What problem does it solve? When a multiplai subsystem misbehaves — a hook times out, a context injection makes no sense, a learning pipeline silently degrades — the evidence is scattered across a dozen log files, and guessing at root causes wastes time. This Skill turns the runtime log directory into an actionable, source-verified fix report. ## Core Features & Use Cases - Passive log scanning: Clusters errors and warnings across subsystems (context_manager, extract_learnings, backfill, dream, session hooks), detects health anomalies, and verifies each root cause against the actual source code before recommending a fix. - Hook timing analysis: Pairs HOOK_ENTRY/HOOK_EXIT lines to show which hook timed out, where its budget went (startup cost, stage breakdown, p95 vs ceiling), and which prompt lost its context. - Injection forensics: Reconstructs context-routing decisions by joining router scores with inject events, explaining why a file was or was not injected (cooldown suppression, low floor, prompt attribution). - Probe mode: Actively exercises a functionality (session lifecycle, deep-research, catalog generation) and asserts its expected log entries appeared, catching missing logging as a bug. - Use Case: A user asks "why is extract_learnings failing?" — the Skill scans recent logs, clusters the recurring traceback, reads the failing code path, and writes a severity-ranked report with a concrete fix target to the workspace INBOX. ## Quick Start Ask the assistant to run log-doctor to analyze the multiplai logs from the last seven days and produce a health report with verified root causes.
