Handoff Button vs Manual Prompt | Generated by AI

Home 2026.06

Now I have the full picture. Here’s the difference:

== “Start Implementation” Button vs Manual Prompt ==

THE HANDOFF BUTTON (planAgentProvider.ts:225-231):

startImplementationHandoff = {
    label: 'Start Implementation',
    agent: 'agent',          // switches to main coding agent
    prompt: 'Start implementation',
    send: true,              // auto-sends the prompt
    model: implementAgentModelOverride  // optional model switch
}

When you click “Start Implementation”:

  1. MODE SWITCH — It calls workbench.action.chat.toggleAgentMode to switch from the Plan agent to the main Agent (coding) mode.

  2. AUTO-SEND — It sends “Start implementation” as the user message automatically (send: true). You don’t type anything.

  3. FULL CONVERSATION HISTORY CARRIES OVER — This is the critical difference. The implementation agent receives the ENTIRE planning conversation: all the Discovery research, Alignment Q&A, Design iterations, and Refinement rounds. The agent has context of:
    • What the Explore/Search subagents found
    • What questions were asked and answered
    • What alternatives were considered and rejected
    • The final approved plan
  4. MODEL OVERRIDE — Can use a different model for implementation via ConfigKey.ImplementAgentModel setting. Plan might use a cheaper model for research, implementation uses a stronger one.

MANUALLY ADDING plan.md + PROMPTING “do it”:

  1. NO MODE SWITCH — You stay in whatever mode you’re already in. If you’re still in Plan agent mode, it will refuse to implement (“Your SOLE responsibility is planning. NEVER start implementation.”)

  2. YOU MUST BE IN AGENT MODE — You need to manually switch to Agent mode first, then attach plan.md and type your prompt.

  3. PLAN.md ONLY — The agent only sees the plan document itself. It does NOT see:
    • The research Discovery phase findings
    • The Q&A Alignment discussion
    • What was explored and rejected
    • The reasoning behind decisions

    The plan.md is a summary; the conversation is the full context.

  4. YOUR OWN PROMPT — You write whatever you want. Could be “implement step 1 only” or “do this but skip the tests”. More control, but less automatic.

  5. NO MODEL OVERRIDE — Uses whatever model the Agent mode is configured with.

== PRACTICAL IMPACT ==

The handoff is richer because the implementation agent gets the full planning transcript as conversation history. With manual prompt, it only gets the plan.md file — a compressed artifact. For complex plans where the “why” matters (edge cases discovered during research, tradeoffs discussed), the handoff preserves that context. For simple plans where the plan.md is self-contained, the difference is minimal.

The handoff also has a second option: “Open in Editor” which creates an untitled file with the plan for further editing before you commit to implementation.

Back Donate