gentle-sdd-apply

If the native sdd-apply sub-agent is available, delegate this command to it. Otherwise, read the skill file at ~/.claude/skills/sdd-apply/SKILL.md FIRST, then follow its instructions exactly inline.

The sdd-apply skill (v2.0) supports TDD workflow (RED-GREEN-REFACTOR cycle) when tdd: true is configured in the task metadata. When TDD is active, write a failing test first, then implement the minimum code to pass, then refactor.

CONTEXT:

• Working directory: Detect agent-side before proceeding by running git rev-parse --show-toplevel with the Bash tool; if that fails, run pwd with the Bash tool.
• Current project: Derive agent-side from the detected working directory basename. Do not use slash-command shell interpolation for this value.
• Artifact store mode: engram

TASK: Implement the remaining incomplete tasks for the active SDD change.

STATUS GATE: Read ~/.claude/skills/_shared/sdd-status-contract.md and produce structured status before acting. If $ARGUMENTS is missing or ambiguous, ask the user to choose and STOP. Do not guess. Continue only when status says apply is ready, spec/design/tasks exist, and actionContext allows implementation edits. If status reports workspace-planning with no allowed edit roots, STOP before launching apply or editing inline. Carry contextFiles, task progress, dependency states, and actionContext into the native sub-agent prompt when delegating.

ENGRAM PERSISTENCE (artifact store mode: engram): CRITICAL: mem_search returns 300-char PREVIEWS, not full content. You MUST call mem_get_observation(id) for EVERY artifact. STEP A — SEARCH (get IDs only): mem_search(query: "sdd/{change-name}/spec", project: "{project}") → save spec_id mem_search(query: "sdd/{change-name}/design", project: "{project}") → save design_id mem_search(query: "sdd/{change-name}/tasks", project: "{project}") → save tasks_id STEP A2 — CHECK PREVIOUS PROGRESS (before starting work): mem_search(query: "sdd/{change-name}/apply-progress", project: "{project}") → if found, save progress_id

• Previous apply-progress (if exists): mem_search(query: "sdd/{change-name}/apply-progress", project: "{project}") → read and merge STEP B — RETRIEVE FULL CONTENT (mandatory): mem_get_observation(id: spec_id) → full spec mem_get_observation(id: design_id) → full design mem_get_observation(id: tasks_id) → full tasks (keep tasks_id for updates) IF progress_id exists: mem_get_observation(id: progress_id) → read previous progress, skip completed tasks, MERGE when saving Update tasks as you complete them: mem_update(id: {tasks-observation-id}, content: "{updated tasks with [x] marks}") Save progress: mem_save(title: "sdd/{change-name}/apply-progress", topic_key: "sdd/{change-name}/apply-progress", type: "architecture", project: "{project}", capture_prompt: false, content: "{progress report}") Set capture_prompt: false when the Engram tool schema supports it; if an older schema rejects or does not expose the field, omit it rather than failing.

For each task:

1. Read the relevant spec scenarios (acceptance criteria)
2. Read the design decisions (technical approach)
3. Read existing code patterns in the project
4. Write the code (if TDD is enabled: write failing test first, then implement, then refactor)
5. Mark the task as complete [x]

Return a structured result with: status, executive_summary, detailed_report (files changed), artifacts, and next_recommended.

REVIEW ROUTING (post-verify, not post-apply): Return control to the parent orchestrator. Apply itself never routes to review — nextRecommended proceeds to verify once tasks are complete. If the parent later observes a fresh reviewOffer block, it may present and run only its exact invocation. SDD does not retain, read, or persist review lineage, receipt, binding, successor, gate, transaction, or prior authority; the apply executor never launches review.

{{GENTLE_AI_AUTHORITY_FIRST_TERMINAL_PROCEDURE}}

Approval returns one exact pending acknowledgement continuation. Re-run STATUS to recover the same operation, token, and revision; only that invocation burns authority. Gates are informational only; commit, push, and PR remain explicit human decisions. Never auto-launch Judgment Day.