菜单
Initialize, upgrade, or migrate the claude-harness plugin in a project. Use for first-time setup, plugin installation, version migration, memory directory configuration, legacy file cleanup, and project harness initialization.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| name | setup |
| description | Initialize, upgrade, or migrate the claude-harness plugin in a project. Use for first-time setup, plugin installation, version migration, memory directory configuration, legacy file cleanup, and project harness initialization. |
| disable-model-invocation | true |
Initialize or upgrade claude-harness in the current project directory.
This command automatically detects what needs to be done:
.claude/commands/ (commands are now served from plugin cache)Note: Commands and hooks are served from the plugin cache. setup.sh only initializes project-level state (memory directories, CLAUDE.md, .gitignore, migrations).
setup.sh handles fresh installs, v2.x migrations, and cleanup in a single script. Always run it.
Steps:
bash {plugin-root}/setup.sh
.gitignore patterns, CLAUDE.md.claude/commands/ and legacy hooks.gitignore patterns are added if missingsetup.sh handles itFallback: If the plugin root path is not available in the session context, fall through to Phase 1 -> Phase 4 (manual setup).
Plugin updates: Run claude plugin update claude-harness to update the plugin itself.
CRITICAL: You MUST update the project's .gitignore to exclude harness ephemeral files. This prevents uncommitted file clutter after /checkpoint.
Execute these steps:
Read the current .gitignore file (create if missing)
Check if .claude-harness/sessions/ pattern exists in the file
If the pattern is NOT present, append these lines to .gitignore:
# Claude Harness - Ephemeral/Per-Session State
.claude-harness/sessions/
.claude-harness/memory/compaction-backups/
# Claude Code - Local settings
.claude/settings.local.json
Use the Edit tool to append these patterns to .gitignore
Report: "Updated .gitignore with harness ephemeral patterns"
DO NOT SKIP THIS PHASE - it is required for proper harness operation.
Canonical schemas are defined in the plugin's schemas/ directory (JSON Schema format). Key state files:
| File | Schema | Created By |
|---|---|---|
sessions/{id}/context.json | schemas/context.schema.json (v3) | Phase 1 context compilation |
sessions/{id}/loop-state.json | schemas/loop-state.schema.json (v9) | Phase 4 implementation |
sessions/{id}/autonomous-state.json | schemas/autonomous-state.schema.json (v4) | --autonomous mode |
features/active.json | schemas/active-features.schema.json (v3) | Phase 2 feature creation |
Memory knowledge layers (memory/decisions/, failures/, successes/, patterns/, rules/) are an OKF v0.1 bundle -- markdown concept files with YAML frontmatter, not JSON. Their conformance rules are documented in schemas/okf-memory.md and checked by scripts/check-okf.py; setup.sh migrates legacy JSON memory files to the bundle automatically (stamp-gated). Semantic memory (memory/semantic/*.json) and runtime state stay JSON with v3 schemas as created by setup.sh.
Report:
/claude-harness:start to compile context and sync GitHub/claude-harness:prd-breakdown @./prd.md to analyze PRD and extract features/claude-harness:flow "description" for end-to-end automated workflow (recommended)/claude-harness:flow --no-merge "description" for step-by-step control/claude-harness:flow --fix feature-XXX "bug" to create bug fixes基于 SOC 职业分类
Save session progress by committing changes, pushing to remote, creating or updating a pull request, persisting decisions and patterns to memory layers, compiling session briefing, and archiving completed features. Use when saving work, creating a PR, preserving session state, or manually checkpointing progress.
Unified end-to-end development workflow - creates a feature (GitHub issue, branch, Gherkin acceptance criteria), implements it in an isolated subagent via ATDD, verifies, checkpoints (commit/push/PR), and merges. Use for implementing features, fixing bugs, batch-processing a feature backlog, or planning implementations.
Start a coding session with full context compilation, GitHub sync, and memory loading. Use for session start, project status check, context restoration, loop state recovery, GitHub issue synchronization, and feature priority review.
Merge all open pull requests, close linked issues, delete feature branches, and clean up local and remote refs. Use when completing features, merging PRs after review, or finalizing a release cycle.
Analyze a PRD (product requirements document) and perform a breakdown into atomic features with Gherkin acceptance criteria. Use for PRD analysis, feature extraction, requirement decomposition, acceptance test generation, dependency graphing, and project planning from product specifications.
