Skip to content

fix(skills): settle the ci-speed review follow-ups - #328

Merged
YuukiFST merged 2 commits into
mainfrom
fix/ci-speed-followups
Nov 16, 2026
Merged

YuukiFST merged 2 commits into
mainfrom
fix/ci-speed-followups

Conversation

@YuukiFST

Copy link
Copy Markdown
Owner

Follow-up to #327. Its review left three findings open, and this PR settles them.

Risk: medium. GLOBAL.md is the payload every session on every harness loads, and this PR adds one routing line to it. The other two changes are one line each in a skill and in a playbook.

What changed

  • GLOBAL.md: slow CI goes to yuu-mode/playbooks/ci-speed.md, with its path on each harness. Before, the performance route caught every "make it faster", so the playbook was reachable only from inside /yuu-mode. yuu-mode is user-invoked, so the line says the playbook may be read by path.
  • ui-polish/SKILL.md: the minimum browser pass asked for touch emulation to show sticky hover, tap highlight and input zoom. mobile-native/SKILL.md 11. Right in Chrome, wrong on phone says emulation reproduces none of them. The pass now reads the fixes in mobile-native's The Symptom Table in code, and uses a real device when one is connected.
  • yuu-mode/playbooks/ci-speed.md: the tests bullet used to put cleanup rules in AGENTS.md prose. It now relies on Vitest config:
    • restoreMocks, unstubGlobals and unstubEnvs on;
    • app singletons reset through their own reset exports in a setupFiles entry;
    • isolate: false only once a shuffled run is green;
    • a pinned --sequence.seed, so an order failure replays.

Evidence

Sources read this session:

  • Vitest docs/config/setupfiles.md: "If isolation is disabled, imported modules are cached, but the setup file itself is executed again before each test file".
  • Vitest restoremocks.md, unstubglobals.md and unstubenvs.md: each option calls its vi.* reset "before each test".
  • Vitest sequence.md: sequence.seed defaults to Date.now(), and shuffling loses the cached longest-first order.
  • mobile-native/SKILL.md has both cited headings verbatim: ## The Symptom Table and ### 11. Right in Chrome, wrong on phone.

Which code reads GLOBAL.md: the sync scripts copy it whole, or compare it whole. The one branch that depends on its content, the OpenCode backup at sync-config.ps1:321, runs only when AGENTS.local.md.bak is missing. That file exists on this machine.

Local CI (bin/agent-cli check) on 3686490:

PASS coding-standards-gate 11.4s
PASS sync-config-windows 73.6s
PASS file-length 11.3s
PASS ai-tells-lint 6.2s
PASS agent-cli 50.2s
PASS skill-routes-and-links 47.3s
summary: 6 pass, 0 fail, 0 skip (73.6s)

Review (interrogate)

Risky, because the diff touches GLOBAL.md, a sensitive path in the base AGENTS.md. Both lenses ran, plus the blast-radius check. The reviewer (opus, high) returned 6 findings, all judgment or grep-backed, with complete coverage.

  • Act on, fixed in 3686490:
    • The first version of the GLOBAL.md line named the playbook without a path, so a model could not find a user-invoked skill's file.
    • The first ui-polish text restated mobile-native's fixes and already got one wrong: the hover query also needs (pointer: fine).
    • The first playbook text presented a setupFiles reset as complete. It only reaches state it can import.
    • The shuffle had no seed and no note of its speed cost.
  • Consider, left open: perf-ratchet's description still says "Use for anything slow" and does not exclude CI. Changing a model-invoked skill's description owes the skill-diet before/after eval, and the GLOBAL.md route now fires first.
  • Noted: no eval ran for the playbook or the ui-polish line. Both changes correct facts against the cited sources, with no behavior variant to compare. Skipping is my call, recorded here.

Not verified

  • The routing was not tried in a fresh session on pi or OpenCode.
  • skills/ui-polish/update-refs.sh ran once on the first commit. It printed no UNROUTED or MISSING line for mobile-native. It was not rerun after the citation changed to The Symptom Table.

ELI5

Asking an agent to speed up CI now leads to the CI checklist instead of the app-performance tool.
The phone checks now say what a desktop browser cannot fake, and the test advice uses settings the tool enforces.

- ui-polish: the browser pass asked for touch emulation to show sticky
  hover, tap highlight and input zoom, but mobile-native section 11
  says emulation reproduces none of them; read those fixes in code
  instead.
- ci-speed playbook: put the per-file state reset in a Vitest
  `setupFiles` entry (rerun before every test file even with
  `isolate: false`, per the Vitest setupFiles docs) and use
  `sequence.shuffle`, instead of AGENTS.md prose a tool can enforce.
- GLOBAL.md: send slow CI to the ci-speed playbook, since the
  performance route caught every "make it faster" first.
- GLOBAL.md: yuu-mode is user-invoked, so name the playbook's path on
  each harness and say it may be read without running /yuu-mode.
- ui-polish: point at mobile-native's symptom table instead of
  restating its fixes, which already got the hover query wrong (it
  needs `(pointer: fine)` too), and add a real device when connected.
- ci-speed: a setupFiles entry only resets what it can import; turn
  on restoreMocks, unstubGlobals and unstubEnvs, gate isolate: false
  on a green shuffled run, pin the seed so failures replay, and note
  shuffle drops Vitest's longest-first ordering (Vitest config docs).
@YuukiFST
YuukiFST merged commit 2ba7dfa into main Nov 16, 2026
24 checks passed
@YuukiFST
YuukiFST deleted the fix/ci-speed-followups branch November 16, 2026 14:24
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant