msd-core/commands/gsd/plan-review-convergence.md at dc139b38e052d1b8cd8cb98d3f53fa47b4df9c9c

Files

Tom Boucher e4e8a9fcb8 fix(#936 ): run plan-phase inline in convergence; guard against nested spawner wraps (#939 )

Both sites in plan-review-convergence.md that wrapped gsd-plan-phase in
Agent() (initial planning + replan loop) are now bare Skill() calls at depth 0.
On Claude Code, a depth-1 Agent has no Agent tool so wrapped plan-phase could
never spawn gsd-planner/gsd-plan-checker — the replan loop silently produced no
revised plan when HIGHs were found. Running plan-phase inline from the depth-0
orchestrator (which retains the Agent tool) restores the full sub-agent chain.

A full audit of all workflow files confirmed these two sites were the only
instances of the anti-pattern (no other workflow wraps a spawner orchestrator
in Agent() without a RUNTIME carve-out).

Added structural guard test bug-936-no-nested-spawner-wrap.test.cjs that
dynamically derives the spawner set (workflows containing subagent_type=) and
asserts no workflow wraps a spawner inside Agent() without a RUNTIME != claude
carve-out — prevents silent regression. Test passes on fixed code, would fail
on pre-fix code at the two de-wrapped sites.

Also applied two low-severity prose nits flagged in review:
- commands/gsd/plan-review-convergence.md: orchestrator role updated to
  describe inline plan-phase + Agent for review (was generic "spawn Agents")
- gsd-core/workflows/plan-review-convergence.md success_criteria: narrowed
  "Each Agent fully completes" to the review Agent (plan-phase is inline now)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

2026-06-09 12:37:32 -04:00

3.2 KiB

Raw Blame History

name, description, argument-hint, allowed-tools, requires

name

description

argument-hint

allowed-tools

requires

gsd:plan-review-convergence

Cross-AI plan convergence loop — replan with review feedback until no HIGH concerns remain.

<phase> [--codex] [--gemini] [--claude] [--opencode] [--ollama] [--lm-studio] [--llama-cpp] [--text] [--ws <name>] [--all] [--max-cycles N]

Read

Write

Bash

Glob

Grep

Agent

AskUserQuestion

phase

review

Cross-AI plan convergence loop — an outer revision gate around gsd-review and gsd-planner. Repeatedly: review plans with external AI CLIs → if HIGH concerns found → replan with --reviews feedback → re-review. Stops when no HIGH concerns remain or max cycles reached.

Flow: Skill("gsd-plan-phase") → Agent→Skill("gsd-review") → check HIGHs → Skill("gsd-plan-phase --reviews") → Agent→Skill("gsd-review") → ... → Converge or escalate

Replaces gsd-plan-phase's internal gsd-plan-checker with external AI reviewers (codex, gemini, etc.). Plan-phase runs inline (bare Skill at depth 0) so it can spawn gsd-planner/gsd-plan-checker at depth 1. Review runs inside an isolated Agent (gsd-review is a Bash leaf — no sub-agents needed). Orchestrator only does loop control.

Orchestrator role: Parse arguments, validate phase, run plan-phase inline (Skill at depth 0), spawn an Agent for gsd-review, check HIGHs, stall detection, escalation gate.

<execution_context> @$HOME/.claude/gsd-core/workflows/plan-review-convergence.md @$HOME/.claude/gsd-core/references/revision-loop.md @$HOME/.claude/gsd-core/references/gates.md @$HOME/.claude/gsd-core/references/agent-contracts.md </execution_context>

<runtime_note> Copilot (VS Code): Use vscode_askquestions wherever this workflow calls AskUserQuestion. They are equivalent — vscode_askquestions is the VS Code Copilot implementation of the same interactive question API. Do not skip questioning steps because AskUserQuestion appears unavailable; use vscode_askquestions instead. </runtime_note>

Phase number: extracted from $ARGUMENTS (required)

Flags:

--codex — Use Codex CLI as reviewer (default if no reviewer specified)
--gemini — Use Gemini CLI as reviewer
--claude — Use Claude CLI as reviewer (separate session)
--opencode — Use OpenCode as reviewer
--ollama — Use local Ollama server as reviewer (OpenAI-compatible, default host http://localhost:11434; configure model via review.models.ollama)
--lm-studio — Use local LM Studio server as reviewer (OpenAI-compatible, default host http://localhost:1234; configure model via review.models.lm_studio)
--llama-cpp — Use local llama.cpp server as reviewer (OpenAI-compatible, default host http://localhost:8080; configure model via review.models.llama_cpp)
--all — Use all available CLIs and running local model servers
--max-cycles N — Maximum replan→review cycles (default: 3)

Feature gate: This command requires workflow.plan_review_convergence=true. Enable with: gsd config-set workflow.plan_review_convergence true

Execute end-to-end. Preserve all workflow gates (pre-flight, revision loop, stall detection, escalation).

3.2 KiB Raw Blame History

3.2 KiB

Raw Blame History