Skip to content

v0.4.0

Choose a tag to compare

@github-actions github-actions released this 01 Nov 17:15
· 4 commits to main since this release
d59e1a0

Minor Changes

  • #38 1bb2a68 Thanks @Paalt! - New blast-radius lens in review-pr, for changes that ship to production. It names the one safety fact the change rests on, proves it as far as reading allows on an evidence ladder (asserted → documented → traced → command-settled → observed, the same names as verifiedBy), writes the probe that settles it by execution, walks the change's reach past grep (other readers of the same bytes, the library at its pinned version, the path production actually runs, timing, rollback), and reports what it cleared as first-class output. On your own checkout the orchestrator runs the probe, so the fact arrives proven or contradicted. The report leads with a Blast radius card, publish.ts puts the fact at the top of the GitHub review body, and resolve-review re-runs the probe after its fixes. Runs from the medium tier up; trivial and light diffs skip it. The method adapts the blast-radius skill in Lauren Tan's (poteto) MIT-licensed pstack, credited in the toolkit README.

  • #41 b5389a5 Thanks @Paalt! - New create-verify-skill skill: generates — or refreshes — a repo-local verify-<app> skill that drives the real app the way a user does (launch, doctor, personas, drive, evidence, cleanup) plus a feature map with one file per user-facing feature, its source Paths and the seeded Personas it needs, proven by an observable end state. The generator interviews the repo, writes the skill into the host's project skill directory, and drives one feature end to end before handing over; refresh mode runs source readers per feature and one live pass and leaves only proven corrections in the working tree (a PR when asked). Both modes run on your own checkout against a local or disposable instance. The toolkit consumes the skill by shape: review-pr runs experiment claims and the blast-radius probe through its recipes on owned checkouts, and resolve-review drives the features a review touched after its fixes land, with the outcome in the resolution ledger. The method adapts the create-verification-skill and maintain-verification-skill skills in Lauren Tan's (poteto) MIT-licensed pstack, credited in the toolkit README.

  • #40 fef7ba8 Thanks @Paalt! - New interrogate[=N] arg for review-pr: the pipeline's one offensive pass. N adversaries (default 2), each in its own scratch worktree, are briefed to break the change by running it — three focuses, inputs and boundaries / state and time / environment and integration, assigned round-robin — and a finding survives only when the orchestrator's rerun in a fresh worktree reproduces it; the repro command, patch, observed and expected output travel with the finding and render on its card. Adversaries execute code, so the arg is honored only on an owned checkout or under REVIEW.md's ## Trust. Runners are detected, never assumed: seats run on the host's subagents, and when the Codex CLI is on PATH one seat runs there so the adversaries span model families; with one family the report says diversity came from the briefs. An Interrogation card shows each seat's found / kept / dropped ledger and the interrogate-only findings. Off by default, like holistic. The method adapts the interrogate and arena skills in Lauren Tan's (poteto) MIT-licensed pstack, credited in the toolkit README.

Patch Changes

  • #43 5dcaca0 Thanks @Paalt! - Field round from a five-PR queue run. review-pr: a repeated mechanical substitution (the same literal, rename, or import swap in every hunk) takes the light tier however many files it touches, and the light-tier text now defers to the ladder row instead of restating it; when the conflict probe hits most of a PR's files and base replaced what the PR edits, the review asks whether the PR is superseded — goal met on base by reading and measurement — before planning any merge. The intent lens compares equivalence and duplication claims (keys, resolved values, exports) instead of accepting them. recap.ts emits a one-line classification-only block when no mapped primitive was touched, instead of an empty diagram and table. reset-workspace.sh leaves review-report/verify/ in place when it archives stale findings — evidence dirs are keyed by phase and SHA, and moving them dangled every link the findings recorded. A running verify instance and the review scripts no longer collide: the generator now mandates runtime state (pids, port, head, log, registered clients) in the git dir — $(git rev-parse --git-dir)/verify/, shared registrations one level up in the common dir — so review-report/verify/ holds evidence only; check-clean.sh refuses while those pids are alive and fails loudly, leaving the original where it names, instead of nesting review-report/review-report when a launch recreated the directory mid-check; a contract test pins the pid-file path across generator and script. Existing generated skills pick the layout up on their next refresh. create-verify-skill: identifiers a generated Launch derives from the checkout path are sanitized to the target's charset, Doctor prints them, and refresh mode re-reads Launch, Doctor, and Cleanup against the generator's contract so existing skills pick the rule up. Review-fix on this round added: a superseded verdict is always comment, never approve; the review-pr README defers to the ladder row and a contract test keeps the light-tier predicate stated once; the light generalist owes a substituted token the hidden-reader grep that blast-radius would have done at a higher tier.