v0.7.0

From one repo to a fleet, from a session to one command

Four features, each with its own commit for traceability:

• One-command onboarding. node <plugin-root>/tools/cdc-bootstrap.mjs . runs the whole setup headlessly from a plain terminal under a stated budget; --no-agent scaffolds everything for $0 (manifest, a blank starter case via the official plugin eval init --bare, .cdc.yml pinned to your installed Claude Code, the workflow). Idempotent, never overwrites.
• Fleet mode. tools/fleet.mjs pulls every repo's published results into one dashboard with a pin policy: status, pins, policy skew, month spend, coverage, last run per repo. Repos with errored or missing data turn the exit red, so the fleet view is itself a check. ci/fleet.yml is the drop-in scheduled workflow.
• Org rollout, serverless. ci/org-reusable.yml goes in your org's .github repo, the secret is set once at org level, and every member repo installs the check with a three-line caller (ci/org-caller.yml). A hosted GitHub App is deliberately out: it would put a server of ours between your code and your key.
• Community suites. The first one is live: the Spring Boot conventions setup with a published with/without measurement (the guard hook is worth +0.75 on its case, the conventions skill +0.25 to +0.60). The run is public and unedited: https://jameskomo.github.io/config-drift-checker/worth/report.html — and docs/community-suites.md says how to contribute the next stack.

88 tests. Includes the intermediate automatic patches from this batch.

v0.6.4

Full Changelog: v0.6.3...v0.6.4

v0.6.3

Full Changelog: v0.6.2...v0.6.3

v0.6.2

Full Changelog: v0.6.1...v0.6.2

v0.6.1

Full Changelog: v0.6.0...v0.6.1

v0.6.0

The red case now tells you which repair it needs

Four additions, all shipped hours after being asked for:

• Discovered vs invoked. Every shim run records which skills were actually discoverable (every SKILL.md that parses, malformed ones flagged). When a case whose grader watches skill invocation goes red, the diff now says which of two very different repairs is needed: "discovered but never invoked: suspect the trigger description" or "not discovered: packaging". This came directly from a user question, thank you.
• trace-keeper.mjs. The official claude plugin eval deletes its transcripts when it exits. Run it with --keep-temp, then node tools/trace-keeper.mjs out.json --clean: each run's trace is copied next to the JSON and the tool calls, response, turn count and model are written inline. Refusal labelling and the model stamp then work on official-runner results too.
• The drift index is now an observatory. The page states what it is (this suite re-runs on every Claude Code release), how many versions it has covered, when it last updated, and where to get alerts or run it on your own setup.
• One-command setup in the README: marketplace add, install, and /config-drift-checker:setup chained into a single line.

78 tests. Includes v0.5.1, the automatic patch that shipped the same changes earlier today.

v0.5.1

Full Changelog: v0.5.0...v0.5.1

v0.5.0

Runs under claude plugin eval, the day it shipped

Claude Code 2.1.269 made the official eval runner public. This release makes the two sides fit:

• Your suite loads under the official runner. The coverage ids cases carry moved from a covers: frontmatter key (which the official runner rejects as unknown) to a covers.yaml sidecar next to prompt.md. The shim and the coverage tool prefer the sidecar and still read the old form, with a warning. Case descriptions with inline colons are quoted for the stricter parser.
• Official-runner results flow through everything. claude plugin eval . --json out.json output now diffs against any baseline and renders in the HTML report and the drift index directly; mixed shim/official history works. Refusal labelling now requires inline transcript evidence, so it never mislabels official-runner results (their trace file is temporary).
• README rewritten around the new landscape: proof and install above the fold, and an honest comparison table of what the built-in command does and what this adds on top (baseline, history, noise bands, canary, bump PRs, budgets, coverage, repair).

74 tests, zero runtime dependencies, verified against a real 2.1.269 run end to end.

Includes v0.4.1 and v0.4.2, which shipped the same changes as automatic patches earlier today.

v0.4.2

Full Changelog: v0.4.1...v0.4.2

v0.4.1

Full Changelog: v0.4.0...v0.4.1