Skip to content

feat: /work skill with Codex delegation mode#580

Open
mvanhorn wants to merge 1 commit intogarrytan:mainfrom
mvanhorn:osc/feat-work-skill-codex-delegation
Open

feat: /work skill with Codex delegation mode#580
mvanhorn wants to merge 1 commit intogarrytan:mainfrom
mvanhorn:osc/feat-work-skill-codex-delegation

Conversation

@mvanhorn
Copy link
Copy Markdown
Contributor

Summary

  • Adds /work, a new skill that bridges the gap between plan approval and code review
  • Three modes: standard (Claude implements), Codex delegation (Codex implements via codex exec -s workspace-write, Claude orchestrates), hybrid (try Codex, fall back to Claude)
  • Reads plan artifacts from /plan-ceo-review, /plan-eng-review, /plan-design-review
  • Persists build log to reviews.jsonl so /ship's review dashboard tracks it

Video Demo

Demo

Why this matters

gstack's sprint cycle is Think > Plan > Build > Review > Test > Ship > Reflect. The Build step is missing. Users drop out of gstack after plan approval and implement manually.

Source Signal Engagement
PR #218 /build skill proposed, same workflow gap Open 8 days, 0 reviews
PR #549 Codex repo root resolution fix Merged same day
PR #570 Codex sandbox boundary hardening Merged same day

Codex delegation is the differentiator. gstack already uses codex exec -s read-only in 6 skills for reviews. This extends that to workspace-write for implementation. The workflow: Claude reads the plan, writes a focused prompt per task, pipes it to Codex, reviews the diff, runs tests, commits. On failure, falls back to Claude.

Token economics: Codex uses 3-4x fewer tokens per task. For users hitting Claude's billing ceiling, this means 3-4x more features per cycle.

How it connects

/office-hours       -> design doc
/plan-ceo-review    -> scope decisions
/plan-eng-review    -> architecture + test plan     -> /work -> /review -> /qa -> /ship
/plan-design-review -> UI constraints

Changes

  • work/SKILL.md.tmpl (303 lines) - new skill template with three execution modes
  • work/SKILL.md (632 lines) - generated output from template

No changes to gen-skill-docs.ts needed. The template auto-discovery in scripts/discover-skills.ts picks up new */SKILL.md.tmpl files automatically.

Testing

  • bun run gen:skill-docs --dry-run shows FRESH for work/SKILL.md
  • bun test passes (1 pre-existing failure in test-2, unrelated to this PR, fails on clean main too)
  • Template follows all existing conventions: frontmatter, {{PREAMBLE}}, {{BASE_BRANCH_DETECT}}, self-contained bash blocks

This contribution was developed with AI assistance (Claude Code).

Adds /work, a plan-to-implementation orchestrator that bridges the gap
between plan approval and code review. Three execution modes: standard
(Claude implements), codex (Codex writes code via codex exec -s
workspace-write, Claude orchestrates), and hybrid (try Codex, fall back
to Claude after 3 consecutive failures).

Reads plan artifacts from /plan-eng-review, /plan-ceo-review, and
/plan-design-review. Decomposes plans into tasks with dependency graphs.
Persists build logs to reviews.jsonl for /ship dashboard integration.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant