AGENTS.md · git:20261019.913c756 · 2026-10-19 · sha256 ada9141bb8cd9961
AGENTS.md git:20261019.913c756A
Immutable. This exact content never changes and is served at /api/v1/blob/ada9141bb8cd9961.
# AGENTS.md This repository contains shared skills for GitHub Copilot, Claude Code, and Codex, and this file defines the workspace-specific rules that should apply to any AI agent operating in `C:\Users\LOQ\.copilot\skills`. ## Required Session Start Rule - Every new session in this workspace must begin by reading `LESSON.md`. - Treat `LESSON.md` as required startup context before analysis, planning, edits, validation, reviews, or advisory work. - If `LESSON.md` is missing or unreadable, stop and report that blocker before continuing. ## Required Completion, Sync, And Publish Rule - For every user-requested mutation task in this workspace, complete the requested work in `C:\Users\LOQ\.copilot\skills` first. - After the work is complete, run the repo validation, then sync outward to the downstream skill folders every time. - If the AI agent judges the result satisfactory, commit and push to GitHub without asking for another confirmation. - Treat work as satisfactory only when validation passes, sync completes, the task is complete, no requested step was skipped, no required command was rejected, no unresolved secret/security/privacy issue remains, and the final diff matches the user's request. - Elevate to the user before commit or push when there are security concerns, incomplete work, skipped steps, rejected or blocked required commands, validation/sync failures, unexpected unrelated dirty files that make staging unsafe, or any other reason the work is not satisfactory. - For read-only or advisory tasks with no file changes, do not create empty sync, commit, or push churn; report that no mutation workflow was needed. ## Workspace Role - Main branch: `C:\Users\LOQ\.copilot\skills` - New maintained skills must be added or imported here first. - Maintained skills live here and are synced outward to downstream targets. - Copied official superpowers are tracked here for discovery and cross-client sync, but they are not maintained the same way as the catalog's maintained skills. ## Codex-Only Blender Skills Overlay - The `arjun988/blender-skills` pack is an explicit exception to normal child promotion. - Its 94 upstream skills plus the separately protected local `raw-scan-to-aaa-preserve-texture` entry (95 protected names total) must remain installed only under `C:\Users\LOQ\.codex\skills`, with its source checkout under `C:\Users\LOQ\.codex\vendor\blender-skills`. - Never promote these skill names into this parent catalog and never sync them to `C:\Users\LOQ\.agents\skills` or `C:\Users\LOQ\.claude\skills`. - `scripts/skill-registry.json` records the protected names and the Codex-only source configuration; generic promotion and sync tooling must honor that boundary. - During parent source-maintenance or "update all skills" work, run `scripts/update-codex-local-blender-skills.ps1`. It fetches upstream, refreshes only the owned Codex copies and shared Blender references, updates the ownership manifest and source commit, and verifies that no Blender skill escaped to a forbidden root. ## Current Counts Snapshot date: `2026-10-19`. Local overlay totals can differ by machine. - Git-tracked catalog in this repository: - `253` tracked skill folders - `221` tracked maintained skills - `32` tracked copied official Superpowers - Live local workspace snapshot, including local-only overlays such as `gws-*` and `recipe-*` when present: - `311` local skill folders detected - `279` local maintained skills detected - `32` local copied official Superpowers detected Copied official superpowers are identified by the explicit `copied_official_superpowers` list in `scripts/skill-registry.json`, not by whether a skill folder has a `CHANGELOG.md`. All `253` tracked skills use catalog `version: "2.0"`. The `166` pre-existing tracked skills retain their prior catalog baselines; the 66 platform skills retain their import provenance, and five Codex Router skills were promoted from the personal Codex root, and one standalone Codex plugin scanner was vendored after review. Three official Playwright workflows are maintained from pinned Microsoft sources, and five scientific workflows were added. The 37 refreshed and five newly imported entrypoints use `last_updated: 2026-10-19`; unchanged entries retain their previous dates. The Playwright entries use `last_updated: 2026-09-12`, and the imported `humanizer` entry uses `last_updated: 2026-09-14`. The `58` local-only Google Workspace overlays keep their upstream `version: "0.22.5"` while receiving the same retained-client sections and maintenance date. The tracked imports `docx`, `jupyter-notebook`, `pptx`, and `xlsx` now have finalized canonical provenance. Child-path promotion is handled by `scripts/promote-child-skills.py`, while `scripts/update-skill-registry.py` refreshes provenance and the reference-source report. The Tavily suite is sourced from `tavily-ai/skills` at commit `ea5e8201b0d3ed9c10b70b71187589bd761fe2d2`. Keep its eight skills self-contained, retain the local secret and prompt-injection safeguards, and do not reintroduce removed-client integrations from upstream references. The selected Matt Pocock skills are sourced from `mattpocock/skills` at the current audited commit recorded in `scripts/skill-registry.json`. Seven current upstream paths remain mapped; `resolving-merge-conflicts` is retained as a catalog-local snapshot because the current upstream removed that path. Existing TDD, debugging, review, implementation, planning, and skill-authoring equivalents remain canonical. The compact `humanizer` workflow is sourced from `blader/humanizer` v3.0.0 at revision `9862685f575c65a8247f90369951df1b3416e3d6`. It is a direct, facts-preserving 25-pattern rewrite workflow; route detection, file edits, preservation checks, and iterative convergence to the broader `avoid-ai-writing` suite. The Playwright package refresh is pinned in `scripts/skill-registry.json`: `playwright-cli` comes from `@playwright/cli@0.1.22` at the tagged `microsoft/playwright-cli` revision, while `playwright-component-testing` and `playwright-trace` come from stable `microsoft/playwright@v1.64.1`. The old broad `playwright` folder is retired; its wrapper and practical catalog guides are retained under `playwright-cli`. The 2026-08-14 source refresh audited current upstream heads and updated the mapped `avoid-ai-writing`, Stitch, Xquik, and Matt Pocock domain-modeling workflows, plus the affected copied Superpowers workflows. Exact-path audits left unchanged mapped skills untouched, and imported support material was reviewed for removed-client paths, credential handling, and no-MCP fallbacks. The 2026-08-16 child reconciliation compared the eleven newly installed `.codex` skill trees byte-for-byte with their official upstream paths and promoted them into the parent: Supabase, Gemini API, Vercel React performance, and the web-quality audit router plus five focused leaves. The web-quality leaves remain distinct; the analyzer documents a Windows/manual fallback, and Supabase/Gemini MCP configuration remains opt-in and authorization-gated. `frontend-design` is the only general frontend creation and art-direction skill. The retired names `frontend-skill` and `premium-frontend-ui` redirect conceptually to `frontend-design`; `web-design-reviewer` remains the separate post-implementation visual QA skill. Keep framework, Figma, and Stitch skills separate because they own narrower implementation or tool-specific workflows. The 2026-08-16 related-skill consolidation audit found no safe content merges. `supabase` and `supabase-postgres-best-practices`, the Gemini API/Live/Omni workflows, and the React/web-quality groups remain separate with explicit routing links because their activation boundaries and evidence paths differ. Plugin-managed Supabase and React copies remain external; the parent catalog is canonical for maintained cross-client content. ## 2026-10-19 Catalog Refresh And Scientific Workflows - Audited 30 pinned source repositories and refreshed only exact mapped paths: 37 skill entrypoints and 54 supporting-file mappings changed across the current source set. Fifteen source heads moved; unrelated path movement was recorded as provenance rather than imported broadly. - Added five scientific workflows from K-Dense's `scientific-agent-skills` at commit `593377bb048752abb0152deab9fbb1083a6087df`: experimental design, hypothesis generation, paper lookup, scientific writing, and peer review. The `literature-review` and `citation-management` candidates were excluded as standalone imports after overlap review against existing research, scholarly-retrieval, and manuscript-writing workflows. - Re-audited only the personal `.codex`, `.agents`, and `.claude` roots; no eligible child-only skills remained. Protected Codex `.system`, Blender, copied Superpowers, and project-specific paths remained outside promotion. - The current Playwright pins are `@playwright/cli@0.1.22` and stable `microsoft/playwright@v1.64.1`. Both mapped Playwright workflow paths were unchanged at the newer stable release. - Xquik now prefers its current MCP server and retains a direct REST fallback with explicit write authorization and credential-destination safeguards. - No new Codex plugin skill was imported after reviewing the current plugin-creator package; its create, update, and submission workflows are narrower than the global catalog's remit. The existing `plugin-creator` snapshot is tracked as a local import, not a Codex system-managed skill. - The Gemini API skill may document Google's currently published Antigravity managed agent, while catalog portability still excludes retired CLI or standalone-client integrations. ## 2026-08-20 VoltAgent Platform Import The VoltAgent repository is a discovery index; canonical vendor repositories were pinned before import. The 66-skill selection is recorded in `scripts/platform_skill_manifest.py`: 8 Vercel, 15 Netlify, 7 MongoDB, the existing current 2-skill Supabase import, 12 Figma, and 24 non-CLI Hugging Face skills. Source commits and paths are regenerated into `REFERENCE_SOURCES.md`. CLI gating was based on commands present on this laptop: `vercel`, `netlify`, and `supabase` were present; `hf`, `huggingface-cli`, `mongosh`, `mongo`, and `figma` were absent. Therefore Vercel, Netlify, and existing Supabase CLI guidance is included, while no Hugging Face, MongoDB, or Figma CLI skill was installed. The importer is repeatable with `python scripts/import-platform-skills.py --source-root <pinned-clones>`. ## 2026-09-14 Humanizer Import And Catalog Routing - Audited the current `blader/humanizer` main revision `9862685f575c65a8247f90369951df1b3416e3d6` (v3.0.0) and passed its upstream package validator before importing the root skill. - Retained only the runtime-relevant `SKILL.md`, OpenAI-compatible `agents/openai.yaml`, MIT license, and attribution notice. Upstream Claude marketplace metadata, CI files, README duplication, and packaging validator were intentionally omitted; this catalog's validator is authoritative. - Kept `humanizer` separate from `avoid-ai-writing` because the former is a compact direct-rewrite prompt while the latter owns detection, preservation, edit-in-place, and iterative routing workflows. - Re-audited only the three approved personal child roots. The scan found 914 skill files: 819 already represented in the parent, 101 excluded as protected/system/Superpowers material, and zero eligible child-only skills; project-specific paths were not scanned. - No additional Codex plugin skill was vendored in this pass after review for cross-client usefulness and bloat. ## 2026-09-14 Mapped Source Refresh - Checked all `27` recorded upstream heads and refreshed exactly `13` mapped entrypoints whose exact source paths changed: Avoid AI Writing and its detector/router/preservation leaves, NeMo Retriever, MongoDB Search and AI, four Figma workflows, and three Hugging Face cloud workflows. - Updated provenance pins for the remaining `68` mapped skills without rewriting unchanged paths. Avoid AI Writing retains the current runtime CLI/gate and focused local regression helpers; upstream marketplace, CI, package-publishing metadata, and the large evaluation corpus remain omitted to avoid bloat. Its final live pin is `aa4da8b255eb9821f0dae2a059762de900bb5d1f` (v3.35.0); the Awesome Copilot head moved outside mapped paths and is recorded provenance-only. ## 2026-09-12 Playwright CLI Skills And Canonical Consolidation - Installed and verified the latest global `@playwright/cli@0.1.19` package on Node 22, then ran the official `playwright-cli install --skills=agents --global` and `--skills=claude --global` installers. - Imported the package's `playwright-cli` workflow and the stable Playwright component-testing and trace workflows into the parent first. Component testing retains the stable templates and `references/typing.md`; its `mount` fixture correctly requires project-local `@playwright/test`. - Kept the three Playwright activation boundaries separate and retired the old broad `playwright` folder. The sync script prunes only that exact known name from approved mirrors and preserves unknown personal folders. - Reviewed available Claude marketplace and Codex plugin Playwright candidates; none met the cross-client usefulness-without-bloat bar, so no additional plugin skill was vendored. ## 2026-09-08 Catalog Refresh, CodeGraph Index, And Mirror Preparation - Rechecked all 24 recorded upstream heads and refreshed only the mapped paths that changed: nine Awesome Copilot workflows and Avoid AI Writing v3.33.2 with its six focused leaves. - Added the current Avoid AI Writing `references/patterns.md` catalog, detector updates, and quote-normalization helpers while preserving the flattened-parent router validator/test adaptations. Removed seven support files that the current upstream no longer ships. - Re-audited only personal `.codex`, `.agents`, and `.claude` roots. No eligible child-only skills remained after the protected `.system`, Blender, Superpowers, and project-path exclusions. No new plugin skill met the catalog's usefulness-without-bloat bar beyond the existing `agent-skillguard` import. - Installed a local CodeGraph index in `.codegraph`; commit only its generated `.gitignore`, keep `codegraph.db` ignored, and use `codegraph explore` before broad text searches for repository-structure questions. CodeGraph is local navigation evidence, not a substitute for validation or mirror sync. - Modernized all `303` live folders and regenerated provenance at the `2026-09-08` baseline. ## 2026-09-05 Catalog Freshness, Plugin Review, And Mirror Preparation - Rechecked all 24 recorded upstream heads and refreshed only exact mapped paths that changed: Avoid AI Writing v3.29 (including its six-skill detector/router suite), both NVIDIA DeepStream workflows, seven Tavily workflows, Gemini API/Live/Omni, and the affected Figma support trees. Head movement outside mapped paths was recorded in provenance without broad imports. - Retired `gemini-interactions-api` after confirming its migration guidance is carried by `gemini-api-dev`; the sync retirement list removes only that exact catalog-owned folder from approved mirrors. - Re-audited only `.codex`, `.agents`, and `.claude`. No eligible child-only skills remained after the protected Blender/local-only, Codex `.system`, copied Superpowers, and project-path exclusions. Reviewed Codex plugin skills and vendored only `agent-skillguard`, a self-contained read-only scanner whose missing upstream `tools/verify_rule_corpus.py` helper is documented rather than invented. - The current inventory is `245` tracked folders (`213` maintained and `32` copied Superpowers) and `303` live folders (`271` maintained and `58` local-only overlays). Validation, helper checks, Blender refresh, and the three-root sync remain required before publication. - Nineteen CRLF shell helpers were normalized to LF after WSL parsing exposed portability failures; the logic was not changed. ## 2026-08-31 Catalog Freshness And Corpus Refresh - Rechecked all 24 recorded source heads against live remotes. Only `avoid-ai-writing` had a new installed-path change: its corpus manifest now records small documentation and conversational pre-LLM seeds, and its selector-aware extraction helper and tests were refreshed. - `awesome-copilot` and NVIDIA moved only outside the installed mappings, so their provenance pins were updated without broad content rewrites. - Re-audited `.codex`, `.agents`, and `.claude`; no eligible child-only skills remained. The protected Blender/local-only set, Codex `.system`, Superpowers, and project-specific paths were not promoted or overwritten. ## 2026-08-29 Catalog Refresh And Mirror Repair - Re-audited all 24 recorded source heads against live remotes. Refreshed the exact mapped paths for `avoid-ai-writing` 3.28.0, `x-twitter-scraper`, both Gemini workflows, `react-view-transitions`, and the web-quality support trees. Head movement outside installed paths was recorded without broad rewrites. - Updated the catalog to the current 69-category detector, rendered-Markdown masking, Gemini transcription/Omni/embedding guidance, current React view transition troubleshooting, and field/agentic web-quality references. - The current Xquik source removed its MCP setup surface, so the registry now uses the honest no-MCP fallback instead of advertising an obsolete preferred server. - Re-audited only `.codex`, `.agents`, and `.claude` skill roots. No eligible child-only skills remained. The approved sync restored a missing top-level Codex `doc` copy without touching `.system`, Blender, Superpowers, or any project-specific path. - The required Blender refresh remains at commit `8f778d2405a214b508d4c7d80742be8e43acdd52`: 94 upstream skills plus one separately protected local entry. ## 2026-08-24 Catalog Refresh And Child Promotion - Compared every recorded upstream source head with its exact mapped skill path. Refreshed material changes in `avoid-ai-writing`, the eight selected Matt Pocock workflows, and `x-twitter-scraper`; unrelated source head movement was recorded without rewriting unchanged mapped paths. - Promoted five eligible personal-Codex child skills: `codex-app-threads`, `codex-computer-use`, `codex-in-app-browser`, `codex-router`, and `codex-router-media`. Their host marker files remain outside the parent; package and tree-digest provenance is recorded in the registry. - Child reconciliation scanned only `.codex`, `.agents`, and `.claude` skill roots. It excluded Codex `.system`, the 94-skill Blender overlay plus the separately protected local entry, copied official Superpowers, and all project-specific `C:\Assumption University` paths. No additional eligible skills remained in `.agents` or `.claude`. - The required Blender refresh completed at upstream commit `8f778d2405a214b508d4c7d80742be8e43acdd52` with 94 upstream skills plus one separately protected local entry and no promotion to the parent, shared, or Claude roots. ## Source Of Truth - Edit maintained skill content in `C:\Users\LOQ\.copilot\skills` first. - Treat `SKILL.md` as the source of truth for skill behavior and wording. - Treat downstream skill roots as synced mirrors or deployment targets, not as the place to author maintained content. ## Downstream Sync Targets The only approved downstream sync destinations are these three personal-global roots: - `C:\Users\LOQ\.agents\skills` - `C:\Users\LOQ\.codex\skills` - `C:\Users\LOQ\.claude\skills` There must be no downstream sync to any other path. The sync script enforces this allowlist and refuses to write anywhere else. Per-target routing: - Maintained skills sync to `C:\Users\LOQ\.codex\skills`, `C:\Users\LOQ\.agents\skills`, and `C:\Users\LOQ\.claude\skills`. - Copied official superpowers sync only to the `superpowers` subfolder of the shared mirror: `C:\Users\LOQ\.agents\skills\superpowers` (inside the approved `.agents\skills` root, not a separate destination). - Skills listed in `codex_system_managed_skills` are excluded from top-level Codex mirror writes because Codex owns their `.system` copies. Their normalized parent copies still deploy to the shared and Claude roots. - Sync removes known catalog-owned top-level shadows that violate those routes, including copied Superpowers outside the shared `superpowers` subfolder. It must preserve unknown personal skills and Codex `.system`. - Sync also prunes the exact retired catalog folders `frontend-skill` and `premium-frontend-ui` from the three approved roots after their content and provenance have been consolidated into `frontend-design`. Host-provided or plugin-managed skills that are not part of this maintained catalog should stay external unless you intentionally vendor them into this repo. ## Upstream-Only Skill Sources Normal child discovery is limited to the personal `.codex`, `.agents`, and `.claude` skill roots. Project-local roots under paths such as `C:\Assumption University` must not be scanned or used as sync destinations unless a later user request explicitly puts them in scope. When a child or categorized skill root contains a skill absent from this parent, audit its activation boundary and provenance, then promote it upstream with `scripts/promote-child-skills.py`. Flatten nested folders by the normalized lowercase hyphen-case skill name; do not copy invalid underscore or title-style names into the parent unchanged. Then refresh provenance with `scripts/update-skill-registry.py`. The only downstream sync call is to the three approved personal-global roots: ```powershell powershell -ExecutionPolicy Bypass -File .\scripts\sync-skills.ps1 ``` ## CodeGraph Repository Index When `.codegraph/` exists, use the local CodeGraph index before broad `grep`, `find`, or file reads for code-structure questions. It exposes symbols and call paths for the maintenance scripts without replacing direct inspection or execution evidence. - Commit only `.codegraph/.gitignore`; the generated `codegraph.db` is ignored, machine-local, and must never be synced to a downstream skill root. - Check freshness with `codegraph status`, query with `codegraph explore "<repository question>"`, and refresh after source edits with `codegraph sync .`. - Initialize only this repository with `codegraph init -y .`; never initialize project-specific `C:\Assumption University` paths during catalog maintenance. - Set `CODEGRAPH_TELEMETRY=0` when a no-telemetry local run is required. CodeGraph output is navigation evidence. The validator, real helper tests, source-pin checks, and downstream mirror audit remain the completion gates. ## Skill Catalog Expectations Every maintained skill folder in this catalog should have: - `SKILL.md` - `CHANGELOG.md` Recommended support folders: - `references/` - `scripts/` Optional: - `examples/` - `LICENSE.txt` Every `SKILL.md` in this repo should: - use valid YAML frontmatter - keep the `name` aligned with the folder name - include the catalog frontmatter fields: `name`, `version`, `last_updated`, `tags`, and `description` - use only approved extra top-level metadata fields when needed: `license`, `compatibility`, and `metadata` - use activation-focused descriptions - include the generated portability section - include the MCP or no-MCP fallback section - include `## Anti-Patterns` - include `## Verification Protocol` - end with `## Related Skills` ## MCP Rules When editing MCP-aware skills: 1. Name the preferred MCP server explicitly. 2. Add a practical fallback path for environments without that MCP surface. 3. Avoid claiming a host-specific tool wrapper exists unless you verified it. 4. Prefer local scripts, CLIs, or browser workflows as the fallback evidence path. The MCP mapping source lives in `scripts/skill-registry.json`. ## Validation Workflow After meaningful changes: 1. Run `python scripts/validate-skills.py`. 2. Sync outward if the repo is in a good state. For a catalog-wide documentation refresh, treat validation and sync as required even when the inventory counts stay the same. The validator now expects the catalog frontmatter fields plus the portability, MCP, Anti-Patterns, Related Skills, and `CHANGELOG.md` baseline. Catalog policy also expects each `SKILL.md` to include `## Verification Protocol` immediately after `## Anti-Patterns`. Changelog entries should use `Added`, `Changed`, and `Fixed` sections only; the validator rejects both `### Tested` and `### Verified` headings. The tracked imports `docx`, `jupyter-notebook`, `pptx`, and `xlsx` now validate against the shared catalog structure and have finalized provenance metadata. After adding a new maintained skill: 1. Install or import it into this repo first. 2. Prefer the canonical upstream source when a discovery list points to a stronger maintained original. 3. Update `REFERENCE_SOURCES.md` and `scripts/skill-registry.json` when the source was external. 4. Smoke-test any bundled helper scripts or local fallback workflow. 5. Update root docs and the relevant changelogs. 6. For source maintenance, run `scripts/update-codex-local-blender-skills.ps1` as well. 7. Then sync it to the downstream targets. ## Documentation Rules When repo behavior, counts, sync flow, portability, supported clients, or workspace startup rules change: - update `README.md` - update `AGENTS.md` - update `CHANGELOG.md` - update `CLAUDE.md` - update `LESSON.md` - update `MIGRATION.md` when a breaking client or sync boundary changes ## Agent-Specific Notes - GitHub Copilot should treat this file and the workspace root docs as the portable instruction source when folder-based skill discovery is limited. - Claude Code should follow this file alongside `CLAUDE.md`, with the narrower local instruction taking precedence if they differ. - Claude Code sessions using the GLM Coding Plan endpoint must not assume Anthropic's native Chrome integration or Codex-only tools are available. Use only active, healthy external MCP/browser tools and preserve manual fallbacks. - Codex should treat `C:\Users\LOQ\.codex\skills` as the primary Codex install root and `C:\Users\LOQ\.agents\skills` as a shared mirror for cross-client reuse. - Codex system-managed skill folders remain authoritative in `C:\Users\LOQ\.codex\skills\.system`; never overwrite them through the top-level mirror sync. ## Related Repo Files - `README.md`: catalog and maintenance commands - `CHANGELOG.md`: repo-wide change history - `CLAUDE.md`: Claude Code usage guidance - `CONTRIBUTING.md`: contribution workflow and repo validation expectations - `LESSON.md`: maintenance lessons and gotchas - `MIGRATION.md`: version 2.0 breaking migration and rollback guidance - `SECURITY.md`: vulnerability reporting and sensitive-disclosure guidance