Skip to content

Releases

v1.50.0

Choose a tag to compare

@sblattj sblattj released this 21 Nov 07:12

Changed

  • Encrypted backups no longer name objects by their content hash. Names are an HMAC under a random key kept inside the encrypted manifest, so a storage host that already has a copy of one of your files can't confirm it is in the backup. Every recipient can read the key, which leaves room for shared backups (#238).
  • Encrypted objects and manifests are zero-padded to Padme size buckets (at most about 12% larger, 4 KiB minimum) before encryption, so exact file sizes don't show. Restore and verify trim to the recorded size (#238).
  • The plaintext _whosaid-backup.recipients file is gone. The recipient fingerprint moved into the encrypted manifest, and a run deletes the old file. Changing recipients now also rotates the names key (#238).
  • Backups written by earlier versions still restore. Their next run re-uploads every object under a keyed name, and backup gc removes the old ones. Without local state or --identity, a run against an existing encrypted backup now warns that it will upload everything again (#238).
  • <ws>/.backup/ state files are written 0600, since they cache the manifest and its names key (#238).

v1.49.1

Choose a tag to compare

@sblattj sblattj released this 21 Nov 05:54

Fixed

  • whosaid drafts audit: the "cache has no messages in this chat since DAY" caveat named the UTC day, so an evening message in the Americas showed as the next day; it now names the local day, like the chat-day folders (#237).
  • Five test suites that failed on main pass again: shell suites no longer lose a grep -q match to SIGPIPE under pipefail (which could also let assert_not_contains pass wrongly), the app network gate recognizes the Drafts page preview's LocalOnlyNavigation coordinator, and the audit caveat test no longer depends on the time of day (#237).

Added

  • ./test/run_all.sh runs every suite with the dependencies it needs and prints one PASS/FAIL line per file; the test-only dependencies (numpy, mcp[cli], websockets) are declared in a test dependency group in pyproject.toml (#237).

Fixes #237

v1.49.0

Choose a tag to compare

@sblattj sblattj released this 21 Nov 05:28

Added

  • whosaid backup init <ws> --to TARGET sets up a trusted backup in one step: checks age, creates a daily identity (0600), prints a recovery identity once for offline storage, writes [backup] with both recipients, then runs a dry run, the first backup and a read-back of every object before turning on the daily schedule and a weekly verify (#234).
  • A state backup tier: the drafts outbox, .watch_state.json and a snapshot of the watcher's launchd settings. Encrypted only (#234).
  • whosaid backup verify <ws> reads objects back from the target and checks their hashes, counts files per tier (backed up, local, waiting), and with --full restores into a temp dir and runs a search that must find a hit (#234).
  • whosaid backup status <ws>: last good backup, last failure, size per tier, last verify and full verify, dated manifests, alarms (#234).
  • Dated manifests (manifests/<UTC>.age, newest 30 kept, --keep), backup restore --at STAMP, and whosaid backup gc, which deletes only objects that no kept manifest references and refuses if any kept manifest cannot be read (#234).
  • The watcher alarms (notification and an ALARM line in .watch.log) when the last good backup is older than twice the interval, when no full verify passed in 8 days, or when age is missing (#234).

Changed

  • A personal backup without the biometric tier now strips voiceprint keys (cluster_emb, centroid, ...) from every JSON file, as a share does, so voice vectors are only backed up when you choose the biometric tier. Restoring a stripped sidecar over a local one that still has its vectors leaves the local file alone (#234).
  • --prune-older-than on an encrypted backup drops the entries from the new manifest and leaves the objects for backup gc, since older kept manifests still name them (#234).
  • backup restore writes files that hold keys or voiceprints with mode 0600, records every restored recording by content hash so the watcher does not transcribe it again under a new name, and re-roots restored watcher settings to the new home and workspace path; it never loads a launchd job (#234).

Fixed

  • A scheduled backup with age missing no longer says "will retry" on every pass: it is recorded as a failure, alarmed once and shown in backup status (#234).

Fixes #234

v1.48.2

Choose a tag to compare

@sblattj sblattj released this 21 Nov 01:36

Changed

  • Docs: the live Parakeet speed and vocabulary-boosting notes now quote only numbers measured on synthetic speech; the boost default is explained without publishing the private benchmark it came from. A test term was renamed. No behavior change.

v1.48.1

Choose a tag to compare

@sblattj sblattj released this 21 Nov 01:32

Fixes #236

Fixed

  • whosaid live --asr parakeet no longer writes words over digital silence (#236). Parakeet makes up a one-word "Yeah." or "Mm-hmm." on stretches of pure zeros, such as the masked mic leg or a re-transcribed sliver. A segment whose own audio is below the -48 dBFS silence floor is now dropped. On the real model, 20 of 30 silent inputs gave a kept line on 1.48.0 and none do now; real speech keeps every word.

v1.48.0

Choose a tag to compare

@sblattj sblattj released this 21 Nov 01:26

Opt-in NVIDIA Parakeet speech-to-text for whosaid live. Whisper stays the default for live and batch.

Added

  • #235: whosaid live --asr parakeet (or WHOSAID_LIVE_ASR=parakeet) transcribes live chunks with parakeet-mlx (mlx-community/parakeet-tdt-0.6b-v2, or WHOSAID_LIVE_PARAKEET_MODEL). Faster per chunk than turbo, and word times always on, so a speaker change inside a line splits at a word. The workspace's [transcribe] vocabulary (--ws, else $WHOSAID_WORKSPACE) plus --vocabulary TERM is boosted on every chunk (beam 5, weight 1.0; on by default when there are terms, --no-boost / WHOSAID_LIVE_BOOST=0 for greedy). The launcher adds parakeet-mlx==0.7.0 to the live uv env only when Parakeet is picked; pip installs use the whosaid[parakeet] extra, and a missing package exits 2 naming it. --model is the chosen backend's model; WHOSAID_LIVE_MODEL stays Whisper's.
  • test/asr_bench/bench_asr.py: Whisper vs Parakeet on your own recording (whosaid-shaped transcript, term spelling, live chunk timing).

Fixes #235

v1.47.0

Choose a tag to compare

@sblattj sblattj released this 20 Nov 23:09

Drafts sent checks catch sends that went out edited, into a new group chat, or into a different chat.

Added

  • #229: a staged new-chat group draft learns its chat's thread id from the Teams desktop cache (exactly one fitting chat created after staging; several: nothing learned, the check says why). A rewritten send soon after staging is reported as probably_sent and stays staged until mark-sent ID confirms it. drafts list, the MCP list and the app flag drafts staged over [drafts] stale_staged_hours (48) with no message of yours since.
  • #232: drafts audit [ID|--staged-before 48h] [--json] and MCP whosaid_draft_audit search every chat for your messages since staging and give one verdict per draft: sent, probably sent, edited, sent elsewhere or not sent (with chats searched and a cache-coverage caveat). Another draft's recorded send is excluded and named. stage on a new-chat draft whose text already went out refuses unless --force/--replace. mark-sent ID --message MID records the chosen message.

Fixes #229
Fixes #232

v1.46.0

Choose a tag to compare

@sblattj sblattj released this 20 Nov 15:28

Added

  • The watcher closes sent drafts (#228). Each pass runs the drafts sent check after the recordings and logs every draft it marks sent (drafts: <id> sent at ...), so a staged draft that went out no longer stays staged until someone runs mark-sent --check. After a Teams cache sync it reads only the workspace, where the sync just put the messages. [drafts] auto_sent_check = false turns it off. It never raises.
  • mark-sent --check --backfill (MCP whosaid_draft_sent_check backfill=true): records the message for sent drafts closed by hand without one, adding a backfilled event and changing nothing else.
  • A stage-time sent guard (#228). Before drafts stage (with or without --replace, and through whosaid_draft_stage) pastes anything, it syncs the Teams cache into the workspace, then checks the draft and every other draft recorded staged in the same chat against it. If one already went out, the stage is refused with already sent at <time> and that draft is marked sent with its message id (how: stage guard). A dry run reports the same and changes no draft; it still runs the sync. WHOSAID_TEAMS_CACHE=off skips the sync, not the check.

Changed

  • mark-sent ID and whosaid_draft_close status=sent look the message up in the chat (--ws / workspace, else WHOSAID_WORKSPACE) and keep sent_message, so a later stage --edit finds it by id. With no workspace or no match, they close as before.
  • drafts new runs the sent check on the recipient's staged drafts before joining one: a draft that already went out is closed (closed_sent in the result) and the new draft starts fresh.
  • sent, dropped, edited and backfilled events record by: the agent session, or watcher.
  • The sent match tolerates an edited send (#229's edited-send case). A name expanded into a full-name mention, a link rendered as label (url) or left bare, and a sentence added before sending no longer hide a sent draft: when a long draft's body (greeting and mention marks aside) is at least 90% inside one of your messages, it counts. The guard and the sent check both depend on this. A draft that is already in a chat as your message is now refused at stage time instead of skipped.

v1.45.1

Choose a tag to compare

@sblattj sblattj released this 19 Nov 23:21

Fixed

  • whosaid backfill no longer rewrites a transcript's # Speakers (N): header with every cluster in the diarization sidecar (#233).
    • Before: the tiny uncounted fragments went into the list too, and N stayed at the old value. One real 87-minute meeting went from 7 names to 110.
    • Now: the header keeps its own membership with renames applied. N moves only by the change, since it counts clusters and two clusters can share a name.
    • No rename: with nothing renamed, the line is left untouched.
    • Uncounted clusters: these stay on the # Uncounted line, except one the registry has now named. A named cluster always counts, so it moves to the header and the sidecar's uncounted and num_speakers follow.

v1.45.0

Choose a tag to compare

@sblattj sblattj released this 19 Nov 15:20

Added

  • whosaid live export-items <stem> [--format json|csv] [--since <stem>] and the read-only whosaid_live_export_items MCP tool (#227). Each ended explainer becomes one row per ask, question, suggestion, decision or next step, with:

    • a stable id (<stem>-<kind>-<12 hex of sha1(normalized text)>);
    • section id and time window;
    • asker, plus an owner when the text names one;
    • the closest verbatim transcript line (clock and speaker);
    • every source (NNN-*.json:asks[i], recap.json:owes[i], <stem>.ai.json:nextSteps[i]).

    Only top-level arrays are read, never revisions[]. An item a correction marked as a mis-hear stays out. Near-duplicates within a call are merged, and each merged row lists all of its sources. Calls whose transcript meta has ended: false are skipped. --since exports every ended call after the given stem. The MCP tool is listed through search_tools; the default listing size is unchanged.

  • <stem>.explain/items.status.json: an open/done overlay keyed by row id. whosaid live mark-item STEM ID --status open|done [--note T] writes it, and so do the served page's checkboxes (POST /item-status, token-checked, same-origin only, at most 4 KB). The page reads it back from /items.json, so done state survives outside the browser's localStorage. export-items merges the status into each row.

Fixed

  • An ended explainer always gets a final recap now (#227). Before, one was written only when the model or an agent sent it. When the call ends with no recap, whosaid writes recap.json and recap.html marked "derived": true from the call's own sections, with no model call: open questions, what the user owes, and agreed decisions. This runs after add/correct, in serve (including a watcher when no engine is configured), at the engine's final pass, and when capture ends. A recap from the engine or an agent is never replaced.