feat(ce-work): reduce token usage by extracting late-sequence references

[tmchow]

Summary

Extract Phase 3-4 shipping content from ce:work and ce:work-beta SKILL.md files into on-demand references/shipping-workflow.md files. This late-sequence content (quality check, code review, PR creation, quality checklist, code review tiers) rode in context for the entire Phase 2 execution loop — 20-60+ tool calls — without being used until the very end. Follows the same extraction pattern established by ce:plan (PR #489) and ce:brainstorm.

Token savings

Skill	Before	After	Reduction
ce:work	445 lines (~6,500 tokens)	316 lines (~4,600 tokens)	~29%
ce:work-beta	518 lines (~7,600 tokens)	389 lines (~5,700 tokens)	~25%

Per-session savings: ~2,200 tokens x ~32 messages before shipping content is needed = ~70,400 context tokens per session for each skill.

What changed

• New reference files: plugins/compound-engineering/skills/ce-work/references/shipping-workflow.md and the identical file under ce-work-beta/references/. Each contains Phase 3 (quality check, code review, final validation, operational validation plan), Phase 4 (screenshots, commit/PR, plan status, notify), Quality Checklist, and Code Review Tiers.
• SKILL.md stubs: Both SKILL.md files replace Phase 3-4 with a 2-line stub that instructs the agent to read the reference file when all Phase 2 tasks are complete.
• Contract tests: Updated to use dual-read pattern — positive assertions read from reference files, negative assertions stay on SKILL.md to confirm extraction completeness. Two new stub pointer tests verify both SKILL.md files reference shipping-workflow.md and don't contain extracted content.

Key decisions

• One reference file per extraction, not four: Phase 3, Phase 4, Quality Checklist, and Code Review Tiers are all used at the same workflow stage (after Phase 2 completes). Bundling into one file matches ce:brainstorm's approach.
• Independent files for stable and beta: Per AGENTS.md skill self-containment rules, each skill directory owns its references. The files are currently identical; beta's sits alongside its existing codex-delegation-workflow.md.
• Key Principles and Common Pitfalls stay inline: Small (~40 lines combined) and provide behavioral guardrails throughout execution.

Test plan

• bun test — 636 pass, 0 fail
• Contract tests verify extracted content is in reference files and absent from SKILL.md
• New stub pointer tests confirm both SKILL.md files reference the shipping-workflow file
• Reference files are byte-identical between stable and beta
• Manual: run ce:work on a plan and verify shipping-workflow.md loads at Phase 3 transition (not prematurely during Phase 2)

Post-Deploy Monitoring & Validation

No additional operational monitoring required — this is a content restructuring within plugin skill files with no runtime or infrastructure impact.

---