Plugin Maintenance

Use the current user mandate and checkout as authority. Installed plugin files and a remembered model name are leads; the source tree, official provider documentation and registry metadata determine what can change.

1. Read AGENTS.md, scoped instructions and the active tracker issue. Verify HEAD, dirty paths, remote and scripts/detect_review_scope.py --json. Preserve other work; isolate a new change when ownership requires it. Record the outcome, protected behavior, allowed writes and publication authority in the repository's tracker. If no tracker is configured, record that limitation.
2. Inventory runtime model IDs, dependency constraints and locks, installed file mappings, public tools and front-door docs. Verify newer stable releases and exact API contracts using primary provider docs and PyPI/npm metadata. Record URLs, inspection date and any deliberately retained supported API.
3. Choose one evidenced defect or upgrade. Implement the smallest complete correction with the affected tests and docs. Independent workers need disjoint write paths and a single integration owner; inherit the configured model. A newer model string is insufficient if its thinking, cache, response or session contract differs.
4. Run the affected check. Let fresh results select the next correction. Preserve failures and skipped checks in the evidence. After one controlled intervention repeats the same infrastructure failure, stop that path and record the gap. Continue independent authorized work. Stop when the accepted outcome is achieved, evidence shows no useful next action, access is missing, resources are exhausted, or the user stops the work.
5. Join required workers, then run the final gates once: locked tests and lint for all three Python packages; node --test tests/installer.test.js; security smoke and offline tool-security checks; uv run python scripts/check_release.py; Python builds and npm pack --dry-run. Re-run affected checks only after a new change or a concrete unresolved failure. Verify a built MCP artifact through tool discovery and a read-only configuration call before claiming packaging acceptance. Provider execution and generated-media quality need separate authorized evidence.
6. Record exact commits, checks, limitations and next eligible work in the tracker and an audit receipt. Commit explicit owned paths. Push, merge, registry upload and GitHub release creation only within the user's stated authority. Keep review slices single-purpose; distinguish a source release from a registry install and from a running provider-backed server.

For later runs, resume the issue and verify any active worker/job identity before dispatching. One coordinator owns canonical writes. A scheduled prompt alone is not self-improvement: promote a workflow change only after a recorded recurring failure and a later independent check of protected capabilities. Use the host's supported automation facility only when recurring continuation is authorized; do not launch an unbounded paid loop from this skill.