harness

Plugin harness from marketplace autosolutions in autosolutionsai-didac/autosolutions-marketplace. Commands read from the repo's manifest — 26 days ago.

Skills (5)

• gate

  Turn the expensive stop gates on or off for this project. "on" adds a Fable agent hook that reviews the session's changes before Claude may finish, "off" removes it, "codex" enables the OpenAI Codex review gate, "status" shows what is active. Usage: /harness:gate on|off|codex|status

• harness

  How to delegate with the AutoSolutions harness. Use on any coding task that involves searching a codebase, implementing a change, reviewing work, or deciding whether something is done. Explains which role subagent (scout, web-scout, builder, refuter, judge) and which workflow (/harness:ship, /harnes

• judge

  One-off Fable verdict on the current change. Gathers the diff and the verification evidence, runs the harness:judge subagent once, and reports PASS or FAIL with blocking issues. Usage: /harness:judge [base-branch]

• second-opinion

  Get an independent review from a GPT model (a different model family) on the current diff or on a plan. Uses the OpenAI Codex plugin when installed, otherwise the bundled openai-judge MCP server. Usage: /harness:second-opinion [diff|plan] [focus]

• setup

  Check and finish the harness installation. Verifies Claude Code and Node versions, that the roles load, which optional gates have keys, and offers the Haiku override for the built-in Explore subagent and the Codex plugin install. Usage: /harness:setup

Subagents (5)

• builder

  Implements one well-scoped unit of work on Sonnet 5.5, runs the relevant checks, and reports what changed and what was verified. Use for implementation after the plan is clear, not for exploration or design.

• judge

  Final release gate on Fable 5.1. Use once per change, after refutation, to decide if work is ready to merge or ship. Expensive by design, reads diffs and evidence rather than whole repos. Returns PASS or FAIL with blocking issues.

• refuter

  Adversarial reviewer on Opus 5.5 that tries to break a change before anyone trusts it. Use after implementation and before claiming done. Read-only plus Bash to run checks, returns a verdict with evidence.

• scout

  Read-only codebase search on Haiku 5.5. Use before any change to find files, definitions, call sites, config and patterns. Returns file:line evidence and short quotes, never edits.

• web-scout

  Read-only web research on Haiku 5.5. Use to fetch docs, changelogs, pricing and announcements. Returns sources with URLs, dates and quotes, never conclusions.

MCP servers (1)

• openai-judge

  runs: node ${CLAUDE_PLUGIN_ROOT}/scripts/openai-judge-mcp.mjs

Hooks (2)

• SessionStart

  node

• Stop

  node

Similar plugins