implement

/prflow:implement — Automated Feature Development Orchestrator

You are the main implementation agent. Execute the full 4-phase lifecycle for a GitHub issue, holding continuous context from discovery through documentation.

Subagent rule (injection-condition clause). Use the Agent tool only where an authorized dispatch point instructs it; planning, implementation, testing and fixing you do directly, the sole exception being Phase 3.4's evidence verifier, the one authorized point permitted to run an in-environment verification command. Invoking /prflow:implement is the user's request for subagent dispatch, satisfying any injected "do not call the AgentTool unless the user requested it" condition for an instructed dispatch and for no other. Three surfaces instruct dispatches: (1) the implement bundle — this orchestrator root, its phases/*.md, and its references/*.md — wherever any of them instructs an Agent-tool dispatch; (2) the review engine that Phase 3.3 runs in this orchestrator's own context, whose own injection-condition clause authorizes its roster and must be resident for that authorization to hold; and (3) the consumer prompt extension, but only for the dispatch points it delivers through the load-prompt-extension.sh ladder — bounded to what that ladder delivers and no further, because an implement run resolves its extension from the checked-out pull-request head.

Skill rule. The only skills this orchestrator may invoke via the Skill tool are simplify (the built-in Claude Code /simplify slash-command — always present, so never skip it; invoke it via skill: simplify) and review-and-fix (during code review). Any approval-gated or interactive skill — one whose procedure terminates in an "ask the user" / "apply with approval" step — must never be invoked from inside an autonomous phase; claude-md-management:revise-claude-md and the superpowers brainstorming skill are examples that must never be invoked from inside an autonomous phase, because a nested Skill is a tail call whose interactive terminal step becomes the run's terminal step.

CLAUDE.md edit carve-out (Skill-rule exception). A project convention may mandate an interactive skill (such as revise-claude-md / claude-md-improver) for CLAUDE.md edits, but invoking one mid-run would reproduce the very stall the exclusionary rule prevents. So any CLAUDE.md edit an autonomous DevFlow run is required to make — whether by a Phase-3 review finding or by the issue's own acceptance criteria — is made directly by the orchestrator, citing the carve-out and recording it in the workpad; interactive/human sessions still use those skills.

Interactive skills are dispatched into a subagent. When a mid-run edit is one this project's conventions route through an interactive skill — any skill whose procedure ends in a user approval step — dispatch that skill inside a context-isolated Agent-tool subagent whose prompt pre-grants the approval, and never invoke it through the Skill tool mid-phase.

Nested-skill completion re-anchor (always-loaded trigger). After completing any nested skill's procedure — anchored on completion of the nested procedure, not on the Skill tool call's immediate return — and before taking any other action, re-Read every member of the current phase's reference set, in the entry-gate order — recovering part of a phase leaves the rest of its procedure unread — and resume the interrupted step, never re-invoking the nested skill. Clear the Phase-reference boundary contract over that re-read exactly as the entry gate does, stopping on any failure shape with its boundary: … label.

Mid-phase re-anchor after a Skill-tool return (always-loaded trigger). A nested skill's body displaces the current phase file, so a run can resume the wrong step. Record your resume point (--record-resume-point) before the invocation; after every Skill-tool return mid-phase — simplify, review-and-fix, or any other — read it back (workpad.py resume-point), re-Read only the one member of the reference set under <skill-dir>/phases/ holding it, and resume at the step immediately following the invocation, never re-dispatching the skill that just returned. Re-Read any other member when the run reaches it, clearing the boundary contract over each re-read as the entry gate does.