A user-friendly adaptive second-opinion skill for Claude Code. For important decisions, plans, strategies, proposals, architecture, and recommendations: frame the real goal, verify decision-critical evidence, pressure-test the biggest risk, compare alternatives, and stop when more reasoning will not change the action.
Install Cruciblev1.0.0
Crucible is both a Claude Code plugin (.claude-plugin/plugin.json) and a marketplace (.claude-plugin/marketplace.json). No external database, service, or model API key is required for normal plugin use.
Option A — Install from Marketplace (Recommended)
Step 1 — Add the marketplace repository:
/plugin marketplace add smshahbaj/crucible
Step 2 — Install the plugin:
/plugin install crucible@crucible-marketplace
Step 3 — Run Crucible on any decision:
/crucible Pressure-test this decision before I commit: should we choose architecture A or B?
Option B — Local checkout with --plugin-dir
Clone and point Claude Code to the repository root:
git clone https://github.com/smshahbaj/crucible.git ~/crucible && claude --plugin-dir ~/crucible
Requirements
Claude Code installed and operational.
Python 3.10+ only for optional tests, benchmarks, or ledger tools.
Zero external databases or API keys required for regular plugin use.
Troubleshooting
Symptom
Probable Cause
Resolution
Plugin not found
Marketplace not registered
Run /plugin marketplace add smshahbaj/crucible first
Command not found
Claude Code session needs refresh
Restart Claude Code session after installation
Ledger script error
Python version below 3.10
Ensure python3 --version is 3.10 or newer
Adaptive Scrutiny Architecture
“More reasoning is not automatically better. The right amount of reasoning is better.”
Crucible does not turn every request into a large multi-agent debate. It chooses a review depth based on stakes, uncertainty, reversibility, and evidence needs.
🟢 QUICKLow-Stakes, Clear & Reversible
Lightweight decisions stay lightweight with a single-pass review for clear, easily reversible choices.
🟡 REVIEWMeaningful Trade-offs & Uncertainty
Balanced multi-perspective evaluation using two independent lenses for non-trivial decisions.
🔴 DEEPHigh Downside, Conflicting Evidence or Irreversible
Three to five heterogeneous lenses plus targeted verification, failure-first analysis, and counterfactual pressure-testing.
Review Pipeline
Decision → Frame real goal → Route by stakes, uncertainty & reversibility → Targeted specialist lenses → Evidence & verification → Challenge / failure test → Quality control → Stable recommendation.
Why "More Agents ≠ More Truth"
Having 14 agents echo the same unverified premise is not consensus; it is anchored bias. Crucible enforces independence before collaboration, subjects critical claims to verification gates, and halts when additional reasoning will not change the chosen action.
14 Specialist Agents
Crucible conditionally routes across 14 dedicated agent lenses to challenge assumptions without redundant bloat:
claim-compressor
Isolates atomic factual claims and assertions from context.
decision-router
Assigns adaptive depth (QUICK, REVIEW, DEEP) based on stakes.
decision-strategist
Structures strategic trade-offs, opportunity costs, and roadmaps.
evidence-auditor
Verifies claims against codebase invariants and empirical evidence.
evidence-gate
Subjects decision-critical claims to strict validation gates.
execution-specialist
Audits implementation hurdles, migrations, and delivery feasibility.
failure-analyst
Pre-mortem specialist hunting edge cases and cascading failure modes.
final-judge
Synthesizes specialist views into an auditable recommendation.
option-analyst
Discovers and evaluates viable alternative architectures.
quality-controller
Enforces stop-when-stable rules and convergence quality.
quantitative-analyst
Calculates expected value, bounds, and probabilistic risks.
risk-red-team
Adversarially probes blind spots and optimistic assumptions.
source-auditor
Audits licenses, dependency health, and supply-chain safety.
verification-planner
Creates concrete, falsifiable tests to validate assertions.
Documentation & References
Crucible's architecture is fully documented across reference modules in skills/crucible/references/:
Adaptive Routing (routing.md)
Determines whether to execute QUICK (single-pass), REVIEW (two lenses), or DEEP (3–5 lenses + verification) based on reversibility, downside impact, and uncertainty.
Evidence Gates (evidence.md)
Enforces distinction between hard codebase facts and speculative assumptions. Critical claims require empirical verification.
Adversarial Red-Teaming (adversarial.md)
Anti-anchoring methodology that forces independent lens generation before leading conclusions become bias anchors.
Convergence monitoring applying stopping rules when further deliberation is unlikely to change the action.
Decision Ledger (ledger.md)
Local durable storage schema and retrospective calibration routines in .crucible/ledger.jsonl.
Frequently Asked Questions
What is Crucible?
Crucible is an adaptive decision-review skill for Claude Code. It pressure-tests important decisions, plans, proposals, architecture choices, and recommendations by framing the real goal, verifying decision-critical evidence, challenging the weakest point, and stopping once the decision is stable.
How do I install Crucible?
Run /plugin marketplace add smshahbaj/crucible in Claude Code, then /plugin install crucible@crucible-marketplace. Alternatively, clone the repository and start Claude Code with claude --plugin-dir ~/crucible for a local install.
Does Crucible require an API key?
No. Normal Crucible plugin use requires no database and no separate model API key. An API credential is only needed for the optional live-model evaluation/benchmark path.
Does Crucible run all 14 agents on every question?
No. Crucible routes each decision to one of three depths — QUICK, REVIEW, or DEEP — based on stakes, uncertainty, reversibility, and evidence needs. Low-stakes reversible questions stay lightweight; only high-stakes or conflicting-evidence decisions receive a deeper multi-agent review.
Is Crucible free and open source?
Yes. Crucible is released under the MIT license and is free to use, modify, and self-host. The source code is available at github.com/smshahbaj/crucible.
What is the difference between QUICK, REVIEW, and DEEP mode in Crucible?
QUICK is a single-pass review for low-stakes, clear, reversible choices. REVIEW uses two independent lenses for decisions with meaningful trade-offs or uncertainty. DEEP uses three to five heterogeneous lenses plus targeted verification and challenge, reserved for high downside, conflicting evidence, or irreversible decisions.
What is Crucible not?
Crucible is not a promise of perfect answers, not "run 14 agents on everything," not a majority-vote truth machine, and not a replacement for primary evidence or qualified professional advice.
Does it expose hidden reasoning?
No. Crucible does not expose private internal model reasoning or hidden agent scratchpads; it provides clean, auditable, structured decision summaries.
What license is it under?
Crucible is distributed under the open-source MIT License.
Local Decision Ledger
The optional Decision Ledger lets you record decisions, reasoning, risks, and real-world outcomes locally in durable .crucible/ledger.jsonl format.