Make several changes to guide the model to request absolute paths,
reducing frequent accidental relative path tool call failures.
- Switch the parameter name: path --> absolute_path.
- Update the tool definition to strongly require an absolute path.
- Update the system prompt to indicate absolute paths are required.
- Update the system prompt tool use examples to use absolute paths.
Test case:
Open GC in GC: "Locate the primary file calling genai"
- Expected: Model opens files with absolute path, successfully.
- Actual (pre-patch): Failure, attempts to read with relative path.
- Actual (post-patch): Success, attempts to read with absolute path.
Related to https://b.corp.google.com/issues/423605555 - I figured this might be a simpler solution to start with, while still also being useful on its own even if we do implement that.
Similar to ctrl+c, ctrl+d can now be used to exit the program. To avoid accidental exit, ctrl+d must be pressed twice in relatively quick succession (same as ctrl+c).
Following common UX pattern, ctrl+d will be ignored when the input prompt is non-empty. This behavior is similar to how most shell (bash/zsh) behaves. To support this, I had to refactor so that text buffer is initialized outside of the InputPrompt component and instead do it on the main App component to allow input controller to have access to check the content of the text buffer.
This change simplifies telemetry setup for users who want to integrate with GCP, as they no longer need to configure a local collector or an explicit endpoint if their project is already set up in the environment.
This change updates the telemetry system to automatically export traces, logs, and metrics to Google Cloud Platform (GCP) if the `GOOGLE_CLOUD_PROJECT` environment variable is set and no explicit `telemetryOtlpEndpoint` is configured by the user.
Key changes:
- The default `telemetryOtlpEndpoint` in `Config` is now an empty string.
- The `initializeTelemetry` SDK logic now prioritizes:
- User-defined `telemetryOtlpEndpoint`.
- `GOOGLE_CLOUD_PROJECT` for direct GCP export.
- Console exporters as a fallback.
- If an invalid `telemetryOtlpEndpoint` is provided, it falls back to console exporters with a warning.
#750
Changes:
- Ensure proper shutdown in non-interactive mode
- Ensures the initial user prompt is logged in non-interactive mode
- Improve telemetry for streaming - handle chunks and input token count is now alongside other token counts in response
To test:
- Follow instructions in https://github.com/google-gemini/gemini-cli/blob/main/docs/core/telemetry.md#google-cloud
- Run CLI in non-interactive mode and observe logs/metrics in GCP Logs Explorer and Metrics Explorer
#750
Refactor OpenTelemetry logging for API requests, responses, and errors. Moved logging responsibility from GeminiClient to GeminiChat for more detailed logging.
#750
Add the user's decision (accept, reject, modify) to tool call telemetry to better understand user intent. The decision provides crucial context to the `success` metric, as a user can reject a call that would have succeeded or accept one that fails.
Also prettify the arguments json.
Example:
#750
CCPA uses a different format than GenAi. This adds conversion code to get it to the right format.
Note that this doesn't work against the current ccpa staging server, The changes it needs are in cl/770266927
Keir Mierle