v3.0.3

Added

• Detached journal log failures now leave a readable trail. The
  recording daemon and capture pipeline run with stdout/stderr on
  /dev/null, so a transcription, land, or recording failure was previously
  undiagnosable after the fact beyond a generic notification. Failures now
  append a timestamped stage\terror line to
  <log.audio.tmp_dir>/pipeline.log (trivially truncated past 1MB), and the
  ✕ journal log failed notification includes the underlying error text
  (truncated for readability) plus the log path. See
  DATA-FLOWS.md and USAGE.md.

v3.0.2

Added

• GUI launchers can now find the journal without shell environment. Two
  changes close the class of silent failures where a Hammerspoon hotkey or
  Claude Desktop MCP config ran journal in a launchd environment (no
  ~/.zshrc exports, minimal PATH) and died invisibly:
  • Global default_repo — journal config set-default-repo <path>
    records the repo in ~/.config/journal/config.yaml (XDG-aware); repo
    resolution falls back to it when no --journal-dir flag, $JOURNAL_DIR,
    or enclosing .journal resolves. Explicit flag/env still wins and still
    fails loudly; the "not inside a journal repo" error now names every
    resolution method. journal config shows the current setting. Applies
    to journal mcp too — Claude Desktop configs no longer need env blocks
    for the single-repo case.
  • Homebrew fallback for external tools (internal/toolpath) — ffmpeg,
    whisper-cli/whisper-cpp/whisper, and terminal-notifier now resolve via
    PATH or well-known install locations (/opt/homebrew/bin,
    /usr/local/bin, Linuxbrew), so GUI-launched recordings work without
    PATH gymnastics. PATH hits always win.

v3.0.1

Fixed

• journal models pull now actually provisions the model journal log
  needs. The pull command's transcriber slot downloads a faster-whisper
  (CTranslate2) model for the WhisperX meeting pipeline — an artifact the
  whisper.cpp backend behind journal log <audio.wav> cannot read — yet every
  missing-model error and doctor hint said "run journal models pull", which
  succeeded while fixing nothing. A new log_transcriber pull slot (config
  log.transcriber.{model_id,revision,checksum}, default
  ggerganov/whisper.cpp) fetches the ggml artifact (ggml-<model>.bin), and
  model resolution — shared by journal log, its record-start preflight, and
  journal doctor — now checks the manual-install path
  (<model_dir>/<model>.bin, which always wins) and the pulled layout.
  models list/verify/MODELS.md cover the log model too. Set
  log.transcriber.model_id: "" to keep manual installs authoritative.
• journal doctor no longer reports failed optional checks as [ok ].
  v3.0.0's audio-toolchain checks (and the older reranker/Quill informational
  checks) printed failure text — "whisper.cpp binary not found", "model missing
  — run journal models pull" — under an [ok ] mark and still concluded
  "all checks passed". Optional-but-broken checks now render as [warn], the
  summary reads "all required checks passed — N warning(s) above", and the
  --json report gains a per-check warn field (additive; ok semantics and
  the exit code are unchanged — warnings still never fail doctor).

v3.0.0

Added

• Linux recording support for journal log. The mic-recording toggle
  (bare journal log) now works on Linux, not just macOS: it captures via
  ffmpeg -f pulse by default, or -f alsa on ALSA-only boxes (set
  log.audio.backend: alsa), selected automatically by OS
  (log.audio.backend: "", the default) with an explicit override available.
  Desktop notifications on Linux go through notify-send, degrading silently
  (same as macOS's osascript/terminal-notifier) when it's absent. An
  unsupported platform (anything other than macOS/Linux) fails fast on the
  starting press, before the recording daemon spawns. macOS behavior
  (avfoundation, osascript/terminal-notifier) is unchanged.
• log.audio.{silence_duration,silence_noise_db} config keys. The
  silence_autostop watchdog's silence-interval length (default 30 seconds)
  and silencedetect noise floor (default -35 dB) are now configurable
  instead of hardcoded, so a noisy room or a different mic can tune when the
  safety-net stop fires. Existing behavior is unchanged unless overridden.
• Meeting pipeline migrated onto the shared land/index core (Phase 5c).
  journal quill-sync and journal transcribe now write and index through the
  same internal/log.Land/IndexTranscript primitives the voice-note pipeline
  (journal log) uses, replacing bespoke os.WriteFile glue in each command.
  Frontmatter (source: quill/source: whisperx), filenames, and chunking
  behavior are byte-identical to before — this is an internal refactor, not a
  behavior change. journal models pull gains a second, independently-optional
  gated model slot (diarization.*, same shape as transcriber.*) so the
  meeting pipeline's pyannote diarization model can be provisioned as a fast
  credential preflight — see docs/TRANSCRIBE.md and
  docs/CONFIGURATION.md. diarization.model_id is
  empty (disabled) by default; journal models pull skips it entirely until
  configured, and attempts every configured model even if one fails, exiting
  non-zero only if any of them did. internal/models.Pull gained a
  PullFile variant supporting a configurable remote/local filename (needed
  since pyannote's repo has no model.bin) — Pull is now a thin wrapper
  around it with no behavior change for existing whisper-model call sites.
• journal models pull gated/HuggingFace-token path (Phase 5b). Extends
  Phase 2a's model provisioning with support for gated HuggingFace repos (the
  first being pyannote's speaker-diarization models, arriving with the meeting
  pipeline migration). Set transcriber.gated: true and transcriber.accept_url
  in config; pull sends HF_TOKEN (read from the environment only, never
  config) as a Bearer token. A gated model with no/invalid token fails with an
  explicit "accept terms at <url>, set HF_TOKEN" message instead of a raw
  401. MODELS.md now records each installed model's gated status and
  acceptance-page link. Ungated pulls are unaffected — no token, no behavior
  change.
• MCP tools journal_log_text / journal_log_audio (Phase 5a). The MCP server
  (journal mcp) now exposes the journal log --text and journal log <audio.wav>
  pipelines as tools: journal_log_text(text) runs shape→assemble→land→index,
  journal_log_audio(audio_path) runs transcribe→shape→assemble→land→index for a
  server-local audio file. Both return {path, title, landed} and honor
  local_only/local_only_mcp exactly like the rest of the MCP surface. The
  mic-recording stage stays CLI-only and is never exposed over MCP — an MCP server
  must not seize the user's microphone.
• Start/finish desktop notifications for journal log (Phase 4, macOS only).
  The recording toggle now pairs the existing terminal output with a real desktop
  notification: starting a recording pops "● recording", and the async pipeline
  landing the note pops "✓ logged: <title>" with the note's relative path. Sent
  via osascript (display notification), falling back to terminal-notifier if
  osascript is unavailable; a missing/failing notifier degrades silently (logged,
  not surfaced) and never blocks or fails the recording or pipeline.
• internal/audio.Notifier: an injectable desktop-notification boundary
  (DefaultNotifier for production, FakeNotifier for tests) mirroring the
  existing Recorder pattern, so tests never pop a real OS notification.
• Hammerspoon hotkey binding documented (docs/USAGE.md) — an init.lua
  snippet binding a single hotkey to bare journal log, plus an optional
  menubar-dot add-on, so the recording toggle is fully usable from one key press.
• journal log recording toggle — mic capture (Phase 3, macOS only). The bare
  journal log command now toggles mic recording: the first press starts a detached
  background recorder (ffmpeg -f avfoundation, 16 kHz/mono/16-bit PCM) and prints
  "● recording"; the second press stops it and hands the finalized WAV off to the
  existing transcribe→shape→assemble→land→index pipeline asynchronously, so both
  presses return immediately. --start/--stop/--cancel/--status give explicit
  control (--cancel discards the recording — no note is produced). The toggle state
  lives in a lockfile ($XDG_RUNTIME_DIR/journal-log.lock or
  <tmp>/journal-log/journal-log.lock) holding {pid, wav_path, started_at}; a dead
  PID is detected and cleaned up automatically on the next press. log.audio.max_duration
  (default 900s) self-finalizes a long recording; log.audio.silence_autostop (default
  off) is an optional safety-net stop after sustained silence. The recorded WAV is
  deleted after a successful run unless log.audio.keep_wav: true, in which case its
  path is recorded in the landed note's audio: frontmatter field. A WAV passed
  directly (journal log <file>.wav) is never auto-deleted, regardless of this
  setting — only recorder-produced scratch files are.
• internal/audio package: lockfile primitives (ReadLock/WriteLock/RemoveLock,
  injectable PIDAlive) and an injectable Recorder interface (FfmpegRecorder for
  production, FakeRecorder for tests) so the mic/ffmpeg dependency never runs in tests.
• log.audio.{tmp_dir,max_duration,silence_autostop,keep_wav} config keys and the
  LogAudioTmpDirAbs() config accessor, additive to the existing log.audio.{device, sample_rate,channels} keys.
• journal log <audio.wav> — transcribe stage (Phase 2b). Pass a WAV file as
  a positional argument to transcribe it locally via whisper.cpp and run the full
  transcribe→shape→assemble→land→index pipeline. No network access: a missing model
  fails fast with "run journal models pull". Silent recordings skip the pipeline;
  transcription errors are retryable (WAV is kept). The transcriber frontmatter
  records the backend/model used; duration_sec is derived from the WAV header.
• log.transcriber.{backend,model,model_dir} config keys replace the previous
  stub log.transcriber.{engine,model} keys. backend selects the engine
  (whisper.cpp, the only built-in); model_dir defaults to the same path as
  transcriber.model_dir so a single journal models pull serves both paths.
• LogTranscriberModelDirAbs() helper on *config.Config mirrors
  TranscriberModelDirAbs() for the log transcriber path.
• internal/log.Transcriber interface with a FakeTranscriber for tests and a
  WhisperCPP default implementation. The boundary is injectable so commands never
  exec a real binary in tests.
• journal log --text "..." — voice note capture (Phase 1). The full
  shape→assemble→land→index pipeline is now available for typed text. The LLM
  shaping step (configurable via log.shaping.enabled) cleans disfluencies,
  generates a title and summary, extracts @todo/@decision/@question markers,
  and tags the note. Notes always land to logs/YYYY-MM-DD-HHMM-<slug>.md even
  when shaping is unavailable (raw fallback). Index failure is non-fatal.
• SourceVoice = "voice" source constant in internal/store; voice chunks
  are indexed separately from notes and transcripts. No schema migration required —
  the source TEXT column already accepts any value.
• journal search --source voice (aliases log/logs) scopes search results
  to voice-note chunks.
• log: config namespace with shaping, landing, audio, and transcriber
  keys; LogAbsPath()/LogRelPath() helpers mirror TranscriptsAbsPath().

Fixed

• journal log hotkey/mic-toggle no longer fails silently. When the
  background transcribe→land pipeline can't run (most commonly whisper.cpp or the
  model isn't installed), the detached pipeline process has no visible stdout, so
  the only feedback the user gets is a desktop notification — and previously only
  the success notification (✓ logged) ever fired. journal log now pops a
  ✕ journal log failed notification naming the retained WAV to retry with, and a
  low-key "empty recording — nothing to log" notification when a recording
  transcribes to silence, so a hotkey press always resolves to a visible outcome.

v2.7.1

Added

• journal dismiss — bulk-dismiss open todos in one commit. A new dismiss
  subcommand selects all open @todos matching a --project filter, an
  --before/--older-than age window, or both, and rewrites each to @done YYYY-MM-DD (same rewrite as journal done). Requires explicit confirmation
  (--yes or an interactive y prompt). All file edits and re-indexing land in a
  single auto-commit whose message records the filter used (e.g. dismissed 5 todo(s) (project=acme, before=4w)). The optional --resolution flag appends a
  Resolution: line to each dismissed block. Single journal done behavior is
  unchanged.

Fixed

• journal today now aggregates all of a day's notes, not just the daily file.
  The Notes section previously read only daily/YYYY/MM/YYYY-MM-DD.md, so a day
  with notes captured via --project (or any project note) was reported empty.
  It now queries the index for all note chunks with source=note and
  created_at >= midnight, grouping project note sections with a source label.
  Meeting transcripts are excluded from Notes and continue to appear under
  Today's meetings. The MCP today tool inherits the fix (it routes through
  gatherToday). The journal://today MCP resource is unchanged — it remains
  the literal daily file; the aggregated view is the today tool.

• Embed retry now rides through transient Ollama runner restarts. The retry
  window for transient embed-runner crashes (400 with EOF / do embedding request body — a known llama.cpp/Metal SIGTRAP flake on Apple Silicon) has
  been extended from ~1.4 s to a 45 s wall-clock budget, sized to outlast a
  model reload. Backoff is exponential (n²×100 ms), capped per-sleep at 5 s,
  and jittered ±25%. Transport-level dial failures (ErrUnreachable) still fail
  fast via maxRetries; non-retryable 4xx (model not found, bad dimensions)
  still fail immediately.

v2.7.0

This release rounds out the MCP surface (the agent-facing side of journal),
makes decisions & todos first-class, and adds project tooling — a docs site,
a curated release pipeline, and the agentic-workflow layer.

Added

• MCP synth tool. The MCP server now exposes synthesis as a synth tool so
  MCP clients (e.g. an agent drafting a weekly Slack summary) can run
  weekly|daily|meetings|decisions|stale synthesis jobs without shelling out to
  the CLI. By default the tool calls the synthesis provider and returns the
  generated text without writing a file (persist: false); set persist: true to
  also write the draft note to disk, mirroring journal synth --write. Honors
  synth_provider, local_only, and voice_profile exactly as the CLI does;
  returns a clean {"error":"…"} when synthesis is unavailable.
  Optional scoping params: kind (default weekly), days, project, date.
• MCP ask tool — runs retrieve→synthesize and returns
  {answer, citations} (grounded text + path:line references) so clients get a
  direct "what did I decide about X" answer instead of raw chunks. When no chunks
  match, returns "No relevant notes found." rather than calling the model
  ungrounded. Honors local_only / provider availability with a clear error
  (#32).
• MCP stats and today tools — the same stable JSON as
  journal stats --json / journal today --json, for "how's my note volume" and
  "what does my day look like" (#31).
• MCP resources — journal://today, journal://recent, and
  journal://projects/{slug}/index expose raw Markdown as addressable context via
  resources/list / resources/read, so clients can pull journal context without
  orchestrating tool calls. URIs are stable across server runs
  (#37).
• MCP prompts — weekly reflection, decisions review, and project status
  are exposed as first-class, one-click prompts with journal context pre-assembled
  (reusing synth.AssembleWeekly / AssembleDecisions)
  (#44).
• journal tags — list distinct #tags with usage counts, and
  journal tags rename <old> <new> rewrites a tag across all notes, re-indexes the
  affected files, and auto-commits (--dry-run previews). Boundary-aware so #foo
  isn't matched inside #foobar (#33).
• First-class decisions & todos. Dedicated capture commands, crisp
  date/statement/citation rendering, proactive surfacing in journal today, and
  resolution notes on journal done. Fully backward compatible — existing
  @decision/@todo/@done blocks parse and surface unchanged
  (#47).
• Documentation site at https://journal.eamann.com — an mdBook site
  (Catppuccin Latte) with warm, second-person prose across 14 chapters (install,
  capture, search, synthesis, meetings, configuration, integrations). Additive: the
  docs/ tree is untouched (#38).
• One-click releases. A prepare-release workflow (Actions → Prepare Release →
  version) finalizes the ## [Unreleased] CHANGELOG section to the version, tags,
  and triggers the Release workflow; and the GitHub Release notes now come from the
  curated CHANGELOG section (scripts/changelog-section.sh + GoReleaser
  --release-notes) instead of an auto-generated commit list. See
  docs/RELEASING.md. Deterministic, project-specific tooling.
• Agentic-workflow layer + CLAUDE.md. GitHub Actions for agent-ready issues
  (auto-label → trigger → PR), a planning-approval gate for high-complexity work,
  @claude PR feedback (gated to write-access collaborators), and a retro flow;
  plus issue/PR templates, a label set, and a contributor guide grounded in the
  real layout. Contributor/CI tooling — no change to the shipped binary
  (#25).

Changed

• Better LLM-as-reranker quality. The rerank prompt is now a structured
  relevance rubric (0/5/10 anchors) instead of "reply with only the number", and
  score parsing is a robust multi-strategy parser (N/10 fraction → labelled
  Score: N → last in-range number, skipping negatives). More reliable precision
  on the optional reranking path (#36).

Fixed

• journal index --watch survives transient embed failures. A per-file embed
  error during a watch pass is now logged and skipped (continue) instead of
  aborting the whole batch; only context cancellation still stops the run.
  Debounce coalescing of rapid edits is now covered by tests
  (#35).
• Indexing retries transient Ollama embed-server failures (sporadic 400 /
  EOF) instead of failing the run, hardening large journal index passes
  (#7).

v2.6.1

Changelog

• 520da68: feat: section-aware transcript chunking (summary is its own chunk) (@ericmann)

v2.6.0

Changelog

• 60a38bc: feat: journal transcribe — ingest non-Quill recordings (WhisperX → summarized, indexed transcript) (@ericmann)

v2.5.0

Changelog

• 20aa6fd: feat: pluggable OpenAI-compatible providers for synth and embeddings (@ericmann)

v2.4.5

Changelog

• 78405a9: fix: bump wazero v1.8.2 -> v1.12.0 for Go 1.26 compatibility (@ericmann)