## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - Paperclip agents can load repository and catalog skills, and Codex renders skill names and frontmatter descriptions into startup context. > - Long descriptions consume the fixed skill metadata budget before Codex can use the progressively disclosed skill bodies. > - The repo `.agents/skills` descriptions and a few shipped catalog descriptions had grown into operational documentation instead of short trigger metadata. > - This pull request keeps the strongest trigger language in frontmatter while leaving detailed procedures in each skill body. > - The benefit is lower prompt overhead, more reliable skill triggering, and a regression guard that prevents description drift from returning. ## Linked Issues or Issue Description No public GitHub issue found for this maintenance item. ### Pre-submission checklist - [x] I have searched existing open and closed issues and this is not a duplicate. - [x] I am working against `master`. - [x] I have confirmed the issue originates in Paperclip's shipped skill metadata, not in a local agent adapter or provider. ### What happened? Codex startup renders discovered skill names and frontmatter descriptions into a fixed skill metadata budget. Several repository skill descriptions and one shipped catalog description had grown into long-form operational guidance, which can force Codex to truncate descriptions before the model has enough trigger signal to select the right skill. ### Expected behavior Skill frontmatter descriptions should stay short trigger summaries: one capability sentence plus a “use when” clause. Detailed procedures should stay in the skill body and load only after the skill triggers. ### Steps to reproduce 1. Inspect `.agents/skills/*/SKILL.md` and `packages/skills-catalog/catalog/**/SKILL.md` frontmatter descriptions. 2. Measure folded YAML `description` values. 3. Observe descriptions above the intended short-trigger range, including descriptions above 300 characters. 4. Run the new shipped catalog test to verify future descriptions stay capped. ### Paperclip version or commit Reproduced on `master` at `cc81eefb6047d8eaf57faf785f421c03dc97073c`. ### Deployment mode Local dev / source checkout metadata inspection. This is not database-related. ### Installation method Built from source. ### Agent adapter(s) involved Codex, because Codex startup uses the skill metadata prompt budget. The metadata source itself is core repository/catalog content. ### Database mode Not database-related. ### Access context Not applicable; this is static repository metadata. ### Node.js version `v22.22.2` in the verification environment. ### Operating system Linux container environment. ### Relevant logs or output Final measurement after this PR: 29 source `SKILL.md` files, max description length 215 chars, 5,449 total description chars, estimated 1,363 description tokens at 4 chars/token. ### Relevant config None. ### Additional context The shipped catalog manifest was regenerated so the generated package metadata matches the edited catalog `SKILL.md` sources. ### Privacy checklist - [x] I have reviewed all pasted output for PII and redacted where necessary. ## What Changed - Shortened long `.agents/skills/*/SKILL.md` frontmatter descriptions to concise capability plus use-when trigger clauses. - Shortened the over-budget shipped skills catalog descriptions for wireframe, Paperclip capsules, and reflection coach. - Regenerated `packages/skills-catalog/generated/catalog.json` so shipped metadata matches source skill frontmatter. - Added a Vitest regression guard that caps repo skill source descriptions and generated catalog descriptions at 300 characters. ## Verification - `pnpm --filter @paperclipai/skills-catalog build:manifest` - `pnpm --filter @paperclipai/skills-catalog test` — 5 files passed, 19 tests passed - `pnpm --filter @paperclipai/skills-catalog typecheck` - Final measurement: 29 source `SKILL.md` files, max description length 215 chars, 5,449 total description chars, estimated 1,363 description tokens at 4 chars/token. Note: the clean PR worktree was created from `origin/master` and contains only this commit, but it does not have `node_modules`; running `pnpm --filter @paperclipai/skills-catalog test` there failed at tool/package resolution (`vitest`, `tsc`, `@paperclipai/shared`). The dependency-equipped workspace passed the commands above before the commit was cherry-picked onto the clean branch. ## Risks Low risk. This changes skill metadata and tests only. The main risk is over-trimming a useful trigger phrase, mitigated by keeping explicit “use when” clauses and leaving detailed guidance in the skill bodies. > For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and discuss it in `#dev` before opening the PR. Feature PRs that overlap with planned core work may need to be redirected — check the roadmap first. See `CONTRIBUTING.md`. ## Model Used OpenAI GPT-5 Codex coding agent via Paperclip/Codex, with shell and file-edit tool use. Exact API model ID and context window were not exposed in the runtime. ## Checklist - [x] I have included a thinking path that traces from project context to this change - [x] I have specified the model used (with version and capability details) - [x] I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work - [x] I have searched GitHub for duplicate or related PRs and linked them above - [x] I have either (a) linked existing issues with `Fixes: #` / `Closes #` / `Refs #` OR (b) described the issue in-PR following the relevant issue template - [x] I have not referenced internal/instance-local Paperclip issues or links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip` URLs) - [x] My branch name describes the change (e.g. `docs/...`, `fix/...`) and contains no internal Paperclip ticket id or instance-derived details - [x] I have run tests locally and they pass - [x] I have added or updated tests where applicable - [x] I have updated relevant documentation to reflect my changes - [x] I have considered and documented any risks above - [ ] All Paperclip CI gates are green - [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing> |
||
|---|---|---|
| .. | ||
| agents/wiki-maintainer | ||
| fixtures/basic-root | ||
| migrations | ||
| skills | ||
| src | ||
| templates | ||
| tests | ||
| .gitignore | ||
| README.md | ||
| esbuild.config.mjs | ||
| package.json | ||
| rollup.config.mjs | ||
| tsconfig.json | ||
| vitest.config.ts | ||
README.md
LLM Wiki
Local-file LLM Wiki plugin for source ingestion, wiki browsing, query, lint, and maintenance workflows.
Scope
This package is the standalone home for LLM Wiki behavior. Wiki-specific routes, UI, prompts, tools, local-folder templates, migrations, fixtures, and tests live here rather than in Paperclip core.
The alpha surface includes:
- manifest-declared Wiki page, sidebar entry, and settings page
- trusted local folder declaration for
raw/,wiki/,AGENTS.md,IDEA.md,wiki/index.md, andwiki/log.md - plugin database namespace migration for wiki instances, sources, pages, operations, query sessions, and resource bindings
- managed
Wiki Maintaineragent, managedLLM Wikiproject, and paused managed routines for wiki update processing, lint, and index refresh - plugin-operation issue creation using
surfaceVisibility: "plugin_operation" - local source capture into
raw/with metadata rows in the plugin DB namespace - opt-in company-scoped Paperclip event ingestion controls for issues, comments, and documents; event ingestion is disabled by default and routes captured raw provenance into the default space only
- manual Paperclip project/root issue distillation and bounded backfill actions with explicit work items, operation issues, source caps, and estimated cost recording
- Paperclip-derived distillation (cursor windows, manual
distill-now, backfill) always writes into the default wiki space in Phase 1; non-default spaces remain on manual / raw-file ingest until per-space Paperclip ingestion profiles ship - Paperclip-derived distillation maintains
wiki/projects/<slug>/standup.mdas the executive current-state view for each represented project, alongside durablewiki/projects/<slug>/index.mdknowledge pages - wiki page writes with plugin path validation, atomic local-folder writes, metadata/revision rows, backlink extraction, and optional stale-hash protection
- wiki tools for search/read/write/propose patch/source/log/index/backlinks workflows
Phase 5 Security Gate
Paperclip-derived text ingestion stays limited to issue titles/descriptions, issue comments, and issue documents.
- Issue attachments/assets are metadata-only in Phase 5.
- Issue work products are metadata-only in Phase 5.
- The wiki must not fetch
/api/assets/:id/content, dereference work-producturlfields, or store those capability-bearing links in source bundles/snapshots.
The accepted policy lives in doc/plans/2026-05-06-llm-wiki-paperclip-asset-security-gate.md.
Development
pnpm install
pnpm dev # watch builds
pnpm dev:ui # local dev server with hot-reload events
pnpm test
From the Paperclip repo root:
pnpm --filter @paperclipai/plugin-llm-wiki typecheck
pnpm --filter @paperclipai/plugin-llm-wiki test
pnpm --filter @paperclipai/plugin-llm-wiki build
Alpha Verification
Run these commands from the Paperclip repo root before handing off alpha plugin changes:
pnpm --filter @paperclipai/plugin-llm-wiki typecheck
pnpm --filter @paperclipai/plugin-llm-wiki test
pnpm --filter @paperclipai/plugin-llm-wiki build
The focused Vitest suite covers:
- standalone package boundaries and package-local harness dependencies
- required local folder bootstrap writes
- raw source capture plus ingest metadata persistence
- hidden plugin-operation issue creation for ingest/query/file-as-page workflows
- disabled and enabled Paperclip event ingestion paths
- managed routine declarations, manual distill/backfill work items, source cap handling, and backfill project/date scoping
- atomic page writes, metadata/revision rows, backlinks, and stale-hash refusal
- query session creation, run-id recording, stream event forwarding, and completion updates
- filing a streamed query answer back into the wiki through a hidden operation
Remaining alpha gaps:
- Browser screenshot capture is maintained separately under
tests/screenshots; generatedscreenshots/outputs are local artifacts and are ignored by git. - Host-level plugin install and live agent invocation still need Paperclip server/runtime smoke coverage when preparing a release candidate.
Install Into Paperclip
curl -X POST http://127.0.0.1:3100/api/plugins/install \
-H "Content-Type: application/json" \
-d '{"packageName":"/Users/dotta/paperclip/.paperclip/worktrees/PAP-3179-design-a-llm-wiki-plugin/packages/plugins/plugin-llm-wiki","isLocalPath":true}'
Build Options
pnpm builduses esbuild presets from@paperclipai/plugin-sdk/bundlers.pnpm build:rollupuses rollup presets from the same SDK.
After changing manifest-loaded assets such as skills, agent instructions, or templates, recompile the local plugin before re-enabling it:
pnpm --filter @paperclipai/plugin-llm-wiki build
The package-local dist/ directory is ignored by git, but local Paperclip
installs load the compiled dist/manifest.js and dist/worker.js files at
runtime. If activation failed before the rebuild, re-enable the plugin or
restart the Paperclip dev server so the host imports the fresh bundle.
Local File Layout
<configured-wiki-root>/
AGENTS.md
IDEA.md
.gitignore
raw/
.gitkeep
wiki/
index.md
log.md
sources/
.gitkeep
projects/
.gitkeep
<project-slug>/
index.md
standup.md
decisions.md
history.md
entities/
.gitkeep
concepts/
.gitkeep
synthesis/
.gitkeep
Use the settings page or bootstrap-root action to configure the folder and
write the starter files. The plugin uses Paperclip's local folder API for path
containment, symlink checks, read/write validation, and atomic writes.
Bootstrap preserves existing files rather than overwriting operator edits. The
default first-install skeleton is copied from the vanilla LLM Wiki layout, with
CLAUDE.md renamed to AGENTS.md and Paperclip project overviews, standups,
decisions, and history kept together under wiki/projects/<slug>/.
Managed Agent Instructions
Plugin-managed agent instruction bundles live under:
agents/<agent-key>/AGENTS.md
For this plugin the Wiki Maintainer source bundle is agents/wiki-maintainer/AGENTS.md.
Any additional files in that folder are installed as sibling instruction files
for the managed agent. The settings health check reports drift from these
defaults, and resetting the managed agent asks for confirmation before replacing
customized instructions.