llm-context-router 0.6.0

Scope your AI context before it explodes

context-router

Problem

Working with AI tools on large codebases causes context explosion. Every task loads the entire codebase into the prompt.

What context-router does

Routes a task to the agent that owns it. Scopes the prompt to only the relevant files. Keeps a lightweight map of who owns what.

What context-router does NOT do

Does not call any LLM. Does not automate anything. Does not execute code.

Without context-router:

"fix the auth bug" → loads 847 files → model touches 6 modules it shouldn't

With context-router:

context-router ask "fix the auth bug" --file src/auth/token.py → Agent: backend-engineer → Touch only: src/auth/token.py, src/auth/session.py → Do not modify: src/api/, src/db/, src/frontend/

Install

pip install context-router

Local dev:

pip install -e .

Usage

1. Define your agents

context-router init creates 3 core agents automatically:

agents/
├── dispatcher.agent.yaml      ← routes all tasks (do not edit)
├── memory-keeper.agent.yaml   ← syncs memory after commits (do not edit)
├── validator.agent.yaml       ← validates diffs before applying (do not edit)
└── default.agent.yaml         ← template — copy and rename for your domains

Add one agent per domain in your project.

Create an agents/ directory in your project and add one YAML file per agent:

# agents/backend-engineer.agent.yaml
name: backend-engineer
owned_paths:
  - src/auth/*.py
  - src/api/*.py
  - src/db/*.py
description: Handles backend logic including authentication, API routes, and database models.

2. Initialize

context-router init

Specify your AI tool:

context-router init --tool claude    # Claude Code
context-router init --tool cursor    # Cursor
context-router init --tool copilot   # GitHub Copilot
context-router init --tool codex     # Codex / OpenAI
context-router init --tool windsurf  # Windsurf
context-router init --tool aider     # Aider
context-router init --tool all       # generate all

If you skip --tool, context-router detects your tool automatically or asks you to choose.

Tool	Config file generated
Claude Code	CLAUDE.md + AGENTS.md
Cursor	.cursorrules
GitHub Copilot	.github/copilot-instructions.md
Codex	AGENTS.md
Windsurf	.windsurfrules
Aider	.aider.conf.yml

Scans all *.agent.yaml files in agents/, writes .context-router/project_map.json, and scaffolds the full memory structure.

After context-router init your project contains:

.context-router/
├── CORE.md                    ← global invariants (always loaded)
├── CORE_LITE.md               ← minimal version for quick tasks
├── project_map.json           ← agent ownership map (auto-generated)
└── memory/
    ├── MEMORY.md              ← status, architecture, recent fixes
    ├── MEMORY_MAP.md          ← file/function ownership
    ├── MEMORY_CHANGELOG.md    ← full fix history (load on demand)
    ├── MEMORY_INTELLIGENCE.md ← domain knowledge (load on demand)
    └── MEMORY_REFERENCE.md   ← lessons learned, roadmap (load on demand)

Session start prompt for your AI tool:

Read: .context-router/CORE.md
Read: .context-router/memory/MEMORY.md
Read: .context-router/memory/MEMORY_MAP.md
Then paste: context-router ask "your task" --file path/to/file

For quick/simple tasks use CORE_LITE.md instead of CORE.md. Load MEMORY_CHANGELOG.md only when debugging old behavior. Load MEMORY_INTELLIGENCE.md only for domain-specialist tasks.

Initialized. 3 agents registered.

3. Route a task

context-router ask "fix the auth token expiry bug" --file src/auth/token.py

Agent:   backend-engineer
Files:   src/auth/token.py, src/auth/session.py
Prompt:
  Read: agents/backend-engineer.agent.yaml
  Task: fix the auth token expiry bug
  Touch only: src/auth/token.py, src/auth/session.py
  Do not modify: src/api/routes.py, src/db/models.py

Without --file, context-router keyword-matches your task against agent descriptions:

context-router ask "authentication is broken"

Agent:   backend-engineer
Files:   src/auth/*.py, src/api/*.py, src/db/*.py
Prompt:
  Read: agents/backend-engineer.agent.yaml
  Task: authentication is broken
  Touch only: src/auth/*.py, src/api/*.py, src/db/*.py
  Do not modify: src/ui/*.tsx, src/infra/*.sh

4. Validate a diff

context-router validate --staged          # checks git diff --cached
context-router validate --diff my.patch   # checks a saved diff file

Loads .context-router/validator.yaml and runs all matching rules against the diff. Exits 0 on PASS, 1 on FAIL.

STATUS: FAIL
REASON: 'reader.readtext' used outside `with _ocr_lock:` block
RETRY_WITH:
FIX_HINT: Wrap reader.readtext() inside `with _ocr_lock:`

Rule types: lock_guard, pair_call, no_output, line_count, threshold_drift. See .context-router/validator.yaml (generated by context-router init) for full field reference.

Usage — run command

Manual mode (no API key needed)

context-router run "fix the login form" --file src/frontend/login.js

─────────────────────────────────────
  Agent:   frontend
  Prompt:  .context-router/current-task.md
─────────────────────────────────────
  1. Paste the prompt into your AI tool
  2. Save the diff output as:
     .context-router/output.diff
  3. Run: context-router run --continue
─────────────────────────────────────

Auto mode (API key in .env)

cp .env.example .env
# add your key to .env
context-router run "fix the login form" --file src/frontend/login.js

Auto mode — using anthropic (claude-sonnet-4-6)
Attempt 1/3...
STATUS: PASS
Apply the diff to your codebase.

Override model

context-router run "task" --file path --model claude-opus-4-6

Meta-controller retry loop

MetaController wraps any agent/validator pair with a failure-aware retry loop.

from context_router.meta_controller import MetaController
from pathlib import Path

mc = MetaController(
    task="fix the token expiry bug",
    root=Path("."),
    config_path=Path(".context-router/retry_config.yaml"),
)
result = mc.run(agent_runner, validator)
# result = {"passed": True/False, "attempts": N, "report_path": None | "logs/..."}

agent_runner(task, prompt_patches, param_patches) → str validator(output) → (bool passed, str fail_text)

On FAIL, the controller:

1. Stashes working tree (git stash)
2. Builds a retry plan from retry_config.yaml (prompt constraints + param overrides)
3. Re-runs the agent with accumulated patches
4. On retry_with: human — escalates immediately, writes failure report
5. After MAX_RETRIES (3) — restores working tree, writes report to logs/

retry_config.yaml example:

categories: [system, logic, output, cross_agent]

prompt_patches:
  logic:
    2: "add inline comments explaining each change"
    3: "minimal fix only — change one thing at a time"

category_map:
  my_rule_id: logic

human_hints:
  logic: "Logic error — review the change manually."

See examples/reel-engine/retry_config.yaml for a full real-world example.

Agent YAML format

name: string                    # unique identifier, used in output
owned_paths:                    # glob patterns — files this agent is responsible for
  - src/auth/*.py
  - src/api/*.py
description: one sentence       # used for keyword fallback matching when no --file given

Adding a new agent

1. Create agents/<name>.agent.yaml
2. Run context-router init

That's it.

How it works

context-router init
  └── scans agents/*.agent.yaml
  └── writes .context-router/project_map.json  {glob → agent_name}

context-router ask "task" --file src/auth/token.py
  └── loads project_map.json
  └── matches file path against globs
  └── prints structured prompt block
  └── no network, no LLM, no side effects

Optional: graphify

graphify builds an AST-based knowledge graph of your codebase. Zero API cost. Run it after every code change:

graphify update .

Output: graphify-out/GRAPH_REPORT.md — node/edge/community structure for your codebase.

How graphify integrates with context-router

Data flow:

code change
  → graphify update .
  → graphify-out/GRAPH_REPORT.md   (always regenerated)
  → @memory-keeper (post-commit)
      reads GRAPH_REPORT.md
      updates MEMORY_MAP.md        (always, after every commit)

agents during a task
  → read MEMORY_MAP.md             (session start, always)
  → read GRAPH_REPORT.md           (on demand only — unknown deps)

Rules:

• graphify update . is always a plain shell command — no API calls
• GRAPH_REPORT.md is read on demand only (not at session start)
• MEMORY_MAP.md is the agent-readable ownership table
• graphify-out/ files are never loaded at session start
• graphify never writes to .context-router/ directly

Trigger graphify after every code change:

graphify update .

Your AI tool reads the output:

Read graphify-out/GRAPH_REPORT.md only when tracing
unknown cross-file dependencies.

Success criteria

• Install in under 2 minutes
• Understand it fully in under 5 minutes
• Add a new agent by editing one YAML file
• Reduces context size when working with any AI coding tool