Agent silently drops context ("ghosts") immediately after approving a /plan

[singhc7]

What happened?

When using the /plan command, the agent successfully generates the markdown plan and prompts for approval. Upon approval, it registers the "Plan approved" event and saves the file. However, instead of transitioning into the execution phase, the agent silently drops its context loop and returns to the standard interactive prompt > without producing any code or error messages.

Steps to Reproduce

1. Launch gemini-cli.
2. Enter plan mode by typing > /plan.
3. Provide a task prompt (e.g., Make a tmux config file focusing on improving usability...).
4. Wait for the agent to generate the plan and approve it.
5. Observe the output: the CLI outputs Plan approved: [file path] but immediately drops back to the empty > prompt.

What did you expect to happen?

After the plan is approved and saved, the agent should automatically transition into the execution phase, following the steps outlined in the generated markdown plan to write the code.

Client information

Client Information

Run gemini to enter the interactive CLI, then run the /about command.

> /about
│ About Gemini CLI                                                                                                                                                          │
│                                                                                                                                                                           │
│ CLI Version                                                0.33.1                                                                                                         │
│ Git Commit                                                 f59f9e931                                                                                                      │
│ Model                                                      Auto (Gemini 3)                                                                                                │
│ Sandbox                                                    no sandbox                                                                                                     │
│ OS                                                         linux                                                                                                          │
│ Auth Method                                                Logged in with Google (b######t@gmail.com)                                                                     │
│ Tier                                                       Gemini Code Assist in Google One AI Pro                                                                        │                                                                                                    

Login information

Google Account

Anything else we need to know?

Environment:

OS: Fedora Linux

CLI Version: v0.33.1

Workaround:

Currently, the generated plan is saved but the agent drops the context. To proceed, I have to manually feed the generated plan back into the prompt to force execution using:

@/path/to/generated/plan.md Implement the steps outlined in this plan.

[Adib234]

What happens if you tell the agent "Please continue the plan you approved", does it not continue executing the plan?

[singhc7]

I didn't test that exact phrase on an independent prompt because of how the failure occurred. Here is the exact sequence of events:

1. The agent generated the plan and asked for approval.
2. I approved it, specifically selecting the option for step-by-step changes (where it should present each edit for my manual approval).
3. It started processing (showing the loading indicator).
4. It never finished processing, never asked for a single change's approval, and made absolutely zero edits to any files.
5. The loading screen just went away, and it silently dropped back to an empty prompt without throwing an error.

Basically, I think it completely failed the transition from "plan mode" to "manual mode".

Because it ghosted entirely and seemed to lose the active context, my immediate workaround for the next prompt was to manually attach the generated file (@/path/to/the/plans/tmux_config.md) and instruct it to implement the steps.

[Adib234]

can you share the chat if you still have the session open by running /chat share debug.json?

[singhc7]

I don't have the active session open anymore to run /chat share, but I grabbed the raw session.json log from ~/.gemini/tmp/ for that exact run.

Looking at the end of the JSON, you can clearly see the bug in action:

1. The agent calls exit_plan_mode and gets the "Plan approved" response.
2. In the final message block, the agent begins "thinking" about "Proceeding with the Next Step" and "Implementing the Write Task".
3. It then immediately returns "output": 0 with an empty content string and drops the loop entirely without making any tool calls to actually edit the files.

I've attached the log file here for your team to review.

(Note: I ran a quick sed replacement on the attached JSON to scrub my local system username from the file paths for privacy, but otherwise the log is untouched).

session-2026-03-13T02-52-55ec1c2d.json

[Adib234]

Is it possible to resume the session and then run /introspect to understand why the model didn't call any tool?