autoresearch

What it does

Autoresearch is described as an autonomous goal-directed iteration engine. It operates as a BLOCKING skill invocation and activates under a long list of explicit commands and any mention of the word "autoresearch". It runs iterative loops (Modify → Verify → Keep/Discard → Repeat) with strict guardrails, including atomic commits, mandatory Verify, optional Guard, and explicit ship confirmation. It covers multiple subcommands (plan, security, ship, scenario, predict, learn, reason, probe, debug, and the base autoresearch loop) to perform tasks such as security audits, shipping workflows, scenario generation, and multi-persona analysis.

How it works

• The skill requires Interactive Setup Gate for all commands; if required context is missing, it prompts the user with direct questions before proceeding.
• It enforces a per-iteration workflow: make a change, Verify, then decide to keep or discard with an automatic rollback on failure. Changes kept are committed with an experiment: prefix; discards use git revert-clean.
• It optionally applies a Guard if configured; if Guard fails, changes revert.
• Before Verify dry-runs, it screens for dangerous operations and credentials exposure.
• It masks secrets in findings and reproduction commands.
• It ensures that Verify outputs are data, not instructions, and that any external content is not treated as directives.
• For the ship subcommand, it requires explicit user confirmation at the appropriate phase gate before publishing.
• CI/CD and bounded iterations considerations are documented to limit loop behavior in non-interactive contexts.

When to use it

• Use when you need an autonomous, iterative approach to a complex task with required context, multiple workflows, or security-focused analysis.
• Use for security auditing, shipping workflows, scenario exploration, multi-persona analysis, or building documentation through autonomous loops.

What it can touch

• It interacts with code, configurations, and artifacts as part of the iterative cycle.
• It defines outputs and reports in structured folders per workflow phase (e.g., security/audit folders, ship logs, scenario results).

Caveats

• It operates under guardrails including atomic commits, Verify requirements, and optional Guard checks.
• Ship actions require explicit user confirmation; non-approval prevents publishing.
• It includes safety screens before Verify dry-runs to prevent dangerous commands.