response subscription
• fix(compaction): auto-compaction no longer spins identical no-op passes to the max after a compaction: a pass with nothing left to cut stops the loop, and the post-pass transcript restamp clears the kept messages' stale generation-time usage so the estimate reflects the compacted size (before, a compacted session kept reporting its pre-compaction ~200k and retriggered compaction on every idle wake)
• fix(tui): the status row is padded to the terminal width — switching from a long model id to a shorter one no longer leaves the old tail on screen
• fix(cli): the status line is live again — ctx% shows the current context pressure (provider-reported usage plus the estimated tail, i.e. what the next request carries, not the last turn's frozen prompt size) and the tok counter grows during streaming (settled turns + in-flight estimate)
• fix(providers): inline <think>…</think> tags in the content stream (kimi k3 via openai-completions and other endpoints without a reasoning field) are extracted into the thinking block instead of rendering as raw tags — streaming-safe (tags may split across deltas)
• fix(tui): history wrapping no longer breaks markdown styling — the wrap is now word-aware (no more mid-word cuts) and active SGR styles are closed at the cut and re-opened on the continuation row, so a wrapped bold/code span keeps its style instead of going plain (visible mostly after resizes)
• fix(tui): restored sessions no longer lose markdown formatting after an unclosed code fence — replay truncation (20-row cap) and the replay budget cut could leave the fence state dangling, rendering everything after as verbatim text; truncated messages now close their fence synthetically and a budget cut starting mid-fence prepends a balancing opener
• fix(tui): restored sessions no longer dump the raw <summary>/<read-files> block into the history — a projected compaction/branch summary replays as one dim marker row with the summary's first line as a hint
• feat(config): provider watchdogs are configurable — providerTimeouts: section in ~/.fah/config.yaml (connectTimeoutMs, streamIdleTimeoutMs; strict parsing) for slow endpoints whose first byte takes minutes; the connect default itself is raised 60s → 180s (loaded reasoning endpoints hold big requests before the first byte) and the stream-idle default 120s → 5min (reasoning models think for minutes between chunks — pi and kimi-cli carry no such watchdog at all / SDK's 600s total)
• fix(tui): Cyrillic/CJK paste no longer arrives as mojibake — dart_tui 2.0.0's bracketed-paste decoder maps every pasted byte to a Latin-1 char code; fa re-decodes the (lossless) mis-decoded text as UTF-8 at the PasteMsg boundary
• fix(tui): Ctrl+S no longer freezes the terminal — dart_tui's raw mode left termios IXON on, so Ctrl+S was the tty driver's VSTOP (XOFF: output froze, the keypress never reached the app); the TUI now disables software flow control for its lifetime and restores the saved termios on exit
• feat(tools): background shell jobs — bash background: true runs detached (log in .fah/bash_jobs/<id>.log), bash_job {status|output|stop} manages them, completions re-enter the conversation as system-notices (steered mid-run, fresh turn while idle)
• feat(agent): steer soft-yield — a message arriving mid tool-call phase (Ctrl+S, subagent completion, inbox mail) asks yield-aware tools to finish the call early WITHOUT stopping the work: a running foreground bash moves to a background job untouched, a blocking task converts still-running children to background jobs; the user message is delivered at the next step boundary instead of after the whole tool call
• feat(task): model-facing task_cancel aborts a running background subagent job; /tasks lists background agents AND shell jobs, /tasks cancel <id> routes by id
• feat(app): background shell jobs on every platform — the desktop app forwards the host shell's capability through ProjectMountEnv; mobile's WasiSandboxShell and web's MemoryShell run jobs as detached script Futures on job-local interpreter clones (own cwd/vars/output capture, shared filesystem), so bash background: true + steer-yield work in the sandbox too

0.1.187

• refactor(ui): sane Settings structure — providers include on-device, one Models group

0.1.186

• fix(crap): decompose FaTuiModel._submit/_wrappedInput (CRAP 12.14/12.01 -> in budget)
• refactor(cli): split provider key helpers out of provider_commands.dart (2800-line gate)
• fix(providers): roles-mode switches preserve the scoped key; messaging presence
• test(cli): visual integration coverage for the new TUI UX
• feat(tui): shell-style input history on ↑/↓
• feat(messaging): live cross-instance chat — discovery, self-address, idle wake
• test(app): web-safety guard — no Platform.is without a kIsWeb guard
• test(messaging): visual integration for the agents inbox + terminal-safe marker
• fix(app): no crash on CodeMie/ChatGPT taps in the web build

0.1.185

• feat(tui): leave the mouse to the terminal by default (FA_TUI_MOUSE=1 to capture)
• fix(app): iOS CodeMie SSO — run the real loopback callback server

0.1.182

• test(messaging): drop unused import (dart analyze warning)

0.1.180

• feat(memory): keyword-only search fallback without an LLM provider
• fix(subagents): fire-and-forget registry persistence + per-session rehydrate
• fix(providers): connect + idle watchdogs on provider streams
• feat(ui): brand provider icons in onboarding, file split, iOS icon sync
• feat(session): never keep an untouched session file
• chore: drop unused import
• fix(cli): provider wizard asks for the key in roles mode; switches accept it
• feat(subagents): crash-resilient child sessions — spawn-time async creation + incremental turn flush (serialized, fire-and-forget)
• fix(memory): inject the
• test(cli): drop flaky real-model live badge visual test (unit-covered badge logic); document why

0.1.179

• fix(crap): simplify pickAgentAction dispatch; broaden badge visual soft-skip catch
• chore: drop unused session_repo imports
• feat(cli): extract active-agents badge to pure helper + unit tests; soft-skip live badge visual test
• fix(tui): soft-wrap long input lines instead of horizontal clipping
• feat(subagents): real JSONL child sessions at completion + /agents open
• feat(providers): list-first model pickers everywhere — quick search, manual escape, agent models flow
• feat(cli): /agents child → Open session action — switch into the subagent's session

0.1.178

• feat(app): live agents badge in FaWorkBar (CLI bg: parity) + fix cli_visual tests for new settings hub order and /agents tree

0.1.177

• feat(ui): onboarding provider step is real and mandatory
• test(crap): decompose + cover /agents panel methods (pure helpers in agent_tree.dart), public subagentManager getter
• fix(analyze): drop redundant null guard in agents section
• feat(app): agents panel in settings — live subagent tree with observe/send (CLI parity)
• test(cli): /agents tree panel integration test (keyless)
• feat(cli): agents viz A+B — live agents badge in status line, /agents tree panel with observe/send
• fix(ui): saturated glyph colors for the light brand icon
• feat(task): Claude Code agent compat — .claude/agents roots + model: frontmatter alias
• fix(ui): macOS traffic-light clearance for pushed AppBar routes
• fix(ui): pin the onboarding Privacy link to the footer center
• feat(ui): single brand tile everywhere, glyph rebalanced in the icon
• fix(analyze): drop unnecessary non-null assertion
• feat(task): subagent model role — settings-picked delegation model (TaskModelsStore + roles: config), smol/explore precedence

0.1.176

• fix(analyze): await in try block, drop unused imports
• test(crap): decompose + cover new subagents/a2a/memory methods (CRAP ≤ 12)
• feat(ui): brand icon in dark + light forms, styleguide colors in onboarding
• chore: drop unused imports
• fix(providers): commit the FA_PROVIDERS filter + enabledProviders definitions unbreaking main
• style: dart format task_executor
• docs(subagents): mark all phases done + AGENTS.md bullets for memory/a2a
• test(subagents): real-model test skips without ZAI key, longer timeout
• feat(a2a): phase 5b — fah serve --a2a HTTP mount + full client↔server loop verified live
• feat(ui): onboarding top padding, privacy link, centered wide layout + light macOS icon
• feat(a2a): phase 5a — a2a: config, A2aManager lazy connect, task tool a2a:
• feat(memory): phase 2 — compaction extraction hook, maintain() pipeline, /memory command
• feat(ui): onboarding redesigned pixel-close to the reference prototype
• feat(subagents): reply tool, agent_message sibling messaging, completed_without_reply notice + pending-queue guards
• fix(ci): remove stray file "flutter_app/" breaking the windows checkout
• feat(ui): Focus Timer as standalone dark card with circular progress ring
• feat(ui): onboarding rewritten to match reference — 3-col mockups, circular Focus Timer, colorful icons

0.1.175

• feat(providers): DIAL provider kind — {baseUrl}/openai/deployments/{model}/chat/completions with Api-Key auth, optional DIAL_API_VERSION query, /openai/models listing; --provider dial --model <deployment> headless

0.1.173

• fix(ci): restrict auto-release tag matching to v*..; fix pubspec version
• feat(ui): auto-focus composer input on session open/switch
• fix(testflight): fail on submit errors; fix framework bundle IDs; cleanup publish job
• feat: switch to hosted flutter_agent_memory dep, remove publish_to: none
• chore: add LICENSE to fa_llm for pub.dev
• feat: prepare fa_llm for pub.dev publishing

fa_llm-v0.1.1

• feat: switch to hosted flutter_agent_memory dep, remove publish_to: none

0.1.172

• fix(version): use Platform.resolvedExecutable so version works regardless of invocation path

0.1.171

• chore: trigger CI + auto-release for v0.1.171
• fix(ci): quote sed command to fix YAML syntax
• fix(ci): keep publish_to: none for analyzer; strip it only in the publish job
• fix(installer): rm -f target before cp — break symlinks so version.txt lives next to the binary

0.1.169

• fix: fa update now copies version.txt alongside the new binary
• feat(ui): permission cards with working action buttons
• ci: CodeQL workflow — Dart + JS only (no Java/Kotlin, no Gradle in root)
• fix(security): exact hostname match for testflight.apple.com (CodeQL #7)
• fix(ui): settings/files/model picker as popup dialogs on wide screens
• fix: disable Impeller on macOS to prevent resize crash
• fix(installer): copy version.txt next to binary + fix version lookup path
• feat(ui): FaMark sparkle brand icon (no background) + files goldens update
• fix(security): URL sanitization in analytics.js + workflow permissions
• fix(ui): full-height dividers — panels extend to window top on macOS
• fix(ui): address prototype feedback — tabs, calendar, timer, model switch
• Create SECURITY.md for security policy

0.1.168

• fix(pages): web demo build is optional (dart:ffi from sqlite3 breaks it)
• fix(installer): $zip_asset… unbound variable — brace the var before ellipsis
• feat(ui): 'Add app' tile in Created-by-you section (prototype style)
• feat(ui): system app tiles grid in AppsPanel (Calendar/Files/Notes/Maps/…)
• feat(ui): tool tiles show display names + dropdown arrows (prototype style)
• feat(ui): user profile section in sidebar bottom (matching prototype)
• fix: suppress dead-code warning on new sidebar null check (line 198)
• feat(ui): permission-denied card for tool errors (prototype style)
• fix: pub.dev publish_to removed, Windows zip uses 7z instead of zip
• feat(ui): Customize label in AppsPanel header (matching prototype)
• feat(ui): composer matches prototype — star icon, Ask anything, up-arrow send
• test(goldens): AppsPanel golden coverage — dark + light variants
• feat(ui): session date grouping + 3-dot menu + subtitle timestamps
• feat(ui): workspace header in wide layout + session tile 3-dot menu

0.1.167

• chore: trigger auto-release for CLI binaries + subagents 2.0
• feat(ui): AppsPanel with search/filters/sections for wide-layout right panel
• fix: remove unused test class + imports causing CI analyze warning
• fix(crap): decompose + cover all new methods to pass CRAP ratchet (12.0)

0.1.164

• test(cli): avoid real network in /provider custom default URL test
• refactor(self_manage): lower fallbackZipUpdate CRAP and cover zip path
• ci: pin crap4dart to 0.2.1 to match pre-commit ratchet
• fix(installer): fallback to .zip extraction when raw binary not in release

0.1.162

• fix(app): key field no longer prefills OPENROUTER_API_KEY for non-OpenRouter providers
• fix(oauth): native iOS/macOS OAuth via HTTPS callback + custom scheme redirect
• fix(oauth): iOS web redirect flow with state + verifier

0.1.157

• fix(oauth): capture OpenRouter web callback via JS object postMessage

0.1.153

• feat(providers): Google Gemini media provider + MediaModelsSection in fa_ui

0.1.152

• fix(providers): voice sample URLs are case-sensitive on the CDN

0.1.149

• feat(apps): Language Tutor rewrite + fitness-trainer device-path probes

0.1.148

• Gate JS apps and skills by platform

0.1.147

• fix(macos): surface EventKit authorization failures
• feat(macos): allow explicit calendar permission bootstrap

0.1.146

• fix(macos): split Debug entitlements for local flutter run

0.1.143

• feat(macos): privacy prompts, configurable signing, and no-sandbox release for Fa

0.1.142

• macOS: no-sandbox release flavor, HealthKit support, privacy entitlements
• Local models heading, Gemma 128k context, context-fit budget fix
• feat(flutter_app): feature-gate WebLLM, expand Gemma context window, filter BYOK picker

0.1.141

• fix(macos): bundle LiteRT-LM companion dylibs for flutter_gemma

0.1.140

• feat(settings): show on-device providers in the Providers section
• feat(settings): voice selection for the TTS media slot

0.1.139

• docs(gemma): update platform comments for macOS support

0.1.138

• feat(gemma): enable on-device Gemma provider on macOS
• feat(fa_ui): wire fa_llm/fa_llm_flutter into provider config

0.1.137

• chore(deps): bump flutter_gemma to latest official releases
• feat(fa_llm_flutter): add FlutterGemma on-device provider

0.1.136

• Add fa_llm package extracted from flutter_agent_memory llm layer

0.1.135

• feat(fa_ui): providerId through the connect flow

0.1.134

• feat(fa_ui): userBubble/userBubbleBorder tokens in FaUiTheme

0.1.133

• feat(fa_ui): avatar builder + theme-driven chat surfaces
• feat(fa_ui): host surface tokens + optional app bar in FaChatScreen
• fix(fa_ui): FaChatScreen honors the host FaUiTheme in the chat theme

0.1.132

• feat(fa_ui): extract the agent chat into the shared fa_ui package

0.1.130

• feat(launcher): 'Restore reference version' tile menu item for demo apps

0.1.128

• fix(apps): jscore multi-instance crash override + seed-error surface + map top inset

0.1.125

• fix(example): pin js_widget_runtime@9498d0c — revert the native-release grace that defeated the lifecycle serialization (tf-6 SIGSEGV); drop the test-only grace config

0.1.123

• feat(site): TestFlight public beta link in the hero CTA row

0.1.120

• feat(fa_ui): present editor/picker pages as constrained dialogs on wide canvases

0.1.119

• feat(yoclip): Fa promo video workspace — 19s promo in 3 aspects x en/ru (App Preview + social + YouTube), creative treatment, VO, music bed, frame QA
• test(example): realistic providers in the store_providers frame; pre-commit format gate scopes to package dirs (yoclip/ is a standalone workspace)

0.1.117

• feat(example): launcher home on all layouts (legacy session sidebar removed), App Store shots v2 ('your own apps, built by chat'), golden orphan gate

0.1.116

• fix(example): close action for full-chrome JS apps (map was unclosable), store copyright name

0.1.114

• ci(ios): scope codesign rewrite to Runner, auto-sign the FaLiveActivity extension (bundle-id collision 90685)
• fix(example): steer button interrupts the run, queued steers run after stop, sheet opens at the latest message

0.1.113

• test(example): real-agent E2E on the macOS host + store promo artwork

0.1.112

• feat(example): iOS background execution + Live Activity, key resolution fix, crash-churn guard, mini drag pill

0.1.111

• fix(example): TestFlight SIGSEGV root cause — serialized engine lifecycle

0.1.110

• fix(example): visible run errors, mini last-message strip, iOS-grade drag&drop, weather timeouts

0.1.109

• feat: CRAP green zone (max ≤ 8), app integration tests, tool-dup fix

0.1.108

• fix(example): TestFlight JSC crash, ownership-aware demo sync, CRAP yellow zone

0.1.107

• docs: privacy policy — PRIVACY.md + published site page, onboarding links it

0.1.106

• feat(example): app content respects the bottom safe area + onboarding replay

0.1.105

• feat(example): icons-per-row setting, tight row gap, pager bounce fix

0.1.104

• fix(example): reliable tile drops, full-width grid, widget drag cards

0.1.103

• feat(example): first-launch onboarding, scene3d wiring + 3D game demo, sheet/tile polish

0.1.102

• feat(example): iOS-style home grid — icon-unit alignment, live reflow, resizable tiles
• fix(example): drop the border on the floating chat bar/icon — shadow only
• feat(example): tile span sizes + floating mini chat bar, directional sheet swipes
• feat(example): live app tiles on the launcher + chat sheet mini-by-default

0.1.101

• fix(example): sheet respects the top safe area + light-theme golden

0.1.100

• refactor(example): drop unused members left by the sheet v3 rewrite
• feat(example): sheet v3 — ONE panel: round icon ↔ mini bar ↔ full sheet

0.1.99

• fix(example): sheet UX — full-bleed, one surface, ghost panel gone

0.1.98

• feat(example): session chat sheet v2 — mini bar with input, smooth physics

0.1.97

• docs(example): AGENTS.md notes for the launcher home, chat sheet and shared composer
• feat(example): session chat bottom sheet over the launcher (pager, shared composer)
• feat(example): apps launcher home on narrow layouts (grid, folders, system tiles)
• fix(example): home control disambiguation (room/UUID) + duplicate-bridge-id write routing
• fix(example): preset carousel is full-bleed — cards slide behind the edges

0.1.96

• fix(example): Home + Health apps scroll — root column → listView

0.1.95

• fix(example): pin js_widget_runtime to git fix for JSC use-after-free (TestFlight crash)
• refactor(example): share chat message rendering with the in-app Fa overlay + full state goldens
• fix(example): Fa panel is one bottom sheet, never two stacked cards

0.1.94

• refactor(test): split agent_cli_test.dart into support + provider/model topical files
• fix(cli): spec env names resolve only for the default hosted endpoint
• fix: video download auth on own-origin urls + provider key name dedupe + keychain preflight
• fix(cli): fa update misdetects a native binary in pub-cache as pub-global

0.1.93

• fix(cli): empty Enter submits in guided flows; parse models[]/alias /models dialect
• docs: commit identity policy — ai.teammate for contributors

0.1.92

• feat(example): generate_video tool — async /videos job on the videoGeneration slot
• feat: request_secret tool — agent asks the user for missing credentials
• fix(example): theme-aware FaWorkBar — one component with the chat overlay
• fix(example): HomeKit entitlement + longer homes wait + notify probe
• feat(example): resume the day's session at boot instead of stacking empties

0.1.91

• feat(example): model presets wizard in settings
• feat(example): audio/video playback in the file preview
• feat(example): story-driven App Store screenshots with real photos

0.1.90

• fix(example): set the App Store copyright field in the deliver lanes

0.1.89

• fix(deps): revert sqlite3 to ^2.9.4 — 3.x build hooks break dart compile

0.1.88

• fix(example): retry deliver on Apple's bursty Connect API 500s
• fix(example): preflight the macOS store version before deliver
• fix(example): tolerate deliver's first-version 'No data' review-detail crash
• fix(example): shrink RU promotional text under App Store's 170-char limit
• ci: store-metadata workflow — App Store content upload on demand
• feat(example): App Store content pipeline — store goldens, metadata, fastlane lanes

0.1.87

• fix(example): iOS build — HMHomeManagerDelegate members shadow the homeManager global
• feat(example): HomeKit maximum API, empty-homes race fix, shareable debug logs
• feat(example): privacy-first analytics facade (Firebase Analytics)
• fix(example): contacts — system back steps out of detail, transient call hint
• feat(example): rename sessions, arbitrary agent keys, persist approval mode

0.1.86

• fix(example): regenerate iOS Podfile.lock (Firebase 12.x + media players)
• feat(example): calendar recurrence, alarms, calendars, span, and url
• fix(providers): thinking — dedupe reasoning vs reasoning_details + tail collapse
• feat(example): inline audio/video playback for sandbox media in chat

0.1.85

• feat(example): render sandbox images in chat — markdown imageBuilder + inline generate_image tiles
• feat(example): expand Fa chat in place inside JS app views
• fix(example): macOS pods — platform 14.0 + regenerated lock (Firebase 12.x)
• fix(example): in-app Fa stays in the bottom sheet on first contact
• chore(deps): update AI integration deps (firebase, flutter_gemma, js_widget_runtime) and sqlite3; migrate sqlite3 dispose -> close

0.1.84

• feat(example): preset default-model override + two-step media slot flow

0.1.83

• fix(example): drop the robot icon + Model header from the sidebar
• fix(example): macOS keychain read + dead platform channels
• feat(example): providers-first settings — provider editor page, default chat model flow, provider-based media slots
• feat(example): providers-first settings — provider editor page, default chat model flow, provider-based media slots
• fix(example): contacts list scroll + live search; paged full-list search with phone matching for dedup
• fix(example): composer stop button while streaming; abort drains steer queue into transcript
• chore(example): pin js_widget_runtime ^0.4.13 (image UA fix)
• fix(providers): dedupe overlapping/cumulative reasoning chunks in openai-completions thinking stream

0.1.82

• fix(example): browser-ish UA for network images (runtime 0.4.13) + url image probes

0.1.79

• fix: commit the missing modifications from models-config and media UI (autostash unstaging)
• feat(example): per-run date refresh, media models full-screen editor with /models picker
• fix(example): dead onPressed buttons (runtime 0.4.12), calendar date-labeled lists + ±7d match, mic e2e probes
• fix(example): current date in system prompt, foreground notification banners, contacts openUrl errors
• feat: CLI models-config (models: config + /models set|remove|config) + media models settings UI

0.1.78

• feat: Keychain key persistence (iOS/macOS), /models endpoint listing, model marks, site updates
• refactor(example): dedupe cache sections into shared model_cache_section; bump runtime 0.4.11
• feat(example): transcription slot wiring + read_video (frames → vision)
• feat(example): media models config + generate_image/speak/generate_music tools + js bridge
• feat(example): iCloud sync for sessions/apps (iOS; macOS pending signing)
• feat(example): local push notifications — channel, notify tool, js bridge, reminders demo
• feat(example): HomeKit control (iOS) + mic/ASR voice input
• feat(example): HealthKit read (iOS), scene3d dep (0.4.10), Android readiness doc

0.1.77

• feat(example): contacts domain — channel, agent tools, js bridge, demo app
• feat(example): jsr.fa.llm chat (multi-turn) + stream (delta events)
• feat(example): calendar write — channel, agent tools, js bridge, demo editing
• feat(example): back-swipe contract (jsr.onBack), privacy manifests + usage descriptions

0.1.76

• feat(example): chrome modes, branding sweep (fah→Fa), textArea+scrolling docs (0.4.8), system-API design doc

0.1.75

• feat(example): animation nodes demo (entrance stagger, animatedSwitcher), theme+reply-sheet sources
• feat(example): jsr.theme plumbing (light/dark live), Fa mini reply sheet, map-app golden
• fix(example): chart node API alignment (0.4.7), gridView docs, bar chart demo

0.1.72

• feat(example): orbit work-bar, light theme, secrets UI, open_app + calendar tools, map node demos

0.1.70

• docs: mandate golden tests for all UI work in AGENTS.md
• feat(example): brand fonts (Inter/JetBrainsMono), marketing-grade full-screen goldens, app quality gates
• fix(example): absolute path for upload_picker_web conditional import
• fix(example): update imports for the sandbox/services/ui layout
• feat(cli): name the key source (environment vs secure store) in the 401 hint
• feat(cli): diagnose auth failures with the key source (env vs store shadowing)
• feat(cli): print the session resume command on exit
• test(example): golden tests for every UI widget + pipeline golden gate
• fix(agent): repair orphaned tool calls in the request payload
• feat(cli): replay the full restored transcript with collapsed tool runs

0.1.69

• fix(example): unblind hosted models — vision detection + settings checkbox
• feat(example): localize the UI (en/ru) with a hardcoded-string guard test
• perf(cli): coalesce streamed output deltas to keep typing responsive
• fix(cli): keep the ctx gauge at the last real usage after a failed run
• fix(example): bump js_widget_runtime to ^0.4.3 — renderer no longer crashes on array borderRadius
• perf(cli): memoize the markdown wrap pass so scrolling stays O(1)
• fix(cli): re-attach follow on submit so the sticky echo pins again

0.1.68

• feat(cli): render provider error lines in red
• fix(cli): hide the physical cursor while a run streams
• feat(example): follow-tail auto-scroll + collapse long thinking blocks
• fix(example): work bar for grid-opened apps; prove permission persistence

0.1.67

• feat(example): UX batch — collapsible tool output, Fa mark, in-app work bar
• test(example): textField onChange delivers typed text to JS
• fix(example): render attached-image messages — add the missing imageMessageBuilder

0.1.66

• feat(example): stream model thinking live into the chat
• fix(example): replace the whole-run timeout with an idle watchdog
• feat(example): SVG app icons for JS apps
• feat(example): restore persisted sessions in the sidebar after restart

0.1.65

• feat(example): teach the js-apps skill how to test apps before handover
• fix(example): fit the on-device Gemma context instead of engine overflow
• test(example): tap test — calculator key reaches the JS engine
• fix(cli): never hang on a keychain system modal
• fix(example): bundle demo app assets — nested asset dirs need explicit entries

0.1.64

• fix(ios,macos): keep -exported_symbol out of Debug link flags

0.1.63

• test(example): end-to-end render test for the calculator demo app
• feat(example): JS apps platform in the Fa app (js_widget_runtime)
• docs: document the wasm_run symbol gate, strip-style pitfall, and new CI secrets/caches

0.1.60

• fix(ios): export wasm_run FFI symbols so Release/TestFlight builds keep them
• ci: whitelist the tracked Firebase config for pub.dev's leak scanner

0.1.59

• fix(ios): use development provisioning profile for Debug builds
• ci: unblock releases — vendor gitignore rule, pubspec catch-up, tag-ahead release
• test(example): drop the unused accessGranted param (CI fatal-warnings)
• feat: widen shell PATH for GUI apps (Homebrew python/node)

0.1.53

• ci(macos): fix provisioning profile entitlement extraction
• fix(cli): rewind context crash and sticky echo duplication
• ci(macos): fix provisioning profile entitlement key
• fix(cli): enable mouse-wheel scrolling in TUI transcript
• ci(macos): fix entitlements heredoc syntax
• fix(cli): long user messages in TUI — ellipsis marker, calm scroll hint
• ci(macos): embed provisioning profile, use git tags for version
• fix(cli): degrade keychain write failures to session-only, never crash
• fix(macos): raise deployment target to 12.0 for TestFlight
• fix(macos): add LSApplicationCategoryType for TestFlight validation

0.1.52

• ci(ios): use absolute IPA path for TestFlight submit
• ci(ios,macos): fix submit artifact path and macOS Ruby PATH

0.1.51

• ci(ios,macos): fix artifact downloads and macOS keychain password
• ci(ios,macos): fix artifact path, macOS framework restore, action versions

0.1.50

• ci(ios,macos): fallback to GitHub release for WasmRun iOS framework
• ci(ios,macos): restore WasmRun frameworks from local runner copy
• ci(ios): restore WasmRun.xcframework from pub cache before build
• fix(ios): use SRCROOT-relative path for wasm_run force_load
• ci(macos): set working-directory for flutter steps and create .env placeholder
• fix(ios): use absolute path for wasm_run force_load in Podfile
• fix(ios): force-load wasm_run via Podfile post_install for Runner target
• test(cli): wizard/registry coverage, help keyword, docs
• fix(ios): force-load wasm_run in pod target only
• fix(ios): use PODS_ROOT path for wasm_run force-load
• fix(ios): force-load wasm_run static lib for arm64
• ci(ios): create placeholder .env asset for build
• chore(ios): track firebase_options.dart for CI builds
• feat(cli): custom provider registry, guided wizard menus, TUI follow latch
• ci(ios): download missing iOS platform before build
• ci(ios): use simulator build to avoid missing device sdk
• ci(ios): boot simulator before flutter build to avoid attached device
• ci(ios): fix simctl invocation in fastlane build_only

0.1.49

• ci(ios): fix signing identity extraction
• ci(ios): use fastlane build_only with temporary keychain
• ci(ios): use persistent ci.keychain on self-hosted runner
• ci(ios): use only build.keychain as default/search list, no OTHER_CODE_SIGN_FLAGS
• ci(ios): import distribution cert into login.keychain and build without isolated keychain
• ci(ios): download Apple WWDR G3 intermediate into build.keychain
• ci(ios): import WWDR intermediate into build.keychain, keep it unlocked, pass --keychain
• ci(ios): pass --keychain to codesign via OTHER_CODE_SIGN_FLAGS and add debug output

0.1.48

• feat(cli): guided custom provider setup (/provider custom): api type (openai/anthropic/google-like), base URL, optional key (saved to the OS secure store), then the model — picked from the endpoint's /models list or typed manually; the TUI provider picker gains + custom provider…. (Code landed inside 7082bc8, swept up by a parallel commit.)

0.1.47

• ci: explicit export-options plist for iOS builds (UUID + full identity)
• fix(example): pin the full signing identity name for iOS CI builds
• fix(example): pin CODE_SIGN_IDENTITY iPhone Distribution for CI builds

0.1.46

• fix(example): manual code signing with Fa Profile for CI iOS builds
• ci: placeholder firebase_options.dart for the repo-wide analyze
• chore(example): untrack leftover Firebase configs from the pre-migration path
• ci: cd /tmp before wiping the workspace in mirror checkout
• ci(pages): tracked firebase_options template instead of git history
• ci: self-updating mirror checkout in build-mobile.yml (same as build-macos)
• ci(pages): placeholder firebase_options.dart for the web build
• fix(example): keep Firebase Analytics from killing web startup
• ci: quote pwsh run line breaking the ci.yml YAML parse
• feat(example): unify bundle id to dev.fa1.app for a single App Store record
• feat(cli): /provider runtime switching and OS secure key storage (/key)
• fix(install): POSIX-clean install.sh and setup.sh for Ubuntu dash
• docs: document app build/TestFlight workflows and secrets in AGENTS.md
• ci: TestFlight submission for iOS and macOS (learn.ai pattern)
• refactor(example): migrate example/flutter_example to flutter_app (fa package)

0.1.45

• feat(cli): fa update and fa uninstall quick commands
• fix(ci): quote pwsh run line — leading & parsed as a YAML anchor, breaking the whole workflow
• fix(windows): fa crash after TUI exit + installer mojibake
• ci: installer-smoke job runs the one-line installers on every tag

0.1.44

• ci: fix Windows binary build + installer mojibake

0.1.43

• feat: agent skills + project context files (all platforms)
• feat(cli): background subagents via the task tool
• fix(cli): keep cursor pinned to input while the spinner ticks
• ci: create GitHub Release before binary upload + embed version

0.1.42

• feat(cli): dart_tui interactive TUI with markdown rendering
• ci: add build-mobile.yml (APK/iOS) and build-macos.yml (DMG) workflows
• feat: multi-session support — AgentSessionManager (core) + FlutterSessionManager (app)
• fix(example): hide empty assistant bubbles in chat
• feat(example): debug-log system prompt platform and WASM runtime setup
• docs(example): drop stale no-WASM-on-iOS comments after static linking fix
• fix(example): iOS gets the full WASM sandbox command set in the system prompt

0.1.41

• fix(site): quote install URLs for zsh glob safety; refine iOS wasm_run static-library flags
• fix(cli): avoid double stdin subscription in TUI REPL
• fix(example): use DynamicLibrary.process for iOS wasm_run static linking

0.1.40

• fix(vendor): force-load wasm_run static lib via podspec and refresh Podfile.lock
• refactor(site): centralize installer banner/recipe in install-config.yaml and use DMTools-style Windows PATH
• fix(vendor): apply iOS wasm_run_flutter static-library linker flags in Podfile
• feat(cli): numbered line-mode slash menu and guard TUI to interactive TTYs
• fix(vendor): iOS wasm_run_flutter static library fallback
• fix(install): use github releases/latest/download direct URLs, avoid API rate limits
• feat(ci): build native fa binaries for win/mac/linux and download them in installers
• feat(ios): enable WASM shell via statically linked executable
• fix(site): repair install dropdown visibility and bust cache; make CLI raw-mode fallback graceful
• feat(cli): add named session management via --session and /session commands
• feat(cli): raw-mode TUI with slash menu, model picker, and dynamic version
• feat(site): add Windows cmd.exe installer wrapper (install.bat)

0.1.39

• feat(cli): Pi-style terminal banner, status bar, and /help filtering
• fix(site,install): remove DMTools from install dropdown and reword PATH symlink comment
• fix(install): make fa available immediately after install without shell reload
• feat(site): add DMTools install options to site dropdown
• fix(example): split SandboxPlatform.mobile into android/ios and disable shell command ads on iOS
• refactor(install): split installer into non-interactive install + interactive setup wizard
• fix(cli,install): primary command is fa, auto-add pub-cache to PATH
• fix(site): cache-bust web demo assets on every deploy
• fix(example): render user messages through the harness loop
• fix(site): mktemp compatibility on macOS

0.1.38

• feat(site): Windows PowerShell installer + generated menus from install-config.yaml
• chore(macos): set bundle identifier to dev.fa1.macos and update copyright
• fix(site): use correct GA4 measurement ID (G-0Z3SW38FYC) and Fa mobile app label
• fix(ios): graceful WASM fallback — app starts without wasm shell on iOS
• feat(prompt-tools): slim on-device system prompt — compact schemas + fewer tools
• feat(cli): modern TUI pack — ! shell commands, /models filter, status line
• chore(site): switch GA measurement ID to Firebase web stream
• feat(cli): interactive installer with progress bar, provider/model picker, and config setup
• fix(example): readable ONNX/WebGPU crash messages + verified engine recovery

0.1.37

• fix(example): halve ONNX Gemma context window to 2048 (WebGPU OOM mitigation)
• feat: optional API token for custom providers (local servers need no key)
• feat(cli): banner shows baseUrl+key status, connection-refused hint, version in --help
• chore(example): ignore Firebase config files with real API keys
• fix(site): full-width header background and Fa branding
• feat(example): release prep — Fa branding, icons, bundle IDs, Firebase Analytics
• feat(cli): prompt overrides (config prompts: + --system-prompt[-file]) and full --help reference

0.1.36

• fix(example): WebLLM context windows sized for the Fa system prompt + compaction scales with model window
• feat(cli): headless mode — fa "prompt", -p alias, file-as-prompt (md/txt content, binary path ref)

0.1.35

• feat(example): persist last connection + downloaded-models quick start on setup screen
• feat(tools): lsp tool backed by the Dart analysis server (diagnostics/definition/references/rename)

0.1.34

• feat(tools): task tool — parallel subagents with schema-validated results (omp port)
• feat(agent): TTSR stream rules — abort, inject, retry mid-generation (omp port)
• feat(providers): model roles (default/smol/slow/plan) with fallback chains, key rotation, path overrides

0.1.33

• feat(tools): read selector grammar (:A-B, :A+C, multi-range, :raw) + zip inner paths + SQLite reads
• feat(tools): image read parity with pi (byte cap, pass-through, EXIF, placeholders) + transcribe_audio tool
• feat(site): set GA4 measurement ID

0.1.32

• feat(tools): web_search with provider chain (DDG keyless first, Brave/Tavily behind secrets) + web_fetch markdown extraction with a pub.dev site handler

0.1.31

• feat(tools): hashline edit format with content-hash anchors (omp port)
• feat(core): approval tiers with per-tool policy, bash interceptor, CLI/app prompt UIs
• feat(example): model lineup — drop <1.5GB presets, add Gemma 4 E4B ONNX (~5.2GB)

0.1.30

• feat(example): WebLLM presets refresh — Qwen3.5 + Qwen2.5-Coder (web-llm 0.2.84)
• feat(example): visible app name is Fa (assistant label, AppBar, transcript, system prompt)

0.1.29

• fix(example): transformers.js download filter+progress, SVG/upload/attach UX, provider-error robustness

0.1.28

• feat(example): central sandbox command registry drives the Fa system prompt
• fix(example): web upload fix + chat uploads→uploads/ + light HTML preview + session delete

0.1.27

• feat(example): transformers.js Gemma provider on web (ONNX q4f16, tools via prompt wrapper)
• feat(brand): rename visible brand to Fa + app favicon matches the site

0.1.26

• fix(example): Gemma web uses -web.litertlm builds + Gemma cache management in settings

0.1.25

• ci: coalesce auto-releases to <=1 per 2h + scheduled catch-up

0.1.24

• feat(example): brand app icon for all platforms (gradient >_ mark)

0.1.23

• feat(example): markdown/HTML file previews + auto-refresh on agent file mutations

0.1.22

• refactor(example): WebLLM goes chat-only + universal prompt-tools wrapper

0.1.21

• feat(example): Gemma provider on web via flutter_gemma litert-lm web (Gemma 4 tools in-browser)

0.1.20

• feat(core): prompt-based tool-calling wrapper (universal chat-model tools)

0.1.19

• fix(example): settings dialogs adapt to narrow phone screens

0.1.18

• feat(example): Gemma 4 on-device provider via flutter_gemma (iOS/Android)

0.1.17

• feat(example): custom provider management + WebLLM model cache management in settings

0.1.16

• feat(example): left sidebar (model picker + sessions), files move right

0.1.15

• feat(example): WebLLM function calling (tools for Hermes-3 FC preset)

0.1.14

• feat(example): branded web loading splash (first-frame fade)

0.1.13

• feat(example): dark theme matching the landing (terminal aesthetic)

0.1.12

• feat(example): full WebLLM preset list matching flutter_agent_memory (22 models)

0.1.11

• feat(example): WebLLM on-device provider for the web demo (no API key needed)

0.1.10

• feat(site): capability comparison table (Browser/macOS/iOS/Android/Windows)

0.1.9

• feat(example): web file upload + IndexedDB-persisted sandbox FS

0.1.8

• feat(site): SEO/GEO pack + OG share image

0.1.7

• fix(example): sharpen Ollama Cloud CORS guidance in BYOK notes

0.1.6

• feat(site): GitHub Pages landing + live web demo with BYOK
• feat(example): BYOK connection settings with provider presets

0.1.5

• refactor(prompts): extract LLM prompts to prompts/*.md + codegen (AGENTS.md convention)

0.1.4

• chore: mark fake PEM stubs as false_secrets for pub validation

0.1.3

• ci: fix auto-release tag push — annotated tag + --atomic (lightweight tags are not sent by --follow-tags)
• test(providers): Ollama Cloud live integration tests (gpt-oss:20b default, OLLAMA_MODEL override)
• ci: create placeholder .env for the example app (asset_does_not_exist)
• fix(example): mobile sessions no longer land in a doubled host path
• ci: fix quality gate — install Flutter SDK + example pub get for repo-wide analyze
• ci: auto-release on push to main (patch bump + tag, OIDC publish) + OLLAMA_API_KEY in integration env
• feat(example): file browser panel (tree + preview, collapsible on wide screens)
• test(providers): live integration tests (OpenRouter live, Anthropic/Google key-gated)
• feat(sandbox): pip-lite for sandbox python (pure-python wheels)
• feat(sandbox): lua interpreter (WASI) in the mobile shell
• feat(sandbox): small utils batch (tree, file, xz/bzip2 -d, base64+hashes on web)
• feat(sandbox): ssh/scp/sftp exec builtins via dartssh2
• feat(sandbox): nslookup/dig + whois network diag builtins
• feat(sandbox): diff/patch builtins (Dart, iOS+web)
• feat(secrets): env injection + redaction (SecretsStore)
• feat(sandbox): web command parity with iOS shell
• feat(example): python3/qjs on web via CDN interpreters + copy-session button
• chore: remove ssh debug script
• feat(sandbox): sqlite3 CLI (WASI build from official amalgamation)
• style: curly braces in web_git remote add (lint info)
• feat(web): local git in the browser sandbox (MemoryShell)
• feat(sandbox): QuickJS JavaScript engine (qjs/js) + web parity checks
• feat(sandbox): python3 (CPython 3.14 WASI) in the mobile shell
• feat(tools): edit (str_replace) tool + sandbox path mapping + coding system prompt
• feat(git): push over smart HTTP (receive-pack) + SSH transport (dartssh2)
• feat(git): remote/fetch subcommands, checkout -b, branch -r, clone fixes
• fix(mobile shell): curl/wget --version, --help, and no-URL error message
• feat(git): smart HTTP git-upload-pack clone for any public remote
• feat(web): pure-Dart MemoryShell for the browser + flutter build web fixed
• feat(mobile shell): cd/export/unset, $VAR expansion, grep/wget, du/stat/tac/expr/id/relpath builtins
• feat(mobile shell): add git support via dart-git + GitHub archive clone
• feat(mobile shell): add dart-native curl/jq/yq builtins
• feat(mobile shell): add WASM sed/awk/tar/gzip/zip+unzip, env builtin, redirect capture fix, POSIX double-quote escapes
• feat(mobile shell): add shell builtins (test, which, whoami, xargs, command -v)
• chore: remove stray temp files accidentally committed
• test: add shell command integration tests (host + WASM sandbox catalog)
• fix(ls tool): return basename when path points to a file
• fix(ios): get WASM shell working on iOS simulator
• feat(example): replace busybox with permissive uutils/ripgrep WASM sandbox
• feat(example): sandboxed WASM bash shell for mobile/web via busybox+wasm_run
• fix(example): cache streaming/error state in ChatScreen for immediate UI updates and add multi-turn test
• fix(example): notify UI before persisting so streaming indicator hides immediately
• fix(example): throttle and incrementally sync messages to avoid SliverAnimatedList crash
• fix(example): move input bar outside Chat widget to fix layout and semantics
• fix(example): replace package Composer with custom input bar to fix ParentDataWidget crash
• feat(flutter_example): integrate flutter_chat_ui with markdown and tool cards
• feat(flutter_example): load API key from .env for simulator runs
• fix(flutter): typing indicator, error banner, 90s timeout; feat(cli): persist last model/provider/mode in ~/.fah/config.yaml
• Add Flutter mobile example with path_provider + LocalExecutionEnv
• Add pi-style agent modes and prompt templates to CLI
• Clean lint info in plugin tests
• Update GOAL.md with plugin/package extension API
• Add plugin/package extension system with built-in inspect_image plugin
• Add inspect_image tool: dedicated vision model analysis like pi-inspect-image
• Add fah/fa executables, rebrand system prompt, image support in read tool
• Add CLI harness: bin/fah REPL with builtin tools, sessions, compaction
• Format codebase, fix lint info, shorten pubspec description (pana 160/160); add format gate to pre-commit
• CI: GitHub Actions quality gates + OIDC pub.dev publish on version tags
• Phase 3: token estimation and LLM compaction pipeline
• Phase 3: ExecutionEnv abstraction and append-only JSONL session tree
• Phase 2: AgentTool registry with JSON-schema param validation
• Phase 2: stateful Agent with steering/follow-up queues and hooks
• GOAL.md: TDD for new code, coverage target >90%, push after every card
• Phase 2: port low-level agent loop with AgentEvent stream and CancelToken abort
• Phase 1: context-overflow detection, Retry-After parsing, sealed exception hierarchy
• Phase 1: port Google provider adapter with native functionCalling streaming
• Phase 1: port Anthropic provider adapter with native tool_use/thinking streaming
• Phase 0: port openai-completions provider adapter (OpenRouter-ready) with errors-as-events and CancelToken abort
• Phase 0: port AssistantMessageEventStream contract and SSE line decoder from pi-mono
• GOAL.md: allow agent publishing on explicit user instruction; OIDC for tagged releases

0.1.2

• test(providers): Ollama Cloud live integration tests (gpt-oss:20b default, OLLAMA_MODEL override)
• ci: create placeholder .env for the example app (asset_does_not_exist)
• fix(example): mobile sessions no longer land in a doubled host path
• ci: fix quality gate — install Flutter SDK + example pub get for repo-wide analyze
• ci: auto-release on push to main (patch bump + tag, OIDC publish) + OLLAMA_API_KEY in integration env
• feat(example): file browser panel (tree + preview, collapsible on wide screens)
• test(providers): live integration tests (OpenRouter live, Anthropic/Google key-gated)
• feat(sandbox): pip-lite for sandbox python (pure-python wheels)
• feat(sandbox): lua interpreter (WASI) in the mobile shell
• feat(sandbox): small utils batch (tree, file, xz/bzip2 -d, base64+hashes on web)
• feat(sandbox): ssh/scp/sftp exec builtins via dartssh2
• feat(sandbox): nslookup/dig + whois network diag builtins
• feat(sandbox): diff/patch builtins (Dart, iOS+web)
• feat(secrets): env injection + redaction (SecretsStore)
• feat(sandbox): web command parity with iOS shell
• feat(example): python3/qjs on web via CDN interpreters + copy-session button
• chore: remove ssh debug script
• feat(sandbox): sqlite3 CLI (WASI build from official amalgamation)
• style: curly braces in web_git remote add (lint info)
• feat(web): local git in the browser sandbox (MemoryShell)
• feat(sandbox): QuickJS JavaScript engine (qjs/js) + web parity checks
• feat(sandbox): python3 (CPython 3.14 WASI) in the mobile shell
• feat(tools): edit (str_replace) tool + sandbox path mapping + coding system prompt
• feat(git): push over smart HTTP (receive-pack) + SSH transport (dartssh2)
• feat(git): remote/fetch subcommands, checkout -b, branch -r, clone fixes
• fix(mobile shell): curl/wget --version, --help, and no-URL error message
• feat(git): smart HTTP git-upload-pack clone for any public remote
• feat(web): pure-Dart MemoryShell for the browser + flutter build web fixed
• feat(mobile shell): cd/export/unset, $VAR expansion, grep/wget, du/stat/tac/expr/id/relpath builtins
• feat(mobile shell): add git support via dart-git + GitHub archive clone
• feat(mobile shell): add dart-native curl/jq/yq builtins
• feat(mobile shell): add WASM sed/awk/tar/gzip/zip+unzip, env builtin, redirect capture fix, POSIX double-quote escapes
• feat(mobile shell): add shell builtins (test, which, whoami, xargs, command -v)
• chore: remove stray temp files accidentally committed
• test: add shell command integration tests (host + WASM sandbox catalog)
• fix(ls tool): return basename when path points to a file
• fix(ios): get WASM shell working on iOS simulator
• feat(example): replace busybox with permissive uutils/ripgrep WASM sandbox
• feat(example): sandboxed WASM bash shell for mobile/web via busybox+wasm_run
• fix(example): cache streaming/error state in ChatScreen for immediate UI updates and add multi-turn test
• fix(example): notify UI before persisting so streaming indicator hides immediately
• fix(example): throttle and incrementally sync messages to avoid SliverAnimatedList crash
• fix(example): move input bar outside Chat widget to fix layout and semantics
• fix(example): replace package Composer with custom input bar to fix ParentDataWidget crash
• feat(flutter_example): integrate flutter_chat_ui with markdown and tool cards
• feat(flutter_example): load API key from .env for simulator runs
• fix(flutter): typing indicator, error banner, 90s timeout; feat(cli): persist last model/provider/mode in ~/.fah/config.yaml
• Add Flutter mobile example with path_provider + LocalExecutionEnv
• Add pi-style agent modes and prompt templates to CLI
• Clean lint info in plugin tests
• Update GOAL.md with plugin/package extension API
• Add plugin/package extension system with built-in inspect_image plugin
• Add inspect_image tool: dedicated vision model analysis like pi-inspect-image
• Add fah/fa executables, rebrand system prompt, image support in read tool
• Add CLI harness: bin/fah REPL with builtin tools, sessions, compaction
• Format codebase, fix lint info, shorten pubspec description (pana 160/160); add format gate to pre-commit
• CI: GitHub Actions quality gates + OIDC pub.dev publish on version tags
• Phase 3: token estimation and LLM compaction pipeline
• Phase 3: ExecutionEnv abstraction and append-only JSONL session tree
• Phase 2: AgentTool registry with JSON-schema param validation
• Phase 2: stateful Agent with steering/follow-up queues and hooks
• GOAL.md: TDD for new code, coverage target >90%, push after every card
• Phase 2: port low-level agent loop with AgentEvent stream and CancelToken abort
• Phase 1: context-overflow detection, Retry-After parsing, sealed exception hierarchy
• Phase 1: port Google provider adapter with native functionCalling streaming
• Phase 1: port Anthropic provider adapter with native tool_use/thinking streaming
• Phase 0: port openai-completions provider adapter (OpenRouter-ready) with errors-as-events and CancelToken abort
• Phase 0: port AssistantMessageEventStream contract and SSE line decoder from pi-mono
• GOAL.md: allow agent publishing on explicit user instruction; OIDC for tagged releases

0.1.1

• Ported pi-mono packages/ai: EventStream contract (partial-first deltas, errors-as-events), SSE line decoder, openai-completions (OpenRouter-ready), Anthropic and Google provider adapters, usage/cost accounting, context-overflow detection, Retry-After parsing.
• Ported pi-mono packages/agent: low-level agent loop, stateful Agent with steering/follow-up queues and hooks, AgentTool registry with JSON-schema param validation.
• Sessions and context management: ExecutionEnv abstraction (pure-Dart memory impl + dart:io impl in lib/io.dart), append-only JSONL session tree with branching/labels, token estimation and LLM compaction pipeline.
• CLI harness (bin/fah.dart): a pi-like terminal agent with built-in read/write/ls/bash tools on the ExecutionEnv abstraction (lib/src/tools/builtin_tools.dart), a pure-Dart REPL core with injectable IO (lib/src/cli/agent_cli.dart) — live streaming output, slash commands (/exit, /reset, /compact, /stats, /model, /help), steering, Ctrl-C abort, JSONL session persistence, and auto-compaction.

0.1.0

• Initial project setup: package skeleton, quality gates (analyze, tests, coverage ≥ 80%, duplication < 1%), GOAL.md with the pi-mono port roadmap.
• Seeded CancelToken / CancelTokenSource / CancelledException — the universal cancellation primitive (Dart counterpart of web AbortSignal).

0.1.268

• memory: flutter_agent_memory roadmap hints (LLM role, graph screen, multi-root)
• feat(memory): flutter_agent_memory 0.2.0 — merge-friendly ids, git support, policy-driven memory_add (0.1.267)

0.1.269

• memory: maintain() leveling pass (level: 2 on 13 notes)
• fix(tui): immortal Working… spinner wedge — 100% CPU for 8h after run end (0.1.268)

0.1.275

• fix(compaction): bounded compactor budgets + attempt progress; loop over-window guard (0.1.274)
• feat(app): copilot connect picks the model explicitly; restart hydrates the entry-scoped key
• fix(agent): run watchdog disarms on abort; AgentService.dispose aborts in-flight runs (0.1.273)
• fix(providers): explicit model picks everywhere — no default models, copilot picker filter (0.1.272)

0.1.279

• fix(providers): reject Copilot fine-grained PATs at connect time (0.1.278)
• refactor(cli): split banner/key-status out of agent_cli.dart, untangle _runPrompt

0.1.281

• fix(providers): correct the Copilot token guidance — fine-grained PATs need the Copilot Requests permission (0.1.280)
• memory: copilot fine-grained PAT 404 root cause + flutter_app flame_3d env breakage

0.1.283

• fix(app): shared DapHubSnapshot, probe dispose ordering, scope reverts (#15 review)
• refactor(cli): one DapHubSnapshot type shared by CLI and app (#15 review)
• fix(app): compile fixes after main merge — barrel import hides, l10n key for widgets catalog note
• refactor(cli): extract packages.yaml loader into lib/ — keeps bin/ out of test coverage (CRAP gate)
• fix(cli): make dap opt-out test teardown race-tolerant
• fix(app): harden DAP hub page error paths and test determinism (#6)
• feat(app): DAP/Hub settings section (#6)
• fix(cli): review fixes for DAP/Hub /settings entry (#5)
• feat(cli): DAP/Hub entry in /settings (#5)

0.1.284

• memory: session access-count sync from trajectory work
• docs(trajectory): AGENTS.md sections for core, fa_ui widgets, CLI commands (#10)
• test(trajectory): fa_ui golden baselines (57 PNGs) + real icon glyphs (#10 phase 11)
• refactor(trajectory): CLI CRAP ratchet — split inspect renderer, cover tail/parse arms (#10)
• fix(trajectory): mirrored live-tail rows keep real-record durations (#10)
• feat(trajectory): Flutter host — service stream, feature flag, AppBar icon, panel (#10 phase 10)
• feat(trajectory): CLI /trajectory family + headless fa trajectory + TUI fallback (#10 phase 9)
• feat(trajectory): wire view, toolbar strings, barrel exports (#10 phases 6-8 integration)
• feat(trajectory): TrajectoryDetails tabbed sheet (#10 phase 8)
• feat(trajectory): TrajectoryTimeline painter + gestures (#10 phase 7)
• feat(trajectory): TrajectoryTable, per-kind cells, virtualised ledger (#10 phase 6)
• refactor(trajectory): split CRAP-heavy layout fold and timed timeline (#10)
• feat(trajectory): fa_ui controller, view skeleton, toolbar, strings (#10 phase 5)
• feat(trajectory): incremental full-text search index (#10 phase 4)
• feat(trajectory): timeline projection — sequence/duration/time/actual modes (#10 phase 3)
• feat(trajectory): event projection, request numbering, live tail, layout fold (#10 phase 2)
• feat(trajectory): core record model, snapshot contract, JSONL walker (#10 phase 1)

0.1.285

• fix(trajectory): timeline lane labels inherit theme font
• test(trajectory): real fonts in fa_ui goldens — Inter/JetBrainsMono + monospace alias

0.1.286

• fix(trajectory): thread ToolCall.parentCallId so subtool rows replay from sessions

0.1.287

• feat(tools): capability-gated tool availability (issue #19) — the pure decision layer (lib/src/tools/availability.dart: ToolsConfig yaml/JSON parsing with mcp:<server> flattening, resolveToolAvailability merging the global < project < session < runtime scope stack over the host's hard capability floor — absent tools can never be force-enabled, unknown ids warn once) and the gate (availability_gate.dart: idempotent registry hide/restore + prompt rebuild, executor tombstones for calls to disabled tools, noteHiddenNames covering late MCP registrations).
• feat(cli): the /tools family (bare list, enable|disable <id> [global|project|session], reload), a Tools entry in the /settings hub, and the runtime scope: --tools 'id=on|off,...' with the FA_TOOLS env twin (flag wins; a malformed spec is a hard startup error). Scopes re-read and re-applied live — no restart; a broken scope file keeps the last good one with a warning.
• feat(dap): the dap_* tools register only when a hub is actually configured (env > hub: section > ~/.dap/config.json > default) — the zero-config install hands the model no dead-end tools; /dap <host> still connects on demand and the tools appear at the next launch. dap: false in any tools: scope turns the family off.
• feat(app): a Tools section in the app settings — one live switch per known tool id (ids the app cannot wire render disabled with the capability's reason), applied to the running agent without a restart and persisted as tools_availability.json via the new ToolsAvailabilityStore (the same ToolsConfig JSON envelope the CLI parses).
• feat(read): the read tool follows the sqlite availability decision — its description carries the SQLite section only while sqlite is enabled, and the variant swap re-registers in place (shared snapshot store, so hashline anchors recorded by either variant validate for edit).

0.1.290

• release: 0.1.289
• release: 0.1.282
• memory: supersede stale CPU-storm notes, record hub-test flakiness
• feat(roles): retry transient transport failures instead of killing the turn
• fix(app): onboarding dark theme — themed provider cards, dots, badges
• deps: switch to fa_hub_client 0.2.8 (hosted) — the published backoff fix
• deps: fah_hub_client 0.2.8 from the IstiN fork — backoff overflow fix (CPU storm)
• feat(cli): tell the agent about local ! commands via a steering notice
• fix(cli): steer file-prefixed busy input; run/drop leftover steering loudly
• fix(tools): media slots never inherit the main provider key; generate_image surfaces MiniMax base_resp errors
• memory: CPU burn investigation notes
• feat(cube): fa1.dev registry client, /cube templates + /cube install
• feat(cli): cube sandbox settings picker lists the built-in security presets
• feat(cube): resolver falls back to built-in security presets by id
• feat(cube): built-in security-level presets (L1-L3 x core/full) with tests
• fix(tui): shift+enter via legacy ESC CR wire (alt+enter); pin keyboard-protocol contract
• fix(tui): kitty shift+enter newline; cap streamed tail line growth
• memory: note that origin is a local file mirror (no GitHub remote)
• fix(cli): approval note Ctrl+U; hoist per-keystroke regexes
• test(cli): pin line-mode /exit-during-run interleaving
• fix(providers): strip a leading UTF-8 BOM in the SSE decoder
• chore(test): drop leftover PTY debug screenshots
• memory: session notes from the tui/gateway/debugging fixes
• feat(memory): hot-reload the memory config at runtime
• fix(cli): folder paths stay messages; session switches re-apply folder model memory
• fix(cli): slash and bang commands execute while a run streams
• fix(tui): backspace erases typed note characters in the approval prompt
• fix(tui): keep the history-cap trim fence-balanced
• test(integration): PTY proof for the approval selector and git-prompt fail-fast
• snapshot: local tree 2026-09-02, grafted onto upstream 30ac68b4 (repo line was rootless)

0.1.291

• ci(publish): honor publish_to: none instead of failing every tag

0.1.292

• feat(widgets): live state sync between board tile and fullscreen app + 1x1 icon tiles

0.1.293

• redact: stage 2 — agent hook wiring for the layered pipeline (issue #24)
• feat(redact): layered redaction pipeline core (issue #24 stage 1)
• feat(pub): publishable again — hosted dart_tui dep, vendor stays local
• feat(app): Copilot provider entries re-auth via the device-code flow
• fix(app): launcher tile labels no longer glued to the icon square
• fix(app): onboarding page 3/4 mockups readable in dark mode + light goldens

0.1.298

• test(ollama): diagnose live forced-tool-call null args instead of a bare cast

0.1.299

• feat(browser): the browser extension (issue #23) — browser_ext/ (Chrome MV3) pairs a local fa with the browser over a loopback WebSocket bridge (fa serve --bridge [--port N] [--token T], /browser connect|status): wire protocol v1 client (hello/welcome, 1s→30s reconnect backoff, offline outbox, msgId dedupe, ping keepalive), one-time 32-byte pairing tokens (.fah/bridge/token, mode 0600, rotated by every connect, constant-time compare, non-chrome-extension:// origins refused, AC15), two-way mail relay over the file messaging fabric (browser-ext/<agentId> mailboxes). Eleven browser_* tools over the bridge (navigate/tabs/switch_tab/click/type/press_key/select/read_dom/ eval/screenshot/wait_for, exec tier, availability-gated under the browser + browser_eval ids — hidden until an extension pairs, docs/tool-availability.md). Two control planes: quiet content-script DOM ops by default, per-call trusted: true chrome.debugger path (E23 denied when DevTools owns the tab, any-tab screenshots via Page.captureScreenshot, AC16) with the debugging infobar as the honesty signal; task tab groups (fa — <task>) close agent-opened tabs on task_end or bridge disconnect (AC17). Self-contained mode: dart2js build of browser_ext/dart/ (scripts/build_browser_ext.sh → sw/agent.js + build/fa-extension.zip) runs the agent core inside the service worker — panel provider form (OpenAI-compatible endpoints; deterministic fake: provider for tests), approval banner (30s timeout = deny), JSONL session
  • compaction in chrome.storage, no shell (shellUnavailable), keys read only inside the SW (AC8). The embedded agent joins the DAP hub with an E2E-encrypted CLI-compatible identity (faDapKey): dap_peers/dap_dm tools, one mail deduper across bridge + hub links (AC18). Headless CI: test/browser_ext/ (real Chrome via --load-extension, --headless=new; integration-tagged) in .github/workflows/ browser-ext.yml; unit layers dart test test/browser/, browser_ext/dart dart test, node --test browser_ext/test/. Docs: docs/browser-extension.md.

0.1.300

• fix(aiin): open the web sign-in popup inside the tap gesture
• fix(app): drop the duplicate aiin_connect_flow golden-guard exemption
• fix(app): drop a duplicate golden-guard exemption key; memory: session notes
• feat(aiin): one-click web sign-in — the OpenRouter-style popup flow
• test(redact): pin the agent file-reading scenarios end to end
• docs(AGENTS.md): name-based addressing in the messaging fabric bullet
• fix(redact): entropy layer stops shredding paths, hashes and lockfile integrity
• test(app): regenerate goldens after AIIN-hosted form changes

0.1.301

• feat(aiin): paste-key fallback — the service now blocks our redirects
• memory: session notes
• feat(skills): bundled create-goal skill — goal-writing discipline as /create-goal

0.1.302

• feat(aiin): official aiin.by mark, identity-provider picker, model search
• test(redact): integration e2e for issue #24 AC5-AC8
• fix(aiin): Safari popup — no awaits before window.open, paint the popup
• test(redact): integration e2e for issue #24 AC5-AC8

0.1.304

• fix(publish): assemble the redaction e2e PEM fixture at runtime — pub.dev's key-leak validator rejected the 0.1.303 upload because the archive carried a literal (fake) private-key block in test/integration/redaction_e2e_test.dart; the file now builds the same byte-identical string from chunks.
• fix(roles): 500-class gateway errors ("Internal network failure, please try again later") now classify as transient transport failures and retry in place with backoff before failing over (previously only 502/503/504 did).
• feat(messaging): agent_directory rows show an 8-char short id, session names, last-activity ("active 2m ago" vs "asleep") and home-shortened cwd; agent_message to an asleep target launches a detached headless run of that session so pending mail is processed immediately (wake: false opts out).
• test(cli): pin session-start memory maintenance off in boot-race tests (CI flake: consolidate() consumed the scripted turn before /exit).

0.1.303

• fix(roles): classify 5xx internal errors as transient transport failures
• feat(messaging): readable agent_directory + auto-wake for asleep mailboxes

0.1.305

• feat(aiin): use the hosted AIIN sign-in page as the connect entry

0.1.306

• ci: workflow_dispatch for ci.yml — manual gates when pull_request can't fire

0.1.309

• fix(cli): skill autocomplete for partial skill names — typing /goal or /skill:goal in the TUI composer now offers /create-goal (the raw-prefix match broke on the embedded slash), the accepted item still inserts the canonical /skill:<name> form.

0.1.308

• fix(roles): provider watchdog timeouts ("TimeoutException after 0:03:00: Future not completed", "request timed out") classify as transient transport failures — retried in place with backoff, then failover to the next chain entry, instead of killing the turn (YoClip agent report).

0.1.307

• fix(widgets): unique router instanceId per engine — frozen tiles and dead buttons

0.1.312

• fix(compaction): the summarizer-down local-trim valve no longer wedges the session — a token-boundary cut could land between an assistant tool call and its result, leaving an orphaned ToolResultMessage that made strict providers reject every following request ("400: tool_call_id is not found", Kimi). The trim now advances past leading tool results.

0.1.311

• fix(messaging): scheduled self-reminders actually arrive — 'self' was never resolved to the agent's mailbox, so schedule_message fired on time and then the mail vanished into a phantom <root>/self inbox; start() now migrates the stranded legacy mail into the real mailbox. Cross-mailbox from attribution fixed; parseDelay ms/fractional units fixed ('90ms' was 90 s).
• feat(cli): terminal visibility for scheduled messages — dim '[sched] in 25m:

0.1.310

• refactor(cli): move trajectory view additions out of agent_cli (file size guard)
• fix(windows): restore generated plugin files stripped in #31 (review)
• fix(trajectory): close audit findings — focus consumption, keyboard copy, a11y, safe-area, CRAP (#25)
• test(trajectory): regenerate toolbar+layout goldens, refresh AGENTS.md (#25)
• feat(trajectory): integration — export menu, request persistence, guard closed (#25)
• feat(trajectory): real-content feed rows, guaranteed details, Gantt timeline (#25 L3-L6)
• feat(trajectory): full-screen adaptive shell + header (#25 L1-L2)
• feat(trajectory): data completeness for the ledger view (#25 L7)

0.1.314

• fix(app): visible section hint on dark theme + device-flow connect by default (#35)
• fix(browser_ext): mirror canvaskit into
• fix(browser_ext): externalize inline web scripts; local canvaskit
• fix(browser_ext): CSP + local canvaskit let the embedded fa app load
• fix(browser_ext): bundle the fa web app where the panel probes it

0.1.315

• fix(browser_ext): panel boots clean — verified by a headless CDP check
• feat(aiin): re-authenticate an existing AIIN entry from the editor
• fix(app): actionable hint when the GitHub token cannot create repos (#35)
• fix(redact): a bare .env filename is no longer masked as a credential

0.1.316

• fix(app): reject repo-rightless GitHub tokens at connect time (#35)

0.1.317

• feat(browser_ext): tools_state/tools_put wire the panel Tools section; seed models from the SW snapshot (#34 phase 2)
• revert(app): restore main.dart and build_browser_ext.sh pre-48f757ba
• revert(app): un-publish support agent's staged WIP accidentally included in 48f757ba
• feat(app): runtime-configurable device-flow OAuth app + fallback warning (#35)
• fix(browser_ext): extension panel boots the relay before any local path (#34 item 1)
• feat(browser_ext): the panel agent moves onto the worker relay (#34 item 1)

0.1.318

• fix(scripts): dart fix e2e_extension_agent — unblock the repo analyze gate
• fix(app): bootstrap empty widget repos via the Contents API (#35)
• fix(browser_ext): issue #41 — unbreak DAP dm replies under tab-context decoration
• fix(app): publish into a freshly created (empty) GitHub repo (#35)
• feat(app): Browser connect tab — GitHub OAuth web flow via fa1.dev (#35)
• fix(app): the relay IS the FaChatConnection the models screens render (#34)
• feat(browser_ext): real-wire e2e for the extension agent + provider merge semantics (#34)

0.1.319

• fix(app): publish widget sources at the repo root, full-snapshot tree (#35)

0.1.320

• feat(js-ext): JavaScript extension system — QuickJS engines, jsr.ext.* bridges, install/trust, fa ext CLI, crap-guard (issue #32) (#37)
• chore: dart fix drift + ignore local worktree/panel build dirs; memory: session notes
• fix(extension): live e2e with a real provider — approvals, user bubbles, yolo mode
• fix(browser_ext): status snapshots must not flip the transport to streaming
• fix(browser_ext): the panel relay never connected — port name + envelope mismatch

0.1.323

• fix(browser_ext): page/app screenshots reach the model as vision image blocks — page_screenshot/app_screenshot and the v1 screenshot op return ImageContent plus compact metadata text; the base64 blob no longer floods the context as invisible text (provider adapters already serialize tool-result images for openai-completions/anthropic/google).
• fix(browser_ext): thinking deltas stream to the panel as thinking_delta UI events (they were dropped at the host) — the relay chat renders reasoning as its own collapsible bubble; pure event→UI mappings extracted to host_event_map.dart (VM-testable, agent_host stays web-only).
• fix(app): a relay client without a mounted approval handler stays silent instead of instantly denying — a bystander surface raced ahead of the real UI and recorded denials the user never chose; the SW's 120s timeout remains the backstop, and a throwing handler still denies.
• test(browser_ext): live e2e grows read_dom (reads example.com's real heading), screenshot (vision block through the loop) and thinking count legs; robust multi-approval auto-allow; scripted fake provider gained think/screenshot directives and an image-seen reply.
• ci: the Chrome extension builds in the release pipeline and ships as a release asset (fa-extension.zip) and on fa1.dev (/extension/), landing page gains the download + load-unpacked card.

0.1.325

• feat(browser_ext): session history in the extension panel — list + open SW archives (#64)
• fix(ci): binary smoke SIGPIPE + smoke gate passes skipped matrix on push/schedule (#63)

0.1.326

• fix(59): schedule_message records survive host reconfigures and always surface (#65)
• fix(browser_ext): explicit boots win over the auto-boot storage snapshot
• feat(27): hub-first agent messaging fabric with file fallback (phase 1) (#58)
• build(browser_ext): strip on-device CDN loaders from the panel bundle
• fix(browser_ext): explicit boot keys persist — auto-boot no longer reverts them
• fix(ui): sharp chevron tip in the web splash and favicon
• feat(browser_ext): durable transcript replay — history survives SW restarts and session_open

0.1.327

• feat(29): config service + fa config verbs + agent config tool (S3) (#70)
• fix(flutter_app): drawer keeps archived sessions after a relay session_open (#72)
• build(browser_ext): mark the on-device stub rejection handled (#69)
• fix(flutter_app): render replay rows whose nested maps come from dartify (#71)

0.1.328

• feat(35): out-of-view status polling + real-GitHub E2E (E2E-1) and AC12 check (#75)
• feat(27): discovery surface in agent_directory - presence, capabilities, name@machine (phase 2) (#68)
• feat(35): timed status polling + offline hint in My publications (#67)
• fix: relay turn-state indicator, yolo-everything, yolo checkbox in the approval dialog (#73)

0.1.329

• feat(27): hub-first agent messaging fabric with file fallback (phase 1) — the CLI composes the DAP hub into the agent messaging fabric (FallbackMessagingRepository over lib/src/messaging/agent_fabric.dart): hub-resolvable targets (16-hex ids, unambiguous display names, #channels) deliver over DAP/1, file inboxes stay the offline fallback with forward-on-reconnect (flushed by the 2s inbox probe), inbound hub mail merges into the main inbox drain only, deduped by id and sender+time+body; recipient guards consult the hub roster so hub peers are deliverable through agent_message. Docs: docs/dap.md §12. Discovery/presence states (busy) and the A2A gateway remain for phases 2-3.

0.1.331

• feat(83): sort session lists current-folder first, each group newest first (#84)

0.1.332

• feat(81): open user requests survive compaction summary-of-summary (#82)

0.1.333

• feat(dap): /dap start one-step bring-up, fa hub serve, DapHubService for web/extension (#90)

0.1.334

• feat(29): AC11/AC12 cross-surface config parity (+AC8/AC10 pins) (#92)
• fix(59): stop cross-instance theft of scheduled self-reminders (#88)
• fix(86): compaction — LLM finds asks itself (no hardcode) + tool-results assessment in checkpoint prompts (#87)

0.1.335

• test(35): make the AC12 E2E catalog-first (#96)
• feat(91): add --log-file to tee the live session trace to a file (#95)

0.1.336

• fix(85): context pair-integrity — validate and repair tool pairing at the request boundary (#93)

0.1.337

• feat(101): add explicit --prompt-file
• fix(97): TUI secret sheet — value-first focus, placeholder name, masked input, Enter explains itself (#98)
• fix(99): clear macOS traffic lights on the trajectory screen (#100)

0.1.338

• build(deps): bump playwright (#106)
• feat(89): Outlook Office Add-in host bridge (fa1.dev/outlook) (#94)

0.1.339

• feat(115): visualize pending scheduled follow-ups in the CLI TUI (#117)

0.1.340

• fix(107): un-red the tag pipeline — publish validation + LLM integration job (#108)

0.1.341

• feat(102): interactive dynamic messages - UI surface (tile, sheet, save-as-app) (#120)
• fix(121): scope the pub.dev publish token to the publish step only (#123)
• feat(113): fa1.dev reflects the Chrome extension, Outlook add-in, and full tool surface (#116)

0.1.342

• feat(125): bench.yml — run fa on terminal-bench on a GitHub runner (#126)
• fix(109): clear stale TUI ask answers and keep the prompt frame inside the terminal width (#112)
• chore(114): terminal-bench adapter and runner for fa (#119)

0.1.343

• fix(122): run dart pub publish --dry-run in PR quality gates (#128)
• fix(118): strip Dart core exception prefixes from model-visible tool errors (#127)

0.1.346

• ci(browser-ext): placeholder .env for the dispatch-only panel-app build (#136)

0.1.349

• docs(143): KB for the Monarch VersionOverrides manifest fix (#156)
• docs: backend agent mode research — fa behind product Go backends
• fix(143): manifest VersionOverrides command surface for Monarch/new OWA (#151)
• docs(144): real Outlook add-in installation guide (#149)

0.1.352

• docs(backend-mode): pin MVP shape with mermaid architecture
• DAP slash-flow + CodeMie SSO expiry detection & auto re-auth (extension/app/CLI) (#145)

0.1.353

• ci(merge-trigger): prefer PAT_TOKEN so machine merges land as the owner (#164)
• ci(159): skip mobile release cleanly when no binaries were built (#162)
• docs(backend-mode): reconnect semantics — turns survive client disconnect
• docs(backend-mode): memory mapping for backend mode
• docs(backend-mode): memory mapping for backend mode

0.1.354

• docs(159): KB — clean no-op for content=none dispatch of build-mobile.yml (#165)

0.1.355

• docs(144): KB — real Outlook add-in installation guide (#154)
• docs(168): KB — mobile trajectory view (parity desktop) (#180)
• feat(166): apps grid custom cell spans (1×2, 2×1, 1×3, 3×1) (#176)
• feat(168): mobile trajectory view — parity with desktop (#175)
• docs(157): KB — DAP e2e family triage (two cases were #158 races) (#163)

0.1.357

• feat(135): windowed session loading for app chat (#139)
• feat(155): backend agent mode — HEP v1 events output, --attach, SIGTERM graceful abort, image registry (#193)
• feat(137): generic chrome.* bridge — browser_api / browser_api_catalog (#172)
• feat(171): session image registry — [Image N] send-once request assembly (#190)
• fix(184): host-portable flutter_app tests — quickjs skip guard + injected-ASR fix (#191)
• ci(161): daily auto-publish — TestFlight + pub.dev + CLI + website + add-in, self-filing fix issues (#170)
• fix(152): root-cause the browser-ext e2e dispatch flake class (#158)
• ci(177): drop the 'Quality gates' alias — protection switched to 'Quality gate' (#192)

0.1.361

• feat(178): honor FA_LOG_FILE env var as the default for --log-file (#216)
• fix(197): four minor windowing defects from the round-4 review (#211)
• fix(196): HEP follow-up from #193 review (#210)
• fix(195): images follow-up - stale citation renumbering guard + F3/F4/F5 + config tests (#205)

0.1.363

• chore(flutter): raise Flutter floor to 3.47 + js_widget_runtime ^0.4.121 (#230)
• fix(apps): unmute JSR media widgets on iOS — audio session category drift (#227)
• feat(apps): widget manifest i18n — additive nameI18n/descriptionI18n keys (#226)
• fix(catalog): single-slash URL join for all release-asset fetches (#219)

0.1.364

• feat(198): tree-grouped session listings (core + CLI) (#220)
• fix(ext): sessions survive reload/update/restart — issue #228 persistence vectors (#236)
• fix(config): merge-before-write + atomic save for ~/.fah/config.yaml (#221) (#235)
• fix(catalog): web installs widgets from CORS-friendly raw URLs (#231)
• fix(ctx): footer meter, over-window guard and compaction share one request-size basis (#217)

0.1.365

• fix(142): unblock tb runs — LPT shards, doubled timeouts, buildable debian task images (#253)
• fix(fa_ui): hide the "Load earlier" banner on empty sessions (#223) (#247)
• test(233): triage + fix the 13+4 integration-failure clusters (#248)
• feat(224): apps collapse toggle - Apps icon in the chat header actions (#245)
• fix(215): split agent_service.dart and settings.dart under the 2800-line gate (#244)
• fix(234): latch the web boot signal; e2e waits on it, not the transient splash class (#242)

0.1.366

• feat(200): session search field pinned atop the session lists (#246)
• fix(238): close migration-ordering data-loss window, stop deleted-session resurrection (#240)
• feat(199): session loading off the UI isolate - background parse, parallel listing, quick names, generation guards (#243)
• feat(239): store automation — daily EXTERNAL TestFlight + manual release-appstore.yml (App Store review) (#252)
• feat(147): Harbor adapter for fa + Terminal-Bench 4.0 run (self-hosted + Modal + Hub upload) (#254)
• fix(app): web IndexedDB session persistence — per-record keys, migration, quota isolation (#237) (#249)

Unreleased

Unreleased

Unreleased

Unreleased

Unreleased

Unreleased

Unreleased

Unreleased