llmdoc

llmdoc

Website · 简体中文

Persistent engineering context that helps coding agents understand a repository without rediscovering its architecture every session.

• Preserve decisions, constraints, and cross-module contracts that source code does not explain cheaply.
• Retrieve only the context a task needs, then verify exact facts against the live repository.
• Recheck knowledge semantically as code evolves instead of accumulating stale implementation notes.

Start in 60 seconds

You need Node.js 18 or newer and a Git repository. Choose the path that matches your repository and agent.

New repository with Claude Code

Add the marketplace and install the plugin:

/plugin marketplace add TokenRollAI/llmdoc
/plugin install llmdoc@llmdoc-plugin

If the install summary says Run /reload-plugins to activate., run that command. If the reload warns about rereading the conversation, rerun it as /reload-plugins --force. Once the plugin is active, initialize the repository:

/llmdoc:init

New repository with Codex

Add the marketplace and start Codex from the repository:

codex plugin marketplace add TokenRollAI/llmdoc
codex

Inside Codex, run /plugins, open the llmdoc-plugin marketplace, and install llmdoc. Review the plugin and its hooks before enabling them. Then start a new Codex session in the repository and ask:

Use the llmdoc:init skill to initialize this repository.

Repository that already has llmdoc/

Open its knowledge map immediately—no plugin is required for direct CLI use:

npx -y @tokenroll/llmdoc tree
npx -y @tokenroll/llmdoc search "revision"

@tokenroll/llmdoc is external tooling. Do not add it to the consumer project's package.json or lockfile, and never use the unrelated bare package name npx llmdoc. For reproducible runs, pin the package spec: npx -y @tokenroll/llmdoc@<version> <command>. The package exposes the llmdoc bin; the scoped npx form keeps it outside the consumer repository.

How it works

flowchart LR
    K[Stable knowledge in llmdoc/] --> R[Task-scoped retrieval]
    R --> S[Source verification]
    S --> U[Semantic maintenance]
    U --> K

llmdoc/ stores durable engineering meaning—not a copy of the repository. An agent first retrieves the smallest useful knowledge set, uses source and tests for current facts, and later verifies affected knowledge. A code change creates a review obligation; it does not automatically create a documentation rewrite.

Two operating layers

• Agent workflows own judgment and safe closeout. Invoke them through the host's command or skill interface.
• Runtime CLI owns retrieval and deterministic mechanics. Call it with the scoped npx command, directly or from a workflow.

The workflows are not four equivalent CLI commands:

• init creates a small, high-value V3 knowledge surface when none exists.
• update semantically verifies affected knowledge; unchanged documents can be recorded as verified without inventing prose changes.
• prune reduces duplicate, fragmented, or cheaply reconstructable knowledge; the CLI only supplies a read-only report.
• upgrade migrates legacy/V2 knowledge. It runs only when the user explicitly asks for it and must never be suggested or folded into another workflow.

Every explicit workflow reports exactly one result state: success, no_change, dry_run, incomplete, or failed.

Daily use

Apply this routing gate before broad exploration and again when entering a new subsystem:

• Concept, contract, term, or “where is X?” → search <query>
• Context for concrete source files → context --files <path...>
• Cold start or unclear scope → tree
• Known topic or kind → index --topic <topic> / index --kind <kind>
• Bodies already identified → show <path...>

These are alternatives, not a fixed sequence. Once llmdoc narrows the working set, use native tools for exact source text, line numbers, test behavior, counts, and Git state.

context --files evaluates every input independently and reports unmappedFiles; a non-empty impacted set never hides unmapped siblings.

Plugin lifecycle hooks invoke the same scoped CLI through an npm package alias. This prevents a same-name local or file: dependency without a built bin from shadowing the hook runtime; normal interactive commands keep the shorter form.

# Map the knowledge surface
npx -y @tokenroll/llmdoc tree --docs

# Find relevant knowledge
npx -y @tokenroll/llmdoc search "revision" --limit 5
npx -y @tokenroll/llmdoc context --files cli/src/cli.ts

# Read only the selected bodies
npx -y @tokenroll/llmdoc show architecture.mdx cli-runtime/retrieval-and-mutation.mdx

# Browse the knowledge surface locally
npx -y @tokenroll/llmdoc serve

Use npx -y @tokenroll/llmdoc --help or npx -y @tokenroll/llmdoc help <command> for the complete, current CLI reference. status and delta assess validity and impact; they are not retrieval steps.

Startup context configuration

Repositories that use the lifecycle hooks can add an optional llmdoc.config.json at the llmdoc workspace root. In a Git repository, this is the nearest Git root that owns llmdoc/:

{
  "$schema": "https://llmdoc.tokenroll.ai/schemas/config.schema.json",
  "schema": "llmdoc.config/v1",
  "startup": {
    "remindSkill": true,
    "preload": [
      "architecture.mdx",
      "plugin-packaging/claude-and-codex.mdx"
    ]
  }
}

• remindSkill controls the SessionStart operating guidance: load the llmdoc skill, use the CLI retrieval gate, and delegate to the llmdoc roles. It defaults to true; set it to false to opt out.
• preload lists exact document IDs, with an optional llmdoc/ prefix. Cold SessionStart inserts their bodies directly in the listed order with no llmdoc character or token budget. A final completion marker confirms that the host supplied the full preload; without it, retrieve only the missing body with show.
• Compact re-entry lists configured document IDs but does not inject their full bodies again. Continue from LLMDOC_STATE and retrieve a body only if needed.
• validate reports malformed config, missing paths, and path escapes. Normalized aliases of the same document are deduplicated with a warning. Lifecycle hooks stay fail-open: unreadable schema/JSON uses the default reminder, while a valid remindSkill choice survives preload-only errors.
• mv rewrites matching preload paths transactionally. prune --report lists preload references that must be updated before a manual merge or deletion.

Without this file, SessionStart emits its state plus the default operating guidance and does not preload documents.

All fixed CLI interface text is English, including help, diagnostics, hook messages, and the local Viewer. Chinese queries and repository document content remain fully supported and are returned unchanged.

Knowledge and safety boundaries

• Stable knowledge belongs in tracked llmdoc/; investigations, caches, and reflection candidates belong in local .llmdoc-tmp/.
• V3 documents are pure Markdown .mdx with YAML front matter and optional <CodeRef> anchors. The path is the document ID, and kind lives in front matter rather than directory names.
• The tree contains root singleton documents and one level of topic folders. It has no index.mdx topic nodes and no nested topic folders.
• llmdoc/meta.json is a Git-revision validity ledger, not documentation. Dirty worktree state is an additional signal, not a second truth system. Never hand-edit the ledger; use the CLI's guarded mutation and commit operations.
• Within agent workflows, investigator gathers temporary evidence, reflector captures privacy-safe lesson candidates, and recorder is the only role that writes tracked knowledge.
• Every workflow authorizes knowledge maintenance only, not source-code edits. Structural writes are validated and confined to the repository's llmdoc/ boundary.
• Hooks emit read-only, fail-open signals through the scoped CLI. SessionStart also injects the configurable operating guidance, so projects do not need to repeat it in CLAUDE.md/AGENTS.md. Review hooks and trust the plugin source before enabling them.
• A delta match means “review this claim,” not “rewrite this document.” Preserve decisions, rationale, boundaries, invariants, contracts, and non-obvious failure semantics; leave reconstructable facts in source, schemas, help, tests, or generated configuration.

Platform integration

Claude Code

The repository-root Claude plugin is the canonical authored surface. It provides the operating skill, four explicit workflows, three roles, and lifecycle hooks. Use the installation flow above, or manage it from Claude Code's plugin UI. See the Claude Code plugin documentation for current installation behavior.

Codex

The Codex plugin is generated from the Claude surface and exposes equivalent skills, roles, and hooks. You can also run /plugins in Codex CLI to browse the Plugins Directory. Start a new session after installation, and inspect third-party hooks before trusting them. Codex IDE extensions do not currently support plugins. See the official Codex plugin documentation for current installation behavior.

Other agents

Agents without a native plugin system can use the same runtime and operating contract. Copy the portable AGENTS.md integration recipe into the consumer repository.

Develop this repository

The repository root is a private development workspace; the public consumer artifact is the @tokenroll/llmdoc CLI. Claude skills and agents are canonical, while the Codex surface is generated—do not hand-edit generated packaging.

npm install
npm run typecheck
npm run lint
npm test
npm run build
npm run validate:dogfood
npm run check:prompts

Install from the repository root so the local llmdoc bin is linked before validation. Changes to CLI semantics must remain synchronized with both host surfaces, the bilingual READMEs, design documentation, and dogfood knowledge.

Reference

• Portable Agent integration recipe
• V3 design notes (currently marked draft)
• Operating protocol
• Workflow contracts: init, update, prune, and upgrade
• Runtime reference: npx -y @tokenroll/llmdoc --help

Project support