Skill upstream maintenance

Last checked: 2026-10-07. Repository base: 163951e (PR #417).

skills/ is canonical. The Amp bundle in configs/amp/plugins/my-ai-tools-skills/skills/ must contain identical files. The Claude marketplace points directly to canonical directories. generate.sh exports installed home configuration; it is not an upstream skill updater.

Refresh policy

• Read local history before replacing an upstream-inspired skill. Preserve local tool names, resource paths, approval rules, and intentional workflow changes.
• Fetch upstream text at a recorded commit. Review scripts, hooks, MCP startup, credential access, network destinations, and external-state changes before import.
• Copy changed canonical bundles to the Amp mirror and run the mirror test.
• Recommendations are live upstream installations, not vendored or locked bundles. A checked commit records the review point, not a promise that a future install uses that revision. Review executable content again before installation or use.

The October 2 refresh was an unpushed commit on another executor. Its thread was read for the method and decisions; its patch was not copied. This refresh compares current files directly with current upstream sources. Later local changes take precedence, including the restored babysit-pr and first-party visual-pr.

Canonical skills: 36 checked

The following 22 skills are local designs or adaptations of ideas/standards, not maintained upstream file copies. Their current repository versions remain canonical:

accountable-engineering, adr, blindspot-pass, capability-experiments, code-quality-review, commit-atomic, context-discovery, doc-search, docs-update, draft-pull-request, git-context, handoffs, implementation-logger, llm-wiki, orchestrating-fusion, pickup, pr-review, qmd-knowledge, quiz-me, security-audit, slop, spec-interview.

The 15 tool-specific SKILL.md files under configs/{amp,cline,kimi-code}/skills/ are local adapters for code-reviewer, test-generator, documentation-writer, ai-slop-remover, and security-audit. They have no independent external upstream; keep their tool-specific guidance. Do not replace the short adapter security skill with the larger canonical audit workflow.

Recommendations: 21 entries, 20 repositories checked

All checked repositories are reachable and unarchived; API trees were not truncated. Named skill selectors were checked against frontmatter, not only directory names. In particular, Engram's plugin/claude-code/skills/memory/SKILL.md declares name: engram-memory; the existing selector is correct despite the directory name. Collection entries remain live collections; this check is not a security audit of every skill they may install. Preserve the local pstack-first order.

Security review notes

• No new executable scripts, hooks, MCP declarations, endpoints, or credentials are imported by this refresh. Existing scripts remain unchanged.
• Portless may change trust stores and hosts files, expose servers, install a root/SYSTEM startup service, or terminate processes. The local skill now makes those approval boundaries explicit. Do not print TLS keys or ngrok tokens.
• Codemap's supplementary secret check prints only filenames. It does not prove absence of all secrets. Invalid scope does not authorize wider exploration.
• Plannotator JSON can contain private goal details. Keep it in the goal directory; browser idleness is not permission to terminate the user's session. Commands use direct stdout redirection so a pipeline cannot hide a CLI failure.
• Engram hooks perform session writes, import Git-synced memory, and may start a local server with cloud autosync enabled. Cloud transmission depends on configured credentials/settings. Review those settings and consent before enabling cloud sync or an external ENGRAM_URL; installing a skill alone is not installing its hooks.
• Visual Recap normally publishes to hosted Plan MCP. Use its supported AGENT_NATIVE_PLANS_MODE=local-files mode when hosted database writes are not authorized. This refresh does not change personal environment settings.
• Autoreview sends captured diffs/source to a reviewer provider and deliberately has no pre-transmission secret scanner. Reviewer findings occur after sending. Run approved local secret checks first and authorize provider transmission. Its executable isolates reviewer credentials/settings; do not weaken isolation.
• Jevgrep and Last 30 Days use external providers and saved credentials. Do not ask for keys in chat, print them, or execute imported setup scripts during a refresh.
• Live collection installs and upstream CLI downloads remain a supply-chain risk. The snapshot above is not a blanket security approval for their current or future scripts. No upstream helper was executed for this refresh.

Validation on this refresh

• Full Bats suite with PyYAML and jsonschema: 589 passed, 1 failed. The failure, validate_config_with_schema fails when schema mismatches, also reproduces on the unchanged base. Its unquoted shell heredoc expands $schema to an empty key, so the fixture does not declare a schema. Leave that unrelated bug unchanged.
• Focused recommendation and Amp bundle tests: 75 passed.
• Bun tests: 48 passed; TypeScript typecheck passed.
• Babysit PR watcher tests: 5 passed.
• Shell syntax, token-efficiency sync, installer/export dry runs, mirror equality, and skill frontmatter checks passed. No real installation was run.
• Changed JSON passes Biome. Full repository Biome reports 427 errors and one informational diagnostic, identical to the unchanged base. No unrelated formatting changes are included.