Batch tool duplicate file reads increase context usage

[MasuRii]

When using batch tools, the antigravity models sometimes duplicate read the same file, causing unnecessary context overhead.

Target branch: dev

[mirrobot-agent]

@MasuRii Thank you for submitting this issue. I am now beginning my analysis and will report back shortly.

[mirrobot-agent]

Initial Analysis Report

Summary: This issue reports that when using batch/parallel tool calls with Antigravity models (Gemini 3, Claude Sonnet/Opus 4.5), the models sometimes redundantly read the same file multiple times within a single batch, causing unnecessary context token overhead.

Issue Validation: Needs More Info

Reproducibility Assessment: Needs More Info

Root Cause Analysis:

The issue appears to describe an upstream model behavior rather than a bug in this proxy codebase. Based on my investigation:

1. What this proxy does: The LLM-API-Key-Proxy is a routing proxy that forwards requests to backend LLM providers (including Antigravity). The antigravity_provider.py handles communication with Google's Antigravity API for Gemini 3 and Claude models.

2. Where duplicate reads would originate: When an LLM model is given tools (like file reading tools from coding agents such as Opencode, Continue, etc.), the model itself decides which tools to call. If the model requests to read the same file multiple times in parallel tool calls, that decision is made by the upstream model - not by this proxy.

3. No deduplication logic exists: The proxy currently does not deduplicate tool results or prevent duplicate tool calls - it faithfully forwards what the model requests and returns results.

4. Potential enhancement area: If this is a consistent pattern, the proxy could potentially implement:

   • Tool call deduplication before forwarding to the client
   • Response caching for identical tool calls within a request
   • Logging/metrics to track duplicate tool call patterns

Suggested Labels:

• enhancement - This appears to be a feature request for optimizing context usage
• needs-reproduction - Specific reproduction steps and examples would help validate the scope

Proposed Next Steps:

1. Clarify scope: Determine if this issue is requesting:

   • Client-side deduplication (in the coding agent/tool using this proxy)
   • Proxy-side deduplication (new feature for this codebase)
   • Upstream model behavior feedback

2. If proxy-side enhancement is desired: Consider implementing a tool call deduplication layer in the response processing that could:

   • Detect identical file read tool calls within a single response
   • Cache and reuse results for duplicate reads
   • Optionally consolidate duplicate calls before returning to the client

3. Gather data: Enable detailed logging to quantify how often this occurs and with which models/scenarios

Missing Information (if any):

• Reproduction steps: Specific example of a prompt/conversation that triggers duplicate file reads
• Model affected: Which specific model(s) exhibit this behavior (e.g., gemini-3-pro-preview, claude-sonnet-4-5, claude-opus-4-5)?
• Client context: Which coding agent/tool is being used with this proxy (Opencode, Continue, etc.)?
• Expected behavior: Should the proxy deduplicate tool calls, or is this feedback about model behavior?
• Frequency: How often does this occur (always, sometimes, under specific conditions)?

This analysis was generated by an AI assistant.

[Mirrowel]

When using batch tools, the antigravity models sometimes duplicate read the same file, causing unnecessary context overhead.

Target branch: dev

This is done via antigravity's model, and they directly do that themselves, sending tool calls. This is opencode-specific and not a proxy bug. This happens rarely, and by models choosing.
I welcome a PR to add opencode-specific fixes like this(can identify it via headers/prompt), but I don't wanna spend my time fixing something that is working correctly here and is a quirk of antigravity models.

Additionally, opencode is smart enough to remove duplicate tool results. So it is not much of a problem.