trajectory-review

# Trajectory Review ## Overview When an agent run fails, the failure mode is rarely "the model is bad". It is usually one of a small set of recurring problems on the trajectory: the agent skipped a planned step, invented a fact, called a tool wrong, misread a tool's output, or pursued the wrong subgoal entirely. Output-only review cannot distinguish these — they all surface as "the answer was wrong". This skill takes a captured trajectory (tool calls, intermediate outputs, final response) and locates the **first unrecoverable step**, classifies it into one of nine categories, and reports the root cause with citations into the trajectory. The taxonomy and four-stage diagnostic procedure are adapted from Microsoft Research's AgentRx framework (2026-04). The classes themselves are stable across agent stacks; the diagnostic stages are how this skill operates inside a Claude Code or Codex session. ## When to Activate - An agent run produced a wrong or incomplete result and you have the trajectory. - A user reports "the agent is broken" and a postmortem is needed. - A regression appeared after a model upgrade and you need to know whether it is a model issue or a harness issue. - A new capability shipped and you want to characterize the failure modes that remain. - A user says "review the trajectory", "diagnose this run", or "why did the agent fail". Do **not** use this skill to evaluate a passing run. For "evals pass but I don't trust it", use the W-18 three-axis evaluation ...