Doc Bridge

doc-bridge

[npm] [CI] [Pages] [OpenSSF Best Practices] [License: MIT] [Node] [TypeScript]

npm: @agentskit/doc-bridge · CLI: ak-docs · Landing: agentskit-io.github.io/doc-bridge

Topics: ai-agents · documentation · developer-experience · mcp · llms-txt · typescript

Compatibility: node >=22 · TypeScript 5.8+ · pnpm, npm, or yarn consumers

Deterministic proof of documentation drift, with evidence-linked handoffs for coding agents.

Turn your docs into executable handoffs: doc-bridge reads your repo docs, ownership map, and human documentation site, then gives humans and agents the same evidence-linked starting point:

• where to start reading
• which files/packages it may edit
• which checks prove the change
• which human docs explain the feature

The core works without any LLM or API key; the documentation portal dogfoods AgentsKit Chat as an optional surface over that deterministic layer.

[doc-bridge maps human docs into structured agent handoffs]

Built for agent-scale repositories

Doc Bridge turns large repository structure and documentation into compact, evidence-linked context that humans and coding agents can query instead of repeatedly traversing the full repository.

Historical context-payload estimate

An earlier anonymized dogfooding cycle estimated up to 99% less serialized context payload. This is a historical payload estimate—not a guarantee of provider-token savings, answer quality, or semantic correctness. It is not the same measure as provider-token usage.

[Estimated context payload reduction]

Historical controlled A/B signal

The published controlled study with 96 anonymized executions reported a directional operational signal of:

• 18.46% fewer paired provider token-equivalent units across 46 token-complete pairs;
• 39.75 seconds lower P95 latency;
• 87.5% operationally completed executions vs. 75.0% with repository-only context.

These are historical, separately defined measures: the 99% figure is an estimated context-payload reduction, while the 18.46% figure uses provider-token-equivalent data from 46 paired observations. The bounded adjudicator recorded zero adjudicator-success outcomes in both arms; because that adjudicator is mechanical and does not independently judge semantic correctness, this result is directional and inconclusive. A newer local pilot is intentionally not promoted here while its semantic evaluation and publication review remain incomplete. See the full methodology and anonymized data. The separate Layer-1 benchmark results measure deterministic documentation findings with native and fixture scores reported separately.

Why teams use it

Agents are powerful, but most repo docs are written for humans. The result is familiar: the agent guesses ownership, edits the sibling package, runs the wrong test, or ignores the human guide that already explained the rule.

doc-bridge works in both directions:

[doc-bridge connects human docs to coding agents and agent memory back to draft docs]

The handoff is a routing contract:

{  "startHere": "docs/for-agents/packages/auth.md",  "editRoots": ["packages/auth"],  "checks": ["pnpm --filter @demo/auth test"],  "humanDoc": "/docs/guides/auth"}

Workflow runs may carry the same optional correlation envelope used by the AgentsKit runtime and Chat protocol. operationId is the cross-repository identity; runId, sessionId, turnId, actionId, and traceId retain local meaning. It is bounded metadata only and must not contain prompts, secrets, or document content.

That contract works from the terminal, MCP, CI, and optional RAG/chat.

Documentation quality and reconciliation

Discovery is only the first step. ak-docs audit documentation compares declared documentation and ownership with the observed project graph and reports evidence-backed findings for missing coverage, stale relations, structured contradictions, exact duplicates, missing examples, and incomplete maintenance metadata.

Natural-language correctness, unnecessary prose, and semantic redundancy remain explicitly not-analyzed until a configured agent or human review evaluates them. Proposed changes stay reviewable and human-approved.

Example finding (anonymized)  CONTRADICTION · high confidence  Documentation declaration differs from the observed project relation  Evidence: 4 source files + 1 documentation declaration  Action: review ownership and update the canonical document

Verify documentation divergence

Compare revision snapshots deterministically: inspect symbol, CLI, configuration and signature changes, historical references and remaining head citations. Broken references require removal evidence; ambiguity and unsupported extraction stay explicit. Changed references require review. Default policy routes findings separately from their evidence status; use --no-policy to inspect raw diagnostics.

From a built checkout, this runnable comparison snapshots a temporary copy of the signature fixture before and after removing a cited export. Use an installed ak-docs, or replace it with node /path/to/checkout/bin/ak-docs.js from a built checkout. The Node step extracts the raw snapshot from the discovery response:

set -ework=$(mktemp -d)cp -R tests/fixtures/signature-api "$work/repo"(cd "$work/repo" && ak-docs discover --json) > "$work/discovery.json"node -e 'const fs = require("node:fs"); const result = JSON.parse(fs.readFileSync(process.argv[1], "utf8")); fs.writeFileSync(process.argv[2], JSON.stringify(result.snapshot))' "$work/discovery.json" "$work/base.json"sed '/^export const describeThing/d' "$work/repo/api.ts" > "$work/api.ts"mv "$work/api.ts" "$work/repo/api.ts"ak-docs diff --base "$work/base.json" --root "$work/repo"ak-docs diff --base "$work/base.json" --root "$work/repo" --no-policyrm -rf "$work"

Capture each revision independently for a real delta. See diff and findings and Action advisory mode for bounded PR reports. Reports never approve documentation corrections.

2.0 workflows

• Vault: opt-in Obsidian discovery reads wikilinks, aliases and tags; ak-docs vault export generates ignored navigation notes and vault diff turns edited generated notes into pending proposals. Vault guide.
• Rationale: enable index.knowledgeEntities.enabled, index decisions/concepts/changes, then run ak-docs why <path|symbol|concept>. MCP exposes knowledge.decision, knowledge.concept and knowledge.whyChanged. Evidence is observed rationale, not inferred causality. Knowledge query contract.
• Remediation: optional ak-docs fix --llm uses scripted fake-provider responses in the key-free guide; real providers use explicit intelligence configuration. Proposals require deterministic rechecking and human review.
• Advisory: CHANGED_REFERENCE concerns a claim about the changed aspect; bare mentions do not trigger it. Unknown syntax remains review-required uncertainty. Action advisory.
• Upgrade: follow the 1.x → 2.0 migration guide. This package ships no Studio UI; use the deterministic StudioGraphV1 export to visualize the graph with other tools.

The latest Layer-1 benchmark and method report native findings at 100% precision (15 TP, 0 FP, 0 FN), historical changed-reference findings at 83.3% precision (10 TP, 2 FP, 0 FN), and 6.25% false-positive rate on historical negative cases (2/32). These purposive controls measure different denominators; they do not establish general semantic accuracy or release readiness.

60-second proof

This README owns the one-command proof; the Getting started guide owns the complete repository setup and first-index workflow.

npm i -D @agentskit/doc-bridgenpx ak-docs demo --text

No config, no docs to read first. Output shows before/after, a real handoff, gate red→green, and the MCP snippet:

After (handoff.resolve / query --agent)  ✓ target:  auth (packages/auth)  ✓ start:   docs/for-agents/packages/auth.md  ✓ edit:    packages/auth  ✓ checks:  pnpm --filter @demo/auth test · pnpm --filter @demo/auth lint  ✓ human guide: /docs/guides/auth
Gate: red → green

Verify the real handoff path

This checked example runs the bundled demo through the public CLI. The repository's scripts/check-readme-standard.mjs gate compares this block byte-for-byte with the executable fixture; .github/workflows/ci.yml runs that gate and its executable tests on every pull request.

The study figures below are not recomputed by the README gate; their protocol, privacy checks, and limitations are documented in the linked study artifacts.

import { execFileSync } from 'node:child_process'
execFileSync(process.execPath, ['bin/ak-docs.js', 'demo', '--text'], {  stdio: 'inherit',})

node examples/verify-handoff.mjs

For the complete repository setup, follow the Getting started guide.

Using Cline? Follow the deterministic llms-install.md setup. It runs the pinned MCP server through pnpm dlx without adding Doc Bridge to your repository dependencies.

What ships

See the surface map for a visual overview of the CLI, MCP, CI, and adapter surfaces.

See Getting started, MCP, Examples, and Why doc-bridge.

Cursor plugin

This repository also contains a Cursor plugin that pairs the Doc Bridge MCP server (read-only except for the explicit proposal tool) with a handoff skill. It resolves startHere, readBeforeEditing, editRoots, and checks before Cursor edits a routed repository. The plugin does not request credentials. Its MCP tools are read-only except docbridge.proposals, which can save a review artifact under .doc-bridge/ and record enrichment approve/reject decisions; no MCP tool edits documentation or source files or applies remediations.

GitHub Copilot plugin

The root Agent Plugins manifest exposes the same portable handoff skill and MCP server to GitHub Copilot CLI. Copilot discovers skills/ and .mcp.json from the standard plugin layout, so the integration stays source-owned instead of copying prompts into another repository.

copilot plugin install AgentsKit-io/doc-bridge

Portable Agent Skill

skills/doc-bridge-handoff packages the same fail-closed routing contract in the open Agent Skills layout for OpenClaw-compatible clients, Hermes Agent, Pi, Cursor, and other runtimes that can execute a local skill script. The skill prefers the read-only MCP tool and falls back to a pinned, zero-credential CLI resolver. It never edits files, runs returned checks, or grants authority outside editRoots.

Install the published skill from ClawHub:

clawhub install doc-bridge-handoff

Pi users can install the same source-owned skill through the npm package:

pi install npm:@agentskit/doc-bridge

Claude Desktop MCP Bundle

Doc Bridge can be packaged as a local MCP Bundle for Claude Desktop. The bundle advertises 14 MCP tools, exercises 8 of them in its smoke test, and marks every tool except docbridge.proposals read-only. It asks the user to select the repository's doc-bridge.config.json; that file defines the project boundary Doc Bridge may read.

From a clean checkout:

pnpm install --frozen-lockfilepnpm mcpb:pack

The command builds Doc Bridge, creates a production-only staging directory, validates the MCPB manifest, packs the extension, checks its file inventory, and writes the local artifact under .mcpb-output/. Generated bundles and staging directories are intentionally excluded from Git.

The current packaged compatibility declaration is macOS-only; this project does not claim bundle support for other operating systems until the exact artifact passes an independent installation test there.

Why this exists

doc-bridge ships AgentHandoff JSON:

{  "type": "agent-handoff",  "startHere": "docs/for-agents/packages/auth.md",  "editRoots": ["packages/auth"],  "checks": ["pnpm --filter @demo/auth test"],  "humanDoc": "/docs/guides/auth",  "bridge": { "humanDoc": "linked" }}

When a human guide is missing, handoffs surface it as a feature:

{  "bridge": {    "humanDoc": "missing",    "action": "ak-docs bootstrap agent-docs"  },  "notes": ["Human guide missing for billing. Run: ak-docs bootstrap agent-docs"]}

Five loops (with real commands)

ak-docs ask "who owns schemas"# Best match: ownership os-core# Handoff preview#   start:  docs/for-agents/packages/os-core.md#   edit:   packages/os-core#   checks: pnpm --filter os-core lint · pnpm --filter os-core test

Coverage your team checks daily

The following is illustrative output from the command, not a measurement of this repository. Run the command locally or in CI for current values.

ak-docs doctor --textak-docs doctor --badge          # shields.io markdown for READMEak-docs index --watch           # keep index fresh while editing docs

Score: 82/100 (B)  Agent docs:      8/10 (80% handoff-ready)  Human guides:    6/10 (60% bridged)  Gates:           3/3 passing
Next actions  → ak-docs bootstrap agent-docs  → ak-docs query package billing --agent

Agent uses it alone

1. MCP auto-wire: ak-docs mcp install --cursor
2. Skill/rule: paste docs/skills/doc-bridge.md into Cursor rules — agents call handoff.resolve before editing packages/*
3. Handoff is the next step: startHere, checks, and bridge are in the JSON/MCP response

CI as first-class citizen

Reuse the bundled GitHub Action on every PR:

Follow the canonical Gate and CI guide, which includes the complete workflow and pins the published Action release.

The Action installs the exact configured package (or the workspace package when dogfooding this repository), then verifies the committed index and configured gates without silently rebuilding them. It rejects non-exact package versions. See the Marketplace guide.

The guide is pinned to the published stable Action release v1.7.45; the checked-in package version is 2.6.0.

Coverage is repository-specific. Run ak-docs doctor --badge locally to emit current handoff and human-bridge badges, or pnpm coverage:badge in CI; this README intentionally avoids publishing a stale static percentage.

For this repository, generate the ignored index before local checks:

node bin/ak-docs.js indexnode bin/ak-docs.js gate run

Never commit this repository's .doc-bridge/index.json. CI builds it twice and compares artifact bytes, then checks reproducibility and documentation conformance. This verifies deterministic generation, not committed-index freshness. Consumer repositories using the Action still verify their committed index before rebuilding.

Product surface

Core — always (no LLM)

Optional AgentsKit peers

npm i -D @agentskit/rag @agentskit/ink @agentskit/adapters @agentskit/memory reactak-docs rag ingest && ak-docs chat

See docs/chat-and-rag.md.

AgentsKit ecosystem

Who uses it (public)

Designed for and dogfooded on open AgentsKit surfaces:

Playbook pattern: docs/playbook/doc-bridge-pattern.md — export with ak-docs playbook pattern --text

Configuration examples

Contract: docs/spec/config-v1.md · CLI: docs/spec/cli.md · MCP: docs/mcp.md · Skill: docs/skills/doc-bridge.md · Pattern: docs/playbook/doc-bridge-pattern.md · Recipes: docs/recipes/index-pipeline.md

Learn loop — memory → draft PR

ak-docs memory ingestak-docs memory classifyak-docs memory promote --pr --dry-run   # preview gh commandsak-docs memory promote --pr              # opens draft PR via gh

Status

Published npm package: v1.7.45 stable. The working-tree package version is 1.8.0 and is not published yet. The Action example below intentionally pins the latest published stable release; the release workflow updates the package and Action version together when a new release is published.

The published package provides portable, fail-closed handoffs through the CLI, MCP server, CI action, and packaged skill; deterministic Documentation Standard v1 conformance; verified release provenance; and documentation-quality audit tooling.

pnpm install && pnpm build && pnpm testpnpm smoke:ollama    # optional — skips if Ollama/peers unavailable

Landing: https://doc-bridge.agentskit.io/

Privacy Policy

The local MCP server reads only the project selected through doc-bridge.config.json. It does not require an API key, send project data to AgentsKit, or collect telemetry. Its MCP tools are read-only except docbridge.proposals, which can save a review artifact under .doc-bridge/ and record enrichment approve/reject decisions; no MCP tool edits documentation or source files or applies remediations. See the complete Privacy Policy for accessed paths, use, storage, sharing, retention, optional integrations, and contact information.

Contributing

Issues and PRs are welcome. Start here:

To improve the evidence base, reproduce the anonymized study, add a language or framework analyzer, contribute a documentation-quality rule, or add a fixture for a real contradiction or stale relation.

License

MIT

Upgrading to 2.0

See the 1.x to 2.0 migration guide for region remediation, explicit index regeneration and configurable memory paths.