v0.3.0

**v0.3.0 — Conversation replay and debugging**

The universal "this saved me hours" feature for voice-agent developers lands here. Open any past conversation in the dashboard, scrub through it like a video timeline, see every STT chunk, every LLM token, every TTS frame, plus the agent's conversation state at every moment — with cost accruing live as you scrub. Replay captures by default for every session at full fidelity, with per-project retention (default 90 days) ageing rows out automatically.

**Highlights**

- **REQ-VG-REPLAY-001 (Open any past conversation)** — Dashboard `Replay` page reachable from every Sessions-row "Open replay" button. Pre-v0.3.0 sessions render a clear "recorded before replay capture existed" banner with a link to the session detail page.
- **REQ-VG-REPLAY-002 (Scrub the timeline)** — Range scrubber + arrow keys step to prev/next event, Shift+Arrow jumps 1s, Home/End jump to call boundaries.
- **REQ-VG-REPLAY-003 (Every captured event)** — Per-session asyncio buffer in `ReplayCapture` records STT chunks, LLM tokens, TTS frames; bounded backpressure drops oldest with a counter that surfaces as "events dropped here" in the dashboard.
- **REQ-VG-REPLAY-004 (Cost accruing live)** — `RunningCostCounter` shows per-modality breakdown + top-3 costliest events tooltip; total updates per scrubber position.
- **REQ-VG-REPLAY-005 (Conversation state at every moment)** — `StateSnapshotter` captures the system prompt + message history + in-flight tool call + structured outputs at LLM message-add, tool-invoke, and tool-resolve boundaries. Rate-capped to one per second; tool-resolve bypasses the cap.
- **REQ-VG-REPLAY-006 (Privacy + retention)** — Per-project `replay.retention_days` (default 90); hourly `RetentionWorker` ages rows out, `replay.enabled: false` opts whole projects out. `DELETE /api/sessions/{id}/replay` cascades across all four `replay_*` tables.

**Storage**

A 60-second conversation captures roughly 100-300 KB on disk under default settings. Per-minute estimates, worked solo-dev (~$0.21/month S3 at 90-day retention) and agency (~$70/month) examples, and the three per-project tuning knobs are documented in [docs/storage/replay-storage-costs.md](https://github.com/mahimailabs/voicegateway/blob/main/docs/storage/replay-storage-costs.md). The OQ1 storage-cost smoke test (`tests/storage/test_replay_storage_smoke.py`) asserts the upper bound at 600 KB/min on a synthetic conversation; well-below at the time of release.

**Foundry Open Questions resolved**

- **OQ1 (storage-cost target)** — RESOLVED AFFIRMATIVE per the smoke. Fallback (`replay.enabled: false`) documented and in place.
- **OQ2-5** — Tuning calls locked: message-add snapshot granularity, pre-fetch over streaming, capture-by-default for short calls, per-character TTS cost distributed in time-weighted slices.

**Migration notes**

No code changes required for v0.2.0 callers. Migration 0004 runs automatically; v0.2.0 sessions are preserved with NULL on `replay_size_bytes` and surface as "recorded before replay capture existed" in the Replay page (linked back to session detail).

The `voicegateway/combined_server.py` re-export shim previously flagged for v0.2.0 removal stayed in place again; bundling a back-compat break with the replay-feature release would be a bad signal. Schedule remains tentatively v0.4.0.

**Refinery and Foundry**

- Refinery: https://linear.app/mahimairaja/document/v030-conversation-replay-and-debugging-refinery-02fdfbb5a241
- Foundry: https://linear.app/mahimairaja/document/v030-conversation-replay-and-debugging-foundry-blueprint-f6e62120f85a