[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M753Z3ENAGP06RM0DFMM54AM":3,"skill-same:01M753Z3ENAGP06RM0DFMM54AM":45},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M753Z3ENAGP06RM0DFMM54AM","toolkit-foundry",null,"Use when creating a new skill for jabworks\u002Fagentic-toolkit, registering it, or shipping the generated trees — \"scaffold a new skill\", \"register this skill\", \"sync dist\", \"bump the toolkit version\" — plus changes to the @jabworks\u002Fcondux npm package. Not for adapting a generic skill to the jabworks stack (adapting-skills), nor for reviewing a SKILL.md's content and trigger wording (toolkit-skill-standards).","jabworks\u002Fagentic-toolkit","jabworks","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F159094119?v=4","01M753P4JBZBTT7WN1PJBPZJZP","JAB Works","01M753P4K5KDGEWGXWT811BZET",2,"MIT","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Fskills\u002Ftoolkit-foundry\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: toolkit-foundry\ndescription: Use when creating a new skill for jabworks\u002Fagentic-toolkit, registering it, or shipping the generated trees — \"scaffold a new skill\", \"register this skill\", \"sync dist\", \"bump the toolkit version\" — plus changes to the @jabworks\u002Fcondux npm package. Not for adapting a generic skill to the jabworks stack (adapting-skills), nor for reviewing a SKILL.md's content and trigger wording (toolkit-skill-standards).\n---\n\n# Toolkit Foundry\n\n## Purpose\n\nCreate and maintain skills in `jabworks\u002Fagentic-toolkit`. Covers scaffolding, marketplace registration, generation of the dist trees, the npm package, and publishing.\n\n## When to Use\n\n- Adding a new skill to the toolkit\n- Updating an existing skill and regenerating the dist trees\n- Registering or modifying a marketplace entry\n- Changing `packages\u002Fcondux-opencode\u002F` (the `@jabworks\u002Fcondux` npm plugin)\n- Bumping the toolkit version\n\n## Skill Structure\n\nSource tree (editable):\n```\nskills\u002F\u003Cname>\u002F\n  SKILL.md              # Required — agent instruction file\n  README.md             # Optional — human docs\n  references\u002F           # Optional — templates, checklists\n    \u003Cfile>.md\n    package.json        # Only if references\u002F ships .js — see \"CommonJS scripts\"\n```\n\nOne source, four generated distributions — `scripts\u002Fsync.sh` produces all of them:\n\n| Generated tree | Channel | Layout |\n|---|---|---|\n| `dist\u002Fplugins\u002F\u003Cname>\u002Fskills\u002F\u003Cname>\u002F` | `\u002Fplugin install` (marketplace) | mirrors `skills\u002F\u003Cname>\u002F` byte-for-byte |\n| `dist\u002Fopencode\u002Fskills\u002F\u003Cname>\u002F` | OpenCode | same files, but `SKILL.md` has `when_to_use` folded into `description` |\n| `dist\u002Fcursor\u002Fskills\u002F\u003Cname>\u002F` | Cursor (2.4+) | the same fold, emitted to its own tree by `scripts\u002Fbuild-cursor.mjs` |\n| `packages\u002Fcondux-opencode\u002Fagents\u002F` | `@jabworks\u002Fcondux` on npm | translated from `skills\u002Fsubagent-execution\u002Fagents\u002F*.md` |\n| `packages\u002Fcondux-opencode\u002Fskills\u002F` | `@jabworks\u002Fcondux` on npm | the 13 condux skills, byte-identical to their `dist\u002Fopencode\u002Fskills\u002F` form |\n\n`npx skills add` is the remaining channel and needs no generation — it installs from\ntop-level `skills\u002F` directly and ignores everything above.\n\n**Every skills tree is flat — one dir per skill, no bundle nesting.** A\nbundle-member skill mirrors to `dist\u002Fplugins\u002F\u003Cbundle>\u002Fskills\u002F\u003Cname>\u002F`, the same\ndepth as a standalone plugin's own skill. This changed on 2026-08-14 with Agent\nPlugins conformance (docket #29): the spec discovers skills only as immediate\nchildren of `skills\u002F`, never recursively, so the old\n`dist\u002Fplugins\u002F\u003Cbundle>\u002Fskills\u002F\u003Cbundle>\u002F\u003Cname>\u002F` nesting hid every bundle member\nfrom spec-conformant clients. `tests\u002Fagent-plugins.test.mjs` now pins depth one.\nThe OpenCode and Cursor trees were always flat, and carry no manifests.\n\n**Never edit a generated tree directly.** Edit `skills\u002F`, then sync.\n\n### CommonJS scripts under `references\u002F`\n\nThe root `package.json` must NOT carry `\"type\": \"module\"` — it would reinterpret\nevery `references\u002F*.js` reference script as ESM and break it. Skills that ship\nsuch scripts also self-protect with their own `references\u002Fpackage.json`:\n\n```json\n{ \"type\": \"commonjs\" }\n```\n\n`discovery`, `plan-review`, and `spec-browser` carry one (`f1e4b53`). Add it to any\nnew skill whose `references\u002F` contains `.js`. CI syntax-checks every one with\n`node --check`.\n\n## Creating a New Skill\n\n### 1. Scaffold\n\n```bash\nmkdir -p skills\u002F\u003Cname>\u002Freferences\nmkdir -p dist\u002Fplugins\u002F\u003Cname>\u002F.codex-plugin\nmkdir -p dist\u002Fplugins\u002F\u003Cname>\u002F.claude-plugin\n```\n\nThe dist skill tree is not in that list on purpose — `sync.sh` creates it. It\nused to be listed, and creating it by hand was load-bearing: rsync cannot make\nnested parents, so a new plugin whose tree was missing failed the copy while\nsync still printed `0 failed` (docket #31).\n\n### 2. Write SKILL.md\n\nRequired frontmatter:\n```yaml\n---\nname: \u003Cname>\ndescription: Use when [triggering conditions only — no workflow summary]\n---\n```\n\n- `name`: kebab-case, letters\u002Fnumbers\u002Fhyphens only\n- `description`: third-person, max ~500 chars, starts with \"Use when...\"\n- Frontmatter total ≤ 1024 chars\n\n### 3. Write plugin manifests\n\nBoth agents read from the same plugin root (`dist\u002Fplugins\u002F\u003Cname>\u002F`) but look in different subdirectories.\n\n**Codex** — `dist\u002Fplugins\u002F\u003Cname>\u002F.codex-plugin\u002Fplugin.json`:\n\n```json\n{\n  \"name\": \"\u003Cname>\",\n  \"version\": \"1.0.0\",\n  \"description\": \"\u003Cshort description>\",\n  \"author\": { \"name\": \"jabworks\" },\n  \"repository\": \"https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\",\n  \"license\": \"MIT\",\n  \"keywords\": [\"keyword1\", \"keyword2\"],\n  \"skills\": \".\u002Fskills\u002F\u003Cname>\",\n  \"interface\": {\n    \"displayName\": \"\u003CDisplay Name>\",\n    \"shortDescription\": \"\u003Cone line>\",\n    \"longDescription\": \"\u003Ctwo sentences>\",\n    \"developerName\": \"jabworks\",\n    \"category\": \"Productivity\",\n    \"capabilities\": [\"Skills\"],\n    \"defaultPrompt\": [\"\u003Cprompt 1>\", \"\u003Cprompt 2>\"]\n  }\n}\n```\n\n**Claude Code** — `dist\u002Fplugins\u002F\u003Cname>\u002F.claude-plugin\u002Fplugin.json`: same structure **minus `interface`** — that field is codex-manifest-only (since 2026-07-29; `manifest-parity` fails if the Claude manifest carries it). `description` may carry the platform prefix (\"Claude Code skill for...\"), while `name`\u002F`version`\u002F`skills` must match the Codex manifest exactly (enforced by tests\u002Fmanifest-parity.test.mjs).\n\nThe `skills` path is relative to the plugin root and must start with `.\u002F`.\n\n### 4. Declare in composition.json\n\n`composition.json` (repo root) is the source of truth for what every plugin is\nmade of. Register the skill there — never hand-edit\n`.claude-plugin\u002Fmarketplace.json` or the catalog tables in README.md \u002F\nCLAUDE.md; sync generates all of them from the declaration\n(`scripts\u002Fgenerate-catalogs.mjs`).\n\n- **Standalone plugin:** add a `plugins.\u003Cname>` entry with\n  `marketplace.description` (the marketplace text — deliberately independent\n  of the SKILL.md description, a ratified divergence).\n- **Bundle member:** add the skill name to the bundle's `skills` array.\n- **Either way:** add a catalog row — `catalogs.readme-skills` (or the\n  bundle's block) and, for a new plugin, a `catalogs.claude-md-skills` row.\n  A plugin with no catalog row fails validation before anything is written.\n- A plugin-level dir loaded from the plugin root (like condux `agents\u002F` \u002F\n  `hooks\u002F` or docket `server\u002F`) is one `pluginDirs` entry — data, not a new\n  sync.sh case.\n\n### 5. Sync the generated trees\n\n```bash\nbash scripts\u002Fsync.sh \u003Cname>\n```\n\nThe script copies every pair `composition.json` declares for the skill —\nbundle membership, standalone target, and any plugin-level dirs — then\nregenerates `marketplace.json` and the doc catalog blocks. Run without\narguments to sync everything. A skill missing from the declaration is a hard\nerror, so step 4 must land first.\n\nEither way it then runs both variant builds. `node scripts\u002Fbuild-opencode.mjs`\nregenerates **all** of `dist\u002Fopencode\u002Fskills\u002F`, `packages\u002Fcondux-opencode\u002Fagents\u002F`,\nand `packages\u002Fcondux-opencode\u002Fskills\u002F` from scratch (the last two are `rm -rf`'d\nfirst); `node scripts\u002Fbuild-cursor.mjs` regenerates `dist\u002Fcursor\u002Fskills\u002F` the same\nway, importing the fold transform from the OpenCode script rather than duplicating\nit. One `sync.sh` covers every generated tree, so never invoke the pieces\nseparately expecting different results, and never hand-edit any of the four.\n\nA brand-new skill that step 4 has not declared fails the sync outright (the\nold behavior — a silent `SKIP` — is exactly the quiet failure the declaration\nreplaced). Neither variant build has such a gate — both pick up every dir under\n`skills\u002F` unconditionally.\n\n### 6. Verify\n\n```bash\nnode --test\n```\n\nRuns the invariant suite. What each file guards:\n\n| Test | Guards |\n|---|---|\n| `dist-mirror` | `dist\u002Fplugins\u002F` skill trees match `skills\u002F` byte-for-byte |\n| `opencode-dist` | `dist\u002Fopencode\u002Fskills\u002F` + `packages\u002Fcondux-opencode\u002Fagents\u002F` match the build script's output; merged descriptions within OpenCode's 1024-char cap; the plugin loads and never clobbers user-defined agents |\n| `cursor-dist` | `dist\u002Fcursor\u002Fskills\u002F` matches `build-cursor.mjs` output; no orphaned dirs; `when_to_use` never survives; frontmatter `name` == folder (Cursor requires it); merged descriptions ≤ 1024 chars |\n| `agent-plugins` | each plugin's root `plugin.json` matches the generator and stays inside the spec's closed schema; every skill sits at `skills\u002F` depth one; docket's spec `mcp.json` keeps the spec dialect |\n| `composition` | every pair `composition.json` declares mirrors byte-for-byte; nothing undeclared in any `dist\u002Fplugins\u002F` root; marketplace.json + doc catalog blocks match the generator |\n| `skill-invariants` | frontmatter budgets, kebab-case `name` == dir, marketplace\u002Fplugin paths resolve, plan-review no-egress |\n| `plugin-manifests` | `plugin.json` \u002F `marketplace.json` validity, `.\u002F`-prefixed paths |\n| `manifest-parity` | the `.claude-plugin`\u002F`.codex-plugin` pair agrees on name\u002Fversion\u002Fskills; `interface` codex-only (absent from the Claude manifest); `hooks` codex-only; trigger contract per skill |\n| `skill-routing-contracts` | mutually-guarded trigger pairs name each other in frontmatter |\n| `scaffold` \u002F `script-safety` \u002F `annotate-server` \u002F `browser-security` \u002F `session-report` | shipped reference scripts: syntax, no-egress, server behavior |\n| `concord-*` (4 files) | the concord plugin's store, rollout sync, paths, budget |\n| `local-hooks` | warn-only — tells you this clone lacks the pre-commit hook |\n\nCI runs the same command — a red suite here is a red build there.\n\nOptionally also run the official validator:\n\n```bash\nclaude plugin validate dist\u002Fplugins\u002F\u003Cname> --strict\n```\n\nUse `--strict` — expect a clean pass, no warnings. (`interface` moved to the Codex\nmanifest on 2026-07-29 precisely so `--strict` passes clean; see\n`toolkit-plugin-reference`.) Anything it reports is a real problem.\n\n### 7. Commit and push\n\n```bash\ngit add skills\u002F\u003Cname>\u002F dist\u002F packages\u002F composition.json .claude-plugin\u002Fmarketplace.json README.md CLAUDE.md\ngit commit -s -m \"feat: add \u003Cname> skill\"\ngit push origin main\n```\n\nStage `dist\u002F` as a whole, not `dist\u002Fplugins\u002F\u003Cname>\u002F` — the sync regenerated\n`dist\u002Fopencode\u002F` and `dist\u002Fcursor\u002F` too, and leaving either out ships a drifted\ntree that `opencode-dist.test.mjs` \u002F `cursor-dist.test.mjs` fails on.\n`packages\u002F` is usually clean for a brand-new\nskill (only agent-source edits dirty it), but staging it costs nothing and keeps\nthis command identical to the update flow below.\n\n## Updating an Existing Skill\n\n```bash\n# 1. Edit skills\u002F\u003Cname>\u002F files\n# 2. Sync (mirrors dist\u002Fplugins, regenerates dist\u002Fopencode + dist\u002Fcursor + package agents)\nbash scripts\u002Fsync.sh \u003Cname>\n# 3. Verify\nnode --test\n# 4. Commit\ngit add skills\u002F\u003Cname>\u002F dist\u002F packages\u002F\ngit commit -s -m \"fix: update \u003Cname> skill\"\ngit push origin main\n```\n\nEditing `skills\u002Fsubagent-execution\u002Fagents\u002F*.md` touches three generated\ndestinations — `dist\u002Fplugins\u002Fcondux\u002Fskills\u002F`, `dist\u002Fplugins\u002Fcondux\u002Fagents\u002F`, and\n`packages\u002Fcondux-opencode\u002Fagents\u002F`. One sync produces all three; stage all three.\n\nAn optional pre-commit hook automates the sync + dist staging — install it with\n`bash scripts\u002Finstall-hooks.sh`. It lives in `.git\u002Fhooks\u002F` (not version-controlled),\nso every fresh clone must install it again. **Never assume the hook is present.**\n\n## Changing the npm Package\n\n`packages\u002Fcondux-opencode\u002F` publishes `@jabworks\u002Fcondux` — the OpenCode plugin that\ninjects the condux specialist agents. Its `agents\u002F` dir is generated (edit\n`skills\u002Fsubagent-execution\u002Fagents\u002F` instead); `index.js` and `package.json` are\nhand-edited.\n\n**Condux changes go on a branch — never straight to `main`.** A push to `main` is\nwhat arms the release, so committing there does the review and the release\ntrigger in one irreversible step. Branch, PR, merge.\n\n```bash\n# 1. Branch first\ngit switch -c \u003Ctype>\u002Fcondux-\u003Cwhat>\n# 2. Edit packages\u002Fcondux-opencode\u002Findex.js (or the agent sources + sync)\n# 3. Describe the release\npnpm changeset          # pick @jabworks\u002Fcondux, pick the bump, write the summary\n# 4. Verify\nnode --test\n# 5. Commit the changeset alongside the change\ngit add packages\u002F .changeset\u002F\ngit commit -s -m \"fix(opencode): \u003Cwhat changed>\"\ngit push -u origin HEAD\ngh pr create --fill\n```\n\nMerging that PR into `main` triggers `.github\u002Fworkflows\u002Frelease.yml`: changesets\nopens (or updates) a **version PR** that consumes the `.changeset\u002F*.md` files and\nbumps `version`. Merging *that* second PR is what actually publishes to npm, via\nOIDC trusted publishing — there is no `NPM_TOKEN`. Two merges, not one.\n**Never hand-edit `version` in `packages\u002Fcondux-opencode\u002Fpackage.json`** —\nchangesets owns it.\n\nA change with no changeset still merges and still ships to the marketplace\nchannel — it just never reaches npm. Nothing warns you. If the change alters what\nthe published tarball contains, it needs a changeset.\n\nThat includes **editing `skills\u002Fsubagent-execution\u002Fagents\u002F*.md`**: those regenerate\n`packages\u002Fcondux-opencode\u002Fagents\u002F`, which is inside the package's `files` array, so\nan agent-wording tweak is a published change like any other.\n\nConstraints that are easy to break:\n\n- `index.js` must export the plugin **named only**. OpenCode calls every export as\n  a plugin function, so a default re-export registers the hooks twice (`08cc554`).\n- Keep generated `agents\u002F` in the package's `files` array, or the published\n  tarball ships a plugin with no agents.\n- Keep `publishConfig.provenance` set to `true`. It is what makes releases carry a\n  signed attestation tying the tarball to the workflow run and commit; the\n  workflow's `id-token: write` permission is necessary but not sufficient on its\n  own. `0.1.0` predates it and has none — verify with\n  `npm view @jabworks\u002Fcondux dist.attestations`.\n\n## Version Bump\n\nTwo independent version schemes — don't confuse them:\n\n| What | Where | How |\n|---|---|---|\n| Plugins (`condux`, `toolkit-ops`, standalone skills) | `version` in **both** paired plugin manifests, kept identical | hand-edited |\n| `@jabworks\u002Fcondux` | `packages\u002Fcondux-opencode\u002Fpackage.json` | changesets only — never by hand |\n\n`marketplace.json` entries carry no version field, and neither variant skill tree\n(`dist\u002Fopencode\u002F`, `dist\u002Fcursor\u002F`) has a version at all — both refresh with the repo.\n\nFor plugins, short form: breaking install path = major, new skill\u002Fcapability =\nminor, anything else that ships = patch. The full bump policy — and the\ncache-refresh rationale for bumping on ANY shipped change — has one home:\n`toolkit-change-control`.\n\n## Marketplace Entry Fields\n\nThe step-4 template above is the complete field set — the schema of record for\nmarketplace entries and both plugin.json variants is `toolkit-plugin-reference`.\nAdd nothing from upstream examples (`strict` \u002F `skills` \u002F `keywords` are not used\nin this repo's marketplace.json).\n\n## Component Types\n\n| Component | Location | When to add |\n|---|---|---|\n| Skill instructions | `skills\u002F\u003Cname>\u002FSKILL.md` | Always |\n| Human docs | `skills\u002F\u003Cname>\u002FREADME.md` | Non-trivial setup |\n| Templates \u002F checklists | `skills\u002F\u003Cname>\u002Freferences\u002F*.md` | Agent needs reusable scaffolds |\n| Reference scripts | `skills\u002F\u003Cname>\u002Freferences\u002F*.js` + `references\u002Fpackage.json` | Skill ships runnable helpers (CommonJS) |\n| Trigger evals | `skills\u002F\u003Cname>\u002Fevals\u002Ftrigger_eval.json` | Skill needs a measured routing baseline |\n\n### Writing trigger-eval oracles\n\nThe oracle is what the corpus asserts is correct. Get it wrong and the harness\nscores a *correct* routing decision as a miss — and since each run is compared\nagainst the last, one wrong oracle poisons every later comparison.\n\ncondux's routing payload states that every implementation request starts at\n`\u002Fworkflow`, and names the skills that execute **within** workflow rather than\ninstead of it (`discovery`, `draft-plan`, `blueprint`, `test-first-development`,\n`preflight`, `finalize`, `code-review`, `root-cause-analysis`,\n`live-verification`, `plan-review`, `technical-spec`, `subagent-*`). For any of\nthose, `workflow` is a legitimate first answer whenever the prompt asks to build,\nfix, or change something.\n\nEncode that by the **shape of the request**, never by which cases happened to\nflake — loosening only the observed failures overfits to model variance and has\nto be re-chased every run:\n\n| Request shape | Oracle |\n|---|---|\n| implementation request aimed at a within-workflow skill | `accept: [\"workflow\"]` |\n| bug-shaped | also `accept` `root-cause-analysis` |\n| question about the practice itself, or ownership of an artifact only that skill owns | strict — no `accept` |\n\nKeep some cases strict on purpose. A corpus where every case accepts `workflow`\nmeasures nothing.\n\nTwo failure signatures, two different fixes:\n\n- **Missing to another skill that doctrine also allows** → too-strict oracle.\n  Widen `accept`. Changes no behaviour, so no re-run is needed.\n- **Missing to `null`** → a reach gap in the description; nothing fired at all.\n  That is a real routing defect, and fixing it *changes behaviour*, so it needs\n  an eval re-run to verify — and watch the neighbouring skills' scores, not just\n  the ones you meant to fix.\n- **An answer the judge was never shown** → out-of-catalog harness\n  contamination. No `accept` alternate can fix one; leave it.\n\n## Common Mistakes\n\n| Mistake | Fix |\n|---|---|\n| Editing `dist\u002F` directly | Edit `skills\u002F`, then `bash scripts\u002Fsync.sh \u003Cname>` |\n| Editing `packages\u002Fcondux-opencode\u002Fagents\u002F` | Generated too — edit `skills\u002Fsubagent-execution\u002Fagents\u002F`, then sync |\n| Forgetting to sync after edits | Run `bash scripts\u002Fsync.sh` — the pre-commit hook does it only if installed (`bash scripts\u002Finstall-hooks.sh`); never assume it's present |\n| Staging only `dist\u002Fplugins\u002F` after a sync | Stage `dist\u002F` and `packages\u002F` — the same sync regenerated the OpenCode tree, the Cursor tree, and the package agents |\n| Pointing a Cursor install at top-level `skills\u002F` | Use `dist\u002Fcursor\u002Fskills\u002F` — Cursor never reads `when_to_use`, so raw source skills load with a thin trigger and fail silently |\n| Committing a condux change straight to `main` | Branch first — a push to `main` arms the release, so main-first collapses review and release into one irreversible step |\n| Shipping a `packages\u002F` change with no changeset | Run `pnpm changeset` and commit the `.changeset\u002F*.md` with the change, or nothing publishes |\n| Assuming an agent-wording tweak needs no changeset | It regenerates `packages\u002Fcondux-opencode\u002Fagents\u002F`, which is published — it does |\n| Hand-bumping `@jabworks\u002Fcondux` version | Changesets owns that field — a manual bump collides with the version PR |\n| Adding `\"type\": \"module\"` to the root `package.json` | Breaks every CommonJS reference script — the package is ESM-scoped at `packages\u002F`, not repo-wide |\n| `references\u002F*.js` with no `references\u002Fpackage.json` | Add `{ \"type\": \"commonjs\" }` beside it |\n| Co-author trailer in commit | Use `-s` (`--signoff`) — no `Co-Authored-By:` |\n| Workflow summary in description | Description = triggering conditions only |\n| `expected_skill: null` on an implementation request | Doctrine routes it to `workflow`, so the model answers correctly and the corpus records a miss — use `accept: [\"workflow\"]` |\n| Widening a description to fix a reach gap, without re-running the eval | A description edit changes real routing; only a re-run shows what it stole from neighbouring skills |\n| Leaving `[FILL]` placeholders in SKILL.md | Run quality check before committing |\n| Missing `.codex-plugin\u002Fplugin.json` | `codex plugin add` will fail with \"missing plugin.json\" |\n| Missing `.claude-plugin\u002Fplugin.json` | Claude Code `\u002Fplugin install` won't find plugin metadata |\n| Wrong `skills` path in plugin.json | Must be `\".\u002Fskills\u002F\u003Cplugin-dir-name>\"` starting with `.\u002F` |\n| Editing only one of the two plugin manifests | Edit the pair together — name\u002Fversion\u002Fskills must match (tests\u002Fmanifest-parity.test.mjs) |\n| Skipping `node --test` before commit | CI runs the same suite and will fail the build — run it locally first |\n","# Toolkit Foundry\n\n## Purpose\n\nCreate and maintain skills in `jabworks\u002Fagentic-toolkit`. Covers scaffolding, marketplace registration, generation of the dist trees, the npm package, and publishing.\n\n## When to Use\n\n- Adding a new skill to the toolkit\n- Updating an existing skill and regenerating the dist trees\n- Registering or modifying a marketplace entry\n- Changing `packages\u002Fcondux-opencode\u002F` (the `@jabworks\u002Fcondux` npm plugin)\n- Bumping the toolkit version\n\n## Skill Structure\n\nSource tree (editable):\n```\nskills\u002F\u003Cname>\u002F\n  SKILL.md              # Required — agent instruction file\n  README.md             # Optional — human docs\n  references\u002F           # Optional — templates, checklists\n    \u003Cfile>.md\n    package.json        # Only if references\u002F ships .js — see \"CommonJS scripts\"\n```\n\nOne source, four generated distributions — `scripts\u002Fsync.sh` produces all of them:\n\n| Generated tree | Channel | Layout |\n|---|---|---|\n| `dist\u002Fplugins\u002F\u003Cname>\u002Fskills\u002F\u003Cname>\u002F` | `\u002Fplugin install` (marketplace) | mirrors `skills\u002F\u003Cname>\u002F` byte-for-byte |\n| `dist\u002Fopencode\u002Fskills\u002F\u003Cname>\u002F` | OpenCode | same files, but `SKILL.md` has `when_to_use` folded into `description` |\n| `dist\u002Fcursor\u002Fskills\u002F\u003Cname>\u002F` | Cursor (2.4+) | the same fold, emitted to its own tree by `scripts\u002Fbuild-cursor.mjs` |\n| `packages\u002Fcondux-opencode\u002Fagents\u002F` | `@jabworks\u002Fcondux` on npm | translated from `skills\u002Fsubagent-execution\u002Fagents\u002F*.md` |\n| `packages\u002Fcondux-opencode\u002Fskills\u002F` | `@jabworks\u002Fcondux` on npm | the 13 condux skills, byte-identical to their `dist\u002Fopencode\u002Fskills\u002F` form |\n\n`npx skills add` is the remaining channel and needs no generation — it installs from\ntop-level `skills\u002F` directly and ignores everything above.\n\n**Every skills tree is flat — one dir per skill, no bundle nesting.** A\nbundle-member skill mirrors to `dist\u002Fplugins\u002F\u003Cbundle>\u002Fskills\u002F\u003Cname>\u002F`, the same\ndepth as a standalone plugin's own skill. This changed on 2026-08-14 with Agent\nPlugins conformance (docket #29): the spec discovers skills only as immediate\nchildren of `skills\u002F`, never recursively, so the old\n`dist\u002Fplugins\u002F\u003Cbundle>\u002Fskills\u002F\u003Cbundle>\u002F\u003Cname>\u002F` nesting hid every bundle member\nfrom spec-conformant clients. `tests\u002Fagent-plugins.test.mjs` now pins depth one.\nThe OpenCode and Cursor trees were always flat, and carry no manifests.\n\n**Never edit a generated tree directly.** Edit `skills\u002F`, then sync.\n\n### CommonJS scripts under `references\u002F`\n\nThe root `package.json` must NOT carry `\"type\": \"module\"` — it would reinterpret\nevery `references\u002F*.js` reference script as ESM and break it. Skills that ship\nsuch scripts also self-protect with their own `references\u002Fpackage.json`:\n\n```json\n{ \"type\": \"commonjs\" }\n```\n\n`discovery`, `plan-review`, and `spec-browser` carry one (`f1e4b53`). Add it to any\nnew skill whose `references\u002F` contains `.js`. CI syntax-checks every one with\n`node --check`.\n\n## Creating a New Skill\n\n### 1. Scaffold\n\n```bash\nmkdir -p skills\u002F\u003Cname>\u002Freferences\nmkdir -p dist\u002Fplugins\u002F\u003Cname>\u002F.codex-plugin\nmkdir -p dist\u002Fplugins\u002F\u003Cname>\u002F.claude-plugin\n```\n\nThe dist skill tree is not in that list on purpose — `sync.sh` creates it. It\nused to be listed, and creating it by hand was load-bearing: rsync cannot make\nnested parents, so a new plugin whose tree was missing failed the copy while\nsync still printed `0 failed` (docket #31).\n\n### 2. Write SKILL.md\n\nRequired frontmatter:\n```yaml\n---\nname: \u003Cname>\ndescription: Use when [triggering conditions only — no workflow summary]\n---\n```\n\n- `name`: kebab-case, letters\u002Fnumbers\u002Fhyphens only\n- `description`: third-person, max ~500 chars, starts with \"Use when...\"\n- Frontmatter total ≤ 1024 chars\n\n### 3. Write plugin manifests\n\nBoth agents read from the same plugin root (`dist\u002Fplugins\u002F\u003Cname>\u002F`) but look in different subdirectories.\n\n**Codex** — `dist\u002Fplugins\u002F\u003Cname>\u002F.codex-plugin\u002Fplugin.json`:\n\n```json\n{\n  \"name\": \"\u003Cname>\",\n  \"version\": \"1.0.0\",\n  \"description\": \"\u003Cshort description>\",\n  \"author\": { \"name\": \"jabworks\" },\n  \"repository\": \"https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\",\n  \"license\": \"MIT\",\n  \"keywords\": [\"keyword1\", \"keyword2\"],\n  \"skills\": \".\u002Fskills\u002F\u003Cname>\",\n  \"interface\": {\n    \"displayName\": \"\u003CDisplay Name>\",\n    \"shortDescription\": \"\u003Cone line>\",\n    \"longDescription\": \"\u003Ctwo sentences>\",\n    \"developerName\": \"jabworks\",\n    \"category\": \"Productivity\",\n    \"capabilities\": [\"Skills\"],\n    \"defaultPrompt\": [\"\u003Cprompt 1>\", \"\u003Cprompt 2>\"]\n  }\n}\n```\n\n**Claude Code** — `dist\u002Fplugins\u002F\u003Cname>\u002F.claude-plugin\u002Fplugin.json`: same structure **minus `interface`** — that field is codex-manifest-only (since 2026-07-29; `manifest-parity` fails if the Claude manifest carries it). `description` may carry the platform prefix (\"Claude Code skill for...\"), while `name`\u002F`version`\u002F`skills` must match the Codex manifest exactly (enforced by tests\u002Fmanifest-parity.test.mjs).\n\nThe `skills` path is relative to the plugin root and must start with `.\u002F`.\n\n### 4. Declare in composition.json\n\n`composition.json` (repo root) is the source of truth for what every plugin is\nmade of. Register the skill there — never hand-edit\n`.claude-plugin\u002Fmarketplace.json` or the catalog tables in README.md \u002F\nCLAUDE.md; sync generates all of them from the declaration\n(`scripts\u002Fgenerate-catalogs.mjs`).\n\n- **Standalone plugin:** add a `plugins.\u003Cname>` entry with\n  `marketplace.description` (the marketplace text — deliberately independent\n  of the SKILL.md description, a ratified divergence).\n- **Bundle member:** add the skill name to the bundle's `skills` array.\n- **Either way:** add a catalog row — `catalogs.readme-skills` (or the\n  bundle's block) and, for a new plugin, a `catalogs.claude-md-skills` row.\n  A plugin with no catalog row fails validation before anything is written.\n- A plugin-level dir loaded from the plugin root (like condux `agents\u002F` \u002F\n  `hooks\u002F` or docket `server\u002F`) is one `pluginDirs` entry — data, not a new\n  sync.sh case.\n\n### 5. Sync the generated trees\n\n```bash\nbash scripts\u002Fsync.sh \u003Cname>\n```\n\nThe script copies every pair `composition.json` declares for the skill —\nbundle membership, standalone target, and any plugin-level dirs — then\nregenerates `marketplace.json` and the doc catalog blocks. Run without\narguments to sync everything. A skill missing from the declaration is a hard\nerror, so step 4 must land first.\n\nEither way it then runs both variant builds. `node scripts\u002Fbuild-opencode.mjs`\nregenerates **all** of `dist\u002Fopencode\u002Fskills\u002F`, `packages\u002Fcondux-opencode\u002Fagents\u002F`,\nand `packages\u002Fcondux-opencode\u002Fskills\u002F` from scratch (the last two are `rm -rf`'d\nfirst); `node scripts\u002Fbuild-cursor.mjs` regenerates `dist\u002Fcursor\u002Fskills\u002F` the same\nway, importing the fold transform from the OpenCode script rather than duplicating\nit. One `sync.sh` covers every generated tree, so never invoke the pieces\nseparately expecting different results, and never hand-edit any of the four.\n\nA brand-new skill that step 4 has not declared fails the sync outright (the\nold behavior — a silent `SKIP` — is exactly the quiet failure the declaration\nreplaced). Neither variant build has such a gate — both pick up every dir under\n`skills\u002F` unconditionally.\n\n### 6. Verify\n\n```bash\nnode --test\n```\n\nRuns the invariant suite. What each file guards:\n\n| Test | Guards |\n|---|---|\n| `dist-mirror` | `dist\u002Fplugins\u002F` skill trees match `skills\u002F` byte-for-byte |\n| `opencode-dist` | `dist\u002Fopencode\u002Fskills\u002F` + `packages\u002Fcondux-opencode\u002Fagents\u002F` match the build script's output; merged descriptions within OpenCode's 1024-char cap; the plugin loads and never clobbers user-defined agents |\n| `cursor-dist` | `dist\u002Fcursor\u002Fskills\u002F` matches `build-cursor.mjs` output; no orphaned dirs; `when_to_use` never survives; frontmatter `name` == folder (Cursor requires it); merged descriptions ≤ 1024 chars |\n| `agent-plugins` | each plugin's root `plugin.json` matches the generator and stays inside the spec's closed schema; every skill sits at `skills\u002F` depth one; docket's spec `mcp.json` keeps the spec dialect |\n| `composition` | every pair `composition.json` declares mirrors byte-for-byte; nothing undeclared in any `dist\u002Fplugins\u002F` root; marketplace.json + doc catalog blocks match the generator |\n| `skill-invariants` | frontmatter budgets, kebab-case `name` == dir, marketplace\u002Fplugin paths resolve, plan-review no-egress |\n| `plugin-manifests` | `plugin.json` \u002F `marketplace.json` validity, `.\u002F`-prefixed paths |\n| `manifest-parity` | the `.claude-plugin`\u002F`.codex-plugin` pair agrees on name\u002Fversion\u002Fskills; `interface` codex-only (absent from the Claude manifest); `hooks` codex-only; trigger contract per skill |\n| `skill-routing-contracts` | mutually-guarded trigger pairs name each other in frontmatter |\n| `scaffold` \u002F `script-safety` \u002F `annotate-server` \u002F `browser-security` \u002F `session-report` | shipped reference scripts: syntax, no-egress, server behavior |\n| `concord-*` (4 files) | the concord plugin's store, rollout sync, paths, budget |\n| `local-hooks` | warn-only — tells you this clone lacks the pre-commit hook |\n\nCI runs the same command — a red suite here is a red build there.\n\nOptionally also run the official validator:\n\n```bash\nclaude plugin validate dist\u002Fplugins\u002F\u003Cname> --strict\n```\n\nUse `--strict` — expect a clean pass, no warnings. (`interface` moved to the Codex\nmanifest on 2026-07-29 precisely so `--strict` passes clean; see\n`toolkit-plugin-reference`.) Anything it reports is a real problem.\n\n### 7. Commit and push\n\n```bash\ngit add skills\u002F\u003Cname>\u002F dist\u002F packages\u002F composition.json .claude-plugin\u002Fmarketplace.json README.md CLAUDE.md\ngit commit -s -m \"feat: add \u003Cname> skill\"\ngit push origin main\n```\n\nStage `dist\u002F` as a whole, not `dist\u002Fplugins\u002F\u003Cname>\u002F` — the sync regenerated\n`dist\u002Fopencode\u002F` and `dist\u002Fcursor\u002F` too, and leaving either out ships a drifted\ntree that `opencode-dist.test.mjs` \u002F `cursor-dist.test.mjs` fails on.\n`packages\u002F` is usually clean for a brand-new\nskill (only agent-source edits dirty it), but staging it costs nothing and keeps\nthis command identical to the update flow below.\n\n## Updating an Existing Skill\n\n```bash\n# 1. Edit skills\u002F\u003Cname>\u002F files\n# 2. Sync (mirrors dist\u002Fplugins, regenerates dist\u002Fopencode + dist\u002Fcursor + package agents)\nbash scripts\u002Fsync.sh \u003Cname>\n# 3. Verify\nnode --test\n# 4. Commit\ngit add skills\u002F\u003Cname>\u002F dist\u002F packages\u002F\ngit commit -s -m \"fix: update \u003Cname> skill\"\ngit push origin main\n```\n\nEditing `skills\u002Fsubagent-execution\u002Fagents\u002F*.md` touches three generated\ndestinations — `dist\u002Fplugins\u002Fcondux\u002Fskills\u002F`, `dist\u002Fplugins\u002Fcondux\u002Fagents\u002F`, and\n`packages\u002Fcondux-opencode\u002Fagents\u002F`. One sync produces all three; stage all three.\n\nAn optional pre-commit hook automates the sync + dist staging — install it with\n`bash scripts\u002Finstall-hooks.sh`. It lives in `.git\u002Fhooks\u002F` (not version-controlled),\nso every fresh clone must install it again. **Never assume the hook is present.**\n\n## Changing the npm Package\n\n`packages\u002Fcondux-opencode\u002F` publishes `@jabworks\u002Fcondux` — the OpenCode plugin that\ninjects the condux specialist agents. Its `agents\u002F` dir is generated (edit\n`skills\u002Fsubagent-execution\u002Fagents\u002F` instead); `index.js` and `package.json` are\nhand-edited.\n\n**Condux changes go on a branch — never straight to `main`.** A push to `main` is\nwhat arms the release, so committing there does the review and the release\ntrigger in one irreversible step. Branch, PR, merge.\n\n```bash\n# 1. Branch first\ngit switch -c \u003Ctype>\u002Fcondux-\u003Cwhat>\n# 2. Edit packages\u002Fcondux-opencode\u002Findex.js (or the agent sources + sync)\n# 3. Describe the release\npnpm changeset          # pick @jabworks\u002Fcondux, pick the bump, write the summary\n# 4. Verify\nnode --test\n# 5. Commit the changeset alongside the change\ngit add packages\u002F .changeset\u002F\ngit commit -s -m \"fix(opencode): \u003Cwhat changed>\"\ngit push -u origin HEAD\ngh pr create --fill\n```\n\nMerging that PR into `main` triggers `.github\u002Fworkflows\u002Frelease.yml`: changesets\nopens (or updates) a **version PR** that consumes the `.changeset\u002F*.md` files and\nbumps `version`. Merging *that* second PR is what actually publishes to npm, via\nOIDC trusted publishing — there is no `NPM_TOKEN`. Two merges, not one.\n**Never hand-edit `version` in `packages\u002Fcondux-opencode\u002Fpackage.json`** —\nchangesets owns it.\n\nA change with no changeset still merges and still ships to the marketplace\nchannel — it just never reaches npm. Nothing warns you. If the change alters what\nthe published tarball contains, it needs a changeset.\n\nThat includes **editing `skills\u002Fsubagent-execution\u002Fagents\u002F*.md`**: those regenerate\n`packages\u002Fcondux-opencode\u002Fagents\u002F`, which is inside the package's `files` array, so\nan agent-wording tweak is a published change like any other.\n\nConstraints that are easy to break:\n\n- `index.js` must export the plugin **named only**. OpenCode calls every export as\n  a plugin function, so a default re-export registers the hooks twice (`08cc554`).\n- Keep generated `agents\u002F` in the package's `files` array, or the published\n  tarball ships a plugin with no agents.\n- Keep `publishConfig.provenance` set to `true`. It is what makes releases carry a\n  signed attestation tying the tarball to the workflow run and commit; the\n  workflow's `id-token: write` permission is necessary but not sufficient on its\n  own. `0.1.0` predates it and has none — verify with\n  `npm view @jabworks\u002Fcondux dist.attestations`.\n\n## Version Bump\n\nTwo independent version schemes — don't confuse them:\n\n| What | Where | How |\n|---|---|---|\n| Plugins (`condux`, `toolkit-ops`, standalone skills) | `version` in **both** paired plugin manifests, kept identical | hand-edited |\n| `@jabworks\u002Fcondux` | `packages\u002Fcondux-opencode\u002Fpackage.json` | changesets only — never by hand |\n\n`marketplace.json` entries carry no version field, and neither variant skill tree\n(`dist\u002Fopencode\u002F`, `dist\u002Fcursor\u002F`) has a version at all — both refresh with the repo.\n\nFor plugins, short form: breaking install path = major, new skill\u002Fcapability =\nminor, anything else that ships = patch. The full bump policy — and the\ncache-refresh rationale for bumping on ANY shipped change — has one home:\n`toolkit-change-control`.\n\n## Marketplace Entry Fields\n\nThe step-4 template above is the complete field set — the schema of record for\nmarketplace entries and both plugin.json variants is `toolkit-plugin-reference`.\nAdd nothing from upstream examples (`strict` \u002F `skills` \u002F `keywords` are not used\nin this repo's marketplace.json).\n\n## Component Types\n\n| Component | Location | When to add |\n|---|---|---|\n| Skill instructions | `skills\u002F\u003Cname>\u002FSKILL.md` | Always |\n| Human docs | `skills\u002F\u003Cname>\u002FREADME.md` | Non-trivial setup |\n| Templates \u002F checklists | `skills\u002F\u003Cname>\u002Freferences\u002F*.md` | Agent needs reusable scaffolds |\n| Reference scripts | `skills\u002F\u003Cname>\u002Freferences\u002F*.js` + `references\u002Fpackage.json` | Skill ships runnable helpers (CommonJS) |\n| Trigger evals | `skills\u002F\u003Cname>\u002Fevals\u002Ftrigger_eval.json` | Skill needs a measured routing baseline |\n\n### Writing trigger-eval oracles\n\nThe oracle is what the corpus asserts is correct. Get it wrong and the harness\nscores a *correct* routing decision as a miss — and since each run is compared\nagainst the last, one wrong oracle poisons every later comparison.\n\ncondux's routing payload states that every implementation request starts at\n`\u002Fworkflow`, and names the skills that execute **within** workflow rather than\ninstead of it (`discovery`, `draft-plan`, `blueprint`, `test-first-development`,\n`preflight`, `finalize`, `code-review`, `root-cause-analysis`,\n`live-verification`, `plan-review`, `technical-spec`, `subagent-*`). For any of\nthose, `workflow` is a legitimate first answer whenever the prompt asks to build,\nfix, or change something.\n\nEncode that by the **shape of the request**, never by which cases happened to\nflake — loosening only the observed failures overfits to model variance and has\nto be re-chased every run:\n\n| Request shape | Oracle |\n|---|---|\n| implementation request aimed at a within-workflow skill | `accept: [\"workflow\"]` |\n| bug-shaped | also `accept` `root-cause-analysis` |\n| question about the practice itself, or ownership of an artifact only that skill owns | strict — no `accept` |\n\nKeep some cases strict on purpose. A corpus where every case accepts `workflow`\nmeasures nothing.\n\nTwo failure signatures, two different fixes:\n\n- **Missing to another skill that doctrine also allows** → too-strict oracle.\n  Widen `accept`. Changes no behaviour, so no re-run is needed.\n- **Missing to `null`** → a reach gap in the description; nothing fired at all.\n  That is a real routing defect, and fixing it *changes behaviour*, so it needs\n  an eval re-run to verify — and watch the neighbouring skills' scores, not just\n  the ones you meant to fix.\n- **An answer the judge was never shown** → out-of-catalog harness\n  contamination. No `accept` alternate can fix one; leave it.\n\n## Common Mistakes\n\n| Mistake | Fix |\n|---|---|\n| Editing `dist\u002F` directly | Edit `skills\u002F`, then `bash scripts\u002Fsync.sh \u003Cname>` |\n| Editing `packages\u002Fcondux-opencode\u002Fagents\u002F` | Generated too — edit `skills\u002Fsubagent-execution\u002Fagents\u002F`, then sync |\n| Forgetting to sync after edits | Run `bash scripts\u002Fsync.sh` — the pre-commit hook does it only if installed (`bash scripts\u002Finstall-hooks.sh`); never assume it's present |\n| Staging only `dist\u002Fplugins\u002F` after a sync | Stage `dist\u002F` and `packages\u002F` — the same sync regenerated the OpenCode tree, the Cursor tree, and the package agents |\n| Pointing a Cursor install at top-level `skills\u002F` | Use `dist\u002Fcursor\u002Fskills\u002F` — Cursor never reads `when_to_use`, so raw source skills load with a thin trigger and fail silently |\n| Committing a condux change straight to `main` | Branch first — a push to `main` arms the release, so main-first collapses review and release into one irreversible step |\n| Shipping a `packages\u002F` change with no changeset | Run `pnpm changeset` and commit the `.changeset\u002F*.md` with the change, or nothing publishes |\n| Assuming an agent-wording tweak needs no changeset | It regenerates `packages\u002Fcondux-opencode\u002Fagents\u002F`, which is published — it does |\n| Hand-bumping `@jabworks\u002Fcondux` version | Changesets owns that field — a manual bump collides with the version PR |\n| Adding `\"type\": \"module\"` to the root `package.json` | Breaks every CommonJS reference script — the package is ESM-scoped at `packages\u002F`, not repo-wide |\n| `references\u002F*.js` with no `references\u002Fpackage.json` | Add `{ \"type\": \"commonjs\" }` beside it |\n| Co-author trailer in commit | Use `-s` (`--signoff`) — no `Co-Authored-By:` |\n| Workflow summary in description | Description = triggering conditions only |\n| `expected_skill: null` on an implementation request | Doctrine routes it to `workflow`, so the model answers correctly and the corpus records a miss — use `accept: [\"workflow\"]` |\n| Widening a description to fix a reach gap, without re-running the eval | A description edit changes real routing; only a re-run shows what it stole from neighbouring skills |\n| Leaving `[FILL]` placeholders in SKILL.md | Run quality check before committing |\n| Missing `.codex-plugin\u002Fplugin.json` | `codex plugin add` will fail with \"missing plugin.json\" |\n| Missing `.claude-plugin\u002Fplugin.json` | Claude Code `\u002Fplugin install` won't find plugin metadata |\n| Wrong `skills` path in plugin.json | Must be `\".\u002Fskills\u002F\u003Cplugin-dir-name>\"` starting with `.\u002F` |\n| Editing only one of the two plugin manifests | Edit the pair together — name\u002Fversion\u002Fskills must match (tests\u002Fmanifest-parity.test.mjs) |\n| Skipping `node --test` before commit | CI runs the same suite and will fail the build — run it locally first |\n",{"name":5,"description":7},"main","5277649b071c1c0cdf4f15eee861b1608fc1bbcb","skills\u002Ftoolkit-foundry",0,"https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit","https:\u002F\u002Fgithub.com\u002Fjabworks","2026-11-03 15:03:51",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add jabworks\u002Fagentic-toolkit --skill toolkit-foundry",{"tool":37,"command":38},"bunx","bunx skills add jabworks\u002Fagentic-toolkit --skill toolkit-foundry",{"tool":40,"command":41},"pnpm","pnpm dlx skills add jabworks\u002Fagentic-toolkit --skill toolkit-foundry","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Fskills\u002Ftoolkit-foundry\u002FSKILL.md",true,[],[46,52,58,65,72,79],{"skillNo":47,"name":48,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":49,"categories":50,"keywords":51,"publishedAt":19},"01M7540BPNTFZH7V41P5E2S692","orphan","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcomposition\u002Fbad\u002Fskills\u002Forphan\u002FSKILL.md",[],[],{"skillNo":53,"name":54,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":55,"categories":56,"keywords":57,"publishedAt":19},"01M7540AG2KFTQ8DJMNV6806MV","no-blurb","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcomposition\u002Fbad\u002Fskills\u002Fno-blurb\u002FSKILL.md",[],[],{"skillNo":59,"name":60,"nameZh":6,"description":61,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":62,"categories":63,"keywords":64,"publishedAt":19},"01M75404N306PY1JFPQ0GGNTYB","workflow","Routes any dev task into the right execution tier (Small\u002FMedium\u002FLarge) and carries condux's operating rules — proportional effort, lazy skill loading, soft gates, implement-yourself-by-default.","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Fskills\u002Fworkflow\u002FSKILL.md",[],[],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M753ZZ6TTA06GRJAG679BQ2H","usage-trace","Checks and manages durable, local usage capture — ingest, health status, and the opt-in limit snapshot — for Claude Code and Codex. Capture itself runs automatically at session start.","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Fskills\u002Fusage-trace\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M753ZSKJSF8WFS7AXR2N65SE","toolkit-skill-standards","Use when writing or reviewing a SKILL.md for jabworks\u002Fagentic-toolkit — frontmatter budgets, the trigger contract (\"Use when…\" description or when_to_use field), progressive disclosure into references\u002F, and keeping a new description out of existing skills' trigger space. Triggers include \"review this skill description\", \"write the frontmatter\", \"does this collide with an existing skill\", \"where should this content live\". Not for scaffold\u002Fsync mechanics; use toolkit-foundry.","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Fskills\u002Ftoolkit-skill-standards\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M753ZKX3KTVYWX15A9A0WZN0","toolkit-research-frontier","Use when planning improvement work on jabworks\u002Fagentic-toolkit itself — open problems (trigger-collision automation, YAML-strict frontmatter checks, docs-catalog enforcement), existing assets, the next three concrete steps, and the four-front library-health campaign with commands and success criteria. Triggers include \"let's improve our toolkit\", \"let's improve this repo\", \"what should we improve next\", \"toolkit roadmap\", \"open problems\", \"run the health campaign\".","https:\u002F\u002Fgithub.com\u002Fjabworks\u002Fagentic-toolkit\u002Fblob\u002Fmain\u002Fskills\u002Ftoolkit-research-frontier\u002FSKILL.md",[],[]]