← Back to .knowledge

// Blog / definition

What is repo-local knowledge routing for coding agents?

Repo-local knowledge routing is the maintained first-read path an agent uses before it spends context crawling a repository from scratch.

// direct answer

Short answer

Repo-local knowledge routing gives a coding agent one versioned, repository-owned map into current modules, evidence, trust status, stale areas, and maintenance commands. It is a routing layer, not a substitute for reading source code and tests.

Understand whether a local routing layer can reduce cold-start context work without replacing source review.

// real failure mode

The failure mode

A new coding-agent session starts by sampling README files, package manifests, source folders, tests, and old notes.

The agent may find useful context, but it has no stable signal for which summary is current, which module is risky, or which files must be checked before behavior changes.

The result is repeated orientation work and a higher chance of trusting prose that should have been rechecked against code.

// repo-local model

The repo-local fix

Repo-local knowledge routing gives the agent one maintained first-read path inside the repository.

The routing bundle points to modules, trust buckets, changed or stale files, wiki/search artifacts, and maintenance commands.

It does not replace source review. It directs the agent toward the source, tests, evidence, and repair queue that matter for the current task.

// concrete example

Concrete example

A task mentions runtime execution. Instead of crawling every top-level directory, the agent reads the routing bundle, sees which module owns runtime execution, checks whether that module is trusted, and then re-reads the current implementation and tests before editing.

If the module is marked low-confidence or stale, the agent treats the knowledge file as a route hint only. It should not quote the summary as truth until source and tests confirm it.

// repo proof

Repo proof to inspect

First-read map

.knowledge/maintenance/routing_bundle.json

This is the compact route into modules, trust state, stale items, and the next files to inspect.

Use it for
  • Choosing the first source files to read
  • Finding the relevant module boundary
Do not use it for
  • Claiming runtime behavior
  • Skipping source and tests
Module cards

.knowledge/modules/*.json

Module cards summarize ownership, confidence, and important paths, but their trust state controls how much weight they deserve.

Use it for
  • Planning a targeted read
  • Spotting low-confidence areas
Do not use it for
  • Treating old summaries as current
  • Making security or billing claims without code
Evidence records

.knowledge/evidence/*.json

Evidence ties claims back to commands, files, or observed checks so the agent can avoid free-floating prose.

Use it for
  • Finding supporting checks
  • Reviewing source-of-truth references
Do not use it for
  • Assuming passing evidence is permanent
  • Ignoring stale timestamps
Repair queue

.knowledge/maintenance/repair_queue.json

Knowledge debt becomes visible work instead of hidden uncertainty inside the prompt.

Use it for
  • Prioritizing rechecks
  • Surfacing stale or suspect areas
Do not use it for
  • Blocking every edit
  • Replacing engineering judgment

// command transcript

Commands and expected checks

node .knowledge/tools/install-agent-integrations.js
expected
Agent entrypoints are installed or refreshed for Codex, Claude Code, OpenCode, and local commands.
inspect next
.agents/skills/*/SKILL.md and .knowledge/agent-integrations/
node .knowledge/tools/flow.js import
expected
The repository is ingested without overwriting curated knowledge.
inspect next
.knowledge/project_index.json and .knowledge/modules/*.json
caution
Use this when bootstrapping or refreshing an existing repository.
node .knowledge/tools/flow.js release --no-color
expected
Routing, trust, freshness, search, metrics, and Inspector outputs are rebuilt.
inspect next
.knowledge/maintenance/trust_report.json and .knowledge/maintenance/routing_bundle.json
First-read path repo-local proof

From cold start to targeted source review

01 Agent task
02 .knowledge/Quick-Start.md
03 maintenance/routing_bundle.json
04 Relevant module cards
05 Current source and tests

// field guide

Routing is / routing is not

Use routing to aim the investigation, not to replace it.
Routing isRouting is not
A first-read map into current code, tests, evidence, and trust state.A hosted memory product that the agent must blindly trust.
A way to mark stale, suspect, low-confidence, or advisory knowledge before use.A shortcut around reading source for critical behavior.
A repo-owned artifact that can be reviewed with the codebase.A private chat-history habit that only one agent can see.

// guardrails

What the agent should not trust blindly

  • Use trusted and near_trusted knowledge for routing and limited planning.
  • Use routing_trusted knowledge only to choose files and boundaries.
  • For auth, billing, runtime execution, queues, storage, signing, secrets, migrations, security, concurrency, or anything stale/suspect/low-confidence, re-read current source and tests.

// common mistakes

Common mistakes

  • Treating the routing bundle as proof of product behavior.
  • Using a wiki page as source of truth when the module is marked stale or low-confidence.
  • Letting every agent invent a new onboarding path instead of sharing one repo-local first read.

// quick FAQ

FAQ

Is repo-local routing the same as RAG?

No. RAG retrieves chunks. Repo-local routing is a maintained first-read map that tells an agent what to inspect and how much to trust each knowledge artifact.

Can an agent edit code based only on routing_bundle.json?

No. The routing bundle should point the agent to source, tests, evidence, and stale areas. Current code and tests still decide behavior.

// page-specific next step

Use this page when you define the first-read contract

Start with the routing bundle as the agent's map, then require source and test reads before behavior-changing edits. That keeps the workflow concrete without pretending the knowledge layer is magic.