What problem does it solve?
Agent task runs often waste time and tokens on environment failures, missing CLI tools, verbose command output, and redundant retries, but these inefficiencies are buried inside tens of megabytes of raw JSONL logs that nobody reviews. This Skill turns a completed run log into a small set of measured, evidence-backed findings filed through the report_insight tool.
Core Features & Use Cases
- Bounded log querying: Detects the log format (pi or ACP) and applies verified jq recipes that cap output with head and string slicing, so multi-megabyte logs never flood the context.
- Finding taxonomy: Classifies inefficiencies into eight categories such as environment_failure, missing_tool, verbose_output, redundant_work, and wasted_retry, with measured wasted effort in tool calls, seconds, tokens, and output bytes.
- Verified reporting: Submits each finding via report_insight with verbatim evidence quotes that the tool checks against the raw log, plus concrete suggested fixes with checkable done-when conditions.
- Use Case: After an agent run felt slow, attach its run-log.jsonl and ask for an analysis; the Skill locates the log, finds that the test suite failed twice because Postgres was not running, measures 14 wasted tool calls and 210 seconds, and files a finding recommending Postgres be pre-started in the sandbox.
Quick Start
Analyze the attached run log for inefficiencies and report each finding with report_insight.
