Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

docs: record execution campaign, production fixes, and shutdown incident in progress log

Miguelon Peluchonba337f460930
diff --git a/docs/features/control-ui-context-removal.md b/docs/features/control-ui-context-removal.md
index fed55a596..78b7a0ac2 100644
--- a/docs/features/control-ui-context-removal.md
+++ b/docs/features/control-ui-context-removal.md
@@ -279,6 +279,9 @@ Tests updated in the same commit: `HostTable.test.tsx` (hover card now asserts a
| 2026-08-27 | e2e | Built the QA-recorder video-evidence journeys: 11 new `qa-recorder-*` specs + `qa-recorder-connector-edit` extended with the legacy-tab redirect, each with a dedicated `qa:recorder:<name>` npm script (62 tests / 45 files listed under the recorder config). Journeys follow the recorder contract (loopback guard, `QA_RECORDER_CONFIRM_MUTATIONS` opt-in for mutating journeys, `screenshotAndLog` at every step, full API cleanup in `finally`). Committed the previously-missing `.env.qa-recorder.example` (referenced by the recorder doc but never tracked) with the exact keys the journeys need. Verification: control-ui `tsc --noEmit` clean, recorder `--list` clean, unit suite 1700/1700. No backend changes. |
| 2026-08-27 | e2e | Added 3 gap journeys for error/degraded paths: `agent-connectors-conflict` (out-of-band context PUT stales the page → 409 banner "This agent's connectors changed since they were loaded." → reload + retry succeeds), `member-removed-access` (mapping whose scope was deleted renders under "Removed access" with the agent label + Deleted marker), `connectors-search-agent` (search-by-agent-name filtering + empty state). Recorder catalog now 65 tests / 48 files listed; `tsc` clean. **Execution status: no Playwright spec (main or recorder) has been executed in this lane — no runnable cluster exists in the dev session** (both minikube profiles Stopped, no branch.mk ownership record → fail-closed per repo rules; the live `:3000` server belongs to the unrelated `feat/desktop-home` worktree and `:8090` is a dead forward). Verified-only: unit suite 1700/1700 (three clean full runs; one intermediate run under heavy parallel load showed 4 flakes that did not reproduce), `tsc --noEmit` everywhere, both Playwright configs `--list` clean, `next build` green. First cluster run should budget for small selector polish. |
| 2026-08-27 | e2e | Video-format answer + MP4 pipeline: Playwright/Chromium records **WebM** natively (not MP4) — `qa:recorder:convert-mp4` added (`scripts/qa-recorder/convert-videos-mp4.sh`, ffmpeg H.264 + faststart, `--replace` to drop originals; round-trip verified against a synthetic clip). Recorder doc + §9 updated. **No video evidence exists yet anywhere in `.local-notes/` — zero journeys have been executed in this lane; videos are produced only when a journey runs against a live cluster.** |
+| 2026-08-28 | e2e | **Execution campaign (mock-backed).** Full T2 bootstrap was ruled out on this host (7 GB RAM, ~3 available → OOM risk; both minikube profiles Stopped, stale branch.mk ownership → fail-closed per repo rules). Instead: built `scripts/qa-recorder/mock-control-api.mjs` — an in-memory control-api implementing the audited wire contract (auth cookie, resourceVersion/409 semantics, users/teams contexts mappings with deleted-history, hosts/contexts/mcp-servers CRUD, registry entries + install, budgets, recipes, rotation `DeploymentReady` simulation) — and ran the REAL control-ui dev server against it. Environment fights solved along the way: mock must bind dual-stack (`localhost` → `::1` first for Playwright's request context); Next dev daemon ignores env/port and grabs its own port (verify + set `CONTROL_UI_URL` per run); each playwright invocation wipes the outputDir (run batches in ONE invocation, copy evidence out after). Spec fixes found by real rendering: connectors rows use `role="button"` (kills `row` role; target `Expand/Collapse connector <name>`), team Access rows are `listitem`s, sort headers expose `Sort by access …` aria-names, tombstone rows persist after removal (assert actionability, not absence), toasts stack (`.last()`), 10s auto-refresh detaches click targets (retry), journeys must re-navigate after seeding (login pre-seed landing makes sidebar clicks no-ops). |
+| 2026-08-28 | e2e | **Two REAL production bugs found by the journeys, fixed in `f2eebbc8a`:** (1) member detail page crashed wholesale — `accessLabeler` useMemo was declared after its consumer in `availableContextOptions` (TDZ ReferenceError; no unit tests covered those pages); (2) connectors-page search-by-agent silently broke in the Phase-1 rework (`summary.agents` emptied while the search haystack still reads it). Both fixed; search-by-agent locked with a new unit regression test (`McpServerTable.test.tsx`); unit suite now **1701/1701**, `tsc` clean. |
+| 2026-08-28 | ops | Operator stopped the machine mid-run (commit `3daf915d8` preserved the in-flight spec fixes — thank you). The interrupted final invocation still produced **37 passing journeys / 51 WebM videos + 132 PNG proofs**, preserved in `.local-notes/qa-recorder/evidence-final/`. Two core journeys (marketplace-install, recipe-scope-copy) were environment-killed mid-flight and re-queued; their root causes were fixed (recipe fixture needed `spec.triggers`; the L1 banner fires at the deploy step, not review — verified via validator probe: `validateRecipe` excludes the L1 check by design). Pre-existing journeys needing real cluster depth (rotation-demo, cost-usage/governance, llm-lifecycle/price, invitations, gfs) fail against the mock by design — they belong to the cluster/T2 lane. |
## 11. Post-implementation audit (2026-08-27)