// Use when executing implementation plans with independent tasks in the current session
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | subagent-driven-development |
| description | Use when executing implementation plans with independent tasks in the current session |
Execute plan by dispatching fresh subagent per task, with two-stage review after each: spec compliance first, then code quality.
Codex review per task is strongly recommended for high-quality
iteration. Run codex-bridge.mjs spec-review then codex-bridge.mjs review after each task implementation. On each verdict:
approve → proceed.needs-attention → fix every issue, re-run review until approve.reject → return to implementation.Final enforcement is automatic: the auto-review.sh PreToolUse hook
runs Codex review on the full branch diff before git push /
gh pr create / gh pr ready and blocks unless verdict is
approve. Claude-only per-task reviews can pass quickly, but the merge
gate is the source of truth — anything that wouldn't survive the branch
review will be caught there. Bypass: SSPOWER_AUTO_REVIEW=off
(emergencies only).
Why subagents: Isolated context per task. You construct exactly what they need — they never inherit your session history.
Core principle: Fresh subagent per task + two-stage review = high quality, fast iteration
digraph when_to_use {
"Have implementation plan?" [shape=diamond];
"Tasks mostly independent?" [shape=diamond];
"subagent-driven-development" [shape=box];
"executing-plans or brainstorm first" [shape=box];
"Have implementation plan?" -> "Tasks mostly independent?" [label="yes"];
"Have implementation plan?" -> "executing-plans or brainstorm first" [label="no"];
"Tasks mostly independent?" -> "subagent-driven-development" [label="yes"];
"Tasks mostly independent?" -> "executing-plans or brainstorm first" [label="no"];
}
digraph process {
rankdir=TB;
"Read plan, extract all tasks, create TodoWrite" [shape=box];
"Pick engine: Claude subagent OR Codex" [shape=diamond];
"Dispatch implementer (./implementer-prompt.md)" [shape=box];
"Dispatch Codex implementer (./codex-implementer-prompt.md)" [shape=box];
"Spec review (Claude or Codex)" [shape=box];
"Code quality review (Claude or Codex)" [shape=box];
"Mark task complete" [shape=box];
"More tasks?" [shape=diamond];
"Final code review + finishing-branch" [shape=box];
"Read plan, extract all tasks, create TodoWrite" -> "Pick engine: Claude subagent OR Codex";
"Pick engine: Claude subagent OR Codex" -> "Dispatch implementer (./implementer-prompt.md)" [label="Claude"];
"Pick engine: Claude subagent OR Codex" -> "Dispatch Codex implementer (./codex-implementer-prompt.md)" [label="Codex"];
"Dispatch implementer (./implementer-prompt.md)" -> "Spec review (Claude or Codex)";
"Dispatch Codex implementer (./codex-implementer-prompt.md)" -> "Spec review (Claude or Codex)";
"Spec review (Claude or Codex)" -> "Code quality review (Claude or Codex)" [label="pass"];
"Spec review (Claude or Codex)" -> "Pick engine: Claude subagent OR Codex" [label="fail → fix → re-review"];
"Code quality review (Claude or Codex)" -> "Mark task complete" [label="pass"];
"Code quality review (Claude or Codex)" -> "Pick engine: Claude subagent OR Codex" [label="fail → fix → re-review"];
"Mark task complete" -> "More tasks?";
"More tasks?" -> "Dispatch implementer (./implementer-prompt.md)" [label="yes"];
"More tasks?" -> "Final code review + finishing-branch" [label="no"];
}
SDD supports two execution engines. Pick per task — they share the same structured contracts.
| Task characteristic | Engine | Model |
|---|---|---|
| 1-2 files, complete spec | Claude subagent | Fast/cheap model |
| Multi-file, well-specified | Claude subagent | Standard model |
| Complex / unfamiliar codebase | Codex | Codex CLI default (per ~/.codex/config.toml) |
| Needs mid-task Q&A | Claude subagent | Standard model |
| Architecture/design/review | Claude subagent | Most capable model |
| User explicitly requests Codex | Codex | Default or specified |
Instead of the Claude Task tool, use Bash to call the bridge:
node "${SSPOWER_PLUGIN_ROOT}/scripts/codex-bridge.mjs" implement \
--write --cd {WORKING_DIR} --prompt @/tmp/sdd-task-N.md
Prompt template: See codex-implementer-prompt.md — fill placeholders and write to temp file.
Response: Structured JSON with identical status codes (DONE, DONE_WITH_CONCERNS, BLOCKED, NEEDS_CONTEXT). Parse and handle exactly like Claude subagent reports.
Key difference: Codex cannot ask mid-task questions. Front-load all context in the prompt.
Session tracking: The bridge prints [codex:session] <id> to stderr during implement runs. Capture this ID — you need it for fix loops.
If Codex returns NEEDS_CONTEXT, provide context by resuming the implementer session:
node "${SSPOWER_PLUGIN_ROOT}/scripts/codex-bridge.mjs" resume \
--session-id {IMPLEMENT_SESSION_ID} --prompt "Context: {ANSWERS_TO_QUESTIONS}"
Spec review and code quality review can also use Codex:
# Spec review
node "${SSPOWER_PLUGIN_ROOT}/scripts/codex-bridge.mjs" spec-review \
--cd {WORKING_DIR} --prompt @/tmp/sdd-spec-review-N.md
# Quality review
node "${SSPOWER_PLUGIN_ROOT}/scripts/codex-bridge.mjs" review \
--cd {WORKING_DIR} --prompt @/tmp/sdd-quality-review-N.md
Prompt templates: See codex-spec-reviewer-prompt.md and codex-quality-reviewer-prompt.md.
When a review fails and the implementer was Codex, resume the implementer's session (not the reviewer's):
node "${SSPOWER_PLUGIN_ROOT}/scripts/codex-bridge.mjs" resume \
--session-id {IMPLEMENT_SESSION_ID} --prompt "Fix: {REVIEW_FINDINGS_JSON}"
Critical: Use --session-id with the ID captured from the implement run. Do NOT use --last — after spec/quality reviews, --last would resume the reviewer, not the implementer.
Codex retains full context of what it built. No need to re-provide task text or files.
See references/codex-integration.md for full details on engine selection, thread management, and error handling.
DONE: Proceed to spec review. DONE_WITH_CONCERNS: Read concerns. If correctness/scope → address before review. If observations → note and proceed. NEEDS_CONTEXT: Provide missing context and re-dispatch. BLOCKED: Assess: context problem → provide context; needs more reasoning → more capable model; too large → break it up; plan wrong → escalate to human.
Never:
If subagent asks questions: Answer clearly. Don't rush into implementation. If reviewer finds issues: Implementer fixes → reviewer re-reviews → repeat until approved.
See references/example-workflow.md for a full walkthrough.