AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
翻译准备中
trace : store : search : analyze local agent interactions
翻译准备中
取自仓库 README
正在加载完整 README。 以下为已保存的摘录。
Trace every agent session on your machine — Claude Code, Codex, Cursor, Copilot —
into one history you and your agents can manage, search, and evaluate.
Thirdeye persists each session as a durable, append-only local event history.
Prompts, responses, tool calls, subagents, usage, and session metadata remain
available after the original agent process exits, giving both people and agents
a consistent record to inspect. Connect thirdeye to Pydantic Logfire to make
full use of that tracing data: explore complete execution trees, investigate
individual turns, and analyze behavior across sessions.查看完整 README ↗Trace every agent session on your machine — Claude Code, Codex, Cursor, Copilot — into one history you and your agents can manage, search, and evaluate.
Thirdeye persists each session as a durable, append-only local event history. Prompts, responses, tool calls, subagents, usage, and session metadata remain available after the original agent process exits, giving both people and agents a consistent record to inspect. Connect thirdeye to Pydantic Logfire to make full use of that tracing data: explore complete execution trees, investigate individual turns, and analyze behavior across sessions.
Visit third3y3.com for guides and the full command reference.
Homebrew is the simplest option on macOS and Linux. On Windows, use pipx
or uv.
brew install duncankmckinnon/tap/thirdeye
Alternatively, install with pipx:
pipx install thrdi
Or use uv or pip:
uv tool install thrdi
pip install thrdi
Run the guided setup after installation:
thirdeye setup
The setup flow:
You can rerun setup whenever you add an agent or change your Logfire configuration.
Pydantic Logfire is where thirdeye's captured data becomes a full observability experience. Once connected during setup, each session appears as an OpenTelemetry trace containing agent turns, model calls, tool executions, subagents, and available token usage and attributed costs. Use Logfire to navigate the complete execution flow, query across sessions, compare agent behavior, and investigate where time and tokens were spent.
Export happens outside the agent hook, so network latency does not interrupt the agent session.
See third3y3.com for Logfire configuration and export details.
Thirdeye includes focused skills that let coding agents work directly with the history it captures:
use-thirdeye — inspect sessions, search events, debug tool calls, and
analyze captured usage.thirdeye-evals — create evaluation rubrics, run evaluators, and interpret
per-turn findings.thirdeye-review — audit agent behavior for inefficiencies and recommend
concrete invocation or skill improvements.thirdeye-filter — translate natural-language questions from the UI into
structured session and turn filters.thirdeye-logfire — configure, verify, and troubleshoot thirdeye's
Logfire integration.The setup flow can install these for you. To add them later, run
thirdeye skills add; use thirdeye skills list to see what is available.
Contributions are welcome. See CONTRIBUTING.md for local development, validation, and pull-request guidance.