Gsx Replan Phase

gsx-replan-phase

Codex Adapter

In Codex, translate this Claude wrapper instead of rewriting it:

• $ARGUMENTS = text after $gsx-*.
• /gsd: ... or Skill("gsd:", args="...") = run .codex/skills/gsd-/SKILL.md with args as {{GSD_ARGS}}.
• AskUserQuestion: only map to request_user_input if your Codex build truly renders an interactive picker; otherwise (the usual Codex case) do not force it. Before any choice among options/plans, first print each option's core content in your reply, then list the choices as a numbered Markdown list and stop and wait for the user's text reply. Never self-answer a prompt the user never saw.
• Claude tool names are intent labels; use equivalent Codex tools.
• Preserve wrapper boundaries; if GSD owns edits, do not edit source here.

Thin orchestrator over /gsd:plan-phase: parse the Phase number → call /gsd:plan-phase {N} --reviews → relay the result. The --reviews flag tells plan-phase to read the phase's REVIEWS.md (the cross-AI peer-review feedback) and re-run the planner so the regenerated PLAN.md answers the reviewers' concerns. The command owns the replanning — spawning gsd-planner, the gsd-plan-checker verify loop, and the PLAN.md write. This skill never drafts or edits the plan itself.

Why it exists / where it sits: this is the third beat of the plan→review→replan convergence loop. gsx-plan-phase produces the first PLAN.md (docs-grounded). gsx-plan-review runs /gsd:review --phase N --codex and writes REVIEWS.md. This skill closes the loop by feeding that critique back in. The value is that the regenerated plan is shaped by an independent reviewer's judgment, not just the original planner's — assumptions get challenged, gaps get filled, before any code is written. The loop can repeat: replan → review again → replan, until no HIGH concerns remain (cf. /gsd:plan-review-convergence).

Boundary — this skill never changes code. It revises PLAN.md (via /gsd:plan-phase --reviews) against existing REVIEWS.md; it does not edit source, run migrations, or commit beyond what plan-phase does. Implementation belongs to /gsd:execute-phase afterward. Reply in Chinese.

---

0 — Parse the Phase number (+ pass-through flags)

Phase dirs: .planning/phases/{NN}-{slug}/, NN zero-padded to 2 digits. --reviews needs the phase's REVIEWS.md to already exist (from /gsx-plan-review / /gsd:review).

• $ARGUMENTS has a number → normalize (ignore case, Phase prefix, whitespace): 4/04/Phase 4 → 4. Keep any trailing plan-phase flags (--tdd, --mvp, --skip-verify, --text, …) — they pass through unchanged.
• Empty → infer: (1) Phase already under discussion in this conversation (e.g. a /gsx-plan-review just finished); else (2) .planning/STATE.md last planned / last reviewed phase; else (3) AskUserQuestion (header Replan Phase) — don't guess.
• No .planning/ → not a GSD project: tell the user "/gsd:plan-phase cannot run (no .planning/)" and stop.
• No REVIEWS.md for the phase yet → there's no feedback to fold in: tell the user to run /gsx-plan-review {N} first (or, for a fresh plan, /gsx-plan-phase {N}) and stop. Don't run --reviews against a missing file.

1 — Invoke

Skill("gsd:plan-phase", args="{N} --reviews {any pass-through flags}")

--reviews makes plan-phase consume REVIEWS.md and replan; it does not re-research, so the grounded RESEARCH.md from gsx-plan-phase stays intact. Let the command own the planner spawn, the gsd-plan-checker verify loop, iteration to pass, and the PLAN.md rewrite. Don't pre-empt it by drafting tasks yourself. Then go to step 2.

2 — Relay the result

Relay the command's outcome as a receipt; don't re-summarize the whole plan:

## gsx-replan-phase — Phase {N}
Command:  /gsd:plan-phase {N} --reviews
Input:    .planning/phases/{NN}-*/{NN}-REVIEWS.md (Codex 评审意见)
Output:   .planning/phases/{NN}-*/{NN}-PLAN.md (已按评审意见重生成)
Addressed:{one line — which concerns the replan folded in / any it explicitly deferred}
下一步：/gsd:execute-phase {N}（若仍有未消化的评审意见，可再跑一轮 /gsx-plan-review 复评）

• If concerns remain unresolved or the reviewer's points conflict with locked decisions, name that plainly rather than implying a clean sweep — the user may want another review round or a manual call.
• If the verify loop didn't converge or the run bailed before PLAN.md was rewritten, say so and name what state the plan is in (old PLAN.md unchanged, or partially written).

---

Guardrails

• Only orchestrate. No drafting or editing PLAN.md / REVIEWS.md yourself. plan-phase and its planner own the replan.
• Needs REVIEWS.md. --reviews consumes existing review feedback — no REVIEWS.md → send the user to /gsx-plan-review {N} first. Don't run --reviews against nothing.
• --reviews doesn't re-research. The grounded RESEARCH.md stays as-is; this is a plan revision, not a fresh research+plan. For a from-scratch plan use /gsx-plan-phase.
• Don't review here. Generating the critique is /gsx-plan-review's job; this skill only folds an existing critique back in. The loop is: plan → review → replan (→ review → replan …).
• No code changes here. This is a planning skill. Implementation belongs to /gsd:execute-phase afterward.
• One Phase per run. Ambiguous → AskUserQuestion. Non-GSD project → stop.
• Reply in Chinese.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JinsYin
• Source: JinsYin/skills
• License: MIT

Install and usage instructions live in the source repository linked above.