yellowjacket/.pi/journal.md

Work log

Temporal memory: what happened and what's next. Structure lives in CLAUDE.md, operational instructions in .pi/skills/yellowjacket-dev/, measured discoveries in .planning/NOTES.md. Don't duplicate those here.

Current state

Plan 005 (agent development harness) is complete — all seven phases. Everything from phase 1 onward is still uncommitted: one large but coherent working-tree diff, nothing pushed.

All four tiers verified green from a cold, cleaned state: make ui-test 313 passed, make lint 0 issues × 3 configurations, make test green × 3 passes, make e2e 19 passed. Both CI jobs verified green in a bare ubuntu:24.04 container, including 19/19 on WebKit.

Committed and pushed as 5ca6cad (the harness) + ccacd67 (a CI fix), and green on the real runner: job check ~4 min, job e2e ~3 min with 19/19 chromium and 19/19 webkit. One commit rather than seven because the working tree was the end state, not per-phase snapshots — Makefile, CLAUDE.md and lefthook.yml are touched by nearly every phase, so a split would have been fabricated history.

Still unverified, because no run has failed yet: the actions/upload-artifact step (continue-on-error, so it cannot mask a real failure) and whether pnpm honours npm_config_store_dir for store caching. Worth checking the next time a spec legitimately fails.

• gitea_ci's job_logs returns 404 on Gitea 1.27.1 — the endpoint is not exposed. Logs come from the VPS instead: zstdcat the file under gitea/actions_log/<owner>/<repo>/<xx>/<task_id>.log.zst, and note zstdcat is not in the gitea container, so docker cp it out first. Job status is action_run_job.status (1 success, 2 failure, 4 skipped, 5 waiting, 6 running). Probably belongs in the gitea skill, not here.

Open items deliberately not fixed: WAV tags are write-only (TestWAVTagsAreNotReadableYet), themeStore.loadFromBackend's failure handler cannot recover, backend/playlist has no CRUD suite.

Log

2026-08-11 — cold skill run, then phase 7 (CI)

• Followed the skill cold first, as the last session asked. It works: app up from a wiped .dev/, an undocumented flow driven (queue panel + shuffle, asserted on QueueModeChanged), stopped — ~1 minute, no dead ends. One real config bug: outputDir in .playwright/cli.config.json resolves against cwd, not the config file's directory (only initScript does that), so snapshots were landing above the repo and a stale one from the previous session answered ls -t instead. That cost a DOM walk to disprove a regression that did not exist. Four smaller doc gaps fixed (sandbox-seed already runs testdata; ui-setup/e2e-setup were undocumented prerequisites; snapshot prints a path; dev-stop leaves the browser open), plus dev-headless.sh's own banner, which was suggesting the bare window.go call its next paragraph warns against.
• Built both CI jobs as container scripts before writing any YAML, then transcribed the YAML back out and re-ran it to prove the transcription. Push-and-see is a bad loop on a self-hosted runner.
• It found a real bug immediately: make lint omitted webkit2_41 on all three passes, so it was linting configurations nothing builds. Invisible on Arch (which still ships webkit2gtk-4.0.pc), fatal on Ubuntu 24.04. Tag sets now match make test.
• Both open decisions settled by measurement: ALSA null PCM for audio (no daemon; the elapsed clock really advances), dead-address stub for the explore artifact (and setting it for the app run, not just seeding, is worth 8x on suite wall clock). WebKit is a required step — it had never been run anywhere, so one throwaway container run replaced a coin flip with 19/19 at +11 s.

2026-08-10 — phase 6, pi affordances

• Added .pi/skills/yellowjacket-dev/ as a directory rather than a flat file: only the description is always in context, so SKILL.md stays short enough that reading it whole is never a decision, and the deeper material sits in references/{harness,fixtures,ui-tier,schema-change}.md.
• Settled the CLAUDE.md-vs-skill split grammatically, not topically, because a topical split is what rots — every new fact gets two plausible homes. Three docs, three tenses: NOTES.md is past (measured, dated, append-only), CLAUDE.md is present (what the system is), the skill is imperative (what to run). A new paragraph's tense decides where it goes.
• The five gotchas (binding timeouts, first-run wizard, pkill -f, seeds-by-running, WebKit-is-CI-only) went inline in SKILL.md, not into a reference: you need them before the failure, not after.
• Trimmed CLAUDE.md's "Fixtures and the headless harness" section by about half — the command sequences and gotchas it was carrying are now the skill's, and leaving both would have created exactly the duplicate description this repo has a standing rule against.
• Added make skill-check / scripts/skill-check.sh + a pre-commit hook: every command in .pi/**/*.md must be a real make target, so the Makefile stays the source of truth for invocation and a renamed target fails a commit instead of misleading an agent later. Verified it fails (it caught its own not-yet-created target) and passes.
• Added the /e2e prompt template: promoting a hand-driven playwright-cli session into a spec is a transcription with four fixed substitutions (refs → testids, sleeps → waitForEvent, raw window.go → callBinding, short fixture → LONG_TRACK), plus three runs — pass, pass again, pass after a DB restore — because the usual failure is a spec depending on state the hand-driving left behind.
• One shell trap: under set -euo pipefail, x="$(make -pqRr | …)" fails the whole assignment, because make -q exits non-zero when a target is out of date and pipefail propagates it.

Earlier

Phases 1–5 of plan 005: fixture generator and manifest, headless launch and seeds, the event bridge + data-testid pass + backend/testctl + e2e/, the Vitest component tier + make bindings-check, and the events.Emit wrapper with its in-process service-event tests. Recaps and the five "verified end to end" blocks are in .planning/plans/active/005-agent-development-harness.md; the lessons are in .planning/NOTES.md.