~/projects/crucible — claude --plugin crucible

Crucible · v0.4.0 · MIT · Evidence-gated

Make Claude Code show its work.
"Looks correct." isn’t a verdict.

Crucible is the plugin that refuses completion without proof. Every plan is reviewed by three independent oracle auditors. Every verdict cites a specific evidence file. Mocks, stubs, and self-review are not advisory — they are structurally blocked. The result is a repeatable forge pipeline where Claude Code can’t mark a task done by saying “looks correct.”

View on GitHub →Install ↓Read the launch post ↗

Skills

Agents

Commands

Hooks

Rules

01What it does

Three properties. Zero compromises.

Crucible is not a linter, not a test runner, not a CI script. It is a plugin that gates completion. The producer is never the reviewer. The reviewer is never one agent. The verdict is never "looks correct."

Refuses without proof

Crucible's completion-gate rejects any verdict that doesn't cite a real evidence file. No screenshots, no curl response, no log line — no PASS. Refusal is a feature, not a failure.

Quorum, not single-agent

Three independent oracle auditors review every plan and every outcome. Two of three must approve. The producer never reviews their own work. Iron rule, not advisory.

Real systems only

No mocks. No stubs. No test files. Validators run real CLIs, hit real HTTP endpoints, inspect real filesystems. Compilation is necessary; it is never sufficient.

02Quick start

Install. Forge. Read the verdict.

Crucible ships as a Claude Code plugin. Add the marketplace, install the plugin, then run /crucible:forge on any task. Two examples below — one greenfield, one bug-hunt — drawn from real session logs.

bashexample 1 · install & first forge run

# 1. Add the marketplace and install the plugin.
$ claude plugin marketplace add krzemienski/crucible
$ claude plugin install crucible@crucible

# 2. Activate Crucible enforcement in this repository.
$ /crucible:enable
→ wrote .crucible/active sentinel
→ Iron Rules in force: no-mocks, cite-or-refuse, cite-paths, no-self-review

# 3. Forge a feature end-to-end. Crucible runs the full 10-phase pipeline.
$ /crucible:forge "add /api/health endpoint with real uptime probe"
# Phase 1  — codebase-analyst ........ evidence/codebase-analysis/SUMMARY.md
# Phase 2  — documentation-researcher . evidence/documentation-research/INDEX.md
# Phase 3  — skill-discoverer ......... evidence/skill-enrichment/INDEX.md
# Phase 4  — planner .................. evidence/.../PLAN.md
# Phase 5  — oracle quorum ............ APPROVE 2/3 (oracle-1, oracle-3)
# Phase 6  — executor ................. evidence/.../step-*.{log,json,png}
# Phase 7  — validator ................ evidence/.../verdict.md
# Phase 8  — final audit .............. APPROVE 3/3
# Phase 9  — outcome .................. evidence/.../OUTCOME.md
→ PASS — verdict cites evidence/.../step-04-curl-health.json:1-22

bashexample 2 · the refusal that saved the deploy

# Run a forge against a vague brief. Crucible doesn't fabricate evidence.
$ /crucible:forge "polish the dashboard"
# Phase 4 — planner emits PLAN.md with 7 MSCs but 3 lack evidence paths.
# Phase 5 — oracle quorum: 0/3 APPROVE.
⨯ REFUSE — oracle-auditor-1: "MSC-3, MSC-5, MSC-7 cite directories, not files (RL-4)."
⨯ REFUSE — oracle-auditor-2: "MSC-2 PASS criterion is 'looks better' — not measurable."
⨯ REFUSE — oracle-auditor-3: "no PASS criterion uses real-system evidence."

# Crucible writes REFUSAL.md and stops. No fabricated PASS. No silent retry.
$ cat evidence/.../REFUSAL.md
## Why this run was refused
3 of 3 oracles cited RL-4 (cite paths) violations on PLAN.md.
Re-scope the brief to a measurable change, then re-run /crucible:forge.

# Iterate the brief, not the verdict. Then forge again.
$ /crucible:forge "reduce dashboard LCP from 4.2s to under 2.5s"
→ PASS — verdict cites evidence/.../step-09-lighthouse.json:lcp=2.31s

03Visual explainer

The forge pipeline. One line, ten phases.

Every /crucible:forge run flows through the same ten phases. Each phase writes evidence to disk. Each gate refuses on missing or invalid evidence. The producer is never the reviewer.

04How it works

Ten phases. Two gates.

The forge command is the canonical entrypoint. Each phase has a single agent role, a single output artifact, and a single gate criterion. The producer of an artifact is structurally barred from approving it (RL-3 No Self-Review).

PH 00

Brief

Capture intent

PH 01

Codebase

Analyst maps repo

PH 02

Docs

Researcher cites upstream

PH 03

Skills

Discoverer ranks tools

PH 04

Plan

Planner emits PLAN.md

PH 05

Plan Review

3 oracles vote

PH 06

Execute

Executor + evidence

PH 07

Validate

Validator emits verdict

PH 08

Final Audit

3 oracles re-review

PH 09

Outcome

OUTCOME.md or REFUSAL.md

05Feature matrix

What ships in v0.4.0.

Counts grounded in the on-disk plugin tree at ~/.claude/plugins/cache/crucible-local/crucible/0.4.0/. No counts are estimated; each row enumerates the actual files.

One line. One marketplace.

Crucible installs as a Claude Code plugin. The plugin ships with all 11 skills, 10 agents, 19 commands, 4 hooks, and the 4 iron-rule fragments. /crucible:enable writes the .crucible/active sentinel that turns enforcement on for this repo.

bashinstall crucible into Claude Code

# Add marketplace + install plugin.
$ claude plugin marketplace add krzemienski/crucible
$ claude plugin install crucible@crucible

# Activate enforcement in the current repository.
$ /crucible:enable

# Run the doctor to verify hooks + skills are loaded.
$ /crucible:doctor
→ hooks.json present, 4 iron-rule fragments composed into CLAUDE.md
→ 11 skills loadable, 10 agents discoverable

# Forge anything.
$ /crucible:forge "<your task>"

07Documentation

Read the source. Then forge.

Crucible has no SaaS, no telemetry, no closed source. Every skill, agent, command, and rule lives in the public repository. Below: the eight pages most worth your time on day one.

01 · Start

Setup & Activation

02 · Forge

End-to-end Forge Pipeline

A verdict without a citation is invalid. A citation to a directory is invalid. A reviewer who is also the author is invalid. This is not advisory.

08Changelog

From single-validator to triple oracle quorum.

Crucible has been on monthly minor versions since January 2026. Major moves: oracle quorum (v0.3), trials and resume (v0.4). The roadmap below v0.5 is being shaped by real forge runs against this very site.

v0.4.0

Apr 2026

Trials + Resume. Reproducible /crucible:trial sandboxes; /crucible:resume picks up halted forge runs without re-running prior phases.

v0.3.0

Mar 2026

Triple Oracle Quorum. Plan-review and final-audit each convene 3 independent auditors. Two of three must APPROVE. Self-review is structurally impossible.

v0.2.0

Feb 2026

Skill Enrichment. Phase 2.5 ranks every plugin skill against the brief and cites the top-N in PLAN.md so the executor never invents skill names.

v0.1.0

Jan 2026

Initial release. Forge command, codebase + docs research phases, single-validator gate, no-mocks rule.

Make Claude Code show its work. "Looks correct." isn’t a verdict.