[pigfyy]

Why

These come from running the orchestrate playbook for about a day on a large personal project: a game-mechanics trainer with dozens of Opus and Codex workers, cross-family reviews, and UE builds that hold a shared lock. Each edit fixes a failure that cost a real retry or a human turn. Each is general, not specific to that project.

What changed

playbooks/hillclimb.md

• Step 1. When the metric matches an external reference whose rule can be read (source, a binary, a spec) and reading it is permitted, read the rule before fitting to it. A fit kept before that stays labelled provisional, and the label never blocks keeping it.
  • Evidence: movement was fitted piece by piece from recordings before anyone read the reference's code. Reading it resolved several rules the fits had missed.
• Step 2. When comparing target and reference, put both sides through one measurement path, and draw reference cases from across the workload dimensions named in step 1.
  • Evidence: the two sides of the comparison were once measured with different filters and lag handling, and a review had to block on it.

playbooks/orchestrate.md

• ACCEPTANCE. A unit built to answer a question is accepted on a ranked list of what the data says is wrong, and "no differences found, with evidence" is a valid result.
• Liveness, retry by mode. On a rate-limit failure, read task_status, and relaunch only when the task is terminal-failed or stalled past a bound. A rate-limited task can resume on its own, and relaunching beside it put two writers in one worktree.
• preferences.md. Each order ends, on its numbered line, with what it supersedes and when it expires. Superseded orders are rewritten in place, with the old text recorded in decisions.tsv. A user-written order is removed only with the user's okay.
  • Evidence: an append-only register grew contradictory orders that workers obeyed.
  • The note stays on the numbered line because orch's parser requires numbered lines.
• Step 2. Before the pilot, create or confirm the project's verify control skill. A single documented verify command counts.

SKILL.md, Subagents

• "Local checkout" is no longer a reason to reuse an agent. A costly checkout alone, such as a cold build or imported assets, hands its worktree path to a fresh agent. The uncommitted-changes and live-process exceptions and the stop-and-confirm rule are unchanged.

tools/forks.json

• hillclimb.md and orchestrate.md are declared policy forks with upstream: not-proposed, so tools/sync.mjs --dry-run accepts the edits.
• The existing SKILL.md entry's why names the narrowed rule.
• CONTRIBUTING prefers changes to what a skill does to go to cursor/plugins upstream. Six of the seven aren't T3-specific; the task_status one is. If you'd rather carry them upstream, these forks can be dropped once that lands.

Verification

• tools/sync.mjs --dry-run passes on the patched tree under WSL with bun 1.3.14, CI's version. On Windows it fails even on an unmodified tree, because of mixed path separators.
• markdownlint reports 0 errors on the three changed Markdown files.
• Not done: the VERSION bump and the CHANGES.md entry. The fork entries assume 2.0.0-rc.19. Adjust them if the next release differs.

🤖 Generated with Claude Code

[pigfyy]

Closing: this was opened by mistake, sorry for the noise. These changes will stay local.