菜单
Generate multi-CLI distribution packages from the Claude Code plugin. Converts shared skills and package assets for OpenCode, Codex CLI, and Pi under dist/. Run after changing plugin components to keep distributions in sync.
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 | sync-tools |
| description | Generate multi-CLI distribution packages from the Claude Code plugin. Converts shared skills and package assets for OpenCode, Codex CLI, and Pi under dist/. Run after changing plugin components to keep distributions in sync. |
| user-invocable | true |
Generate distribution packages for OpenCode, Codex CLI, and Pi from the Claude Code plugin.
Input: $ARGUMENTS — optional CLI name (opencode, codex, pi) to sync one target. Empty = all targets.
| Role | Path |
|---|---|
| Sources (read-only) | claude-plugins/manifest-dev/ and claude-plugins/manifest-dev-tools/ |
| Output | dist/{opencode,codex,pi}/ |
| Conversion rules | .claude/skills/sync-tools/references/{cli}-cli.md |
| Per-CLI sync state | dist/{cli}/.sync-meta.json (records last-synced source SHA — drives diff-first workflow) |
| Pi namespace metadata | dist/pi/component-namespaces.json (component → owning plugin; Pi only) |
| GitHub repo | doodledood/manifest-dev |
Sync only these source payloads:
claude-plugins/manifest-dev/ — core workflow skills.claude-plugins/manifest-dev-tools/ — tools skills.Both plugins are skills-only — manifest-dev ships no agents or hooks on any target (the former functional agents are skills).
Never sync other plugins (e.g., manifest-dev-collab — uses Agent Teams/Slack, inherently incompatible). Never modify source files. Skip sync-tools skill from output (meta-tool).
Namespacing model is per-CLI (see each reference file):
manifest-dev, manifest-dev-tools) each bundle their skills under original names. OpenCode: one plugin entry (dist/opencode/plugin/) registers the whole skills payload via skills.paths; skills keep bare names (native discovery is first-found-wins). No install-time suffixing, no component-namespaces.json, no installer. Plugin-qualified skill-reference handling (manifest-dev:<skill> — strip vs keep) is per-CLI: see each reference file.component-namespaces.json (ownership metadata — Pi keeps package-scoped skill names, no suffixing). Regenerate it on every sync from the discovered source components; never hand-maintain it.For each target CLI, read its reference file first. The reference file is the single source of truth for conversion rules — tool name mappings, frontmatter format, hook protocol, package shape, directory structure, and limitations. Do not duplicate conversion logic here; follow the reference.
Each dist/{cli}/.sync-meta.json records the source commit that dist was last synced from:
{
"source_commit": "<sha>",
"source_path": "claude-plugins/manifest-dev",
"source_paths": [
"claude-plugins/manifest-dev",
"claude-plugins/manifest-dev-tools"
],
"synced_at": "<ISO 8601 UTC>"
}
On invocation, prefer a delta sync over a full re-sync:
dist/{cli}/.sync-meta.json. If missing, malformed, or the recorded SHA is unreachable from HEAD (e.g., rebased away, force-pushed branch), fall back to full sync for that CLI.HEAD (these define the substitution rules — any change can affect every dist file):
.claude/skills/sync-tools/SKILL.md.claude/skills/sync-tools/references/{cli}-cli.mdgit diff --name-status <recorded-sha>..HEAD -- claude-plugins/manifest-dev/ claude-plugins/manifest-dev-tools/ and process each entry:
AGENTS.md) only if the set of skills changed (added/removed/renamed). Body-only edits don't require regenerating these.
claude-plugins/manifest-dev/.claude-plugin/plugin.json's version changed between recorded SHA and HEAD, mirror it into dist/opencode/plugin/package.json per the reference file.dist/pi/component-namespaces.json from the current dist component set and source ownership map. Skip entirely for plugin-native Codex and OpenCode — the plugin is the namespace there; generating namespace metadata would resurrect retired installer concepts.dist/{cli}/.sync-meta.json with the new HEAD sha and a fresh synced_at UTC timestamp. Keep the file even when the diff was empty — the timestamp records "we checked".The metadata is an optimization, not a correctness anchor. When in doubt — unreachable commit, ambiguous rename, mid-rebase repo state, suspicious dist drift — fall back to full sync rather than trusting the recorded SHA.
| Component | Goal |
|---|---|
| Skills | Copy unchanged (Agent Skills Open Standard = universal) from both source payloads. Include all subdirectories. Replace operational CLAUDE.md references (e.g., "write to CLAUDE.md") with CLI context file name per reference file. Replace operational tool-name references in skill body prose with the target CLI's names (e.g., "use the Read tool" → the target's read tool) — the mappings live in each reference file's tool-name lookup table. Leave research/reference content unchanged (teaching documents in references/*.md that explain Claude Code conventions stay Claude-Code-centric; only operational instructions remap). Exception 1: files in references/execution-modes/ — replace Claude-specific model names (haiku, sonnet, opus) with inherit. Model tier routing is Claude Code-only; other CLIs use session model for all tiers. Exception 2: lines that surface a session-file path to the user (e.g. Session: ~/.claude/projects/<dir>/${CLAUDE_SESSION_ID}.jsonl, "Session JSONL files live at ...") — retarget per the reference file's session-line rule. If the target CLI has no per-session file the agent can locate at runtime, omit the line. |
| Agents | None shipped — manifest-dev has no agents on any target; the former functional agents are skills. Do not generate agent files. |
| Hooks | None shipped. If a future hook ships, adapt per the target's reference file — and re-derive that target's hook capability map first (the retired research is stale). |
| Commands | Codex bundles skills without command shims. OpenCode generates no command files, but its plugin config hook registers slash-command wrappers in cfg.command for source skills whose user-invocable is missing/true, preserving existing user/project commands; user-invocable: false helpers stay skill-tool-only. Pi invokes /skill:<name> and ships prompt-template aliases for /do, /auto, and /babysit-pr. |
| Context file | Workflow overview + skill descriptions in the CLI's native context format per reference file. |
| README | Component table, install instructions, feature parity table, required config, link to GitHub repo. |
| Package manifest | Generate only for targets whose reference file declares a package-native install surface. For Pi, repo-root package metadata is source-owned; generated skills/prompts under dist/pi/ are package resources consumed by that package. For OpenCode, the plugin entry (dist/opencode/plugin/package.json + index.js) is generated and versioned per the reference file. |
| Install script | None. No target ships an install script — Codex and OpenCode are plugin-native; Pi installs via its package manager. |
| CLI extras | Extension manifests, plugin configs, execution rules — per reference file. |
| Namespace metadata | Package target (Pi) only. Regenerate component-namespaces.json from source ownership every run; every distributed component appears exactly once under its component type with its owning plugin. Plugin-native Codex and OpenCode have none — skip them (the plugin boundary is the namespace). |
The CLI-native install method from the reference file is the primary method: Codex plugin marketplace add, OpenCode repo clone + plugin config line (with the clone-or-pull update alias), Pi package manager from the repo root. Include other methods from the reference file as alternatives only when they actually work for that target.
| Constraint | Why |
|---|---|
| Shell-based text processing during sync must work in both bash and zsh | macOS default shell is zsh; bash-only constructs break |
| Reference files are authoritative for conversion rules | Avoids two sources of truth — update one place |
| Unmapped tool names in skill prose pass through unchanged | Only operational references remap; names without a lookup-table row are left as-is |
| Empty component sets skip gracefully | Codex has no hooks — note in README, don't error |
| Skill prompt bodies stay faithful to Claude Code originals | Prompts are carefully crafted — don't simplify, rewrite, or truncate for other CLIs |
Always update dist/{cli}/.sync-meta.json at end of run | The recorded SHA is what next run's diff-first path keys on. Skipping the update silently degrades future syncs to full re-syncs. |
Write to /tmp/sync-tools-{timestamp}.md after each CLI: counts, warnings, what was generated. Read the full log before writing the final summary.
Summary table after all CLIs processed:
| CLI | Skills | Agents | Hooks | Commands | Status |
|---|---|---|---|---|---|
| OpenCode | N (1 plugin) | none (all skills) | none | N plugin-registered wrappers (no command files) | Complete |
| Codex | N (2 plugins) | none (reviewers = review-code skill) | none | — | Complete |
| Pi | N compatible | /do, /auto, /babysit-pr aliases | none | prompt-template aliases | Complete |
End-to-end autonomous execution: figure-out → define → do, chained without manual approval gates. Use when you want to define and execute without intervention during planning, when the user asks for autonomous or end-to-end work, says just build it, or asks to tend or babysit a PR.
Figure things out together — any topic, problem, or idea. Presses relentlessly until shared understanding is reached. Use when you need to understand before acting, when figuring it out is the goal, or when the user asks to think through a decision, dig deeper, press an assumption, investigate why something is happening, or work through a problem.
End-to-end autonomous execution: figure-out → define → do, chained without manual approval gates. Use when you want to define and execute without intervention during planning, when the user asks for autonomous or end-to-end work, says just build it, or asks to tend or babysit a PR.
Manifest executor. Works through Deliverables verifying every Acceptance Criterion and Global Invariant. Use when executing a manifest, running a plan, implementing a defined task, or when the user asks to run, execute, implement, or ship a manifest-backed plan.
Completion marker for the /do workflow. Outputs a plain-prose summary of what was built. Called by /do after every Acceptance Criterion and Global Invariant has fresh PASS evidence, when the manifest is complete, all criteria pass, or the workflow needs to wrap up with a completion summary.
Author-side PR lifecycle babysitter and companion to review-pr. Use when the user wants to tend an existing GitHub PR through CI, review threads, description sync, mergeability, auto-fixes, or asks to babysit a PR with or without an existing manifest.