Conclave is a structured development methodology plugin for coding agents. It provides composable "skills" — TDD, design brainstorming, verification, and code review — that activate automatically based on task type and guide the agent through disciplined workflows. Benchmark data from 796 isolated trials shows these skills improve code quality by 10-12 points over unguided development. Optionally includes multi-agent consensus (Claude, Gemini, Codex) for higher-stakes decisions.

Forked from obra/superpowers under MIT license

It starts from the moment you fire up your coding agent. A task classifier automatically selects the right methodology: brainstorming for new features, TDD for implementation, verification before completion. Before any code gets written, the brainstorming skill teases out what you're actually trying to build through structured questions. You can answer interactively, or switch to Autopilot mode where the agent explores design decisions autonomously while you watch and intervene only when needed.

Once the design is set, the system generates an implementation plan broken into bite-sized tasks. Each task follows TDD discipline: write failing test, implement minimal code, verify, commit. A completion gate ensures fresh verification (test + build + lint) and diff review before any work is declared done.

The result: your coding agent works autonomously for hours at a time, following proven methodology that catches issues before they compound. And because the skills trigger automatically, you don't need to do anything special.

Consensus review is available as an opt-in enhancement at key workflow points:

Brainstorming -> Writing Plans -> Execution -> Debugging -> Verification
     |              |              |           |            |
  Design        Architecture    Per-task    Root cause   Final check
  validation    validation      review      validation   before done

7 skills enhanced with consensus:

• brainstorming - Design validation + autopilot mode
• writing-plans - Architecture/risk/scope validation
• subagent-driven-development - Third review stage after code quality
• executing-plans - Consensus review after each batch
• finishing-a-development-branch - Final review before merge
• systematic-debugging - Root cause hypothesis validation
• verification-before-completion - Multi-agent final check

In brainstorming, you can choose Consensus Autopilot mode:

Two modes available:
1. Interactive - I ask questions, you answer
2. Consensus Autopilot - Multi-agent consensus answers questions,
   you watch and can interrupt anytime to override

The council debates each design decision while you watch. Interrupt anytime to override, go back, or take over.

# Review code changes
./skills/multi-agent-consensus/auto-review.sh "Added authentication"

# Review with explicit base
./skills/multi-agent-consensus/auto-review.sh --base=HEAD~5 "Recent fixes"

# General question
./skills/multi-agent-consensus/consensus-synthesis.sh \
  --mode=general-prompt \
  --prompt="What could go wrong with this architecture?" \
  --context="$(cat design.md)"

Note: Installation differs by platform. Claude Code has a built-in plugin system. Codex and OpenCode require manual setup.

In Claude Code, register the marketplace first:

/plugin marketplace add signalnine/honest-gabes-marketplace

Then install the plugin from this marketplace:

/plugin install conclave@honest-gabes-marketplace

Check that commands appear:

/help

# Should see:
# /conclave:brainstorm - Interactive design refinement
# /conclave:write-plan - Create implementation plan
# /conclave:execute-plan - Execute plan in batches

Tell Codex:

Fetch and follow instructions from https://raw.githubusercontent.com/signalnine/conclave/refs/heads/main/.codex/INSTALL.md

Detailed docs: docs/README.codex.md

Tell OpenCode:

Fetch and follow instructions from https://raw.githubusercontent.com/signalnine/conclave/refs/heads/main/.opencode/INSTALL.md

Detailed docs: docs/README.opencode.md

1. brainstorming - Activates before writing code. Refines rough ideas through questions, explores alternatives, presents design in sections for validation. Saves design document.

2. using-git-worktrees - Activates after design approval. Creates isolated workspace on new branch, runs project setup, verifies clean test baseline.

3. writing-plans - Activates with approved design. Breaks work into bite-sized tasks (2-5 minutes each). Every task has exact file paths, complete code, verification steps.

4. subagent-driven-development or executing-plans - Activates with plan. Dispatches fresh subagent per task with two-stage review (spec compliance, then code quality), or executes in batches with human checkpoints.

5. test-driven-development - Activates during implementation. Enforces RED-GREEN-REFACTOR: write failing test, watch it fail, write minimal code, watch it pass, commit. Deletes code written before tests.

6. requesting-code-review - Activates between tasks. Multi-reviewer consensus from Claude, Gemini, and Codex. Groups issues by agreement level (all agree -> high priority, majority -> medium, single -> consider). Critical issues block progress.

7. finishing-a-development-branch - Activates when tasks complete. Verifies tests, presents options (merge/PR/keep/discard), cleans up worktree.

The agent checks for relevant skills before any task. Mandatory workflows, not suggestions.

796 isolated Docker trials across 34 orchestrator variants on 11 standardized coding tasks. All scoring is deterministic (tests, build/lint, coverage, code metrics) — no LLM judges. Full methodology: docs/plans/2026-02-26-thunderdome-findings-and-roadmap.md

Key findings:

Takeaways:

• Structured methodology is the engine. All discipline genes cluster within 1.4 points of each other. The gap between any discipline and no discipline is massive (+10-12pp).
• TDD is the single most effective methodology. Sonnet + TDD beats every Opus variant at half the cost.
• Model capability is secondary. With strong methodology, Sonnet matches or beats Opus. Structure matters more than model capability.
• Multi-agent consensus is optional. The skill text drives all the value; consensus adds noise on benchmarks. It remains available for users who want multi-perspective validation on high-stakes decisions.

Testing

• test-driven-development - RED-GREEN-REFACTOR cycle (includes testing anti-patterns reference)

Debugging

• systematic-debugging - 4-phase root cause process (includes root-cause-tracing, defense-in-depth, condition-based-waiting techniques)
• verification-before-completion - Ensure it's actually fixed

Collaboration

• brainstorming - Socratic design refinement with optional multi-agent validation
• multi-agent-consensus - Get consensus from Claude/Gemini/Codex on any prompt (design validation, architecture decisions, debugging, code review)
• writing-plans - Detailed implementation plans
• executing-plans - Batch execution with checkpoints
• dispatching-parallel-agents - Concurrent subagent workflows
• requesting-code-review - Multi-reviewer code review using consensus framework
• receiving-code-review - Responding to feedback
• using-git-worktrees - Parallel development branches
• finishing-a-development-branch - Merge/PR decision workflow
• subagent-driven-development - Fast iteration with two-stage review (spec compliance, then code quality)
• ralph-loop - Autonomous iteration wrapper - runs tasks until success or iteration cap hit (fresh context per iteration, stuck detection, failure branches)

Meta

• writing-skills - Create new skills following best practices (includes testing methodology)
• using-conclave - Introduction to the skills system

Conclave includes a message bus for inter-agent coordination, supporting two use cases:

Opt-in debate round between agents during consensus review. After independent analysis (Stage 1), agents see each other's thesis summaries and produce rebuttals before the chairman synthesizes (Stage 2).

# Enable debate round
conclave consensus --mode=general-prompt --debate \
  --prompt="Review this architecture" --context="$(cat design.md)"

# Control rounds and timeout
conclave consensus --debate --debate-rounds 2 --debate-timeout 90 ...
conclave auto-review --debate "Review recent changes"

Wave-scoped boards let parallel ralph-run tasks share discoveries. Tasks emit structured markers in their output:

<!-- BUS:discovery -->The API uses cursor-based pagination<!-- /BUS -->
<!-- BUS:warning -->Package X v2 has breaking changes<!-- /BUS -->
<!-- BUS:intent -->Modifying internal/auth/handler.go<!-- /BUS -->

The orchestrator summarizes each wave's board for the next wave, giving later tasks accumulated project knowledge.

Long-running workflows (multi-wave execution, brainstorming sessions, debugging marathons) can exhaust the context window. Conclave skills include /compact guidance at phase transition points to proactively reclaim context before it becomes a problem.

Skills with built-in compact checkpoints:

Each compact point includes a template with the right focus summary so compaction preserves what matters for the next phase. For example, between waves in subagent-driven-development:

/compact Completed waves 0-2. Tasks done: 1-5. Tasks remaining: 6-8. Next wave: tasks 6, 7.

The focus summary tells /compact what to prioritize when summarizing the conversation, ensuring later phases have clean context with the right details preserved.

Validate SKILL.md files and plan filenames against authoring standards:

# Lint everything (skills/ + docs/plans/)
conclave lint

# Lint specific directory or file
conclave lint skills/brainstorming/
conclave lint skills/brainstorming/SKILL.md

# JSON output for CI
conclave lint --json

# Custom word count threshold
conclave lint --word-limit 1000

Rules checked:

Exit code 0 = clean (or warnings only), exit code 1 = errors found.

Conclave includes a transparent reverse proxy that sits between Claude Code and the Anthropic API, counting input/output tokens per session for cost estimation and complexity budgeting.

# Terminal 1: start the proxy
conclave proxy --port 8199

# Terminal 2: point Claude Code at the proxy
ANTHROPIC_BASE_URL=http://localhost:8199 claude

Every API request is logged to stderr with model and token counts. When you Ctrl+C the proxy, it prints a session summary:

Session summary:
  Requests:             47
  Input tokens:         128,432
  Output tokens:        31,208
  Cache creation:       5,000
  Cache read:           10,000
  Total tokens:         159,640

Handles both streaming (SSE) and non-streaming responses. The proxy is fully transparent — all headers, bodies, and status codes pass through unmodified.

Set CONCLAVE_NON_INTERACTIVE=1 to skip all interactive decision points. Skills automatically choose sensible defaults and announce each choice:

CONCLAVE_NON_INTERACTIVE=1 claude

• Test-Driven Development - Write tests first, always
• Systematic over ad-hoc - Process over guessing
• Complexity reduction - Simplicity as primary goal
• Evidence over claims - Verify before declaring success

Skills live directly in this repository. To contribute:

1. Fork the repository
2. Create a branch for your skill
3. Follow the writing-skills skill for creating and testing new skills
4. Submit a PR

See skills/writing-skills/SKILL.md for the complete guide.

Skills update automatically when you update the plugin:

/plugin update conclave

MIT License - see LICENSE file for details

• Issues: https://github.com/signalnine/conclave/issues