Agent Skills: code-grill

  <substep n="1a" title="Track Progress">
    <action if="first iteration">Create todo list from all instruction step titles (TodoWrite)</action>
    <action>Mark current step as in_progress (TodoWrite)</action>
  </substep>

  <substep n="1b" title="Handle Step Attributes">
    <check>If optional="true" and NOT YOLO -> Use Question tool to ask user whether to include (Yes/Skip)</check>
    <check>If if="condition" -> Evaluate condition</check>
    <check>If for-each="item" -> Repeat step for each item</check>
    <check>If repeat="n" -> Repeat step n times</check>
  </substep>

  <substep n="1c" title="Execute Step Content">
    <action>Process step instructions (markdown or XML tags)</action>
    <action>Replace {{variables}} with values (ask user if unknown)</action>
    <execute-tags>
      <tag>action -> Perform the action</tag>
      <tag>action type="script" -> Execute as shell command, capture output</tag>
      <tag>check if="condition" -> Conditional block wrapping actions (requires closing &lt;/check&gt;)</tag>
      <tag>ask -> Prompt user using Question tool and WAIT for response. When options exist, present them as selectable choices. For open-ended questions, provide helpful suggested options with custom input enabled</tag>
      <tag>invoke-skill name="skill-name" -> Execute another skill via OpenCode skill system</tag>
      <tag>invoke-agent name="subagent" prompt="task" -> Delegate to subagent via Task tool</tag>
      <tag>goto step="x" -> Jump to specified step (used for explicit loop-back in Q-R style flows)</tag>
    </execute-tags>
  </substep>

  <substep n="1d" title="Handle template-output Tags">
    <if tag="template-output">
      <mandate>Generate content for this section</mandate>
      <mandate>Save to file (Write first time, Edit subsequent)</mandate>
      <action>Display generated content</action>
      <action>Use Question tool to present checkpoint options. WAIT for response.
        <options>
          <option label="Continue" description="Approve this section and move to the next step" />
          <option label="Advanced Elicitation" description="Explore this section deeper before proceeding" />
          <option label="YOLO" description="Auto-complete the rest of this workflow without stopping" />
        </options>
        <if response="Advanced Elicitation">
          <action>Invoke advanced-elicitation skill</action>
        </if>
        <if response="Continue">
          <action>Continue to next step</action>
        </if>
        <if response="YOLO">
          <action>Enter YOLO mode for the rest of the workflow</action>
        </if>
      </action>
    </if>
  </substep>

  <substep n="1e" title="Step Completion">
    <action>Mark current step as completed (TodoWrite)</action>
    <check>If no special tags and NOT YOLO:</check>
    <action>Use Question tool to ask:
      <options>
        <option label="Continue" description="Proceed to the next step" />
        <option label="Edit" description="Revise this step before moving on" />
      </options>
    </action>
  </substep>
</step>

<step n="2" title="Completion">
  <check>Confirm all instruction steps have been executed</check>
  <output>Workflow complete.</output>
</step>

<inputs>
  Session state from code-orient (planning mode) or from a prior
  code-grill iteration:
  - bundle_dir, ticket_path, ticket_alias, ticket_slug, primary_tag
  - design_path (placeholder, not yet populated)
  - glossary (CONTEXT.md state)
</inputs>

<outputs>
  Updated bundle:
  - research-YYYY-MM-DD-{slug}.md (created on first turn with
    findings; appended per turn)
  - CONTEXT.md (inline updates as terms resolve)
  - DRs at {shared_folder}/decisions/ (when three-condition gate
    fires)

  When readiness gate passes, hands off to code-design.
</outputs>

<key-discipline>
  User-interaction model:
  - Q-phase is OVERWHELMINGLY codebase-directed, not user-directed.
  - User escalation ONLY when: ticket internally contradicts
    itself, AC criterion is genuinely ambiguous, or a foundational
    decision in the ticket appears wrong after research.
  - Otherwise loop Q↔R against the codebase autonomously.

  Hide-the-intent boundary:
  - code-grill dispatches researcher with ONLY questions + scope +
    optional child hints. NOT the ticket. NOT the design.
  - researcher dispatches the 6 research children with the same
    discipline — they receive questions only.

  Q-phase output:
  - Research dispatch INSTRUCTIONS (in-memory structure: questions
    + which patterns + which subagents + scope hints).
  - NOT a written artifact. NOT confused with the P-phase plan
    doc. Lives in session state only.
</key-discipline>

<step n="1" title="Load Bundle State">
  <action>Load and execute: references/steps/step-01-load-bundle.xml</action>
</step>

<step n="2" title="Derive Kipling Questions">
  <action>Load and execute: references/steps/step-02-derive-kipling-questions.xml</action>
</step>

<step n="3" title="Formulate Research Dispatch Instructions">
  <action>Load and execute: references/steps/step-03-formulate-dispatch.xml</action>
</step>

<step n="4" title="Dispatch and Synthesize">
  <action>Load and execute: references/steps/step-04-dispatch-and-synthesize.xml</action>
</step>

<step n="5" title="Update Bundle Files">
  <action>Load and execute: references/steps/step-05-update-bundle.xml</action>
</step>

<step n="6" title="Readiness Gate">
  <action>Load and execute: references/steps/step-06-readiness-gate.xml</action>
</step>