Send meeting processing time and stage timings to PostHog

[r3dbars]

Requested by Justin · project thread

Why

Before: PostHog can't say how long a meeting takes to turn into a transcript. The only number is a guess from pairing the stop and save events, and it can't tell slow speech-to-text from a Mac that slept, a cold model, or a slow Mac.

After: every meeting_transcript_saved says how long the job took, how much of that the Mac slept, and how long each stage took, plus the speech model and Mac class. That's what we need to rank the "faster meeting transcripts" fixes on real users.

Product Impact

• Affects: meetings
• Lane: meeting reliability
• Why this matters: meeting speed work needs real P50/P95/P99 per stage, model and Mac, not a proxy.

What changed

• New MeetingPipelineTimings (Core): a task-local, lock-guarded recorder bound once per job in TranscriptionTaskManager (live, import, and failed-meeting retry paths). Stages add their own time: ensureModelsReadyForPipeline, AudioResampler.loadAndResample, DiarizationService.diarizeOffline (incl. ERes2Net re-embed), and each speech-to-text call via MeetingSTTAdapter. Sleep = wall clock minus systemUptime.
• publishTranscriptSaved takes an optional snapshot and sets lastPipelineTimings before publishing, so the host never reads a previous save's numbers.
• New MeetingProcessingTelemetry (app) rounds them: processing_ms, sleep_ms, models_ready_ms, resample_ms, diarize_ms, stt_ms (10 ms), stt_calls, stt_input_seconds, recording_minutes, plus stt_model, mac_chip, memory_gb_bucket.
• Allowlist + reviewed-property PSVs, privacy doc and PostHog plan updated; tests: fast testMeetingProcessingTelemetry, SPM MeetingPipelineTimingsTests.

How I checked it

• scripts/dev/agent-preflight.sh
• bash scripts/dev/linux-checks.sh (48 passed)
• Selected checks from .agents/test-matrix.yml for the files changed (CI)
• bash build.sh --no-open (CI)
• bash run-tests.sh (CI)
• Performance budget passed
• bash run-integration-smoke.sh (CI)
• swift test (CI)
• bash run-e2e-smoke.sh / swift test --package-path Tools/<Package> if the matrix maps them
• Manual check:

Checks I could not run, and why:

• All Swift builds and tests: cloud session on Linux, no Swift toolchain. Relying on CI for this head.

Mac or hardware test still needed? If yes:

• Optional: record one meeting, then check local events.jsonl/PostHog debug for meeting_transcript_saved carrying processing_ms, stt_ms, stt_calls. On main those keys are absent.

Risk Review

• Privacy / local-first behavior reviewed (durations, counts, model id, coarse Mac class only)
• New analytics properties avoid the sanitizer's drop fragments (check-telemetry-keys.py PASS)
• Checked the text-pin tests for every file I edited (check-source-pins.py --changed-only PASS)
• Storage path or migration impact reviewed (none)
• Public-facing copy stays concrete (no UI change)
• Release/update impact reviewed (none)
• Agent PRs got an independent deep review of the full diff
• UI changes include sanitized visuals (no UI change)
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

The retranscribe-saved-meeting path stays untimed on purpose (sends no timings). Timing overhead is one uptime read per stage and per speech-to-text call.

Agent handoff

COORD_DONE: BRIEF | this PR | meeting speed telemetry | none | none | linux-checks, source pins, telemetry keys | CI + deep review

🤖 Generated with Claude Code

https://claude.ai/code/session_01LcH2PBaqN6yYkxeoaYrj45

---

Generated by Claude Code