Code-review bot

May 16, 2026 · View on GitHub

A pull-request review bot fronting ~24 analysis tools. Demonstrates how the context firewall carries the weight of a code-review workflow: diff dumps and grep output are large, so the firewall and artifact store are heavily exercised while the prompt stays compact.

Run it

python examples/architectures/code_review_bot/main.py

(Or make architectures / make example.)

A captured run of the script lives in OUTPUT.md.

This is a reference architecture, not a tutorial recipe. The cookbook gives you copy-paste snippets for individual primitives (routing, firewall, drilldown, BYO tools); the architecture wires them together around a realistic problem shape so you can see how they interact.

It is mocked: tool implementations return canned strings, no real git / linter / type checker is invoked. The point is to demonstrate the contextweaver glue around a code-review-shaped transcript, not to integrate with a code host.

Setup

The 24-tool catalog lives in catalog.yaml. Loading it:

from contextweaver.routing.catalog import Catalog, load_catalog_yaml

catalog = Catalog()
for item in load_catalog_yaml("catalog.yaml"):
    catalog.register(item)

Namespaces: grep, git, lint, typecheck, test, review. Several tools have side effects (lint.fix, review.post_comment, review.approve, …); the catalog records this on each SelectableItem.side_effects so a real deployment could refuse to call them automatically (Router.route(..., exclude_tags=...)).

The review

The bot walks a six-step review of a refactor that introduces a regression in payments/charge.py:

"show me the diff of this pull request against main" — git.diff (large output → firewall)
"grep for the symbol legacy_charge in the codebase" — grep.symbol (large output → firewall)
"run the test suite for the changed module" — test.run_module
"run mypy on the changed module to surface type errors" — typecheck.module
"run ruff on the changed files and report style violations" — lint.run
"post a review comment requesting changes on the regression" — review.post_comment

See TRANSCRIPT in main.py for the exact text.

What's load-bearing

contextweaver feature	Used	What it does here
`Router.route(query)`	✅	Narrows 24 tools → top-3 shortlist (`top_k=3`)
Bounded choice pattern	✅	Bot picks from the shortlist, not from the whole catalog
`TreeBuilder` DAG	✅	One-shot graph build at startup; routes are sub-millisecond
Context firewall	✅✅	Compacts the ~28 KB diff dump and ~6 KB grep result down to ~500-char summaries before they touch the prompt
Artifact store	✅	Raw bytes stay addressable for drilldown; only summaries land in the prompt
Persistent facts	✅	`pr.target_file`, `pr.test_status`, `pr.type_errors` survive across all six steps
Tight per-phase budgets	✅	`ContextBudget(route=1500, call=2500, interpret=2500, answer=3500)` keeps the answer prompt small even after the firewall externalises the heavy bytes

What's intentionally not here

Real git integration. Mock tool responses keep the example deterministic and CI-friendly. A real deployment would wire git.diff to subprocess.run(["git", "diff", "main"], ...) or to an MCP server.
LLM-based diff summarisation. The review.summarize_diff tool's canned response is a stand-in for what a Summarizer plugin (issue #26) would do.
Multi-PR session state. The fact store survives the in-process review, but to persist across PRs you would swap the default InMemoryFactStore for a SqliteFactStore (issue #174) or a Mem0 / Zep adapter (issue #195).