App-improvement follow-through: action-item done/due metadata, semantic search, silent-failure instrumentation, disk guards

[r3dbars]

Why

Follow-through on the "what else would you improve" review: the highest-leverage gaps that were still real after re-verifying the tree (most of FIX_ROADMAP already landed). The biggest two: list_action_items has advertised a status filter since the cross-meeting tools shipped but nothing ever wrote a status (the filter could never match), and search was lexical-only so paraphrase queries against the meeting library came up empty.

Product Impact

• Affects: meetings / agent artifacts / dictation
• Lane: agent workflow / meeting reliability
• Why this matters: closes the remaining loops on the "ask my history" moat — open/done finally means something on action items, semantic_search finds paraphrases FTS5 can't ("pricing pushback" finds "they balked at the cost") with a fully on-device embedding, and the agent orientation tools (recap/recent_context) now answer "what was this meeting" with the summary instead of greetings. Also turns three classes of silent failure (skipped audio maintenance, dictation buffer truncation, doomed-from-the-start low-disk recordings) into visible, upfront signals.

What changed

• Action items carry real done/due metadata via trailing bullet markers — (due: Friday), (status: done), bare (done):
  • MeetingQuickSummaryExtractor appends (due: ...) when a commitment sentence has a recognizable deadline cue. Extraction never marks anything done — closing an item is a write-back: edit the marker in the saved Markdown, the MCP file watcher reindexes.
  • CaptureSummaryParser parses markers into ActionItem.status/.due (bare-token set is curated so real parentheticals like "call Bob (the vendor)" are never shredded).
  • MCP TranscriptIndex adds status/due columns, list_action_items filters on real values instead of AND 1=0, and digest open-counts become accurate. Unmarked items — including all pre-existing summaries — stay open.
• Local semantic search (semantic_search MCP tool): Apple's on-device NLEmbedding sentence model (ships with macOS — no bundled weights, nothing leaves the machine) embeds ~400-char chunks of consecutive meeting utterances plus dictation entries at index time into a new semantic_chunks table; queries brute-force cosine over normalized Float32 blobs. Model-unavailable systems get an explicit message and lexical search is unaffected. Tests use a deterministic bag-of-words embedding so CI never depends on the model asset. Index schema bumped to v4 (existing indexes rebuild from disk via the existing version gate).
• recap / recent_context previews use the real summary (NEXT_WORK refactor(ui): Split FloatingPanel.swift into 10 organized files #2): recap prefers summary prose + compact Decisions/Action-items/Open-questions lines parsed from the Markdown; recent_context builds a one-line preview from the meeting's indexed summary facts. Both fall back to the old dialogue/first-utterance behavior only when a meeting has no summary.
• Silent failures now report: retained-audio maintenance (catch { continue } in prune / WAV→M4A / playback-mix / stale-temp / failed-audio paths) forwards each skip through a hook to EventReporter as audio_maintenance_failure (operation + NSError domain/code only, no paths). ParakeetEngine's 30-min buffer cap emits a once-per-recording audio_buffer_truncated warning with the dropped duration.
• Disk safety: RecordingValidator.minimumDiskSpace raised 100MB → 1GB (dual-WAV capture burns ~1.4GB/hour and the runtime watchdog only stops at 50MB, so 100MB start floors let doomed recordings begin). New CaptureLibrarySize measures the capture library after launch maintenance and emits a bucketed capture_library_size event with the retention setting — first visibility into unbounded growth under the default never retention.
• docs/NEXT_WORK.md refreshed against the tree — most of the original list had shipped; the doc now says what's actually left (speaker-identity trust work, plus the items blocked on a real Mac).

How I checked it

• scripts/dev/check-build-source-lists.py (source-list contract passes with the new files)
• CI green on this branch as of the semantic-search commit (build + fast tests + swift test + integration/e2e smokes + all four Tools packages); latest commit re-running
• bash build.sh --no-open locally — not run: authored in a Linux container, no macOS toolchain; relying on CI
• New coverage: CaptureLibrarySizeTests, extractor due-marker suite, maintenance-failure reporting suite (fast tests); CaptureSummaryParserTests marker grammar; SummaryRollupTests status filter + digest counts + recap/recent_context preview behavior; SemanticSearchTests chunking/ranking/filters/reindex/fallback
• Manual check: mark a bullet (done) in a saved meeting, confirm list_action_items status:"open" stops returning it after reindex; run semantic_search with a paraphrase query against a real library

Risk Review

• Privacy / local-first behavior reviewed — new events carry operation names, error domain/code, and coarse size buckets only; embeddings are computed and stored locally, never sent anywhere
• Storage path or migration impact reviewed — MCP index schema bump triggers the existing rebuild-from-disk gate (first rebuild pays a one-time embedding cost proportional to library size); no on-disk transcript format change (markers are plain bullet text)
• Public-facing copy stays concrete (disk-space error message now states the real requirement)
• Release/update impact reviewed — none
• Agent PR stays draft until human review
• UI changes — none
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

Still deferred, with reasons (also captured in the refreshed docs/NEXT_WORK.md):

• Onboarding meeting dry-run — the auto-detect copy fix already landed in the calendar step; what remains is a real capture flow inside first-run onboarding, which shouldn't be built blind.
• Hardware-smoke CI lane — needs a self-hosted macOS runner (infrastructure, not code).
• Settings/Home god-object split — large refactor, wrong to attempt without a compiler.
• Transcribe-instead-of-discard pre-sleep dictation audio — touches real-time recovery semantics; needs hardware testing.
• Speaker-identity work (negative exemplars, multi-exemplar voiceprints) — rides the eval harness; should not be changed blind.

Agent handoff

COORD_DONE: BRIEF | this PR | action-item done/due metadata + semantic_search + summary previews + silent-failure instrumentation + disk guards | none | verify on macOS CI + manual semantic_search sanity run | check-build-source-lists.py only (no local toolchain) | run full test matrix on macOS

🤖 Generated with Claude Code

https://claude.ai/code/session_012muZDEbHFBfzsuhR2XHmJC

[r3dbars]

Maestro hold review: HOLD at head f9fc688. Still draft and broad: action-item writeback semantics, MCP schema rebuild to v4, local semantic embeddings/chunks, summary previews, audio-maintenance failure events, buffer truncation warnings, disk-space start gate, and capture-library telemetry. This needs Mac/tool proof and real-library behavior proof, not just green CI. Smallest clear path: run MCP/Tools tests, semantic_search on a real local library, action-item (done)/(due) reindex check, index rebuild/perf sanity on a nontrivial library, and confirm telemetry carries only operation/error buckets.

[r3dbars]

Fair hold — CI green isn't real-library proof, and this branch was authored from a Linux box, so the Mac-side proofs need someone at a Mac. Making that as cheap as possible:

1. MCP/Tools tests (same suites CI ran, locally):

swift test --package-path Tools/TranscriptedCaptureKit
swift test --package-path Tools/TranscriptedMCP

2 + 4. Real-library semantic_search + rebuild/perf sanity, without touching the production index — point only the index at scratch, keep the real capture library as data:

cd Tools/TranscriptedMCP && swift build -c release
TRANSCRIPTED_INDEX_DIR=$(mktemp -d) time ./.build/release/transcripted-mcp --self-test

--self-test runs the full reconcile, so time here IS the v4 rebuild+embed cost on your real library (expect it to be the slowest first run; incremental after that). Then wire the same binary + env into Claude Desktop/Code as a scratch server and run semantic_search with a paraphrase query you know only appears reworded ("pricing pushback" style), plus recap to see summary previews instead of greetings.

3. (done)/(due) reindex check: edit any saved meeting's Action Items bullet to end with (done), bump the file (save), rerun the self-test (or let the file watcher catch it in a live server), then list_action_items with status:"open" → the item should be gone; status:"done" → present with status populated.

5. Telemetry payloads — provable from source without a Mac: both new events are fixed-key literal dictionaries, no interpolated paths/titles:

• audio_maintenance_failure: operation / error_domain / error_code (TranscriptedAppState.swift, handler wiring)
• capture_library_size: size_bucket / audio_retention (bucketed in CaptureLibrarySize.bucketLabel)
• audio_buffer_truncated: dropped_seconds / capacity_seconds (ParakeetEngine.swift)
  LocalObservabilityPayloadSanitizer still applies on top before disk writes.

If the breadth is the concern, happy to split this into 3 PRs (metadata+previews / semantic search / instrumentation+disk) — say the word and I'll cut them.

---

Generated by Claude Code