Skip to main content

Causal traceability SDK for AI agent runtimes — observe, replay, understand agent behavior

Project description

causetrace

中文CIPyPI

causetrace captures tool calls from coding agents (Claude Code, OpenCode, Aider, Continue.dev, Codex CLI, GitHub Copilot) and links them into causal trees and DAGs — not flat timelines. Every event records why it happened, enabling replay, root-cause analysis, and behavior explanation.

Data sources: Claude Code (hooks), OpenCode / Continue.dev / Codex CLI / GitHub Copilot (log tailing), Aider (process wrapper)
Storage: ~/.causetrace/data/<session_id>.jsonl — append-only JSONL, zero external dependencies


Showcase

The same session viewed four ways — flat vs. causal.

Timeline (flat)

$ causetrace timeline ses_10d2f16e
[03:13:37] Read(file_path=src/main.py)
[03:13:37] Grep(pattern=FIXME)
[03:13:37] Read(file_path=src/utils.py)
[03:13:37] Read(file_path=src/utils.py)
[03:13:38] Edit(file_path=src/utils.py)
[03:13:38] Bash(command=python -m pytest tests/ -x)
[03:13:38] Grep(pattern=counter)
[03:13:38] Edit(file_path=docs/api.md)
[03:13:38] Bash(command=python -m pytest tests/)

Chronological order, but no insight into why each event happened.

Causal Tree

$ causetrace tree ses_10d2f16e
[03:13:37] Read(file_path=src/main.py)
    └─ [03:13:37] Grep(pattern=FIXME)
      └─ [03:13:37] Read(file_path=src/utils.py)
[03:13:37] Read(file_path=src/utils.py)  [caused by: need_context]
    └─ [03:13:38] Edit(file_path=src/utils.py)
      └─ [03:13:38] Bash(command=python -m pytest tests/ -x)
[03:13:38] Grep(pattern=counter)
    └─ [03:13:38] Edit(file_path=docs/api.md)
      └─ [03:13:38] Bash(command=python -m pytest tests/)

Parent→child chains reveal the causal structure: each tool call is a direct response to its parent.

Why (causal chain trace)

$ causetrace why ses_10d2f16e <event_id>
[03:13:38] Grep(pattern=counter) ──→
[03:13:38] Edit(file_path=docs/api.md) ──→
[03:13:38] Bash(command=python -m pytest tests/) ◀── TARGET

Trace why a specific event happened — follow the causal chain backward from any event to its root.

Multi-parent DAG

$ causetrace graph ses_3e23bcc8
[02:42:40] Bash(command=python -m pytest tests/)  ← Edit(file_path=docs/api.md)
[02:42:41] Read(file_path=src/main.py)
[02:42:41] Grep(pattern=FIXME)  ← Read(file_path=src/main.py)
[02:42:41] Read(file_path=src/utils.py)  ← Grep(pattern=FIXME)
[02:42:41] Read(file_path=src/utils.py)
[02:42:41] Edit(file_path=src/utils.py)  ← Read(file_path=src/utils.py)
[02:42:41] Bash(command=python -m pytest tests/ -x)  ← Edit(file_path=src/utils.py)
[02:42:42] Grep(pattern=counter)
[02:42:42] Edit(file_path=docs/api.md)  ← Grep(pattern=counter)

Fan-in DAGs visualize convergent causation — one tool consuming multiple prior results. Support for multi-parent causal links via comma-separated parent_event_id.


Supported Agents

Agent Method How it works
Claude Code Hook bridge PreToolUse / PostToolUse hooks via ~/.claude/settings.json
OpenCode Log tailing Parses ~/.local/share/opencode/log/*.log for tool.registry entries
Aider Process wrapper Runs aider as subprocess, parses stdout for tool calls
Continue.dev Log tailing Parses ~/.continue/logs/core.log for JSON tool call entries
Codex CLI Log tailing Parses ~/.codex/sessions/.../rollout-*.jsonl for actions/observations
GitHub Copilot Log tailing Parses ~/.config/Code/logs/ extension host logs for Copilot tool calls
# Claude Code — automatic via hooks
causetrace tale <session_id>

# Log-based agents — scan and save
causetrace opencode --save
causetrace continue --save
causetrace codex --save
causetrace copilot --save

# Aider — run with tracing
causetrace aider -- --model gpt-4 --yes "fix the bug"

Usage notes:

  • Claude Code — most precise, captures full causality via Pre/Post hooks
  • Aidercausetrace aider --save -- [aider args] wraps the CLI; best-effort parsing from output
  • Continue.dev, Codex CLI, Copilot — post-hoc log scanning; causality inferred from temporal proximity via infer_relations()
  • All log-based agents infer causality heuristically — timestamps between events determine parent→child chains

Quick Start

pip install causetrace

# Run a demo with sample data
causetrace timeline ses_10d2f16e
causetrace tree    ses_10d2f16e
causetrace replay  ses_10d2f16e --summary
causetrace why     ses_10d2f16e <event_id>

Hook up Claude Code

Add to ~/.claude/settings.json to start recording every Claude Code session automatically.

{
  "hooks": {
    "PreToolUse": [{ "matcher": "*", "hooks": [{
      "type": "command",
      "command": "python3 /path/to/causetrace/hooks/claude_code.py",
      "timeout": 5
    }]}],
    "PostToolUse": [{ "matcher": "*", "hooks": [{
      "type": "command",
      "command": "python3 /path/to/causetrace/hooks/claude_code.py",
      "timeout": 5
    }]}]
  }
}

Scan OpenCode logs

causetrace opencode --save

Parses OpenCode log files, infers causal relations from temporal proximity, and saves as a causetrace session.


Data Model

Every event is a ToolEvent. The four causal fields (parent_event_id, session_id, event_type, caused_by) distinguish causetrace from flat logging systems.

Field Description
event_id UUID
parent_event_id Causal parent (comma-separated for fan-in)
session_id Owning session
tool_name e.g. Bash, Read, Write
tool_input Serialized input arguments
tool_output Serialized output
timestamp ISO 8601
duration_ms Execution time
event_type tool_call / reasoning / context_update / user_input
caused_by user / reasoning / event_id / semantic tag

CLI Reference

Command Description
causetrace timeline <id> Flat chronological view
causetrace tree <id> Causal parent→child tree
causetrace graph <id> Multi-parent DAG (fan-in)
causetrace sessions List recorded sessions
causetrace export <id> Export as JSON
causetrace replay <id> Replay with provenance
causetrace why <id> <eid> Trace causal chain from event
causetrace opencode [--save] Scan OpenCode logs
causetrace aider [--save] -- [args] Run aider with tracing
causetrace continue [--save] Scan Continue.dev logs
causetrace codex [--save] Scan OpenAI Codex CLI logs
causetrace copilot [--save] Scan GitHub Copilot agent logs

Architecture

┌──────────────┐  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐
│ Claude Code  │  │   OpenCode   │  │    Aider     │  │ Continue.dev │  │  Codex CLI   │  │   Copilot    │
│  (hooks)     │  │ (log tail)   │  │ (subprocess) │  │ (log tail)   │  │ (log tail)   │  │ (log tail)   │
└──────┬───────┘  └──────┬───────┘  └──────┬───────┘  └──────┬───────┘  └──────┬───────┘  └──────┬───────┘
       │                 │                 │                 │                 │                 │
       ▼                 ▼                 ▼                 ▼                 ▼                 ▼
┌────────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                    TraceRecorder                                                    │
│                             (causal linking, storage)                                               │
└────────────────────────────────────────────────────────────────────────────────────────────────────┘
                                                │
                                                ▼
┌────────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                           JSONStore                                                │
│                                (append-only JSONL, no DB)                                          │
└────────────────────────────────────────────────────────────────────────────────────────────────────┘
                                                │
                                                ▼
┌────────────────────────────────────────────────────────────────────────────────────────────────────┐
│                              Tree / DAG Builders                                                    │
│                              Renderers / ReplayEngine                                               │
└────────────────────────────────────────────────────────────────────────────────────────────────────┘
Module Responsibility
causetrace/core.py Data model, TraceRecorder, JSONStore, tree/DAG builders, renderers, ReplayEngine
causetrace/causality.py Temporal causal inference for unstructured logs
causetrace/cli.py argparse CLI dispatching to 12 subcommands
causetrace/hooks/ Agent-specific bridges and tailers
causetrace/hooks/claude_code.py Claude Code hook bridge
causetrace/hooks/opencode_tailer.py OpenCode log tailer
causetrace/hooks/aider_bridge.py Aider subprocess wrapper
causetrace/hooks/continue_tailer.py Continue.dev log tailer
causetrace/hooks/codex_tailer.py Codex CLI log tailer
causetrace/hooks/copilot_tailer.py GitHub Copilot log tailer

Development

git clone https://github.com/milkoor/causetrace.git
cd causetrace
pip install -e .
python -m pytest tests/ -v

License

MIT

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

causetrace-0.1.0.tar.gz (28.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

causetrace-0.1.0-py3-none-any.whl (29.4 kB view details)

Uploaded Python 3

File details

Details for the file causetrace-0.1.0.tar.gz.

File metadata

  • Download URL: causetrace-0.1.0.tar.gz
  • Upload date:
  • Size: 28.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.3

File hashes

Hashes for causetrace-0.1.0.tar.gz
Algorithm Hash digest
SHA256 a725e4431023fa6ee7f018752079959ffdb0099b5a2ebf4d4e71f872944f8aae
MD5 3688a018a03ef4c3004542e337ead843
BLAKE2b-256 f5eb0e9337546779e9d3a30f77cbc6608b8ed1dc34bd1f6997d3c14604d4e3e7

See more details on using hashes here.

File details

Details for the file causetrace-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: causetrace-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 29.4 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.3

File hashes

Hashes for causetrace-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 2d08b4ca9063bf469d5f4724bd8d374fd97b6e07789aba8427c2ee2bbfe7147b
MD5 30f30de1de398dba82474c47381d94d4
BLAKE2b-256 c02253d06f0952524d8e44abcddb29497af3c97a85176033012a965cfa8fdca2

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page