draft-agent-loop（草案 Agent 循环）

Draft Agent Loop Skill (HITRL)

Use this skill to implement a rigorous human-agent collaboration loop. This is the "Human-in-the-Right-Loop" (HITRL) method, designed to eliminate "blind box" agent outcomes by forcing plan approval and result verification.

Scope: This skill orchestrates remote agent workflows using the draft CLI. All persistence is through Draft pages (via the canonical draft-cli dependency). It does not write to local disk or agent memory.

Trigger Guidance

Trigger this skill when:

• "Work on this task and check with me before and after."
• "I want to review your plan first before you do anything."
• "Use HITRL / use draft-agent-loop for this."
• "Apply structured oversight to this task."
• New task received where the user's intent is high-stakes or complex enough to warrant human gates.
• A new iteration is requested after a Phase 3 sign-off.

Do NOT trigger this skill when:

• The user asks to do a task directly with no mention of approval or review gates.
• The user only asks about raw Draft commands or page automation without approval gates (use draft-cli).
• The user wants a local-file authoring workflow (authorship in local markdown).

Core Rules

• Source of Truth: The "Task Journal" Draft page. All plans, logs, and results live there.
• Environment: Always use headless page mode through draft ... --runtime v2.
• Runtime dependency: Follow the startup and page-operation rules from draft-cli.
• Handoff Mode: Blocking. STOP and wait for human approval/sign-off in the chat before proceeding to the next phase.
• No Sensitive Data in Logs: Do NOT include credentials, secrets, tokens, or PII in execution log entries or plan documents. Limit evidence to status indicators and non-sensitive file names.

Phase 0: Setup & Connection

Before doing anything, establish a stable Draft connection:

# 1. Start the daemon in headless runtime v2
draft start-server

# 2. Confirm the session is READY before proceeding
draft status --json

If draft status does not show a healthy v2 headless session, follow the draft-cli recovery pattern:

• DAEMON_OFFLINE → re-run draft start-server
• wrong runtime selected → stop and correct runtime before writing
• Only proceed once draft status --json shows a healthy headless v2 session

Phase 1: Plan (Proposal & Approval)

Before executing any code or changes:

1. Create Journal: Create a new Draft page titled <Task Name> - Task Journal.

    draft page create "<Task Name> - Task Journal" --json

1. Submit Plan: Author a detailed Task Journal using the mandatory template. Append it to the Journal.

    cat << 'EOF' | draft page append <id> --json
    # 📋 Task: [Title]

    ## Context
    [Detailed background and motivation]

    ## Problem / Goal
    [Specific issue being solved or feature being added]

    ## Acceptance Criteria (Outcome-Focused)
    - [ ] [Criterion 1]
    - [ ] [Criterion 2]

    ## Implementation Notes
    [Current state analysis, relevant files, suggested approach, technical considerations]

    ## Test Coverage
    [Test strategy, scenarios to test, existing tests to reference]
    EOF

1. Confirm Before Publishing: Before publishing, explicitly confirm with the user: "I am about to publish the Task Journal for external review. Please confirm."
2. Handoff: On confirmation, publish the page and ask for approval.

    draft page publish <id> --invite-code "${GLOBAL_INVITE_CODE:-innosage}" --json

Handoff Phrase: "I have initialized the Task Journal with the plan and requirements: [URL]. Please review the context and acceptance criteria. Once you are ready for me to proceed, please reply with APPROVED or LGTM here in the chat."

1. Wait: STOP. Do not proceed until the user explicitly tells you to continue in the chat. Note: Do not rely on Draft page comments for approval on remote/public pages.

Phase 2: Execute (Action & Logging)

Once approved:

1. Verify Approval: Confirm the user has provided approval in the chat.
2. Execute: Perform the tasks outlined in the plan.
3. Log Evidence: For every significant action, append a concise, non-sensitive log entry to the Journal under a # 📜 Execution Log section. Do NOT include raw command output, file contents, or credentials.

    cat << 'EOF' | draft page append <id> --json
    ### [Timestamp] Action: [Description]
    - **Status**: Success/Failure
    - **Evidence**: [e.g., "Modified src/components/Button.tsx — added reveal prop"]
    EOF

Phase 3: Verify (Result & Sign-off)

Once the execution is complete:

1. Submit Results: Append a # ✅ Final Results summary to the Journal. Include links to artifacts (e.g., PR URL, Draft page URL). Do not include raw file dumps.
2. Confirm Before Publishing: Explicitly confirm with the user before re-publishing.
3. Handoff: On confirmation, re-publish the page.

    draft page publish <id> --invite-code "${GLOBAL_INVITE_CODE:-innosage}" --json

Handoff Phrase: "I have completed the task. Please verify the results in the Task Journal: [URL]. If satisfied, reply with DONE or ✅ here in the chat."

1. Wait: STOP. If the user provides sign-off (DONE / ✅), proceed to Phase 4. If feedback is received, enter the Iteration Loop.

Phase 4: Archive (Draft Page Only)

After sign-off, append a final summary to the existing Task Journal page. This keeps all persistence within Draft — no local filesystem writes.

cat << 'EOF' | draft page append <id> --json
# 🗂 Task Complete — Summary
- **Outcome**: [Brief description of what was achieved]
- **Key Decisions**: [Any notable trade-offs or design choices]
- **Artifacts**: [Links to PR, relevant files, or external references]
EOF

The Draft page itself (with its page ID and URL) is the durable record. No additional writes to TASK_LOG.md, knowledge/, or other filesystem locations are required or expected.

Iteration Loop

If the user provides feedback or requests changes after Phase 3:

1. Acknowledge: Acknowledge the feedback in the chat.
2. Start New Iteration: Enter Phase 1 again to propose how you will address the feedback.
3. Append to Same Page: Do NOT create a new page. Append the new plan to the existing Task Journal under a new heading: # 📋 Iteration [N]: Addressing Feedback.
4. Template: Use the mandatory Phase 1 template for each new iteration to maintain context and traceability.

Non-Goals

• Do NOT use browser-backed v1 or --mode workspace.
• Do NOT skip the plan approval gate.
• Do NOT execute multiple un-logged steps.
• Do NOT write to local agent filesystem (no TASK_LOG.md, no knowledge/ writes).
• Do NOT include credentials, PII, or sensitive command outputs in Draft page content.