What problem does it solve? When an AI agent session shows suspicious token consumption, unexpected API-equivalent costs, stalled tool calls, or confusing sub-agent behavior, there is no easy way to determine what actually happened. This Skill collects bounded, immutable evidence from local session logs and Tokscale usage data, then produces a structured diagnostic conclusion. ## Core Features & Use Cases - Scoped evidence collection: Diagnoses a single identified client/session at turn, session, or both scope, copying Codex logs into immutable snapshots before analysis. - Multi-surface capability reporting: Reports usage, model activity, tool calls, unfinished calls, sub-agent lifecycle, collaboration, and waits with explicit available/unavailable/failed status per capability. - Privacy-preserving conclusions: Reports only metadata and derived observations, never prompts, responses, tool inputs, or credentials, and always expresses amounts as estimated API-equivalent cost. - Use Case: A developer notices a Codex session consumed far more tokens than expected. They invoke this Skill with the client and session ID, and receive a structured report distinguishing observed facts from inference, ending in exactly one conclusion: anomaly observed, no anomaly observed, or inconclusive. ## Quick Start Ask the agent to diagnose the current Codex session for abnormal token usage and tool activity at both turn and session scope.
