What problem does it solve? When an AI coding agent struggles in a repository, the evidence of why disappears with the session. Teams guess at which codebase problems actually cost time, and static analysis can only infer difficulty rather than measure it. This Skill converts completed session transcripts into structured friction observations with quantified cost, so repo improvements are ranked by real measured waste instead of intuition. ## Core Features & Use Cases - Transcript-Based Signal Detection: Walks each session episode against a catalogue of friction signals (wasted reads, search churn, repair loops, silent failures) and records each as observed, clean, or not-applicable. - Measured Cost Attribution: Counts wasted tool calls, unused file reads, and estimated tokens per signal, and distinguishes repo-caused friction from agent-caused mistakes using explicit attribution tests. - Observation Store and Consolidation: Emits JSONL records per session into a .friction/ store, which a weekly consolidation step groups by fingerprint, ranks by total cost, and promotes only patterns seen across multiple sessions. - Use Case: After a week of agent runs, consolidate the friction store to discover that one ambiguously named type caused nine failed searches across six sessions costing roughly 154k tokens, then file that as a ranked, evidence-backed finding. ## Quick Start Ask the agent to audit the transcript of the session that just finished and record any friction observations with measured cost into the friction store.
