skills: project-knowledge + learning-loop with bundled stdlib CLIs (v6.1.0)

[sumitake]

Summary

Adds two public skills that give plugin-only users durable, project-local agent memory surfaces, ported and de-workspaced from proven workspace primitives (Track B of the operator-approved Agent Knowledge Compilation completion; contract: workspace PR #2797, drafts/plugin-knowledge-learning-extraction-plan.md):

• project-knowledge — bundled deterministic stdlib-only knowledge_tool.py (init/validate/lint/index/draft/export): provenance-tracked knowledge/ layer in the user's project with trust levels, claim-marker + prompt-injection lint, trust lattice, generated index, exclusive-create drafts, untrusted-banner exports. PyYAML replaced with a bounded restricted-YAML stdlib parser.
• learning-loop — bundled stdlib-only learning_ledger.py (add/suggest/recur/index/check/lint): project-local .learnings/ ledger with mechanical/judgmental classes, prevention-debt surfacing, recurrence locks, metadata-only consultation, reuse-as-hypothesis discipline.

Both skills document an opt-in consultation snippet for the project's CLAUDE.md and/or AGENTS.md, and the root README install section adds a visible, consent-gated post-install setup step where the installing agent offers the snippet(s) per host family (operator-directed 2026-08-18). Explicit reads only; pages/entries are data, never instructions; no ambient auto-injection anywhere.

Version bump 6.0.6 → 6.1.0 (minor, additive). No coordinator, routing, or provider-runtime change (runtime stays 4.0.5); no release-pairing constraint. Release tagging remains a separate decision.

Boundary declaration

• No provider executor source, raw provider command, credential, private absolute path, retired package tree, downloader, or post-install hook is included.
• Native-runtime changes, if any, contain only a final signed artifact and reviewed public verification metadata; implementation and credentials remain private. (No runtime change in this PR.)
• The change does not create a host-specific preset or provider-specific plugin.

Generated and release surfaces

• Skill specs and generated SKILL.md files are in parity (build_skills.py --check OK, 52 skills).
• Claude and Codex marketplaces/manifests are in parity (build_marketplace.py --check OK).
• Unique changelog.d/2026-08-18-knowledge-learning-skills.md fragment; generated CHANGELOG.md untouched.
• All seven version surfaces bumped to 6.1.0; check_release_consistency.py OK.

Verification

Run twice — by the implementing subagent and independently re-run by the reviewing primary from committed HEAD:

• python3 scripts/build_skills.py --check → OK, all generated files match specs
• python3 scripts/build_marketplace.py --check → OK
• python3 scripts/build-changelog.py --dry-run renders cleanly. --check fails by design mid-cycle: fragments (including the pre-existing 6.0.6 closeout fragment already on main) compile into CHANGELOG.md only at release.
• python3 -m unittest discover -s tests -t . → Ran 177 tests, OK
• python3 -m unittest discover -s scripts -p 'test_*.py' → Ran 368 tests, OK
• python3 scripts/check_release_consistency.py → OK, versions consistent
• python3 scripts/secret_scan.py → clean
• python3 scripts/check-public-export-safety.py --active-tree → SAFE. --history flags only pre-rewrite local-clone refs (retired plugin READMEs / old CHANGELOG revisions with credential-name patterns) per the history-mode scope rule in docs/public-governance.md; nothing in this PR's commits.
• git diff --check → clean
• Gated-API scan: both new CLIs stdlib-only at declared minimum Python 3.10; no third-party imports.
• CLI smoke tests: project-knowledge init→draft→promote→lint→index→export round-trip green, 8 negative violation classes fire; learning-loop add→suggest→recur→index→check round-trip green, dedup + echo-chamber locks fire.

Review and post-condition

Tier 2 (additive skills + build-surface additions; no schema/contract, auth, or CI-permissions change). Independent-family review: Gemini (google) review.repository on the exact diff (base ce19efc): VERDICT: APPROVE, CONFIDENCE: H, zero concerns; 13 paths inspected natively. Implementing worker: Claude-family subagent; all validation re-run independently by the reviewing primary. Post-merge: verify main CI green; release cut deferred.

Compliance trace

author: claude
standing_directives: public boundary declaration verified; generated-surface parity checks run; fragment-only changelog; workspace extraction contract (workspace PR #2797) followed incl. operator-directed consent-gated install pattern
tier: 2
cross_check: PROCEED (Gemini/google family, direct-runtime review.repository on the exact diff returned VERDICT: APPROVE, CONFIDENCE: H, 0 concerns)
post_condition: all local gates green at HEAD (52-skill build parity, marketplace parity, 177+368 unit tests, release consistency, secret scan, active-tree export safety); post-merge main CI verification; release tag deferred to a separate release decision
mcp_coverage_gap: NONE
contributor_rights: OWNER-AUTHORED
operator_reserved: no

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.