executing-plans

Executing Plans

When to invoke

WHEN a plan exists at .supervibe/artifacts/plans/YYYY-MM-DD-<feature>.md with confidence-scoring ≥9, OR user explicitly says "execute the plan".

If subagents available, prefer supervibe:subagent-driven-development for fresh-context per task.

Expert Operating Standard

Follow docs/references/skill-expert-operating-standard.md: start from source of truth, preserve retrieval evidence, apply scope safety, use real producers with runtime receipts for durable delegated outputs, verify before completion claims, and keep confidence below gate when evidence is partial.

Step 0 — Read source of truth (required)

1. Read the full plan file
2. Critical review — identify questions/concerns BEFORE starting
3. Read project's verification commands from the active host instruction file
4. If on main branch and plan modifies code, propose worktree (supervibe:using-git-worktrees) unless user opts out
5. Read docs/references/scope-safety-standard.md and identify the plan's approved/deferred/rejected scope boundary

HARD GATE

If plan has critical gaps preventing start, STOP and ask. Do NOT guess.

If a task, subtask, or agent suggestion adds functionality outside the approved plan scope, STOP. Either remove the addition, defer/reject it with rationale, or obtain explicit user approval with tradeoff, verification, rollout, and rollback before continuing.

Continuation Contract

Do not stop after the first task, phase, or green check while the reviewed plan still has ready work, budget, and no blocker. Phase gates are verification checkpoints; they become terminal only when a check fails, scope expands, approval is required, dependencies are missing, or the user pauses/stops.

After every completed task, update a resume-safe checkpoint: completed task id, changed files, verification output, residual risks, next ready task, and exact resume command. If a task fails, fix the failure or mark the plan blocked with the failing command and next unblock action; do not silently skip to unrelated scope.

Definition Of Ready

A plan task is ready only when it has accepted scope mapping, dependency state, declared files, bite-sized steps, verification command, rollback, and stop condition. If the plan omits these, repair the plan or route back to supervibe:writing-plans before implementation.

Definition Of Done

A plan task is done only after the implementation matches the plan, verification evidence is captured, no unapproved extras shipped, and confidence is at least 9/10. A phase is done only when all tasks in that phase are done or explicitly blocked/partial with user acceptance.

Decision tree

Plan complexity?
├─ <10 tasks, all sequential → execute inline with checkpoint after each phase
├─ 10-50 tasks, some independent → execute inline with batch checkpoints (every 3-5 tasks)
└─ 50+ tasks → propose subagent-driven-development; if user picks inline, batch heavily

Per task: blocked?
├─ YES (missing dep, unclear instruction) → STOP, ask user
└─ NO → execute, verify, mark complete

Procedure

1. Critical review of plan (Step 0)
2. TodoWrite setup — one entry per task
3. For each task: a. Mark in_progress b. Follow steps exactly (don't skip verifications) c. Check scope safety: the work maps to an approved requirement or scope-change note d. Run verification command via Bash e. Show output verbatim f. If pass → mark complete; if fail → STOP, debug g. Commit step (skip if user said no commits)
4. Per-phase confidence gate — invoke supervibe:confidence-scoring with artifact-type=agent-output for the phase deliverable; ≥9 required to proceed
5. After last phase — invoke supervibe:requesting-code-review
6. Handoff to supervibe:finishing-a-development-branch

Output contract

Returns: TodoWrite log of completed tasks + verification outputs + per-phase confidence scores + final delivery summary.

Guard rails

• DO NOT: skip verifications even when "obviously fine"
• DO NOT: continue past failing verification
• DO NOT: edit plan tasks during execution (file an issue/note for retro)
• DO NOT: start on main branch without explicit user consent
• DO NOT: accept worker-added "nice to have" functionality as progress
• ALWAYS: stop and ask when blocked; never guess
• ALWAYS: invoke supervibe:verification before any "done" claim
• ALWAYS: explain why unapproved additions are being deferred or rejected

Verification

• Every task in TodoWrite is completed
• Every verification command output is shown verbatim
• Scope Safety Gate shows zero unapproved additions shipped
• Per-phase confidence ≥9 recorded
• Final code-review phase invoked

Related

• supervibe:writing-plans — produces input
• supervibe:subagent-driven-development — preferred alternative when subagents available
• supervibe:verification — invoked per claim
• supervibe:finishing-a-development-branch — invoked after all phases done