Span-level replay: fork from the failure

Enables ingesting caller-supplied agent execution traces and comparing them against signed, versioned baselines of normal behaviour to surface silent drift. Reports severity with per-dimension attribution, a replayable exemplar episode, and visible suppression of declared maintenance windows or known change markers.

Exposes an agent skill debugging lab over MCP, letting an agent record model calls at the proxy, find the step where successful and failed runs diverge, fork and resample that step, and A/B test skill wording edits with confidence intervals and significance tests. It also enables live breakpoints, workspace snapshots, end-to-end restart of a recorded run, and token-level analysis of which token decided a step.

Enables coding agents to turn explicit artifact paths or argv commands into bounded runtime evidence by coordinating profilers, benchmark tools, trace processors, and direct local targets through named analysis and capture MCP tools. Results are returned inline with provenance, coverage, limitations, and optional durable evidence IDs for inspection and replay.

Provides structured, sandboxed test and lint feedback for coding agents, returning compact typed verdicts with failure fingerprints instead of raw runner output. It enables impact-selected test execution in isolated containers, distinguishing pre-existing failures from regressions.