[thornbrambles]

Refs #1361

Changes

• Run the existing ladder before app or feature design, planning, assessment, and delegation.
• Ask who needs a proposed mechanism, what already exists, where to fix the source, and who would act on a gate and at what cost. Give subagents the outcome and room to reject the mechanism.
• Sync the compact rules, packaged OpenClaw skill, and fallback instructions.

Worked example

A user asks whether the release-note writer checks for active voice. Before drafting a subagent brief, ask Does a separate check need to exist? Release editors already review drafts, and a false positive from a blocking check would delay publication and consume tokens on retries. The existing writer prompt is the first place to address the style rule.

Help the release-note writer produce active-voice prose. Trace where the notes are generated and see whether the existing prompt can address the issue. You can rule out a separate check if it has no reader or the prompt can solve the problem.

Comparison

I ran GPT-5.6 Sol at low reasoning effort in Codex, with six trials per arm on each task. All arms received identical fixture source in their prompts. The current and patched arms loaded their respective compact rules through AGENTS.md. I scored the delegation replies with the arm labels hidden and checked the generated order handlers with the repository's conv-service scorer.

Result	No Ponytail	Current	Patched
Correct diagnosis, prompt first, separate checker optional	6/6	6/6	6/6
Identified who reviews or acts on the output	1/6	5/6	6/6
Named a concrete cost of a separate check	0/6	0/6	2/6
Median delegation output tokens, including reasoning	136	196	231
Order-count regression checks passed	6/6	6/6	6/6

The patch tied on the main delegation outcome. The patched replies mentioned the cost of a separate check more often, and their median output was longer. This is a Codex comparison of the compact rules. The documented full-plugin Claude Code benchmark is pending, so this PR is a draft.

Local verification

• 77 focused integration tests and 30 Pi tests passed. Rule-copy and version checks passed.
• Full npm test and the offline benchmark self-test showed the same failures on both the patched and untouched Windows checkouts.