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
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
1f84973 — Reference the odere-pro GitHub account in the manifests: add owner.url to marketplace.json, fix the plugin.jsonauthor block (name had held an email) into name/email/url, and point homepage at the repository.
ea809f7 — Make the calibrate self-audit return clean on a correctly-built repo by fixing the defects and false positives dogfooding surfaced: calibrate-claude-md's lint.sh no longer aborts mid-loop under set -e/pipefail (honours the always-exit-0 contract) and only flags genuinely aspirational prose for vague-rules (skips markdown headers, fenced code, and precise modal prohibitions); the subagents/rules/hooksenumerate.sh scripts skip CLAUDE.md/README.md so memory/doc files aren't audited as components; the hooks script-scan no longer mistakes case labels, quoted strings, comments, and break/continue/;; for PATH commands; skill:cli-not-wrapped only fires when a skill has bare unscoped Bash (so it could actually shell out); and general:nested-claude-md-conflict now counts only nested CLAUDE.md the root CLAUDE.md doesn't index, so documented layering isn't flagged as sprawl.
5d489e1 — README badges + power-words glossary: add a shields badge row (incl. live gates + OpenSSF Scorecard), define the project's power-word vocabulary in docs/glossary.md (notably agent vs subagent as distinct terms — a subagent is an agent that runs in its own context window), and add gate G17 asserting the glossary defines that vocabulary.
5d489e1 — Add an author-only deterministic eval harness (tests/eval/): run-eval.sh scores the plugin's own shipped payload (gate floor + correctly-scoped lint + adversarial durability checks) into a versioned JSON snapshot, and compare-eval.sh diffs snapshots to track improvement/regression across runs, with the time series stored in tests/eval/history.jsonl + a blessed baseline.json.
5d489e1 — Harden the plugin via a self-calibration run: add CRITICAL gate G18 (18-changelog-fragment-unique) asserting every changelog/<NN>-<slug>.md carries a distinct <NN> (and renumber the duplicate 02-readme-badges-glossary-gate fragment to 03 so it passes), and make gate G17 skip the gitignored .claude/calibration/ run artifacts so a local self-calibration run no longer trips the glossary scan — closing two unenforced/over-broad spots the dogfood surfaced.
fd18948 — Track whether calibration improves a setup over iterations: add the /calibration-track flow and its snapshot.sh / compare.sh engine — a deterministic config-quality snapshot (doctor floor + signature-keyed lint over all nine features) compared vs a baseline anchored to the last PR merged onto main and vs the previous iteration, persisted in a local gitignored ledger and independent of /calibrate's circular built-in delta.
57d9ecb — Fix a claude-md:contradicts-nested false positive in calibrate-claude-md's lint: it diffed a root CLAUDE.md against itself (relative-path string comparison) and read shell comments inside fenced code blocks as headings — now compares by inode identity (-ef) and strips fenced blocks before extracting headers, so the signature only fires on genuine cross-level header collisions.
1ce0d38 — Add /calibration-flow, a shipped behavioural-flow evaluation capability: it drives a multi-step workflow over a case set of golden fixtures and grades whether the chain delivers intent and keeps its handoffs sound (node recall/precision, edge handoff:* contracts, flow intent), reusing the existing severity/signature/recurrence vocabulary. The verdict comes from a pure, deterministic scorer (score-flow.sh) — no LLM, no network — so it can be wired as both a CI gate (G19, via lint-fixtures.sh) and an on-demand run, with a calibration-flow-evaluator worker producing the findings the scorer judges.