janpeter/scrum4me-mcp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat(codex): Phase 1 mcp slice — codex plan-review prompt + runtime-aware selection #43

Merged
janpeter merged 10 commits from feat/codex-plan-review-phase1 into main 2026-06-09 07:20:25 +02:00
Conversation 0 Commits 10 Files changed 6 +1078 -16
janpeter commented 2026-06-08 23:12:00 +02:00
Owner
Copy link

What

The mcp slice of Phase 1 — Codex as a fleet worker for IDEA_REVIEW_PLAN plan-review. Library capability only; the runner wiring (scrum4me-docker) and the UI (scrum4me-workers) land in follow-up PRs per the dual-GO design.

  • src/prompts/idea/review-plan.codex.md — codex-portable, autonomous IDEA_REVIEW_PLAN prompt: 3-round active rewrite (persists via update_idea_plan_md({ idea_id, markdown })), self-determined verdict (no ask_user_question), review_log.approval.status synced to approval_status, and a plan-write failure blocks approved.
  • src/lib/kind-prompts.ts — runtime-aware getKindPromptText(kind, runtime = 'CLAUDE') + getIdeaPromptText(kind, runtime = 'CLAUDE'); compound cache key ${runtime}:${kind}; (CODEX, IDEA_REVIEW_PLAN) -> review-plan.codex.md; every other (kind, runtime) is byte-identical to before (default CLAUDE).
  • scripts/seed-idea-review-codex-canary.ts — seeds an IDEA_REVIEW_PLAN / CODEX / QUEUED canary against a throwaway PLAN_READY idea with a rich, deliberately-improvable plan_md + grill_md.
  • docs: the dual-GO design spec + the codex-GO implementation plan.

Authoritative prompt source (note for reviewers)

payload.prompt_text (src/tools/wait-for-job.ts:957, via getIdeaPromptText) is intentionally not runtime-threaded in this PR. Per the design (codex plan-review P1-2), the docker runner is the authoritative prompt source: run-one-job.ts:364 calls getKindPromptText(ctx.kind, runtime) (runtime from getWorkerRuntimeFromEnv()) and passes it to buildCodexArgs. The payload field is redundant/informational, threading it is optional (and would also require adding runtime to the getFullJobContext select). The codex prompt reaches codex via the runner, which lands in the separate scrum4me-docker PR (plan Task 5).

Review & verification

  • Design: dual-GO — scrum4me-server:claude (operational) + mac:codex (source, round-2). Plan: mac:codex round-2 GO ("executable as written").
  • Internal: spec-compliance review clean (byte-for-byte, no scope creep); code-quality review — its single "Critical" was a cross-repo architecture misunderstanding (the runner, not the payload, is authoritative; see above).
  • npx vitest run -> 85 files / 682 tests green; npx tsc --noEmit -> clean.

🤖 Generated with Claude Code

## What The **mcp slice** of Phase 1 — Codex as a fleet worker for `IDEA_REVIEW_PLAN` plan-review. Library capability only; the runner wiring (scrum4me-docker) and the UI (scrum4me-workers) land in follow-up PRs per the dual-GO design. - `src/prompts/idea/review-plan.codex.md` — codex-portable, **autonomous** IDEA_REVIEW_PLAN prompt: 3-round active rewrite (persists via `update_idea_plan_md({ idea_id, markdown })`), self-determined verdict (no `ask_user_question`), `review_log.approval.status` synced to `approval_status`, and a plan-write failure blocks `approved`. - `src/lib/kind-prompts.ts` — runtime-aware `getKindPromptText(kind, runtime = 'CLAUDE')` + `getIdeaPromptText(kind, runtime = 'CLAUDE')`; compound cache key `${runtime}:${kind}`; `(CODEX, IDEA_REVIEW_PLAN) -> review-plan.codex.md`; every other `(kind, runtime)` is byte-identical to before (default `CLAUDE`). - `scripts/seed-idea-review-codex-canary.ts` — seeds an `IDEA_REVIEW_PLAN` / `CODEX` / `QUEUED` canary against a throwaway `PLAN_READY` idea with a rich, deliberately-improvable `plan_md` + `grill_md`. - docs: the dual-GO design spec + the codex-GO implementation plan. ## Authoritative prompt source (note for reviewers) `payload.prompt_text` (`src/tools/wait-for-job.ts:957`, via `getIdeaPromptText`) is intentionally **not** runtime-threaded in this PR. Per the design (codex plan-review P1-2), the **docker runner is the authoritative prompt source**: `run-one-job.ts:364` calls `getKindPromptText(ctx.kind, runtime)` (runtime from `getWorkerRuntimeFromEnv()`) and passes it to `buildCodexArgs`. The payload field is redundant/informational, threading it is optional (and would also require adding `runtime` to the `getFullJobContext` select). The codex prompt reaches codex via the runner, which lands in the separate scrum4me-docker PR (plan Task 5). ## Review & verification - Design: dual-GO — scrum4me-server:claude (operational) + mac:codex (source, round-2). Plan: mac:codex round-2 GO ("executable as written"). - Internal: spec-compliance review clean (byte-for-byte, no scope creep); code-quality review — its single "Critical" was a cross-repo architecture misunderstanding (the runner, not the payload, is authoritative; see above). - `npx vitest run` -> **85 files / 682 tests green**; `npx tsc --noEmit` -> clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code) 
Approved brainstorm design for routing IDEA_REVIEW_PLAN to a codex fleet
worker as an autonomous active-improver (3-round rewrite, no ask_user_question
gate), reusing update_idea_plan_reviewed. Full+UI scope; no schema / migration /
shared change; Claude's plan-review untouched. Depends on Phase 0 substrate.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Phase 0 confirmed merged/live on scrum4me-docker master (ee8c647): multi-stage
  Dockerfile + agent-codex service + runtime branch. §1 prerequisite SATISFIED.
- rich seed plan_md (P2-2), claimer-tier note (P2-1), no-op rescale (P3-1),
  UI-gate env (P3-2), build-without-target (P3-3), parameterized claim-filter (P3-4),
  product_id canary recommendation; added review-log section.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
P1-1 workers snapshot overrides requested_model (not a fake model_id);
P1-2 close runner-vs-payload P1: docker runner getKindPromptText is authoritative
     (run-one-job.ts:364) -> getKindPromptText(ctx.kind, runtime);
P2-3 require review_log.approval.status to mirror top-level approval_status;
P2-4 update_idea_plan_md failure must block 'approved' (no false-positive);
P3-5 refresh source anchors (wait-for-job.ts:895-968, run-one-job.ts:364).
All five verified against source before applying.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- §10 risk row: prompt-source RESOLVED (run-one-job.ts:364 authoritative)
- §4 workers: correct model_id usage-write anchor (update-job-status.ts:853-855)
- review-log: codex round-2 GO recorded

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
8 tasks, bite-sized TDD: mcp codex-portable prompt + runtime-aware getKindPromptText
(TDD) + rich seed; docker runtime passthrough (runner authoritative); workers
requested_model override (TDD) + template allowedRuntimes; gated host canary.
Derived from the dual-GO spec (2026-06-08-codex-plan-review-phase1-design.md).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
P1-1 codex prompt calls update_idea_plan_md({ markdown }) not plan_md (schema);
P1-2 Task 6 RED test mock gets launch_preview_json.context.ideaId (idea-binding);
P1-3 run-one-job passes capabilities to registerWorker+startHeartbeat so the workers
     UI gate (capabilities has 'review') counts the codex worker, else review enqueue
     is blocked despite claimable agents;
P2-4 force a canary host via pausing max2 / a semantic tag, not required_capability=LOW_P;
P3-5 tsx-load smoke marked explicitly non-binding.
All verified against source. Spec §4/§8/§14 updated to match.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
(IDEA_REVIEW_PLAN, CODEX) -> review-plan.codex.md; all other (kind, runtime)
combinations unchanged; runtime defaults to CLAUDE so existing call-sites are
byte-identical.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
janpeter/scrum4me-mcp!43
No description provided.