Cairn

The problem

AI coding agents start every session cold. They can read your code, but they don't know what failed last week, which patterns keep recurring, or whether the last fix held. The context window resets. Manual context documents go stale.

What Cairn does

Cairn captures structured signal from your running Python code and distills it into agent context that compounds over time:

• Test signal — not just pass/fail, but assertion values, local variable state at failure, which source lines were implicated
• Execution traces — function calls, return values, exceptions, duration, call depth via sys.monitoring
• Structural patterns — untested branches, missing annotations, deep nesting, common anti-patterns
• Git correlation — commit-to-failure linkage, blame, whether fixes held

Signal gets normalized into TruthRecords, distilled into named patterns, and injected into your agent's context before the next session.

Health analysis

With margin installed (pip install cairn-ai[health]), each pattern cluster gets:

• Drift tracking — is this pattern STABLE, DRIFTING, ACCELERATING, or OSCILLATING?
• Anomaly detection — is this failure EXPECTED or NOVEL?
• Causal discovery — which clusters are correlated? which leads to which?
• Escalation policy — LOG, ALERT, or HALT based on health + trajectory

Agent integration

Cairn exposes an MCP server with tools for Claude Code, Cline, or any MCP-compatible agent:

• cairn_context — full distilled context as Markdown
• cairn_search — semantic search across the truth store
• cairn_recent_failures — latest failures, decay-filtered
• cairn_health — drift, anomaly, causality analysis
• cairn_synthesis — ranked intervention plan

Zero-config pytest plugin. Auto-distillation via git hooks. Per-project SQLite stores. AGPL-3.0.