Clarify Confirm Continue

Clarify-Confirm-Continue (ccc)

Disciplined task intake workflow for multi-step or complex user requests across codebases, document trees, and project workspaces. Ensures the agent and user are fully aligned before any modification or extensive execution begins.

Composability

• Design Interviews: Works seamlessly alongside the grilling skill when an extended multi-turn architectural interrogation is warranted before drafting the plan.

Procedural Workflow

Phase 1: Fact-Checking & Fork-Hunting

1. Fact-Check the Workspace First (Zero-User-Burden)

Before asking any questions, inspect the environment autonomously using read-only tools (grep_search, view_file, list_dir, code_search, moma_search):

• Codebases: Check function signatures, type definitions, conventions, and existing test suites.
• Documents & Project Files: Read relevant Markdown files (*.md), design docs (g3doc/, docs/), notes, specifications, references, and schemas.
• Rule: Never ask the user for verifiable facts that already exist in the workspace files or documentation.

2. Fork-Hunting (Genuine Decisions vs. Fake Dilemmas)

Evaluate the task for unresolved decisions by categorizing them:

• Examples of Fake Dilemmas (Do not ask—execute standard best practices):

  • "Should I write tests for the new feature?" (Always write tests).
  • "Should I fix the syntax error or leave it broken?" (Obvious brownie vs. kick-in-the-face).
  • "Should I format the code to comply with style rules?" (Standard hygiene).

• Examples of Genuine Decision Forks (Clarification required):

  • API & Architecture: Breaking change vs. deprecation shim; synchronous vs. streaming response; choosing between two existing architectural patterns.
  • Documentation & Non-Code: High-level executive summary vs. deep technical breakdown; standalone doc vs. appending to an existing guide; choosing between competing taxonomies.
  • Scope Boundaries: Inclusions vs. exclusions when both are reasonable and defensible.

3. Interrogate Unsettled Forks

If one or more Genuine Decision Forks exist:

• Present independent decision forks in a single focused ask_question call; ask dependent/sequential forks one at a time.
• State the trade-offs clearly.
• Always provide your recommended choice with rationale.
• Wait for user feedback before drafting the final execution summary.
• Anti-Smuggling Rule: Never convert unresolved decision forks into unverified "assumptions" to skip this interrogation step.

Phase 2: In-Modal Structured Summary & Readiness Gate (ask_question)

Once all genuine decision forks are settled (or immediately if zero genuine forks exist), invoke the ask_question tool with a single readiness gate.

[!IMPORTANT] Preventing UI Collapse: In many agentic harnesses (e.g., Jetski, Antigravity), pre-tool chat text emitted in the same turn as a tool call is automatically collapsed into a progress/thought accordion. To ensure the structured summary is immediately visible and prominent, embed the full markdown summary directly inside the question argument of ask_question.

Structure the question argument in ask_question using markdown:

• 🎯 Goals: What will be accomplished.
• 🛡️ Non-Goals / Scope Boundaries: What will explicitly be left untouched.
• 📌 Settled Decisions & Invariants: Agreed choices from Phase 1 and verified workspace facts (no unresolved questions or speculative assumptions).
• 🛠️ Execution Plan: High-level steps to be performed once approved.
• Conclude with: *How would you like to proceed?*

Configure the options array with:

• "(Recommended) Yes, proceed"
• "Open in artifact"
• "No, adjust in chat"

Phase 3: Handling Gate Selection

• If "(Recommended) Yes, proceed": Transition immediately to task execution.
• If "Open in artifact": Use write_to_file to save the structured summary and execution plan as a markdown artifact (.md file) in the session's artifact directory. Configure the file to request interactive approval/feedback (e.g., set ArtifactMetadata with UserFacing: true, RequestFeedback: true, and a concise Summary to render an interactive 'Proceed' button). If the harness does not support interactive gates, instruct the user in chat to review the file and reply 'Proceed' when they are ready. Stop calling tools and go idle to await reactive wakeup.
• If "No, adjust in chat" or Custom Write-In: Apply user feedback, adjust the scope or plan accordingly, and re-confirm if substantial changes were made.

Execution Safety Invariants

• No Pre-Approval Modifications: Do not edit workspace files, create branches, or run mutating commands during Phase 1 or Phase 2.
• Explicit Approval Gate: You must receive confirmation via ask_question (or write-in approval) before transitioning to execution.
• Global Guardrails Preserved: Approval at the confirmation gate authorizes that you understand the task and authorizes you to begin drafting changes. All standard version control, push, PR, and landing approval gates in AGENTS.md remain independently and strictly enforced during execution.