Chisel

Test impact analysis and code intelligence built for LLM agents — especially when several agents or sessions touch the same repo (solo developer, multi-agent workflow).

Chisel maps tests to code, code to git history, and answers: what to run, what’s risky, and where attention should go — including blame-based lineage when you need audit context (not “team roster” features).

Who this is for

• Solo developers using Cursor, Claude Code, or other MCP clients — not a substitute for human code review queues.
• Multi-agent usage: parallel agent runs, background tasks, or sequential sessions that share one project. Chisel keeps one consistent graph (project-local .chisel/ storage, cross-process locks) so agents don’t corrupt analysis mid-write.
• Primary interface: MCP tools and structured responses (next_steps, diagnostic statuses), not dashboards for managers.

Docs: Agent playbook (recommended tool loop, start_job / job_status, source field) · Zero-dependency policy · Custom extractors (register_extractor, CHISEL_BOOTSTRAP — bring your own tree-sitter in your venv).

The Problem

An LLM agent changes engine.py:store_document(). It then either:

• Runs all 287 tests (slow, wasteful), or
• Guesses with -k "test_store" (misses regressions)

When multiple agent runs (or agents plus you) work on the same codebase, changes in one area can break another. Chisel gives each agent test impact, import-aware suggestions, and risk signals so they narrow what to run and where regressions may hide — before you merge or ship.

Install

pip install chisel-test-impact

Or from source:

git clone https://github.com/IronAdamant/Chisel.git
cd Chisel
pip install -e .

Use with Claude Code (MCP)

Add to your Claude Code MCP config (~/.claude/settings.json or project .mcp.json):

{
  "mcpServers": {
    "chisel": {
      "command": "chisel-mcp",
      "env": {
        "CHISEL_PROJECT_DIR": "/path/to/your/project"
      }
    }
  }
}

Or run the HTTP server for any MCP-compatible client:

chisel serve --port 8377

Once connected, agents can call the full tool surface — analyze, diff_impact, suggest_tests, risk_map, triage, and more. Run analyze first to build the project graph, then diff_impact after edits to narrow which tests to run. For long analyses on large repos, prefer chisel analyze / chisel update in a terminal so MCP clients don’t time out.

Use with Cursor / Other MCP Clients

Chisel exposes a standard MCP interface. For stdio-based clients:

pip install chisel-test-impact[mcp]
chisel-mcp

For HTTP-based clients, point them at http://localhost:8377 after running chisel serve.

Quickstart (CLI)

# Analyze a project (builds all graphs)
chisel analyze .

# What tests are impacted by my current changes?
chisel diff-impact

# What tests should I run for this file?
chisel suggest-tests engine.py

# Who owns this code?
chisel ownership engine.py

# What files always change together?
chisel coupling storage.py

# Which tests are stale?
chisel stale-tests

# Risk heatmap across the project
chisel risk-map

# Incremental update (only re-process changed files)
chisel update

# Find code with no test coverage, sorted by risk
chisel test-gaps

Try It on This Repo

git clone https://github.com/IronAdamant/Chisel.git
cd Chisel
pip install -e .

chisel analyze .
chisel risk-map
chisel diff-impact
chisel test-gaps
chisel stats

MCP tools (core)

Core query and write tools below; the MCP server also exposes advisory file-lock helpers for multi-process coordination. See schemas.py / chisel serve for the full list.

Tool	What it does
analyze	Full project scan — code units, tests, git history, edges
start_job	Run analyze or update in a background thread; poll job_status (avoids MCP timeouts)
job_status	Poll a job id from start_job until completed or failed
update	Incremental re-analysis of changed files only
impact	Which tests cover these files/functions?
diff_impact	Auto-detect changes from git diff, return impacted tests
suggest_tests	Rank tests by relevance (edges, co-change, import graph) + failure rate
churn	How often does this file/function change?
ownership	Blame-based authors (useful for audit / “who wrote this line”)
who_reviews	Recent commit activity on the file (heuristic “hot spots”, not org chart)
coupling	Co-change partners + import-graph neighbors and numeric scores
risk_map	Risk scores for all files (churn + coupling + coverage gaps)
stale_tests	Tests pointing at code that no longer exists
test_gaps	Code units with zero test coverage, sorted by risk
history	Commit history for a specific file
record_result	Record test pass/fail for future prioritization
stats	Database summary counts
triage	Composite: top risk + gaps + stale tests in one call

Features

• Zero dependencies — stdlib only, works everywhere Python 3.11+ runs
• Encoding-safe — handles non-UTF-8 content in git history (Latin-1 commits, binary diffs) without crashing
• Multi-language — Python, JavaScript/TypeScript, Go, Rust, C#, Java, Kotlin, C/C++, Swift, PHP, Ruby, Dart
• Framework detection — pytest, Jest, Go test, Rust #[test], Playwright, xUnit/NUnit/MSTest, JUnit, XCTest, PHPUnit, RSpec, Minitest, gtest, Dart test
• Incremental — only re-processes changed files via content hashing
• MCP servers — both stdio and HTTP for LLM agent integration
• Risk scoring — weighted formula: churn, coupling, coverage gaps, author concentration, test instability
• Branch-aware — diff_impact auto-detects feature branch vs main
• Optional custom extractors — register_extractor() + CHISEL_BOOTSTRAP for your own tree-sitter/LSP stack (user-installed; core stays stdlib-only)

Ecosystem

Chisel is designed to sit in the agent loop (MCP): impact → tests → record results → refresh analysis. It works standalone or alongside tools like Stele for semantic code context — Chisel stays focused on test graph, git signals, and static imports for blast-radius reasoning.

Design Notes

Coupling: co-change vs. import-graph

Chisel's coupling tool exposes two coupling sources:

1. Co-change coupling (co_change_partners) — files that often appear in the same git commits. Stronger when history has many small commits (including a solo dev committing often, or multiple agents landing separate commits). Sparse history → thin co-change signal.

2. Import-graph coupling (import_partners, plus numeric import_coupling / effective_coupling) — static import/require edges. Always available after analysis and is the main structural signal for single-author repos.

risk_map and impact tools combine both; import graph also powers transitive test suggestions (e.g. facade tests covering inner modules).

Coverage Gap: Graduated Scoring

Coverage gap is quantized to 4 steps (0.0, 0.25, 0.5, 0.75, 1.0) rather than binary. This graduated scoring provides finer granularity for risk assessment in risk_map.

--verbose Flag

chisel analyze does not accept a --verbose flag. Using it causes the command to silently fail. For diagnostic output after analysis, use chisel stats to verify edge counts.

License

MIT