llm-flow plugin

A Claude Code plugin that ships a multi-model development workflow:

analyze → plan → revise plan → build → review build → rebuild → document

Multi-model fan-out happens at revise plan (validators in parallel) and review build (reviewers in parallel). Each phase is its own slash command; /llm-flow:full-cycle is the orchestrator.

Install

Recommended: as a Claude Code plugin

# From a marketplace (once published):
claude plugin install llm-flow

# From a git URL or local path during development:
claude plugin install <git-url>
claude --plugin-dir /path/to/packages/plugin

This installs the slash commands, subagents, the Stop hook (rebuild-iteration cap), and registers the multi-model MCP server.

Portable: into bare .claude/ directories

For older Claude Code versions or other agent tools that read .claude/skills/ and .claude/agents/:

# Per-user (writes ~/.claude/skills/llm-flow-*/ and ~/.claude/agents/llm-flow-*.md):
./bin/install-portable --user

# Per-project (writes <path>/.claude/...):
./bin/install-portable --project /path/to/project

Portable mode prefixes every skill/agent name with llm-flow- (e.g. /llm-flow-analyze, @llm-flow-explorer) to avoid collisions. It does NOT install the Stop hook or MCP server automatically — see the script's output for manual wiring instructions.

Required: @llm-flow/multi-model

Either install path requires the multi-model CLI on PATH for fan-out to work:

npm install -g @llm-flow/multi-model
multi-model doctor  # confirm provider CLIs are authenticated

Provider CLIs (claude, gemini, codex) must already be installed and authenticated.

Usage

# Run the full cycle on a feature request:
/llm-flow:full-cycle "add a /healthz endpoint with 200 OK"

# Or invoke phases individually:
/llm-flow:analyze "..."
/llm-flow:plan
/llm-flow:revise-plan
/llm-flow:build
/llm-flow:review-build
/llm-flow:document

The orchestrator pauses at four user checkpoints: after analyze, after revise-plan, at the rebuild iteration cap, and before commit.

Configure

Optional .llm-flow/config.json at project root:

{
  "phases": ["analyze","plan","revise-plan","build","review-build","document"],
  "models": {
    "revise-plan": ["claude","gemini","codex"],
    "review-build": ["claude","gemini","codex"]
  },
  "maxRebuildIterations": 3
}

• phases — overrides the default phase list. Names not in the canonical seven are looked up as /llm-flow:<name> (plugin) or .claude/skills/<name>/SKILL.md (project-local).
• models — per-phase provider subset. Defaults to all available.
• maxRebuildIterations — hard cap, enforced by the Stop hook. Defaults to 3.

Customize

Three layers, in order of weight:

1. Config: .llm-flow/config.json (above) — most teams stop here.
2. Drop-in phases: create a project-local skill at .claude/skills/<name>/SKILL.md and add its name to the phases array. Use /llm-flow:new-workflow <name> to scaffold one.
3. Fork: for fundamental topology changes, fork the plugin.

Artifacts

Per-run state lives at .llm-flow/:

• .llm-flow/current-task.md — task description and phase status.
• .llm-flow/exploration.md — analyze output.
• .llm-flow/plan.md — current plan (overwritten by revise-plan if validators recommend changes).
• .llm-flow/code-review.md — synthesized review (overwritten each cycle).
• .llm-flow/summary.md, .llm-flow/commit-message.txt — generated by document.
• .llm-flow/sessions/<id>/ — multi-model fan-out artifacts (one dir per fan-out call).

Add .llm-flow/ to .gitignore unless you want session history checked in.

Phases

| Skill | Purpose | Multi-model | User checkpoint | |---|---|---|---| | analyze | @explorer maps codebase, identifies impact | no | yes | | plan | @planner produces initial plan | no | no | | revise-plan | Multi-model validators, @synthesizer merges | yes | yes | | build | @implementer executes plan | no | no | | review-build | Multi-model reviewers, @synthesizer merges | yes | only at iter ≥ cap | | rebuild | @implementer addresses feedback; loops back to review-build | no | no | | document | @documenter updates docs, @summarizer drafts commit | no | yes (final) |

Subagents

@explorer, @planner, @implementer, @reviewer, @synthesizer, @documenter, @summarizer.

(In portable install they are prefixed llm-flow-.)