v6.2 is a harvest release: edits adopters had made locally to shared toolkit files are folded back in, generalized, reviewed and tested, so you can update without losing them.
It is also the first public release since v5.2. It ships the shared Claude/Codex layout: one canonical .agents/ source (roles, skills, helpers, hooks, adapters, templates) with generated .claude/ and .codex/ projections. If you are coming from v5.x, read the upgrade note below.
What ships
- 49 workflows, 11 roles, 23 knowledge skills.
- GitHub-based learning intake. Disabled by default, GitHub Issues only. Each project sets its own receiver as
OWNER/REPO. Installing or updating never turns delivery on and never contacts GitHub. - An author-neutral
ghostwritingskill andshadowrole. The craft rules stay. The voice comes from a project-supplied fingerprint, and the skill ships with a blank voice-fingerprint template. While it is unfilled, the skill stops instead of inventing a voice. A filled fingerprint lives underreferences/_voice/, which updates treat as project-owned and do not overwrite. - A publish preflight that refuses to copy private names or personal paths into the public tree.
Highlights
Gate
- New opt-in
visualleg. Runs in the full gate only, locally only, and a run that finds zero snapshots fails. - The test-leg summary now reports TIMED OUT with detail, and a killed visual leg is reported as TIMED OUT.
- The build, lint, e2e and visual legs get a default timeout when your config omits one.
Secret scanning
- A source-code reference allowlist and a TypeScript type-member allowlist for the generic secret-assignment patterns. They apply to source files only and are judged on the whole line tail. The scanner and the redactor share them, and the read-output scrubber and the secret-safe writer now pass the file path to the redactor.
- Nothing exempts a credential-shaped quoted literal, a comment, a shell string or an opaque run.
- Runtime-binding references wrapped in inline code are no longer flagged.
- Reference checks are bounded to the assigned expression. A literal fallback is rejected, and redaction covers the whole flagged expression without hiding sibling secrets.
Skill evolution and memory
- Attribution now uses IDF weighting, a distinct-token gate and a confidence margin. Near-ties are reported as ambiguous instead of attributed, and an explicit "none attributable to the agents" section counts as caught.
- The memory promoter no longer suggests the architecture skill's template scaffold as the target for constraints.
Verification
- Published tree: 159 test files, 5036 tests passed, 11 skipped, 0 failed. Run on a clean checkout of the public repository's current HEAD. The generated-projection check (
render-shared-agents.js --check) reports the projections current. - Independent review of the v6.2 harvest: 0 critical, 6 major, 7 minor and 8 nits, all fixed. A closure pass found 4 minor and 5 nits, all fixed.
- One adopter application that took this version ran its own application gate: build passed, typecheck and lint clean, 9768 passed, 0 failed. That result is one adopter's evidence and does not cover other projects.
Upgrading
No migration is required for v6.2. One note: IMPROVE scores from skill-evolution intake are IDF-weighted from this version and are not comparable with earlier ones.
Update through the template updater. Preview first, then apply, then confirm the generated files match their sources. Claude uses .claude/core/; Codex uses .codex/core/ with the same arguments.
node .claude/core/template-updater.js update /path/to/project --merge --dry-run
node .claude/core/template-updater.js update /path/to/project --merge
node .claude/core/render-shared-agents.js --check
--mergekeeps your root instructions, role personas and Project Context, bespoke skills, marked project additions and local settings.- A conflict in a generated file is a conflict, not a signal to force. Move the intended change to its canonical source under
.agents/, then render. - Newly installed or changed native hooks may need review or trust in your runtime before they run.
- Coming from v5.x: read
docs/reference/agents/migration-v6.md. Your application's own build, test and PR gates are unchanged. Toolkit tests run in the toolkit repository, not in your project.
Full release page: https://github.com/Domdhi/Domdhi.Agents/releases/tag/v6.2