You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fewer critique rounds. Session logs showed that most extra rounds came from partial fixes (a fact corrected in one place but left stale elsewhere) and from fix side effects (new content contradicting existing content), not from issues the first review missed.
Every FAIL now cites every location the defect appears.
Open-ended criteria (req-coverage, logic-correct, edge-cases) are swept exhaustively, and the reviewer greps for other copies of each defect.
A new Applying Fixes procedure: fix → propagate → reconcile → verify → record a change list.
Re-critiques receive the previous log (PREVIOUS) and the change list (CHANGES). They check prior FAILs first (marking survivors (recurring)), then the changed regions, then run a fresh full pass.
Severity tags. Every FAIL is tagged [blocker] or [minor] after the binary call. Style and naming rules stay minor even when a standards file documents them. Minor-only results get VERDICT: MINOR FIXES ONLY and skip the re-critique.
Session entries record Target: (so a re-critique can find its previous check) and keep (recurring) on the Failing line. The log skill ranks the worst check by blocker count and surfaces (recurring) criteria.
New deterministic checks (200 total across both suites) guarding the round-reduction rules, the session entry format, and the new dispatch slots. Drift guards now also tie the README's deterministic total and the README and plugin/CLAUDE.md eval counts to the real numbers.
Three new evals (9 total): repeated injection must cite both locations, a re-critique must flag an unfixed sibling as recurring, and a minor-only change must get the minor-only verdict. The planted-bug evals also assert [blocker] tags.
Changed
Dispatch hands the subagent the whole skill. The subagent now reads and follows SKILL.md itself (skipping Step 0). The old template embedded only the criteria and log steps, so the reviewer never received the output format, the Unverified requirement, the scope rules, the context gate, or the anti-manufacturing rule. The evals missed this because they only exercise the inline path.
The reviewer gets the task. A new TASK slot carries the user's intent (not the author's reasoning), which logic-correct and req-coverage need.
The context gate gathers context before refusing. It used to STOP whenever files hadn't been read yet. It now outputs CONTEXT INSUFFICIENT only when tools can't obtain what's missing.
Unverifiable criteria are marked PASS (unverified) and listed under Unverified instead of FAILed, since a FAIL needs evidence of a defect.
tests-pass and no-regressions no longer overlap.no-regressions now covers broken existing contracts (APIs, signatures, config, output formats).
Plan criteria are stack-neutral (dropped "Zod"). api-verified accepts installed source and type definitions as evidence.
README documents severity tags, the minor-only verdict, and the round-reduction loop. Its example output uses the current format, and the check/eval counts are current.
Prompting refreshed for current models: dropped "think step by step", explained the reasons behind rules instead of shouting them, and gave the code-mode target a default (uncommitted changes, otherwise HEAD).
Fixed
A grep that matched nothing inside a $(...) assignment silently aborted the suites under set -euo pipefail: no FAIL line, no summary, and the remaining checks skipped. The :-none, :-0, and "not found" fallbacks were unreachable. All seven such substitutions are guarded now (the count and dimension guards, eval.sh's result parsing, and the hook counters), so a missing value prints a real FAIL. Found by independent critique check #18.
v3.3.1's drift guards only covered SKILL.md — README.md and plugin/CLAUDE.md, two of the three files the dimension bug actually shipped in, were unguarded. Found by the plugin's own independent critique (check #15, tests-exist FAIL). Four new checks verify the dimension/criteria counts stated in both files against the headers counted from SKILL.md. Guard verified by seeded regression: flipping README back to "8 dimensions" fails the suite.