chihlasm/resolutionflow
1
0
Fork 0
Code Issues 23 Pull Requests 1 Actions Packages Projects Releases Wiki Activity

docs(ai): pause Escalation Mode build mid-SSE for Codex review

Browse Source 
Update HANDOFF to reflect:
- Build paused after the WIP SSE commit (87bd0b7)
- What Codex should look at on the SSE bus + endpoint + dispatch wiring
- Resume point post-review: re-run tests with -n auto, then frontend
  SSE subscription, then magic-moment screen
- Test-suite watch-out: per-test DROP SCHEMA fixture means concurrent
  pytest runs on the same DB collide; always one-suite-at-a-time or
  -n auto with conftest's per-worker DB isolation

No code change.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Michael Chihlas
2026-04-27 19:29:16 -04:00
parent 87bd0b7c56
commit ba46fc5644
1 changed files with 36 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
.ai/HANDOFF.md
View File

@@ -2,13 +2,34 @@
# HANDOFF.md
**Last updated:** 2026-04-27 EDT
**Last updated:** 2026-04-27 EDT (paused mid-build for Codex review)
**Active task:** **Escalation Mode** wedge build. See [`CURRENT_TASK.md`](CURRENT_TASK.md) for the full status; this file holds the resume point only.
**Branch:** `feat/escalation-metric-endpoint` — five commits stacked on top of `main` (`c0ed6d9`). Nothing pushed yet.
**Branch:** `feat/escalation-metric-endpoint` — six commits stacked on `main` (`c0ed6d9`). Working tree has UNCOMMITTED WIP for the SSE push.
## Status — paused for Codex review
Build is paused mid-flight on the SSE push. Hand the branch (and the WIP) to Codex for an outside-voice pass before stacking more commits, fixing tests, or pushing. Reasons: local backend test loop got tangled (multiple stale pytest processes contended on the same Postgres test schema; the suite design rebuilds the schema per test which doesn't tolerate concurrent runs well), and the SSE work is the kind of cross-layer surface a second pair of eyes is most valuable on.
What Codex should look at:
1. The new SSE endpoint at [`backend/app/api/endpoints/session_handoffs.py`](../backend/app/api/endpoints/session_handoffs.py) — `stream_escalations` — and the in-memory pub/sub bus at [`backend/app/core/escalation_bus.py`](../backend/app/core/escalation_bus.py).
2. Whether the bus's single-process / non-durable design is acceptable for the v1 pilot (Railway single-replica) and what the swap-to-Redis story should look like.
3. The dispatch wiring in [`backend/app/services/handoff_manager.py`](../backend/app/services/handoff_manager.py) — `dispatch_escalation_notifications` now publishes to the bus before the email fan-out. Race / ordering / failure-mode review.
4. Auth on the SSE stream — same `require_engineer_or_admin` dep as `/queue` and `/claim`. Browsers can't send custom headers via the native `EventSource` API; the planned frontend uses a fetch-based `ReadableStream` reader (matching the existing `streamDocumentation` pattern in [`frontend/src/api/aiSessions.ts`](../frontend/src/api/aiSessions.ts)). Verify that's the right call vs. a query-token scheme.
5. Whether the bus's "drop-on-full-queue" semantic is acceptable, given a stuck subscriber would silently miss live-arrival cards (they'd still see them on next page load via REST `/queue`).
## Resume point (after Codex review)
1. **Get the test suite back to green.** Stale pytest zombies in the container were cleared (PIDs 1790034, 1844996, 1883167, 1916565, 1935830, 2009437, 2009449 — all dead, parent uvicorn-reload didn't reap them; PID slots remain but no live processes). Re-run with `pytest -n auto` to keep wall-clock manageable. Files: `tests/test_escalation_bus.py` (7 tests), the 4 new dispatch + SSE tests in `tests/test_handoff_manager.py` and `tests/test_session_handoffs_api.py`.
2. **Frontend SSE subscription** in `EscalationQueue.tsx` — fetch-based reader, prepend new cards with the locked 200ms slide-in, reconnect with backoff, tab-title flash when backgrounded, respect `prefers-reduced-motion`. Then ship the magic-moment handoff-context screen (4 sections, dissolves into FlowPilot session view).
3. Push the branch + open a draft PR.
## Stack
```
WIP (uncommitted): SSE bus + endpoint + dispatcher publish + 7 bus tests + 1 dispatcher test + 2 SSE endpoint tests
a283d0d docs(ai): refresh handoff state mid-flight on Escalation Mode build
9f0bfd4 feat(escalations): mount time-to-first-action stat-card on /escalations
07d0db9 feat(handoff): email engineer-or-admin teammates on escalation
7a5b853 feat(api): role-gate handoff claim to engineer-or-admin
@@ -16,34 +37,28 @@
d51e95c docs(plans): add escalation-mode wedge design + test plan
```
## Resume point
Pick up the **WebSocket/SSE push** — the live-arrival half of the notification dual-path. Email is already wired (commit `07d0db9`); push is the second channel that makes the demo's "30-second magic moment" undeniable when the receiving senior is online and on the queue page.
Suggested first slice: a thin server-side SSE endpoint scoped to `current_user.account_id`, fan out from `HandoffManager.dispatch_escalation_notifications` (alongside email), and hook the frontend `EscalationQueue` to subscribe and prepend new cards with the locked 200ms slide-in. Reconnect logic, tab-title flash, and `prefers-reduced-motion` respect are part of this slice per the locked UI spec in the design doc.
After the dual-path is feature-complete, the **magic-moment handoff-context screen** is next (4 sections, dissolves into the FlowPilot session view on first action).
## Where things stand
- CI on `main` still healthy. Branch protection: `CI / frontend (pull_request)` required, `CI / backend (pull_request)` required, `CI / e2e (pull_request)` not yet required (ops-only follow-up — two consecutive green runs cleared the threshold).
- 20 backend tests green on this branch (handoff_manager, session_handoffs_api, flowpilot_analytics_escalations). Frontend `tsc -b` clean. Branch has not been pushed; no CI runs yet.
- The plan doc at [`docs/plans/2026-04-27-escalation-mode-wedge-design.md`](../docs/plans/2026-04-27-escalation-mode-wedge-design.md) is the source of truth for every UI / metric / scope decision. The embedded **GSTACK REVIEW REPORT** at the bottom shows Eng + Design CLEARED and Codex INFO with the disposition of all 12 of its findings.
- CI on `main` still healthy. Branch protection: `CI / frontend (pull_request)` required, `CI / backend (pull_request)` required, `CI / e2e (pull_request)` not yet required.
- The 20 tests passing as of `9f0bfd4` are still passing (last green run logged before the SSE work). The newly added SSE tests (7 bus + 1 dispatcher integration + 2 endpoint) HAVE NOT been verified end-to-end this session — they ran clean on the bus suite alone (7/7 in 0.14s) but the DB-backed integration tests were aborted before completing.
- The plan doc at [`docs/plans/2026-04-27-escalation-mode-wedge-design.md`](../docs/plans/2026-04-27-escalation-mode-wedge-design.md) is the source of truth for every UI / metric / scope decision. The embedded **GSTACK REVIEW REPORT** at the bottom shows Eng + Design CLEARED and Codex INFO from the design-stage pass.
## Useful breadcrumbs
- New endpoint: [`backend/app/api/endpoints/flowpilot_analytics.py`](../backend/app/api/endpoints/flowpilot_analytics.py) — `get_escalation_metrics` at the bottom of the file.
- Notification dispatch: [`backend/app/services/handoff_manager.py`](../backend/app/services/handoff_manager.py) — `dispatch_escalation_notifications`. Wired in [`backend/app/api/endpoints/session_handoffs.py`](../backend/app/api/endpoints/session_handoffs.py) **after** `db.commit()` so a rolled-back handoff never emails.
- Metric endpoint: [`backend/app/api/endpoints/flowpilot_analytics.py`](../backend/app/api/endpoints/flowpilot_analytics.py) — `get_escalation_metrics` at the bottom.
- Notification dispatch (email + bus publish): [`backend/app/services/handoff_manager.py`](../backend/app/services/handoff_manager.py) — `dispatch_escalation_notifications`. Wired in [`backend/app/api/endpoints/session_handoffs.py`](../backend/app/api/endpoints/session_handoffs.py) **after** `db.commit()` so a rolled-back handoff never emails or fans out.
- SSE endpoint (WIP): [`backend/app/api/endpoints/session_handoffs.py`](../backend/app/api/endpoints/session_handoffs.py) — `stream_escalations`. Heartbeat every 25s, account-scoped subscribe, role-gated to engineer-or-admin.
- Pub/sub bus (WIP): [`backend/app/core/escalation_bus.py`](../backend/app/core/escalation_bus.py). Module-level singleton, in-memory, `asyncio.Queue` per subscriber with 64-event maxsize and drop-on-full semantics.
- Frontend stat-card: [`frontend/src/components/flowpilot/EscalationMetricCard.tsx`](../frontend/src/components/flowpilot/EscalationMetricCard.tsx). Renders `n_with_action / n_claimed`, avg + median, and the metric_definition disclaimer.
- Two-metric framing — required reading before quoting any number to a pilot. The in-product endpoint measures *post-claim time-to-first-action*; the savings claim is `manual_baseline − in_product`. Manual baseline comes from the founder's stopwatch on the next 5 escalations (The Assignment in the design doc).
- The `notification_sent` boolean is intentionally NOT being written. Per Codex's correction it should be replaced by per-channel delivery records; v1.x story. For now, application logs are the audit trail.
- Two TODOs added during this session: peer-tech escalation (deferred to v2) and the (already moved-in-scope) claim role gate. See [`TODO.md`](TODO.md).
- Two-metric framing — required reading before quoting any number to a pilot. In-product endpoint measures *post-claim time-to-first-action*; the savings claim is `manual_baseline − in_product`. Manual baseline comes from the founder's stopwatch on the next 5 escalations (The Assignment in the design doc).
- The `notification_sent` boolean is intentionally NOT being written. Per Codex's design-stage correction it should be replaced by per-channel delivery records; v1.x story. For now application logs are the audit trail.
## Watch-outs
- `ai_session_step` has NO `user_id` column — the metric query keys "first action by senior" off `session_id + created_at > claimed_at`, which is fine because session activity post-claim IS the senior's activity (the session is reactivated under `escalated_to_id`). If a future change adds `user_id` to `ai_session_step`, the metric query can become more precise.
- `account_id` is denormalized on `ai_session_step` (Phase 4 RLS pattern). The metric query and any new SSE subscription scoping must use it directly, not join through `ai_sessions`.
- POST `/handoff` still requires the session owner to be the escalator (`AISession.user_id == current_user.id`). Peer-tech escalation is captured as a v2 TODO. Don't widen this without a UX decision.
- `ai_session_step` has NO `user_id` column — the metric query keys "first action by senior" off `session_id + created_at > claimed_at`. Fine for v1 because session activity post-claim IS the senior's activity (session reactivates under `escalated_to_id`).
- `account_id` is denormalized on `ai_session_step` (Phase 4 RLS pattern). Use it directly; don't join through `ai_sessions`.
- POST `/handoff` still requires the session owner to be the escalator (`AISession.user_id == current_user.id`). Peer-tech escalation is a v2 TODO.
- The test suite uses `DROP SCHEMA public CASCADE` + `CREATE SCHEMA public` per test (see [`backend/tests/conftest.py:144`](../backend/tests/conftest.py#L144)). Concurrent pytest runs against the same test DB collide. Always run one suite at a time, or via `-n auto` xdist with the per-worker-DB isolation already in conftest.
## Kill-switch (week 8)