菜单
Executes a bd-decomposed implementation plan sub-issue by sub-issue with context isolation and mandatory TDD per behavior
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Adversarial review gate for scenario-based BC work; sole emitter of work_done for assign_scenarios and bugfix-with-scenarios dispatches
Routes inbound lead messages to the correct BC subagent after sufficiency gating; the sole entry point for all BC inbox processing
Determines whether an inbound lead message is sufficient to act on, or whether a clarify response is required before dispatch
Merges the work branch into the BC's main and pushes so the work_id commit is reachable from origin/main before any work_done emission
Situational mechanics for the turn-limited PO<->Architect BC-decomposition collaboration. Load this when you are actively running a decomposition exchange with the PO (or, as the PO, with the Architect) and need the round cap, the extension protocol, and the ship-from-here rule. Per PDR-014.
Use when implementing any assigned scenario or bugfix, before writing implementation code
| name | subagent-driven-development |
| description | Executes a bd-decomposed implementation plan sub-issue by sub-issue with context isolation and mandatory TDD per behavior |
After the router dispatches to the implementer, and after the implementer has decomposed the work into bd sub-issues (writing-plans-bdd), this skill governs how to execute that plan — one sub-issue at a time, with context isolation between sub-issues, and mandatory TDD for every behavior.
The router orchestrates execution of the bd DAG (built by writing-plans-bdd)
using the following loop. The loop runs until the DAG is fully drained.
bd ready
This returns all sub-issues with no open blockers (all dependencies closed). On the first iteration, this will be all RED ("write the failing test for ") sub-issues that have no cross-behavior dependencies.
Dispatch ALL ready sub-issues simultaneously to bc-implementer subagents (one subagent per sub-issue). Independent sub-issues run concurrently — do not serialize work that the DAG says can proceed in parallel.
Each bc-implementer subagent receives:
Wait for all dispatched subagents to complete. Then, gate between dependency layers:
For every sub-issue just dispatched:
Verify it is closed (bd show <sub_id> shows status closed).
For behavior sub-issues (RED or GREEN), verify the commit sequence:
test(red): <behavior> commit exists in the
work-branch history.feat(green): <behavior> commit follows the
corresponding test(red) commit.git log --oneline bc/<work_id> # inspect work-branch history
If the test(red) commit does not precede its feat(green) commit,
the gate fails — do not proceed to the next layer.
This inter-layer gate is what enforces test-first ordering across the parallel dispatch. The gate between layers is where the router verifies the observable artifacts of each completed layer before unblocking the next.
Once the gate passes for the current layer, return to Step 1. The next
bd ready call will return sub-issues whose dependencies were just closed.
Repeat until bd ready returns empty (all sub-issues are closed).
After the DAG is fully drained:
integrating-to-main to land the work branch on origin/main.digraph sdd_loop {
rankdir=TB;
ready [label="bd ready\n→ list unblocked sub-issues", shape=box];
empty [label="DAG drained?", shape=diamond];
parallel [label="Dispatch ALL ready sub-issues\nIN PARALLEL to bc-implementer subagents", shape=box, style=filled, fillcolor="#ccffcc"];
wait [label="Wait for all dispatched\nsubagents to complete", shape=box];
gate [label="Gate between layers:\n• sub-issue closed?\n• test(red) precedes feat(green)?", shape=diamond, style=filled, fillcolor="#ffffcc"];
fail [label="Gate FAILED:\nblocked — investigate", shape=box, style=filled, fillcolor="#ffcccc"];
integrate [label="integrating-to-main", shape=box];
review [label="Dispatch bc-reviewer", shape=box];
ready -> empty;
empty -> integrate [label="yes"];
empty -> parallel [label="no"];
parallel -> wait -> gate;
gate -> ready [label="pass"];
gate -> fail [label="fail"];
integrate -> review;
}
Each sub-issue has a focused scope. Before starting a sub-issue:
This discipline keeps implementation correct sub-issue by sub-issue rather than drifting across a sprawling context.
The implementer executing this skill never emits work_done for scenario-based work. That is the reviewer's gate — and the reviewer's alone.
Specifically:
assign_scenarios → implementer closes all sub-issues, runs the outer Gherkin loop, and hands off to the bc-reviewer via the router.request_bugfix with non-empty scenarios → same.request_bugfix with no scenarios, request_maintenance → implementer emits work_done directly (no reviewer dispatch).If you are the implementer and the work carried scenarios, your job ends at: all sub-issues closed, outer Gherkin scenario(s) pass, working tree clean, hand off to reviewer.
After all sub-issues are closed, before handing off:
# Run the assigned Gherkin scenario(s) explicitly
pytest features/ --tags="@scenario_hash:<hash>" # or your BC's runner
If any assigned scenario fails: reopen the relevant sub-issue, return to the TDD inner loop, do not hand off yet.
bd show <work_id> # see all sub-issues and their status
bd update <sub_id> --claim # claim a sub-issue
bd close <sub_id> # close a completed sub-issue
bd create "behavior" --parent <work_id> # add a newly discovered behavior
If implementation reveals a behavior not captured in the original decomposition, create a new sub-issue for it (bd create --parent <work_id>) rather than silently expanding the scope of an existing one.