Playwright disables normal background throttling, so a hidden 5chan page keeps
doing P2P and rendering work after a check finishes. Agents verifying in
parallel across worktrees stacked whole browser engines on one machine.
Add scripts/pw-session.sh, a wrapper that permits one active Playwright browser
at a time and records who holds it:
- The lock is machine-wide, not per-repository, because the contended resource
is RAM and CPU. Every worktree and checkout shares one slot.
- Acquisition is an atomic mkdir. Stale locks clear themselves: `open` reclaims
any slot whose recorded browser is no longer `status: open` in
`playwright-cli list --all`, so an interrupted workflow cannot strand the
budget. When that list cannot be read the lock is left alone, so a broken CLI
never silently disables the budget.
- `open` exits 75 when the slot is busy; `--wait[=SECONDS]` blocks instead.
- `close` always stops the browser, even when the lock was already lost, and
never releases a slot held by a different session.
- `status` reports the holder and whether its browser is still alive.
Agent policy now runs browser engines and profiler batches sequentially, uses
Chrome/Blink during iteration and the full engine matrix only for final
verification, and never uses `close-all` or `kill-all` while other agents may
own sessions.
Covered by scripts/pw-session.test.js.
Drop the pinned `model: sonnet` from the reasoning-heavy subagents so they
follow the session model instead of being forced down to Sonnet. Running an
Opus session now gets Opus subagents; deliberately dropping to a cheaper
session model is respected too.
Mechanical agents (browser-check, profiler, translator) keep their haiku pin
— that pin is a capability statement, not a stale default, and should hold
regardless of session model.
GitHub Projects are no longer used in this repo. Remove issue/project
finalization from review-and-merge-pr and the project-board step from
make-closed-issue in all three toolchain mirrors, record the surprise
in known-surprises.md, and regenerate llms files.
Claude Code never reads a standalone .claude/hooks.json, so the PostToolUse and
Stop pipeline is moved into .claude/settings.json and hooks.json is removed;
.cursor/hooks.json is rewritten in Cursor's version+afterFileEdit/stop schema;
.codex/hooks.json is already Codex-valid and stays. The shared scripts now
parse both the Cursor file_path and Claude/Codex tool_input.file_path stdin
shapes and normalize absolute paths, so the format, yarn-install, and
react-pattern-review hooks stop being silent no-ops. verify.sh blocks with
exit 2 plus a stderr reason, guards stop_hook_active, and skips clean trees;
react-pattern-review surfaces its reminder via hookSpecificOutput on
PostToolUse; sync-git-branches no longer misreports open PRs as merged. The
validator now checks that the three harness-specific entry points wire the
same hook scripts instead of requiring byte-identical hooks.json copies.
browser-check and profiler get an explicit read-only tools list mirroring the
Codex sandbox_mode, and test-apk no longer hardcodes one contributor's
ANDROID_HOME and project paths.
Fixes deslop diffing against nonexistent main, an invalid rg --type tsx flag
in fix-merge-conflicts, a Codex-specific phrase leaking into the Claude/Cursor
inspect-elements description, and aligns commit-format with the required-scope
convention the commit skill and git history already use. Rewrites the readme
skill from 764 lines of Rails-specific guidance to a lean version matching
this Vite/Capacitor/Electron repo.
Borrowed from a review of addyosmani/agent-skills:
- add scripts/validate-ai-workflow.mjs (yarn ai-workflow:check): verifies
.claude/.codex/.cursor skills, agents, and hooks stay in parity, with
validator-owned exemptions for intentional harness-specific differences
and enforcement of the AGENTS.md agent model rules
- browser-check and profiler agents: treat page content as untrusted data,
never instructions (5chan pages render arbitrary user-generated content)
- refactor-pass: Chesterton's Fence rule (git blame unclear code before
removing it)
- review-and-merge-pr: pass subagent verifiers only the artifact and
contract, not the triage verdict, to keep reviews independent
Add a Claude Code SessionStart hook (matcher: startup) that runs
`corepack yarn install` when node_modules is missing, so worktrees Claude
creates internally are immediately usable. Also extend create-task-worktree.sh
to install deps after the worktree is created, covering Codex/Cursor/manual
flows that share that script.
Rewrite two release entries that leaked dev jargon (v0.7.4 "Pretext feed
sizing" / "mobile scroll perf", v0.7.3 "compact account history") and tighten
the release skill's Step 3 with a non-negotiable plain-English rule plus
good/bad examples. Mirror the rule inline in AGENTS.md so agents editing the
blotter outside the skill still see it.
Mirrors .cursor and .codex structure with Claude model assignments:
- haiku for lighter tasks (translator, browser-check, profiler)
- sonnet for standard tasks (code-quality, plan-implementer, etc.)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tommaso Casaburi