Scope And Execute

Scope and Execute

Replaces the "should I /plan first or /loop?" decision with one workflow that handles the whole arc: understand → plan → execute → merge.

Auto-invocation triggers

• User asks to "fix X", "implement Y", "build the Z feature" with non-trivial scope
• User asks to "look at this repo and figure out what to do"
• Work spans multiple files or services
• Estimated >1h of effort

Workflow

Phase 1 — Analysis (always)

• New repo / unfamiliar code: invoke adt-repo-intake
• Multi-repo or workspace context: invoke ecosystem-health
• Already in active work: skip Phase 1
• Output: scope brief — which files, which contracts, which dependencies

Phase 2 — Context loading (always for non-trivial work)

• Invoke context-pack (or auto-context-pack if it has already fired this prompt)
• Pulls relevant code, standards, prior ADRs, related decisions
• Cap at 2k tokens

Phase 3 — Plan (skip for trivial single-file changes)

• Invoke plan to write a phased plan with validation per phase
• Output: .claude/plans/<task>.md
• Skip if scope is genuinely 1-2 files with obvious edits

Phase 4 — Execute (always)

• For independent parallel tracks: invoke dispatch
• For sequential inspect→act→verify: invoke loop
• For complex 3+ phase work: invoke three-man-team
• Implement against Phase 3's plan

Phase 5 — Pre-merge (always for shippable work)

• Invoke merge-confidently (or pr-merge-readiness directly)
• Resolve any blockers
• Invoke ship when verdict is MERGE

Phase 6 — Capture (always)

• If non-trivial decision was made: invoke adr-write
• Invoke knowledge-loop to update memory + RAG

Reconciliation

Output a scope summary at start, status updates between phases, final summary:

SCOPE AND EXECUTE — <task>
  Phase 1 Analysis:    <files in scope>, <key dependencies>
  Phase 2 Context:     <token count loaded>
  Phase 3 Plan:        <plan path, N phases>
  Phase 4 Execute:     <commits, files changed>
  Phase 5 Ship:        <PR number, merge SHA>
  Phase 6 Capture:     <ADR / memory paths>

Outputs / Evidence

• Scope brief
• Plan file
• Implementation commits
• Merged PR
• Decision capture

Failure / Stop Conditions

• Analysis reveals scope >1 week of work → stop, recommend breaking into phases
• Plan rejected by user → restart Phase 3 with their feedback
• Phase 5 verdict is FIX → stop and surface blocker, do not push through
• Never auto-merge without Phase 5 returning MERGE