[Hyper-AI-Lab]

What

PAACM 3.2.0 acts on an efficiency review of 3.1.0. The review measured the method's context cost (about 700 tokens always loaded, roughly 8,000–10,000 for a full DEEP task), compared the kit with its own study, and searched for evidence the study does not cite.

• Reasoning effort by role. paacm-reviewer runs at effort: high and paacm-implementer at effort: medium, as ADR 0010 prescribed but the agents did not set. Cursor ignores the field.
• Code intelligence (ADR 0017). The explorer and the reviewer list the read-only LSP tool and use it for definitions, references and implementations when a code-intelligence plugin is installed; the tool stays inactive otherwise. The setup guide explains Anthropic's official plugins (pyright-lsp, typescript-lsp, …), and paacm doctor warns when a project enables one whose language server is not on PATH.
• Test strength (ADR 0018). Phase 5 adds a property-based test for rule, parsing and transformation logic when the project has a property-testing library. A project may add a mutation check to the full tier; when it fails, the agent lists the surviving mutants and stops with BLOCKED: trigger 4 instead of editing the frozen tests.
• Evaluation guide. How to read outcome, tokens, cost and time from a claude -p --output-format json trial for eval grade. Measured on Claude Code 2.1.292: in a resumed run total_cost_usd covers the whole conversation, while usage, duration_ms and num_turns cover only that run.
• Research addendum (17 October 2026) in 04-verification.md and 07-claude-code-and-cursor.md, with the new sources (MutGen, SWE-Mutation, agentic property-based testing, PBT-Bench, a minimal-pair study of clean code, the code-intelligence docs).

Considered and not done: an agent runner for paacm eval (an L1 arm stops for human approvals, and the guide leaves running trials to the team's harness), and affected-tests tiers (no measurement yet shows gate time matters).

Verification

• node bin/paacm.mjs validate: exit 0
• node bin/paacm.mjs build --check: exit 0
• npm test: 88 of 88 pass (2 new: doctor and the code-intelligence plugins)
• markdownlint-cli2@0.23.3 "**/*.md": 0 issues
• claude plugin validate plugins/paacm --strict and claude plugin validate .: passed
• The CI step that installs into the four fixture projects: exit 0, doctor 0 problems in each

After merging

Tag v3.2.0 on the merge commit. v3.1.0 still goes on 089cd0c, the current main, independently of this PR.