pr-cycle - SKILL.md Agent Skill

name: pr-cycle description: "Automated PR review cycle for codesmith when PR mode applies. Opens the PR, waits for Gemini Code Assist review via the `await_pr_review` tool, dispatches `codesmith-validator` to assess each unresolved thread against the codebase, implements approved fixes via `codesmith-worker`, responds to threads, requests re-review, and loops until Gemini approves or a context-handoff signal fires. Escalates to human for the final merge decision. NEVER auto-merges. Used at V-phase PR-review gate (Lock 11 gate 5) when PR mode applies; pairs with `local-merge` for the non-PR alternative."

Execute workflow by following instructions and producing output Always read COMPLETE files - NEVER use offset/limit when reading any workflow related files Instructions are MANDATORY - either as file path, steps or embedded list in YAML, XML or markdown Execute ALL steps in instructions IN EXACT ORDER Save to template output file after EVERY "template-output" tag NEVER skip a step - YOU are responsible for every step's execution without fail or excuse

Steps execute in exact numerical order (1, 2, 3...) unless an explicit goto directs otherwise Optional steps: Ask user unless YOLO mode active Template-output tags: Save content, discuss with the user the section completed, and NEVER proceed until the user indicates to proceed (unless YOLO mode has been activated)

For each step in instructions:

  <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>

Full user interaction and confirmation of EVERY step at EVERY template output - NO EXCEPTIONS except YOLO mode Skip all confirmations and elicitation, minimize prompts and try to produce all of the workflow automatically by simulating the remaining discussions with a simulated expert user

step n="X" goal="..." - Define step with number and goal substep n="Xa" title="..." - Sub-step within a step phase n="X" - Phase within a substep optional="true" - Step can be skipped if="condition" - Conditional execution for-each="collection" - Iterate over items repeat="n" - Repeat n times action - Required action to perform action type="script" - Execute shell command action if="condition" - Single conditional action (inline, no closing tag needed) check if="condition">...</check> - Conditional block wrapping multiple items (closing tag required) ask - Get user input via Question tool (ALWAYS wait for response before continuing). Present defined options as selectable choices. For open-ended questions, offer helpful suggestions with custom input enabled goto - Jump to another step (explicit loop-back) invoke-skill - Call another skill invoke-agent - Delegate to subagent via Task tool invoke-protocol - Execute a reusable protocol (e.g., discover_inputs) template-output - Save content checkpoint critical - Cannot be skipped example - Show example output note - Informational content output - Display to user without saving halt-conditions - Define conditions that MUST stop execution condition - Single halt condition validation - Pre-condition rules to check

- This is the complete workflow execution engine - You MUST follow instructions exactly as written - This workflow uses INTENT-DRIVEN PLANNING - adapt organically to product type and context - YOU ARE FACILITATING A CONVERSATION with a user to produce a final document step by step - The whole process is meant to be collaborative - helping the user flesh out their ideas - Do not rush or optimize and skip any section

Run the PR review cycle for codesmith: open PR → wait for Gemini → assess threads → implement fixes → respond → re-review → loop until done → escalate to human for merge.

<inputs>
  - bundle_dir       Implementation bundle directory
  - ticket_slug      Parent ticket slug
  - ticket_alias     Parent ticket canonical alias
  - branch           Implementation branch (from worktree-setup)
  - target_branch    Target to merge into
  - worktree_path    Worktree path
</inputs>

<outputs>
  Returns to caller (JSON):
  {
    "status": "READY_FOR_MERGE" | "CONTEXT_HANDOFF" | "TIMEOUT" | "ERROR",
    "pr_number": <n>,
    "pr_url": "https://...",
    "iterations": <n>,
    "threads_addressed": <n>,
    "threads_declined": <n>,
    "threads_deferred": <n>,
    "commits_added": <n>
  }
</outputs>

<hard-constraints>
  NEVER merge the PR. The final merge is always a human decision.
  Violation breaks the trust model.

  NEVER skip the `await_pr_review` tool. Polling Gemini state via
  `gh pr view` directly is forbidden — the tool handles settling
  windows and timeouts.

  NEVER trust Gemini blindly. Every thread requires independent
  verification via `codesmith-validator` before action.

  All threads must be either Addressed (with a commit SHA) +
  resolved, Declined + resolved, or Deferred (with a tracking
  issue) + resolved. No unresolved threads at exit.
</hard-constraints>

<step n="1" title="Initialize PR Cycle State">
  <action>Load and execute: references/steps/step-01-initialize.xml</action>
</step>

<step n="2" title="Open PR (first iteration only)">
  <action>Load and execute: references/steps/step-02-open-pr.xml</action>
</step>

<step n="3" title="Wait for Gemini Review">
  <action>Load and execute: references/steps/step-03-wait-for-gemini.xml</action>
</step>

<step n="4" title="Assess Unresolved Threads">
  <action>Load and execute: references/steps/step-04-assess-threads.xml</action>
</step>

<step n="5" title="Validate Assessment">
  <action>Load and execute: references/steps/step-05-validate-assessment.xml</action>
</step>

<step n="6" title="Implement Approved Fixes">
  <action>Load and execute: references/steps/step-06-implement-fixes.xml</action>
</step>

<step n="7" title="Respond and Resolve Threads">
  <action>Load and execute: references/steps/step-07-respond-threads.xml</action>
</step>

<step n="8" title="Request Re-Review and Loop">
  <action>Load and execute: references/steps/step-08-request-re-review.xml</action>
  <action>Increment iteration count; update state file.</action>
  <action>If context warning received: exit with CONTEXT_HANDOFF.</action>
  <action>Else: goto step 3 (wait for Gemini's response to the re-review request).</action>
</step>

<step n="9" title="Escalate to Human for Merge">
  <action>Load and execute: references/steps/step-09-escalate-merge.xml</action>
</step>

On successful path: PR is approved by Gemini, human has been shown the assessment summary, human makes the merge decision (not this skill). After merge, the caller (code-verify) runs `wt remove` to clean up the worktree, then transitions the ticket to `verified`.

  On context-handoff path: state is persisted at
  {bundle_dir}/pr-cycle-state.md; a fresh pr-cycle instance can
  resume from the after-timestamp captured in state.
</action>