synthesis-meeting-transcripts

Synthesis Meeting Transcripts

A protocol for fetching AI-generated meeting transcripts from the user's Gmail and Google Drive into a local markdown archive. Designed for teams where Google Meet + Gemini (or equivalents) produce meeting notes + word-for-word transcripts that live in Google Docs, and the user wants them mirrored locally alongside their project context.

The earlier release notes and rationale are retained in references/earlier-version-history.md. Unknown authentication, tab enumeration, or archive coverage is a visible gap, never a no-source attestation.

Binding rules

1. Declared means fetched. A ritual sync runs Step 0 below: enumerate the declared set, fetch every member, and account for each as fetched or as an unclosed gap. Relevance is judged after fetching, never before.
2. Unknown coverage is a visible gap. Unknown authentication, tab enumeration or archive coverage never becomes a no-source attestation, and the meetings watermark never moves past an unsaved doc or a meeting that may still be under way.
3. A Drive search listing cannot prove a window complete, because it omits the provider completeness flag. A full page, a next-page token, an incompleteSearch warning or a failed search leaves the window open; say so in the report.
4. The verbatim transcript is the only primary source. Fetch it in full whenever the source has one. The Gemini email is for discovery only, never for content.
5. Select the transcript by the document's stable tab ID, never by order or a regular expression; a title only when exactly one tab in a complete tab list carries it. Only a complete tab list establishes transcript-tab-absent.
6. Verify both halves landed with verify_transcripts.py --file <exact saved file> --json before calling a meeting fetched. A directory scan does not substitute.
7. Use the <!-- VERIFIER: no-source-transcript --> marker only when the source truly has no transcript, with a one-line reason beside it.
8. Never derive attribution from a tool summary. Who warned, decided, approved or owns comes from the verbatim transcript; if it does not say, keep the passive form.
9. A scanned commitment is a candidate until the principal confirms it in the same turn. Stamp each confirmed one with exactly one owner; nothing is created from scanner output alone.
10. Cross-check every date against at least two independent signals before writing a dated file.
11. Never leave transcripts in ~/Downloads/. Commit them to transcripts_repo, whose pre-commit gate refuses a summary saved as a transcript.

Contents

• references/protocol-steps.md: Steps 1 to 6 with every command (verify_transcripts.py, extract_commitments.py), the email warning, the no-source marker, the attribution rule, date verification and the daily-rituals integration. Read it before fetching a meeting, and for each member of a Step 0 sweep.
• references/transcript-primary-and-commit-gate.md: the v0.4.0 hierarchy of evidence and the v0.5.0 pre-commit gate. Read it before citing anything from a meeting record, and when setting up a transcripts repo.
• references/setup.md: the .agents/meeting-transcripts.yaml schema (required: workspace, google_account, transcripts_repo, transcripts_path), prerequisites and multi-account options. Read it when a project has no config, a config is refused, or an account is unreachable.
• references/release-notes.md and references/earlier-version-history.md: release notes from v0.2.0 to 1.0.1. Read them when a result differs from what an earlier version did.
• optional-workspace-mcp/README.md: the optional self-hosted multi-account server helper: fetch-meeting.py (one meeting, or --window FROM THROUGH for saved and unsaved docs), mcp_client.py, the service scripts and doctor.sh.
• references/coverage-map.md: where each part of the earlier 0.14.0 text now lives, and the M3 script changes.
• references/preserved.md: the retired acquisition and attribution-receipt text, verbatim. Read it only to review the cut.
• Protocol, Step 0: below.

Protocol

Step 0: Declared-window sweep (v0.9.0 — how "declared means fetched" executes)

The agent does not get a vote on which declared items are interesting. Relevance is judged after fetching, never before. Every declared item is fetched or retained as an unclosed gap.

Inventory the window first: with the optional workspace-mcp helper, python3 <synthesis-meeting-transcripts-root>/optional-workspace-mcp/fetch-meeting.py --window FROM THROUGH lists every doc the declared patterns find, each marked saved or unsaved, and the moment the watermark may advance to; its scope names the only patterns searched, and undeclared lists Gemini notes docs no declared pattern finds (declare a recurring one to cover it). With hosted connectors, run the same Drive search per declared pattern with the window's modifiedTime bounds, follow every page, and compare each doc ID with the IDs the saved files name: a **Source ID:** header, a Google Doc link, or the first backticked ID on a > Source: line (> Source: Gemini doc `<id>`), the three forms --window reads.

HTTP health alone is not sign-in: a read as the declared account shows which account answers. Never filter by meeting title or perceived relevance. Then advance with python3 <synthesis-daily-rituals-root>/scripts/sync_watermark.py advance --workspace <W> --surface meetings --through <T>, where T, written with its offset, is the earliest of the window's end (Drive reads its days as UTC), two hours before the check ran (a meeting under way may have no doc yet) and one second before each unsaved doc was created. A failed search, a full page, a next-page token, Drive's incompleteSearch warning or an unknown tab list keeps the watermark where it is.

The single-meeting path below serves an explicit user request. A ritual sync (Day-Start 3c, Day-End Step 1) runs this sweep instead, because the policy above has to have an execution path or it is prose:

1. Enumerate the declared set for the window from .agents/meeting-transcripts.yaml: every declared meeting pattern crossed with every occurrence that ended inside the window. The enumeration is the complete decision — no per-item judgment about which meetings look interesting stands between the list and the fetch.
2. Run Steps 2–5 for every member. Relevance is judged after fetching, from content.
3. Account for every member. The sync report carries one line per declared member: fetched (with path), or an unclosed gap naming the member, the window, and the failure ("no doc found", "fetch failed", "verbatim half missing"). A member missing from the report is the defect this step exists to prevent; a gap is closed this run or carried as a blocking item, never dropped.

Steps 1 to 6, the single-meeting path whose Steps 2 to 5 the sweep runs for every member, are in references/protocol-steps.md.