self-critique

name	self-critique
description	Audit a completed task against the user's original request before declaring done. Catches omitted constraints, misread scope, partially-met requirements, and unsupported "it's done" claims after long multi-tool loops. Opt-in quality gate — run it, report the verdict, never silently re-loop.
version	1.0.0
author	Nous Research (proposed by @dimokru, issue
license	Apache-2.0
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["quality","self-critique","reflection","verification","audit","review","post-task"],"requires_toolsets":["terminal"]}}

Self-Critique

Audit a finished task against what was originally asked, not against how good the final answer looks. Fluent, well-formatted responses routinely omit a requested constraint, misread scope, or stop one step short — especially after long multi-tool loops where the final message drifts from the initial ask.

When to use

• Right before telling the user a non-trivial task is complete.
• After a long multi-tool run, a hand-off, or a plan with several steps.
• When invoked explicitly (CLI / cron / a /self-critique style request).

Skip it for routine short answers — it adds noise without value there.

Output shape

Always one JSON object:

{
  "verdict": "satisfied | partial | missing | unknown",
  "missing_items": ["concise unmet requirement", "..."],
  "suggested_follow_up": "one short actionable sentence, or empty"
}

• satisfied — every explicit requirement met (missing_items empty).
• partial — some requirements met, some not.
• missing — the core ask is unmet.
• unknown — the audit could not run (no auditor available); never a guess.

How to run

Feed the original request, the final response, and (optionally) the tool trace to the script. It uses Hermes' shared auxiliary client for a cheap audit:

echo '{"original_request": "<the user ask>", "final_response": "<your answer>", "tool_trace": "<optional>"}' \
  | python optional-skills/quality/self-critique/scripts/self_critique.py

Or from a file:

python optional-skills/quality/self-critique/scripts/self_critique.py --input audit.json

You can also call it in-process and inject your own LLM function (used in tests and when embedding). Put the scripts/ dir on sys.path first (or run from inside it):

import sys; sys.path.insert(0, ".../optional-skills/quality/self-critique/scripts")
from self_critique import critique
result = critique(original_request, final_response, tool_trace_json)

Hard rules

• Report only. This skill never edits conversation history and never re-enters the agent loop on its own. Surface the verdict; let the user decide whether to act.
• Opt-in. It is not part of the default tool schema and must not run on every short turn.
• No false confidence. If the auditor is unavailable, return unknown — do not fabricate a satisfied.

What it checks

• Omitted or partially-satisfied explicit constraints.
• Scope drift (answered a narrower/different question than asked).
• Completion claims unsupported by the tool trace.