Add lite mode for fewer intermediate checks

[shinpr]

Summary

The full workflow runs a verifier, a cross-document sync check, a quality fixer after every task, and a separate security review. That is the right default, but some users would rather spend fewer tokens: early-stage projects, changes where they trust a strong model, or work that does not need every check. This PR adds lite mode for them. The user asks for it in the recipe request, and the workflow keeps the same phases and approval stops while running fewer intermediate checks.

Changes

• Added: subagents-orchestration-guide/references/lite-mode.md. When lite mode is on, the orchestrator skips code-verifier, design-sync, and security-reviewer. In a Work Plan task set it also skips the per-task quality fixer and runs one Final Quality Run after the last task commit. Code review and the quality gate on post-review corrections still run. Small, single-cycle flows keep their only quality fixer call.
• Changed: subagents-orchestration-guide/SKILL.md points to that file only when the user asks for lite mode. The recipes are unchanged, and normal-mode runs never read the lite mode rules.
• Removed: the unused type field from governingDocuments in code-reviewer and security-reviewer, and from the recipes that pass it. Neither reviewer branches on it, and callers were filling it with document-reviewer labels such as DesignDoc (see Stop security-reviewer from blocking on governing document type labels #188).
• Changed: the README and its five translations now focus on what a user needs to adopt and run the workflow. Lite mode is described under How It Works and linked from the paragraph about cost. The scale table, template excerpts, post-run checklist, agent list, and repository layout are removed. The scale table also listed a security review for Small changes, which Small does not run.

Also included: the minor version is bumped to 0.28.0, and the plugin copies are synced.

Verification

I ran each phase once with claude -p, --plugin-dir dev-workflows, and user plugins excluded, on a small Node fixture repository (an in-memory todo store and HTTP API with lint and node:test).

Run	Agents called	Lite mode behavior
/recipe-design in lite mode, with one existing Design Doc	codebase-analyzer, technical-designer, document-reviewer	No code-verifier. No design-sync, although a second Design Doc existed. The approval stop came right after document review, and document-reviewer ran without verification_evidence.
/recipe-build in lite mode, with a 2-task Work Plan	task-executor ×2, integration-test-reviewer, quality-fixer ×1, code-reviewer	Each task was committed without a per-task quality fixer. One Final Quality Run took direct_scope, governing_sources, and observable_verification from the Work Plan and task files, and lint plus 20 tests passed. No security-reviewer.

Both runs read lite-mode.md before their first workflow decision. Not covered: a comparison against a normal-mode run, the frontend and fullstack plugins, and the stub_detected path of the Final Quality Run.

claude plugin validate passes for the marketplace and all plugins.

🤖 Generated with Claude Code