Start implementation posts into current Plan session instead of handing off to implementation

[loofiboss-bit]

Summary

Pressing Start implementation in a Plan workflow does not create an implementation session. It sends Start implementation as a normal message into the existing Plan chat, and the next response is still governed by the same Plan-only prompt.

Repro steps

1. Start a task in Plan mode
2. Let the plan be created and approved
3. Press Start implementation
4. Observe that the same chat continues in Plan mode

Expected behavior

The action should create a new implementation/coding handoff session, or otherwise switch the workflow into implementation mode.

Actual behavior

The action stays inside the same Plan session. The model responds as if it is still in Plan mode instead of actually starting implementation.

Why this looks like a product bug

From the session log:

• the Start implementation action was recorded as a normal user message
• the next model call reused the same session
• it also reused the same system prompt file
• there was no new handoff/new session event
• the active prompt still explicitly said the chat was in Plan mode and must never start implementation

This makes it look like the handoff action itself did not create the implementation session.

Environment

• GitHub Copilot Chat: 0.44.2
• VS Code: 1.116.0
• OS: Windows

Workaround

Open a fresh coding/implementation chat and pass the approved plan manually instead of using the current Start implementation action.

[loofiboss-bit]

Stronger log evidence from the same session:

• main.jsonl:297 records Start implementation as a plain user_message with the same session id (c4388bd1-3b83-48c8-bf3d-795fb1b9506c).
• main.jsonl:300 shows the next model call is still in that same session and still uses systemPromptFile":"system_prompt_0.json" and toolsFile":"tools_1.json".
• main.jsonl:301 shows the assistant responding by updating session memory (marking the plan approved) rather than entering implementation.
• main.jsonl:309 shows a second model call that still uses system_prompt_0.json / tools_1.json.
• main.jsonl:310 explicitly responds: "I’m still in planning mode, so I can’t start editing".

The active prompt file was also still the Plan prompt at that point:

• system_prompt_0.json:2 contains "You are currently running in "Plan" mode"
• system_prompt_0.json:2 contains "Your SOLE responsibility is planning. NEVER start implementation."
• system_prompt_0.json:2 also says "Approval given → acknowledge, the user can now use handoff buttons"

So the evidence points to the button/action not creating a new implementation session at all. It looks like it injected text into the existing Plan session, and the model then followed the still-active Plan prompt exactly.

[aeschli]

Likely dup of #311420

[amirhossein92]

I am not sure if it was a coincidence or not, but I was able to bypass it by changing my model from GPT 5.4 to Auto (it went with sonnet 4.6)

VS Code: 1.117.0

[gcannardD2I]

I am not sure if it was a coincidence or not, but I was able to bypass it by changing my model from GPT 5.4 to Auto (it went with sonnet 4.6)

VS Code: 1.117.0

Here's another maybe coincidence. I was using Auto, which was consistently choosing GPT 5.3. The problem has gone away now that I've explicitly chosen Opus 4.6.