Drive plan

drive-plan — execute to acceptance, not endless polish

Read delivery-methodology, the repo's AGENTS.md/CLAUDE.md, and linked delivery/testing docs before implementation. Preserve a supplied Project Root. Use the supplied Plan comm/file; create a separate Plan via worklog only when the initiative's complexity benefits from one.

Source of truth

Keep scope, acceptance, delivery slices, topology, decisions, material status, receipts, and remaining risks current in the Plan/worklog body. Put review batch summaries, steering exchanges, and escalations in comments. Local-file mode keeps both current state and a concise evidence log in the file.

Keep the working Comment.io route/identity. Invoke comment-identity only immediately before an uncredentialed direct-REST write; never switch to an ambient registered or Botlets identity.

Operating rules

• Make reasonable implementation decisions and record only material ones.
• Work one bounded delta at a time with an explicit base and intended head.
• Optimize for the simplest user-useful implementation; do not add speculative enterprise hardening, abstraction, or hypothetical edge-case machinery.
• Run the narrowest useful checks while editing.
• A phase is not automatically a review, staging, steering, or artifact gate.
• Review at a delivery-slice boundary, when a sensitive invariant changes, or when uncertainty makes review useful.
• Update the worklog at material milestones, not every command or minor edit.
• Stop when acceptance/evidence hold and no known severity blocker remains; record optional polish and unrelated cleanup as follow-ups.

Delta loop

For each delivery slice or direct candidate:

1. Normalize scope, acceptance, invariants, base SHA, topology, and evidence.
2. Implement the smallest coherent delta that satisfies that boundary.
3. Run focused checks. Use staging/preview only when deployed behavior adds material evidence.
4. Update the plan with the real outcome and remaining risk.
5. When review is warranted, invoke review-loop with explicit base/head and risk tier. Lock it to the declared subject and delta, fix one complete batch, and record its receipt. Route unrelated discoveries to a GitHub issue per delivery-methodology; do not absorb them into the slice.
6. Poll steer before an irreversible/materially ambiguous decision, after a long autonomous stretch, or at a delivery boundary. Continue when no real decision blocks useful work. Put any human-attention handoff in a comment. A newly added visible body @mention and a comment @mention both notify; use one surface for the handoff so the recipient is not pinged twice. On direct REST, that agent-authored notifying comment must include the structured notify brief.
7. Mark the boundary complete when acceptance and required evidence pass, no known blocker remains, mandatory repo rules hold, and residual risks are explicit.
8. Move to the next slice or hand the completed boundary to ship.

Normal work gets at most two finding-bearing review rounds. Repeated findings trigger redesign, targeted proof, residual-risk recording, or human escalation.

Lift execution

For a controlled lift, keep the ordered receipt ledger in the Project Root, merge slices sequentially into the lift with real merge commits, sync the release base periodically, and use delivery-methodology's freeze/coverage/promotion barrier. Do not let a late slice land after promotion begins.

Stop conditions

Stop for a human only when a new fact materially changes the goal/feasibility, or when an external action, permission, secret, or product decision blocks all useful work. Update the current-state body first and ask one crisp question.

Composes

Read the full SKILL.md before using delivery-methodology, worklog, steer, review-loop, or ship.

Repo config

Run the repo's commands, not hardcoded ones. Read AGENTS.md (else CLAUDE.md) and linked delivery/testing/deploy docs. If missing, infer focused checks and final certification from package/build/CI config and offer comment-init.