What problem does it solve? When an AI agent runs a verify-fix-retry loop for embedded system validation, loop state (attempt counts, failure fingerprints, exit reasons) lives in fragile conversation context and gets lost. This Skill moves that bookkeeping into a persistent session.json state machine so scripts handle mechanical accounting while the AI focuses on failure analysis and code fixes. ## Core Features & Use Cases - Dual-layer counting: patience counter (increments only when the same failure fingerprint freezes, cap 3) plus total_attempts cost guardrail (cap 10), preventing infinite retry loops. - Failure fingerprinting: hashes (failed stage, verify exit code, normalized first error line) with timestamp/path/hex-address normalization to distinguish a frozen problem from an evolving one. - Five-way attribution and exit protocol: classifies each round as pass, task_fail, env_fail, framework_error, or param_error, and terminates with pass, task_unsolvable, cost_cap_exceeded, env_fail, or framework_error, generating a diagnosis report skeleton. - Use Case: After a Raspberry Pi 5 build fails acceptance testing, start a session, run the workspace-verify workflow, record each receipt with done, and let the state machine decide when to stop retrying and emit a diagnosis for handoff. ## Quick Start Start a verification convergence session for target dev with a verify-cases label, then follow the run guidance and record each receipt with done until the session terminates.
