task-autoresearch

Read-only investigator for autoresearch tasks. Use when the parent agent needs to gather pointers from large captures (multi-MB HAR, sprawling logs, unfamiliar codebases) without polluting its main context window. Returns a structured evidence list — file paths with line numbers, log timestamps, HAR entry indices, extracted artifacts — not analysis.

Read-only diagnostic specialist for autoresearch reverse-analysis tasks. Use when the parent agent has gathered evidence and needs to identify the earliest actionable failure point or recover a protocol/contract boundary. Returns a failure-point analysis distinguishing observed evidence, inference, and unknowns — not a fix.

Verification specialist for autoresearch development tasks. Use after a code change to run the narrowest useful check (unit test, type check, lint of changed lines, replay of the originally failing input) and report a clean pass/fail with evidence. Does not modify code.

Use when an autoresearch task needs its final report rendered from the collected evidence — e.g. when the user asks for "the report", "write up the findings", "produce REPORT.md", or when verification has just passed and the task is closing. Renders the framework's standard report shape from the conversation's evidence trail.

Use when an autoresearch task needs the narrowest useful verification step run and its outcome captured into the evidence trail. Trigger after a scoped change has been made, or when the user asks to "verify", "run the verification", "check the change", or passes an explicit verification command.

Use when an agent must run a bounded development task, automation analysis, reverse analysis, HAR/log investigation, or codebase diagnosis with evidence and verification. This is the entry skill — pick the mode, scope the boundary, work the loop, and hand off to the verify and report skills when the task closes.