Publish the official Dispatch plugin for Claude Code and Codex

[selfcontained]

Makes this repo a plugin marketplace for both Claude Code and Codex, shipping eleven narrow skills that teach agents to use Dispatch's own capabilities.

Closes the dispatch-usage-skills-bundle idea. Manual install only — stale-version detection, auto-attach, and marketplace submission stay deferred per Brad's decision.

What ships

.claude-plugin/marketplace.json # Claude Code marketplace
.agents/plugins/marketplace.json # Codex marketplace
plugins/dispatch/
├── .claude-plugin/plugin.json # Claude Code manifest
├── .codex-plugin/plugin.json # Codex manifest
├── skills/<name>/SKILL.md # 11 skills, shared by both platforms
├── evals/ # 4 ablation cases
└── README.md

Tier	Skills
Migration — existing doc-tool payloads become skill bodies	whiteboard, personas
Extraction — the conditional launch-guidance blocks	review-workflow, ui-validation
Discovery — "agents don't know it's an option"	sharing, brain, subagents, repo-tools
Discovery, low-frequency	jobs, templates, personalities

Install:

claude plugin marketplace add selfcontained/dispatch && claude plugin install dispatch@dispatch
codex plugin marketplace add selfcontained/dispatch && codex plugin add dispatch@dispatch

Verified end-to-end on both CLIs

Against claude 2.1.231 and codex-cli 0.147.0, using throwaway CLAUDE_CONFIG_DIR / CODEX_HOME roots — not the machine's real config.

Claude — claude plugin validate passes on both the marketplace and the plugin (clean under --strict too). After marketplace add + install, claude plugin details dispatch@dispatch reports:

Component inventory
Skills (11) brain, jobs, personalities, personas, repo-tools, review-workflow,
sharing, subagents, templates, ui-validation, whiteboard
Agents (0) Hooks (0) MCP servers (0) LSP servers (0)
Projected token cost
Always-on: ~592 tok added to every session

Codex — codex plugin marketplace add + codex plugin add both exit 0 non-interactively and install 0.1.0 from the Codex-native manifest. codex debug prompt-input shows all eleven reaching the model prompt inside <skills_instructions>, namespaced dispatch:<name> with the right descriptions — so this is confirmed model-visible, not just installed.

Design

The description is the product. Descriptions load into every session unconditionally; bodies load only on a match. So they're written as symptom triggers rather than feature labels — an agent that doesn't know a capability exists will never match its name, but will match a description of the situation it's currently in. Total always-on cost is 2,258 chars (~592 tokens), under the 2,891 chars launch guidance already injects.

Narrow skills, not mega-skills. Eleven short descriptions always-on, exactly one body loaded on a match. A single "automations" skill would load four unrelated bodies whenever one of them fired.

Dual manifests. Codex does fall back to reading .claude-plugin/, but that's undocumented by OpenAI, so the Codex-native manifests are carried explicitly rather than relying on it.

Explicit version in both manifests — Claude uses it as the update cache key, so routine commits to main don't register as a plugin update. Omitting it would resolve the version to the commit SHA and make every push look like a new release.

Purely additive. Launch guidance is not trimmed and the documentation-bearing MCP tools (whiteboard_howto, persona_templates) stay in place; content lives in two places for now. Deleting either would regress any repo that hasn't installed the plugin. Both trims are follow-ups gated on a real soak period.

Deliberately not skills:dispatch_event, pin discipline, and session naming. Skills only load on a task match, so always-relevant protocol obligations can't move out of launch guidance.

The vitest guard

apps/server/test/plugin-manifest.test.ts checks the manifests agree on name/version/source shape and that every SKILL.md's frontmatter parses.

This isn't speculative. claude plugin validate caught a real defect mid-build: an unquoted colon in the personalities description broke its YAML frontmatter, and the failure mode is silent — the skill still installs, it just loads with empty metadata and never fires. CI has no claude/codex CLI to catch that. Confirmed the guard fails on the reintroduced bug and passes once fixed.

It also asserts the always-on description budget stays under the 2,891-char launch-guidance ceiling, so the plugin can't quietly grow past the thing it augments.

Evals

Four ablation cases (sharing, brain, subagents, repo-tools) in the documented prompt.md + graders/criteria.md layout, aimed at --ablation with-without so the no-plugin baseline arm gives a score delta.

They have not been executed.claude plugin eval is in early access and refuses to run on 2.1.231 — every invocation returns `plugin eval` is currently in early access and exits. plugins/dispatch/evals/README.md says so plainly; treat the criteria as a first draft to tune once the runner is available, not as a passing suite.

sharing is the sharpest case in the set: the rule it encodes already exists in two always-on places (launch guidance and CLAUDE.md) and still gets ignored, so a measurable delta there would be evidence that arriving at the moment of action beats more background text.

Checks

• pnpm run check — passes
• pnpm run test — 3,635 passed, 9 skipped across all three suites
• pnpm run format — clean
• pnpm run test:e2e — run locally before pushing (to avoid colliding with CI's stack)
• No apps/web/ changes, so finalize:web doesn't apply

The plugin-format spec work in #951 was pulled into this directly rather than linked; that PR can be closed without merging.

🤖 Generated with Claude Code