Composto

Causal memory layer for coding agents. Catches the bug your agent is about to reintroduce.

Composto is a repo-local graph of your git history that your AI coding agent consults before every edit. When a file was reverted recently, has a fix cluster in its history, or was last touched by someone who left the team, Composto surfaces that signal as in-context guidance before the agent writes the code. Hook-enforced on Claude Code, Cursor, and Gemini CLI. Local-first, MIT.

$ composto impact src/memory/signals/hotspot.ts

verdict:    medium
score:      0.52
confidence: 0.50
signals:
  revert_match       ■■■■■■■■■■ strength=1.00 precision=1.00
  hotspot            ■          strength=0.10 precision=0.54
  fix_ratio          ■          strength=0.07 precision=0.54
  author_churn       ·          strength=0.00 precision=0.16

# This file was touched by a Revert commit in history.
# blastradius remembers. Your LLM couldn't.

Quick Start

# Install
npm install -g composto-ai

# One-command setup, wires MCP + PreToolUse hook into your AI client
cd your-project
composto init --client=claude-code     # or cursor, or gemini-cli

# Restart your AI client. Hook fires on every Edit / Write / MultiEdit.
# On medium|high|unknown verdicts, the agent gets a composto_blastradius
# block in context before it acts. Passthrough on low.

# Observe
composto stats              # hook invocations, verdict distribution, latency
composto stats --disable    # local-only opt-out (writes .composto/telemetry-disabled)

# Query on demand
composto impact src/auth/login.ts
composto index --status     # diagnostics: schema, freshness, calibration

Also in the box: AST compression tools

Composto also ships a tree-sitter based AST compressor (about 89% token savings) and a smart context packer for bug-fix tasks. These are separate from the causal layer but live in the same binary.

composto ir src/app.ts                 # compress a file to IR (L0/L1/L2/L3)
composto context src/ --budget 2000    # smart context within a token budget
composto benchmark .                   # see compression stats

See the IR Layers, Health-Aware IR, and Context Budget sections below for details.

MCP plugin (Claude Code, Cursor, Claude Desktop)

The MCP server is bundled inside composto-ai. Install the package globally first, then register the server with your client:

npm install -g composto-ai

Claude Code:

claude mcp add composto -- composto-mcp

Cursor — add to ~/.cursor/mcp.json (or project-local .cursor/mcp.json):

{
  "mcpServers": {
    "composto": {
      "command": "composto-mcp"
    }
  }
}

Then restart Cursor and verify under Settings → MCP that composto is green.

Claude Desktop — add the same block to ~/Library/Application Support/Claude/claude_desktop_config.json.

Composto adds 5 tools to your AI assistant: composto_ir, composto_benchmark, composto_context, composto_scan, and composto_blastradius (the last one gated by COMPOSTO_BLASTRADIUS=1 during beta).

Cursor: one-command setup

Registering the MCP server only exposes the tools — Cursor's agent often defaults to its built-in read_file / codebase_search. To configure both the MCP server and a project rule that tells the agent when to call Composto, run:

cd your-project
composto init

This writes .cursor/mcp.json (project-local MCP registration) and .cursor/rules/composto.mdc (an alwaysApply: true rule that gets injected into every conversation). Existing files are merged, never overwritten. Restart Cursor and check Settings → MCP that composto is green.

Without the rule, hit rate is ~30-50%; with it, ~85-95%. The rule template is embedded in src/cli/init.ts (CURSOR_RULES_MDC) — open the generated .cursor/rules/composto.mdc to customize per-project.

Hook-enforced injection (v0.6.0+)

Instead of asking the agent to remember to call composto_blastradius, wire a hook so it gets consulted automatically before every Edit / Write / MultiEdit. The agent receives a <composto_blastradius> context block in-line when verdict is medium or high — you don't do anything, the warning just shows up where it's needed.

cd your-project
composto init --client=claude-code    # or cursor, or gemini-cli

This writes:

• The platform's MCP config (same as before)
• A PreToolUse hook (Claude Code / Gemini CLI) that invokes composto hook <platform> pretooluse on every file-targeting tool call. The hook extracts the target file, runs composto_blastradius, and injects the verdict as additionalContext. Passthrough on low verdict — no noise.
• For Cursor: a .cursor/hooks.json entry that denies the tool call on verdict: high (Cursor's additional_context is dropped per forum #155689, so hybrid strategy — the existing .cursor/rules/composto.mdc rule carries medium/low, the hook only interrupts on high).

Existing settings are merged, never overwritten. Re-running composto init is idempotent — no duplicate hook entries.

Observe what's happening:

composto stats            # hook invocations, verdict distribution, p50/p95 latency
composto stats --json     # machine-readable
composto stats --disable  # opt out (writes .composto/telemetry-disabled marker)

Telemetry is local-only — writes to .composto/memory.db in your repo, nothing leaves your machine. No user ID, no cloud sync, no account.

Platform matrix:

| Platform | MCP | Hook | Strategy | |---|:---:|:---:|---| | Claude Code | ✅ | ✅ PreToolUse | additionalContext on medium|high|unknown, passthrough on low | | Cursor | ✅ | ✅ preToolUse | Deny-on-high via permissionDecision; medium/low via .mdc rule | | Gemini CLI | ✅ | ✅ BeforeTool | additionalContext on medium|high|unknown | | Claude Desktop | ✅ | — | MCP-only (no hook API yet) |

How It Works

Composto uses tree-sitter to parse your code into an AST, then walks every node and classifies it:

| Tier | Action | What | % of nodes | |------|--------|------|-----------| | Tier 1 | Keep | imports, functions, classes, interfaces, types, enums | 0.8% | | Tier 2 | Summarize | if, for, while, switch, return, throw, try/catch | 0.9% | | Tier 3 | Compress | variable declarations → one-liner, await → kept | 6.9% | | Tier 4 | Drop | string contents, operators, punctuation, comments | 86.6% |

86.6% of your code's AST nodes are noise. Composto drops them.

Commands

# Benchmark token savings across your project
composto benchmark .

# Generate IR at different detail levels
composto ir <file> L0    # Structure map (~10 tokens) — just names
composto ir <file> L1    # Full IR — compressed code + health signals
composto ir <file> L2    # Delta context — only what changed
composto ir <file> L3    # Raw source — original code

# Smart context packing within a token budget
composto context <path> --budget <tokens>
# Fits maximum information into your budget:
# hotspot files get L1 (detailed), rest get L0 (structure)

# Scan for security issues and debug artifacts
composto scan .

# Analyze git history for health trends
composto trends .

# Compare LLM quality: raw code vs IR (requires ANTHROPIC_API_KEY)
composto benchmark-quality <file>

# Historical blast radius — beta, gated by COMPOSTO_BLASTRADIUS=1
composto index                 # bootstrap .composto/memory.db from git history
composto impact <file>         # risk verdict + signals for a file
composto index --status        # diagnostics: schema, freshness, calibration

BlastRadius

Beyond compression, Composto indexes your repo's git history into a local SQLite graph and exposes it as a queryable risk surface. Before your agent edits a file, it can ask: "has this region been reverted? who fixed the last similar bug? is the last author still around?" — signals no LLM can infer from current code alone.

Four signals per query: revert_match, hotspot, fix_ratio, author_churn. Verdict is low / medium / high / unknown; when confidence is low the tool returns unknown rather than guessing. Precision is repo-calibrated (self-validation over the repo's own fix history).

verdict:    high
score:      1.00
confidence: 0.30
signals:
  revert_match       ■■■■■■■■■■ strength=1.00 precision=0.50
  hotspot            ·          strength=0.00 precision=0.30
  ...

Where we are, honestly. The v2.1 time-travel backtest (rewinds the DB to each pre-fix snapshot) shows revert_match carrying most of the product's value — it clears the ship gate on picomatch (precision 0.65, recall 0.78 on the medium|high band). Signal-attributed precision (excluding revert_match) is weaker: the three non-revert signals are alive but need calibration work. See docs/blastradius-proof-v2.md for numbers on all four band combinations across two public repos + the per-signal diagnostic behind them.

The honest framing: BlastRadius is a bug-history memory layer that agents query before editing. "This file was reverted three weeks ago" is the primary promise v1 delivers. The other signals expand the query surface — calibration work on hotspot and fix_ratio is the open follow-on.

Feature-flagged via COMPOSTO_BLASTRADIUS=1 during the beta. Available as both CLI (composto impact, composto index) and MCP tool (composto_blastradius).

Quality Proof

We tested 4 files from simple to hard. Same question, raw code vs IR: "What does this file do?"

| File | Complexity | Raw Tokens | IR Tokens | Savings | Comprehension | |------|-----------|-----------|----------|---------|--------------| | hotspot.ts | Simple | 299 | 77 | 74.2% | Full | | layers.ts | Medium | 765 | 249 | 67.5% | Full | | detector.ts | Medium | 704 | 160 | 77.3% | Full | | ast-walker.ts | Hard (448 lines) | 3,782 | 663 | 82.5% | ~90% |

Even on a 448-line recursive AST walker with nested switches, an LLM can fully explain the architecture, all 12 functions, and the data flow from the IR alone.

What IR preserves: function signatures, parameter types, imports, control flow, return values, class/interface declarations.

What IR drops: string contents, regex patterns, operator details, formatting — things the LLM already knows.

Full benchmark: docs/benchmark-proof.md

IR Layers

| Layer | Tokens | Use case | |-------|--------|----------| | L0 | ~10 | "What's in this file?" — just function/class names | | L1 | ~85 | "What does this file do?" — compressed code + health signals | | L2 | ~65 | "What changed?" — git diff with context | | L3 | variable | "Show me the exact code" — raw source |

When to use which

"Explain the architecture"     → L1 for all files
"Fix this bug"                 → L3 for target file, L1 for context
"Review this PR"               → L2 for changed files, L1 for context
"What files are in this repo?" → L0 for everything

Health-Aware IR

Composto analyzes git history and embeds health signals directly into IR:

FN:handleAuth({credentials}) [HOT:15/30 FIX:73% COV:↓ INCON]
  IF:!session → RET 401
  RET { token, expiresAt }

• [HOT:15/30] — 15 changes in last 30 commits (hotspot)
• [FIX:73%] — 73% of changes were bug fixes
• [COV:↓] — Test coverage declining
• [INCON] — Inconsistent patterns from multiple authors

Only unhealthy code gets annotated. Healthy files stay clean.

Context Budget

Don't guess which files to send. Let Composto decide:

composto context src/ --budget 2000

Output:

== L1 (detailed) ==
[hotspot] src/auth/login.ts
  USE:[./types.js, ./session.js]
  OUT ASYNC FN:login(credentials)
    TRY
      IF:!valid → THROW:AuthError
      RET { token, user }

== L0 (structure) ==
src/utils/helpers.ts
  FN:formatDate L5
  FN:parseQuery L23
...

Budget: 1994/2000 tokens
Files: 9 at L1, 16 at L0

Hotspot files get full detail. Everything else gets structure. Budget is never exceeded.

Stats

Overall compression: 89.2%
L0 compression:      97.5%
AST engine:          51/51 files (0 regex fallback)
Languages:           TypeScript, JavaScript, Python, Go, Rust
Tests:               251 passing
BlastRadius v2.1:    precision 0.65, recall 0.78 (picomatch, time-travel,
                     medium|high band). Honest signal-attributed numbers
                     in docs/blastradius-proof-v2.md.

Configuration

Optional .composto/config.yaml:

watchers:
  security:
    enabled: true
    severity:
      "src/**": warning
      "tests/**": info
  consoleLog:
    enabled: true

trends:
  enabled: true
  hotspotThreshold: 10
  bugFixRatioThreshold: 0.5

All settings have sensible defaults. The config file is optional.

Contributing

git clone https://github.com/mertcanaltin/composto
cd composto
pnpm install
pnpm test        # 145 tests
pnpm build       # builds to dist/
npx composto benchmark .  # see compression stats

License

MIT