fiddle

Fiddle

Claude Code plugin for orchestrating a four-phase development lifecycle with a calibrated evaluator loop and multi-model support.

Orchestrate

/fiddle:orchestrate <topic> chains four phases. Each phase is also a standalone skill.

DISCOVER /fiddle:discover — Scan project docs, research the ecosystem via external providers (Claude + Codex), and challenge scope assumptions until every branch is resolved.

DEFINE /fiddle:define — Brainstorm approaches, run a multi-model adversarial panel (Claude + Codex + Gemini), challenge the chosen design, then produce an implementation plan with sized beans.

DEVELOP /fiddle:develop — Execute beans via the evaluator loop: dispatch implementer → dispatch per-domain, per-provider evaluators → check thresholds → converge or iterate. Multi-domain evaluation with holistic cross-domain review when all tasks complete.

DELIVER /fiddle:deliver — Drift analysis via external providers (Claude + Codex), evaluator evolve step (calibration updates, antipattern capture, threshold tuning), update technical docs, close the epic.

[!NOTE] Any CLI that accepts a prompt on stdin works as a provider (Codex, Gemini, Copilot, etc). Configure per phase in orchestrate.json.

Evaluator Loop

The develop phase uses a calibrated evaluator loop: dispatch implementer → dispatch evaluators → check thresholds → converge or iterate. Script-enforced convergence (two consecutive passes required) with a dispatch budget that prevents infinite iteration.

Per-domain evaluation. Tasks spanning multiple domains (frontend + backend) get independent per-domain scoring. resolve-domains.sh maps task domains to evaluator templates. Domain-specific templates (frontend, backend) replace the general template for typed projects.

Runtime verification. Evaluators launch and interact with the running app via project-configured MCP tools (marionette, curl, go-dev-mcp) to gather evidence beyond static code review.

Multi-provider scoring. Multiple LLM providers evaluate each task for diversity of judgment. Conservative scoring (min across providers per dimension). Disagreements (spread >= 3) surfaced in eval logs.

Holistic review. After all tasks complete, a cross-domain integration review produces a spec coverage matrix and remediation loop for any gaps.

Calibration + evolve. Attended mode shows scorecards to humans before acting — corrections become calibration anchors in project-specific files. Antipattern files loaded by implementer and evaluator. The deliver evolve step encodes improvements for future runs.

Skills

Skill	Description
fiddle:brainstorm	Collaborative design dialogue with calibration anchor extraction.
fiddle:write-plan	Generate implementation plan from a design spec.
fiddle:evaluate	Evaluator protocol — score an implementation against domain template and criteria.
fiddle:challenge	Walk the decision tree on any plan or design until shared understanding.
fiddle:panel	Multi-model adversarial debate — Claude, Codex, Gemini argue positions and cross-review.
fiddle:tdd	Test-driven development workflow.
fiddle:verify	Run checks and verify implementation.
fiddle:discover-docs	Socratic dialogue to bootstrap or review project docs.
fiddle:deliver-docs	Post-ship doc updates — SYSTEM.md, ADRs, BACKLOG.
fiddle:define-beans	Task sizing rules for decomposing plans into beans.
fiddle:adr	Create an architecture decision record.
fiddle:feedback	Append a user feedback signal.
fiddle:backlog	Append an idea or tech debt item.
fiddle:debug	Structured debugging workflow.

Configuration

Orchestrate reads orchestrate.json from the project root. All keys optional — defaults apply when omitted. See fiddle:orchestrate for the full reference.