Implement Next

Purpose

Advance exactly one ready task from the TASKS.md ledger through a single implementation cycle, using its persisted issue brief as context. One invocation advances at most one task; it never iterates. Repeat the command for the next task.

Preconditions

• TASKS.md exists and is an index-only ledger whose rows link to issue briefs.
• If TASKS.md is missing or contains no ready task, stop and report the ledger state; do not invent work.

Flow

1. Read TASKS.md and select the first task in state ready.
   • Stop at a human-gate task and explain the approval it needs; do not proceed past it.
   • Skip blocked and sequenced tasks; they are not ready.
   • If no ready task exists, stop and report.
2. Mark the selected task in-progress in TASKS.md through the client's write-approval flow.
3. Load the linked issue brief and run subagent-driven-change with the brief as Fresh Context. Its owner runs code-quality-reviewer before change-risk-reviewer; initiate or resume the closed change-risk handoff and report its current state, but do not independently invoke another review, redefine budgets, or reset counters.
4. Mark the task done only after the owner returns a validated terminal CLEAN handoff and required tests run green. Validate the closed ChangeRiskOrchestrationStateV1 record; its snapshot matches the current snapshot (modulo docs/review-learning/) and its terminal status is CLEAN. Do not infer completion from free-form prose or an unvalidated status. The next task requires a new invocation.
5. If the validated handoff reports NO_PROGRESS or NEEDS_HUMAN_REVIEW, report that outcome and its required escalation; do not mark the task done or start another review. Keep the task out of done until a human resolves the escalation.

Failure Path

Stop and report BLOCKED when any of these holds:

• implementer returns BLOCKED or NEEDS_CONTEXT,
• a review finding cannot be resolved within the brief's scope,
• GREEN is unreachable within the brief's declared seam.

On failure, mark the task blocked in TASKS.md with a one-line reason; if the declared seam was wrong, include why it failed. Then stop. Do not touch the next task, do not edit the brief, and do not continue to another task. A human decides whether to edit the brief, re-grill, or split the task, and flips the state back to ready.

Output

Report the selected task, the state transitions applied, the subagent results, the tests run, the change-risk handoff status when applicable, and the final state (done or blocked with its one-line reason).

Safety

• Do not upload source code.
• Do not read or print secrets.
• Do not iterate across tasks; one invocation advances at most one task.
• Do not self-approve; all writes, commits, and destructive steps go through the client's write-approval flow.
• Do not edit issue briefs or advance a task that failed.
• Do not propose bypassPermissions, tool pre-approval, dependency auto-installation, hosted execution, or remote MCP behavior.