Skip to content

swarm-flight-recorder.md

Latest commit

 

History

History
31 lines (20 loc) · 2.29 KB

swarm-flight-recorder.md

File metadata and controls

31 lines (20 loc) · 2.29 KB
Raw

Swarm Flight Recorder

The swarm flight recorder is a deterministic E2E evidence harness for multi-agent runs. It records redacted runtime events from real AgentSession execution, built-in tool calls, JS extension hooks, session persistence snapshots, and external coordination markers into JSONL rows with schema pi.swarm.flight_recorder.event.v1.

The harness is designed for replay without live provider credentials. Tests use deterministic in-process providers and real Pi runtime components, so operators can inspect timing and coordination behavior without depending on OpenAI, Anthropic, or other provider accounts.

Artifacts

  • swarm_flight_recorder.jsonl: append-only event rows with correlationId, agentName, component, eventKind, redaction summary, and redacted payload.
  • swarm_flight_recorder_report.json: summary report with schema pi.swarm.flight_recorder.report.v1, replay command, dominant latency components, component counts, and coordination failures.

Every JSONL row is validated by validate_swarm_flight_recorder_jsonl for current schema, monotonic sequence numbers, and required identity fields. Sensitive payload keys such as tokens, prompts, API keys, cookies, secrets, transcripts, and message content are replaced with [REDACTED], and the row records which keys were redacted.

Replay

Run the focused deterministic replay with:

rch exec -- cargo test --test e2e_swarm_flight_recorder -- --exact multi_agent_flight_recorder_bundle_replays_without_credentials --nocapture

The report embeds the same cargo test command without the rch exec -- prefix so local artifact readers can see the underlying replay target. Agents must still use rch exec -- for CPU-heavy validation in this repository.

What The Harness Proves

  • Multiple Pi sessions can run against isolated temp workspaces in one scenario.
  • Built-in tool execution is real, not a synthetic fixture.
  • Session persistence is exercised through real Session state.
  • JS extension lifecycle hooks observe agent, turn, tool call, and tool result activity.
  • Agent Mail or other coordination failures can be captured as non-blocking markers while Beads remains the soft-lock fallback.
  • The summary report identifies the dominant measured latency contributors and coordination failures from the event bundle.