[euwars]

Why

These rules lived only in a local Claude Code plugin cache, where each one came from a repeated correction during real sessions. A plugin update drops them. Moving them into the source gives every install the same guidance.

What changed

• poteto-mode gains four subagent rules: brief an agent on an unknown system from its source instead of your guess, watch a long job by progress instead of the clock, give parallel worktrees their own databases and ports, and treat a stop order that adds work as a fresh task. Three principle entries and their leaf descriptions widen to match.
• The Claude Code adapter adds worktree cleanup steps and says the parent runs the merge, since a worktree worker cannot git -C the main checkout.
• architect adds a learned-surface budget and a blind guessability probe. blast-radius, the investigation playbook, and technical-writing widen their triggers.
• harness/pstack/claude-code is regenerated.

Scope

Prose only. No new skills and no model or budget changes. The version bump is left to the maintainers.

Blast radius

Every pstack reader on Cursor and Claude Code sees these lines. Each addition narrows behavior and none removes a rule, so the risk is wording, not routing.

Verification

• node scripts/generate-pstack-harness-trees.mjs --check reports the tree is up to date.
• node scripts/validate-plugins.mjs passes.
• The generated tree is byte-identical to the local plugin copy these rules ran in, apart from two sentences reworded here to drop a flourish and a mid-sentence colon.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HGD8v9WkwgmcM4JPipW6fe

---

Note

Low Risk
Prose-only skill and harness documentation; it narrows agent behavior but does not change code paths, routing, or model defaults.

Overview
Moves session-hardening guidance from ephemeral plugin cache into pstack skills and the Claude Code harness doc so every install gets the same rules.

poteto-mode adds subagent operating rules: brief unknown systems from primary sources (no planted guesses), monitor long jobs by commits and item progress instead of time estimates, isolate parallel worktrees on their own DBs/ports/migrations, persist model/effort changes into /setup-pstack, and treat a stop order that adds scope as a fresh consolidated task. Related principle leaf descriptions widen (context-window full reads, explain-the-number on tuned scores, encode-lessons for habits and flakes, shared-state separation).

architect Phase B now budgets learned surface (keys, options, verbs, routes) and requires a blind guessability probe with a pre-written answer key before synthesis.

blast-radius, investigation, and technical-writing expand when they apply (global defaults proved by behavior suites, pinned-source grounding, fresh-agent how-to tests, and labeling repo proposals as not built).

docs/harness/claude-code.md documents worktree .env copy, parent-owned merge --ff-only, worktree cleanup, and self-checking background job progress.

^{Reviewed by Cursor Bugbot for commit 6f101d2. Bugbot is set up for automated code reviews on this repo. Configure here.}