Critique proposed changes against project principles before they land.
npx skills add https://github.com/arpitarya/fux --skill fux-critic-arpitarya
Please help me install this Agent Skill. Skill: fux-critic Source: https://github.com/arpitarya/fux/tree/main/archive/v0.1/fux/data/skills/critic Command: npx skills add https://github.com/arpitarya/fux --skill fux-critic-arpitarya
fux critic to evaluate hard invariants (money, PII, numbers, audit) as code at zero cost, exiting with code 2 when violated. - Judgment self-critique: Lists needs-judgment principles so the host agent critiques its own proposal in-session, producing pass/fail verdicts with rationales. - Advisory-first escalation: Judgment failures suggest rather than block by default, with bounded revision rounds (max 3), escalation to humans, and optional debate for contested calls. - Use Case: Before committing a change that touches billing logic, run the critique to verify deterministic invariants pass and self-assess tone, completeness, and grounding principles, with all verdicts recorded to .fux/out/critic.jsonl for audit. ## Quick Start Ask the agent to run /fux critic with a summary of your proposed change before committing it.High-intent search queries and answers about installing and using this skill.
Run `fux critic "<proposed change>"` with a diff summary or commit message before the change lands. It recalls relevant principles, runs deterministic checks, and lists judgment principles for the agent to self-critique.
Deterministic principles (money, PII, numbers, audit) run as code checks that always block and exit with code 2 when violated. Judgment principles (tone, completeness, grounding) are evaluated by the agent in-session and are advisory by default.
No. The deterministic pass runs entirely locally at zero cost with no API call. Judgment critique happens in the host agent's existing session using tokens already being spent, so Fux itself never calls a model.
Yes. Add the principle id to `critic_block_judgment` in `.fux/config.toml`, or set it to true to block all judgment principles. Deterministic principles always block regardless of this setting.
Revision is capped at 3 rounds; after that, escalate to a human with the principle, proposal, and conflict rationale. For borderline judgment calls, the workflow suggests running /fux debate for a two-agent debate rather than deciding alone.
vtexUpdate Spec Kit constitution templates from principle inputs and synchronize dependent files.
5onyyGuides human reviewers through code changes via structured walkthrough, risk, and testing steps.
fryanpanMonitors Claude account token quotas across multiple pools and reports estate status with rotation verdicts.
Open-PawsAutomate layered code review for pull requests in CI pipelines.