Quill meetings integration + after-hours digest processing#27
Conversation
Connect the local Quill note app over MCP to search meetings and ask about them on the local model. Asking about a full transcript costs ~15-17s of prefill on the NPU, so a daemon scheduler pre-computes a digest (summary / goals / action items) per meeting during a configurable idle window (default 17:00-21:00, idle-gated) and caches it in data/meeting_digests.jsonl — daytime reads are then instant. New modules: ffp_quill (stdlib MCP-over-HTTP client + parsing) and ffp_meetings (digest store, batch worker, pure scheduler gate, GetLastInputInfo idle detection). New config block `meetings` (off by default, opt-in; mcp_url validated loopback-only). New dashboard Meetings tab (search / cached digest / Process now / Ask) + a Config card with a "Run batch now" button. Daemon gains a scheduler thread + 8 actions (quill_status, quill_search_meetings, meeting_digest_get, meeting_digests_list, meeting_process, meeting_batch_run, meeting_batch_status, meeting_ask); 59 -> 69 actions. The meeting actions self-lock and write a separate cache file, so a long batch never blocks the global config write-lock. Tests: test_ffp_meetings (scheduler window/idle, digest store upsert, batch skips-cached + max-per-run, ask prefers cached digest), test_ffp_quill (parsing), meetings config-patch whitelist incl. loopback rejection; daemon action-count 69. Validated live: search + process_meeting against real Quill + NPU (33s full-transcript digest, accurate). Gates: ruff clean, 305 pytest passed, node --check app.js. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 28273ec040
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| def _act_quill_search_meetings(args: dict) -> dict: | ||
| import ffp_quill | ||
| url = str(_meetings_cfg().get("mcp_url") or ffp_quill.DEFAULT_MCP_URL) | ||
| return ffp_quill.search_meetings(str(args.get("query") or ""), int(args.get("limit") or 12), url=url) |
There was a problem hiding this comment.
When meetings.enabled is false, opening the Meetings tab still calls searchMeetings(), which reaches this action and queries Quill with the configured/default URL. That violates the off-by-default opt-in behavior and can expose meeting titles/participants even though the integration is disabled; the Quill data actions should refuse unless meetings.enabled is true.
Useful? React with 👍 / 👎.
| if not content: | ||
| content = ffp_quill.get_transcript(meeting_id, client=client) | ||
| used = "transcript" if content else used |
There was a problem hiding this comment.
If the user selects source: "minutes" from the Config UI (Minutes only), meetings without minutes still fall through here and fetch the full transcript. That silently defeats the privacy/performance reason for the setting and can feed a full transcript to the model during batch processing; only auto should fall back to transcript.
Useful? React with 👍 / 👎.
| if not c.connect(): | ||
| return {"ok": False, "error": "Quill MCP server is not reachable"} |
There was a problem hiding this comment.
For an already-processed meeting, meeting_ask should be able to answer from meeting_digests.jsonl, but this early connect() makes it fail whenever Quill is closed or temporarily unavailable. Since the cache is the advertised fast path, check get_digest() first and only connect to Quill when falling back to live minutes/transcript content.
Useful? React with 👍 / 👎.
2 participants
Connect the local Quill note app over MCP to search meetings and ask about them on the local model — with an after-hours scheduler that pre-computes digests so daytime reads are instant.
Why
Asking the NPU about a full meeting transcript costs real prefill: a POC measured ~17 s to first token for a ~7k-token transcript (~25–35 s total). So instead of paying that at read time, a daemon scheduler runs during a configurable idle window (default 17:00–21:00, only when idle) and caches a digest (summary / goals / action items) per meeting in
data/meeting_digests.jsonl. Reads are then instant; ad-hoc "Ask about this meeting" grounds on the cached digest.What's new
ffp_quill— minimal stdlib MCP-over-HTTP client (initialize → tools/call) + parsing of Quill's meeting/transcript/minutes output. Loopback-only.ffp_meetings— digest cache (upsert/list/get), batch worker (pages undigested meetings, idempotent, max-per-run), pure scheduler gate (should_run_batch: enabled + window + idle), andGetLastInputInfoidle detection.meetings.enabled+ in window) + 8 actions (quill_status,quill_search_meetings,meeting_digest_get,meeting_digests_list,meeting_process,meeting_batch_run,meeting_batch_status,meeting_ask). 59 → 69 actions. The meeting actions self-lock + write a separate file, so a long batch never holds the global config write-lock.meetingsblock (off by default, opt-in;mcp_urlvalidated loopback).Validation
ffp_quill.status+search_meetingsagainst real Quill;process_meetingagainst real Quill + NPU produced an accurate digest in ~33 s on a full transcript.node --check app.js.Off by default — enable it in Config → Meetings & after-hours processing.
🤖 Generated with Claude Code