Prompt file imported from sebastianbasadre/tenazas (
.codex/prompts/gsd-execute-phase.md). Fill in{{arguments}}before use. Copyright stays with the author.
{{arguments}}
Objective
Execute all plans in a phase with wave-based parallelization
Compatibility
- Use .codex/skills/get-shit-done-codex semantics.
- Treat upstream workflow as the source of truth.
- Replace user-specific paths with workspace-relative paths (.claude/..., .planning/...).
- Use one of these engine paths (prefer local, fallback global): node .claude/get-shit-done/bin/gsd-tools.js ... node ~/.claude/get-shit-done/bin/gsd-tools.js ...
- If
.jsis unavailable, use the same path with.cjs. - Run engine commands through PowerShell.
- Do not set
node <path> ...as one string variable and invoke& <cmd_var>; run directnode <path> ...or& node <path> .... - Parse JSON with ConvertFrom-Json; parse key/value output when workflow uses KEY=value raw mode.
- No jq / bash-only constructs.
- Accept natural-language command input; do not require an exact literal argument template.
- If a required argument is still missing after extraction, ask one concise clarification question.
Subagent lifecycle (required)
- Translate each upstream
Task(...)intospawn_agent->wait->close_agent. - Spawn only when the upstream workflow defines an agent role.
- Use
.claude/agents/gsd-*.mdas role context for each spawned agent. - Do not advance workflow steps until wait and close complete.
Input contract
- The command argument is authoritative.
- If user runs
gsd-execute-phase 28, execute phase28. Do not auto-discover a different phase. - Parse phase from either:
- the explicit command argument tail, or
- the latest user message (for example: "execute phase 28").
- Accept the first numeric token as phase number (integer or decimal, e.g., 30 or 30.1).
- Parse flags
--gaps-onlyand--autofrom the same input text when present. - Do not claim "phase missing" when the latest user input already contains a valid phase token.
- Only ask for input if no valid phase token is found anywhere in the latest user request.
Execution
- Parse " [--gaps-only] [--auto]" from command argument text, or extract from the latest user message.
- Treat parsed phase as the execution target and keep it unchanged for all downstream steps.
- Run init with the parsed phase: node init execute-phase --raw
- If
--gaps-onlyis present, apply gap-closure filtering exactly as defined in the upstream workflow. - If
--autois present, preserve auto-advance behavior from the upstream workflow. - Load .claude/get-shit-done/workflows/execute-phase.md and execute it step-by-step.
- Translate each Task(...) in workflow into:
- spawn_agent with the matching role file context from .claude/agents/.
- Set
agent_typeto the matching GSD role (gsd-*). - wait for each spawned agent and apply returned output before moving forward.
- Preserve all gates and routing from upstream workflow.
- Preserve commit behavior using node commit "message" --files ....
- If commit preflight fails (no git / no commit flag), proceed in read-only mode and report clearly.
Completion output
- Summarize key artifacts created/updated.
- Next recommended command:
/prompts:gsd-verify-work <phase>(Codex) //gsd:verify-work <phase>(Claude) - Never recommend internal
node ... gsd-tools ...commands to the user.
