Save transcript to file before function call is performed by laurentsimon · Pull Request #17996 · google-gemini/gemini-cli

laurentsimon · 2026-01-30T21:44:31Z

Summary

This PR closes #17922.

It flushes transcript to disk to ensure it's available in a BeforeTool hook.

Details

Before this PR, transcript contains:

{
  "messages": [
    {
      "type": "user",
      "content": "analyze code test.py"
    }
  ]
}

After this PR:

{
   "messages": [
      {
       "type": "user",
       "content": "analyze code test.py"
      },
      {
       "type": "gemini",
       "content": "I'm calling the function...",
       "thoughts": [
        {
          "subject": "Initiating Code Exploration",
          "description": "I'm starting by examining the code located at `test.py`. My immediate plan is to use the `read_file` tool. This will help me acquire the content of the file so I can provide a comprehensive explanation.",
        }
       ]
     }
  ]
}

Related Issues

How to Validate

Pre-Merge Checklist

gemini-code-assist · 2026-01-30T21:44:45Z

Summary of Changes

Hello @laurentsimon, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses an issue where the chat transcript might not be fully available to BeforeTool hooks, particularly when the Gemini model provides internal 'thoughts' alongside or instead of direct text responses. By modifying the GeminiChat logic, it now explicitly tracks and records these 'thoughts' to the transcript, ensuring a complete and accurate record is persisted to disk before any subsequent tool executions.

Highlights

Transcript Persistence: The pull request ensures that the full chat transcript, including the Gemini model's internal 'thoughts', is flushed to disk before any function calls are performed. This guarantees that BeforeTool hooks have access to the complete interaction history.
New hasThoughts Flag: A new boolean flag, hasThoughts, has been introduced in GeminiChat to explicitly track whether a model's streamed response contains 'thought' parts.
Updated Transcript Recording Logic: The condition for recording model responses to the chat transcript has been updated. Messages are now recorded if they contain either explicit responseText or if the hasThoughts flag is true, ensuring 'thoughts' are always captured.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request updates the chat recording logic to save the transcript to a file even when the model's response contains only thoughts and no text, which is useful for BeforeTool hooks. The changes involve adding a hasThoughts flag that is checked before recording the message.

My main feedback is regarding a redundant check for thoughts being added, which could be consolidated with an existing check to improve efficiency and maintainability. Otherwise, the logic seems correct and addresses the intended issue.

packages/core/src/core/geminiChat.ts

gemini-cli · 2026-01-31T02:52:43Z

Hi there! Thank you for your contribution to Gemini CLI. We really appreciate the time and effort you've put into this pull request.

To keep our backlog manageable and ensure we're focusing on current priorities, we are closing pull requests that haven't seen maintainer activity for 30 days. Currently, the team is prioritizing work associated with 🔒 maintainer only or help wanted issues.

If you believe this change is still critical, please feel free to comment with updated details. Otherwise, we encourage contributors to focus on open issues labeled as help wanted. Thank you for your understanding!

bdmorgan · 2026-01-31T19:42:14Z

Fixes #17922

laurentsimon · 2026-02-23T17:19:53Z

Hi there! Thank you for your contribution to Gemini CLI. We really appreciate the time and effort you've put into this pull request.

To keep our backlog manageable and ensure we're focusing on current priorities, we are closing pull requests that haven't seen maintainer activity for 30 days. Currently, the team is prioritizing work associated with 🔒 maintainer only or help wanted issues.

If you believe this change is still critical, please feel free to comment with updated details. Otherwise, we encourage contributors to focus on open issues labeled as help wanted. Thank you for your understanding!

yes this change is "critical"

theshloksschauhan · 2026-02-23T18:12:15Z

Hi @laurentsimon — just a quick follow-up. I’m still working with the CLI and keeping this branch up to date with main. The change has been working well locally and does preserve the session history when a tool call fails, which was the main issue I encountered while testing tool integrations. Whenever you have time, I’d really appreciate your thoughts on whether this approach aligns with the project’s intended behavior, or if you’d prefer a different direction for handling transcripts. Happy to adjust the implementation based on your feedback.

…

On Mon, 23 Feb 2026 at 22:50, laurentsimon ***@***.***> wrote: *laurentsimon* left a comment (google-gemini/gemini-cli#17996) <#17996 (comment)> Hi there! Thank you for your contribution to Gemini CLI. We really appreciate the time and effort you've put into this pull request. To keep our backlog manageable and ensure we're focusing on current priorities, we are closing pull requests that haven't seen maintainer activity for 30 days. Currently, the team is prioritizing work associated with *🔒 maintainer only* or *help wanted* issues. If you believe this change is still critical, please feel free to comment with updated details. Otherwise, we encourage contributors to focus on open issues labeled as *help wanted*. Thank you for your understanding! yes this change is "critical" — Reply to this email directly, view it on GitHub <#17996 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BE4BKIVEM5SQNE57VQHD4MT4NMZFNAVCNFSM6AAAAACTPQQCE2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSNBWGEZTGMZYGU> . You are receiving this because you commented.Message ID: ***@***.***>

laurentsimon · 2026-02-23T18:25:32Z

@theshloksschauhan
where is your code? How does it differ from this PR?

theshloksschauhan · 2026-02-23T18:41:13Z

Hi @laurentsimon — thanks for the question, and sorry for the confusion caused by my wording. When I said “my branch/code,” I meant the branch for this PR (#17996) (i.e., the changes under review here), not a separate implementation. So there isn’t another patch to compare against—what I’ve tested locally is exactly what’s in this PR: flush the transcript to disk before the tool/function call runs, so a BeforeTool hook can reliably read the latest transcript state (including in cases where the tool call fails). If it’d help, I can add clear repro/validation steps to the PR description to make the behavioral difference explicit.

…

On Mon, 23 Feb 2026 at 23:56, laurentsimon ***@***.***> wrote: *laurentsimon* left a comment (google-gemini/gemini-cli#17996) <#17996 (comment)> @theshloksschauhan <https://github.com/theshloksschauhan> where is your code? How does it differ from this PR? — Reply to this email directly, view it on GitHub <#17996 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BE4BKIQPOAIQQOQQYA4HS634NNA3RAVCNFSM6AAAAACTPQQCE2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSNBWGUYTQMBSHA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

krishdef7 · 2026-02-25T10:24:39Z

Hi @bdmorgan @laurentsimon,

This change looks useful, especially for ensuring BeforeTool hooks have access to the latest transcript state.

One thing I was thinking about while reading this — currently the flush happens when handling model output (including thoughts). For cases where tool calls are triggered asynchronously, would it make sense to explicitly flush the transcript at the point where a tool call is dispatched as well?

That could ensure consistency even when execution paths diverge (e.g., streaming responses vs tool-triggered execution).

Happy to explore this further or help test edge cases if useful.

Thanks!
Krish

laurentsimon · 2026-02-25T16:58:35Z

Hi @bdmorgan @laurentsimon,

This change looks useful, especially for ensuring BeforeTool hooks have access to the latest transcript state.

One thing I was thinking about while reading this — currently the flush happens when handling model output (including thoughts). For cases where tool calls are triggered asynchronously, would it make sense to explicitly flush the transcript at the point where a tool call is dispatched as well?

That could ensure consistency even when execution paths diverge (e.g., streaming responses vs tool-triggered execution).

Happy to explore this further or help test edge cases if useful.

Thanks! Krish

Help greatly appreciated. I'm not super familiar with the codebase so I did the best I could to make it work, but like you say there are likely edge cases this PR does not cover

krishdef7 · 2026-02-25T17:23:24Z

Hi @laurentsimon — thanks, that makes sense.

I’ll take a closer look at the tool dispatch path and how it interacts with transcript persistence, especially in async and streaming scenarios.

Initial thought: it might help to explicitly flush the transcript right before tool execution is dispatched, so BeforeTool hooks always see the latest state (including thoughts), even if the execution path diverges.

I’ll validate this against:

tool failures mid-stream
multiple tool calls in a session
partial/streaming responses

If it holds, I’ll follow up with a concrete PR + tests shortly.

Does this direction align with what you had in mind?

laurentsimon · 2026-02-25T18:31:26Z

Hi @laurentsimon — thanks, that makes sense.

I’ll take a closer look at the tool dispatch path and how it interacts with transcript persistence, especially in async and streaming scenarios.

Initial thought: it might help to explicitly flush the transcript right before tool execution is dispatched, so BeforeTool hooks always see the latest state (including thoughts), even if the execution path diverges.

Makes sense, in particular for thoughts. Tool execution metadata is available in BeforeTool inputs, so having the tool info in the transcript is less critical

I’ll validate this against:

tool failures mid-stream

multiple tool calls in a session

partial/streaming responses

If it holds, I’ll follow up with a concrete PR + tests shortly.

Does this direction align with what you had in mind?

That makes sense, yes.

Really appreciate the help, thank you for taking the time.

…esponses Extends the fix in google-gemini#17996: when the model emits a functionCall with no response text and no thoughts, recordMessage was not called, leaving BeforeTool hooks with no gemini entry in the transcript at dispatch time. - Add hasThoughts flag (from google-gemini#17996) to track thought-only responses - Add hasToolCall to flush condition so pure tool calls also trigger a transcript write before tool execution - Add regression test verifying gemini entry is written on pure tool call Fixes the edge case noted in PR google-gemini#17996 discussion.

krishdef7 · 2026-02-26T10:05:52Z

Hi @laurentsimon — I've opened a follow-up PR (#20419) that addresses the remaining edge case. When the model emits a pure functionCall with no text and no thoughts, recordMessage was still not called, so BeforeTool hooks saw no gemini entry at dispatch time. The fix extends the flush condition to responseText || hasThoughts || hasToolCall — a 3-line change in processStreamResponse, no new dependencies. Includes a regression test. Happy to rebase onto your branch if you'd prefer to land both together.

laurentsimon · 2026-02-26T17:04:52Z

Hi @laurentsimon — I've opened a follow-up PR (#20419) that addresses the remaining edge case. When the model emits a pure functionCall with no text and no thoughts, recordMessage was still not called, so BeforeTool hooks saw no gemini entry at dispatch time. The fix extends the flush condition to responseText || hasThoughts || hasToolCall — a 3-line change in processStreamResponse, no new dependencies. Includes a regression test. Happy to rebase onto your branch if you'd prefer to land both together.

Great. Let me close this PR then. Thank you!

laurentsimon requested a review from a team as a code owner January 30, 2026 21:44

gemini-code-assist bot reviewed Jan 30, 2026

View reviewed changes

packages/core/src/core/geminiChat.ts Outdated Show resolved Hide resolved

gemini-cli bot added the area/agent Issues related to Core Agent, Tools, Memory, Sub-Agents, Hooks, Agent Quality label Jan 30, 2026

laurentsimon added 2 commits January 30, 2026 22:01

Save transcript to file before function call is performed

efd1a19

set hasThoughts where it was already checked

509fb36

laurentsimon force-pushed the feat/hook-transcript branch from e8ee17e to 509fb36 Compare January 30, 2026 22:02

gemini-cli bot closed this Jan 31, 2026

bdmorgan added help wanted We will accept PRs from all issues marked as "help wanted". Thanks for your support! priority/p1 Important and should be addressed in the near term. labels Jan 31, 2026

bdmorgan reopened this Jan 31, 2026

Merge branch 'main' into feat/hook-transcript

e8253a0

laurentsimon mentioned this pull request Feb 2, 2026

Context of sub-agent is not available in transcript file and BeforeTool hooks #18128

Open

SandyTao520 force-pushed the main branch from 4a78a96 to 0bf3a01 Compare February 26, 2026 03:03

krishdef7 mentioned this pull request Feb 26, 2026

fix(core): flush transcript for pure tool-call responses to ensure BeforeTool hooks see complete state #20418

Closed

3 tasks

krishdef7 mentioned this pull request Feb 26, 2026

fix(core): flush transcript for pure tool-call responses to ensure BeforeTool hooks see complete state #20419

Merged

3 tasks

laurentsimon closed this Feb 26, 2026

Conversation

laurentsimon commented Jan 30, 2026

Summary

Details

Related Issues

How to Validate

Pre-Merge Checklist

Uh oh!

gemini-code-assist bot commented Jan 30, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

gemini-cli bot commented Jan 31, 2026

Uh oh!

bdmorgan commented Jan 31, 2026

Uh oh!

laurentsimon commented Feb 23, 2026

Uh oh!

theshloksschauhan commented Feb 23, 2026 via email

Uh oh!

laurentsimon commented Feb 23, 2026

Uh oh!

theshloksschauhan commented Feb 23, 2026 via email

Uh oh!

krishdef7 commented Feb 25, 2026

Uh oh!

laurentsimon commented Feb 25, 2026

Uh oh!

krishdef7 commented Feb 25, 2026

Uh oh!

laurentsimon commented Feb 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

krishdef7 commented Feb 26, 2026

Uh oh!

laurentsimon commented Feb 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

laurentsimon commented Feb 25, 2026 •

edited

Loading