gemini-cli ACP mode skips tool output truncation, letting one command blow past the 1M token limit

barney-s · ghdev · 2026-10-16

The bug

Issue #29777 on google-gemini/gemini-cli reports that when gemini-cli runs as an ACP agent (gemini --acp), the tools.truncateToolOutputThreshold setting (default 40,000 characters) is never applied — tool output is sent to the model in full. A single shell command with large output can push the request past the model's input limit, failing with 400: input token count exceeds 1048576. In interactive and -p modes, the same command is truncated as documented.

Root cause

The normal scheduler path calls truncateOutputIfNeeded in tool-executor.ts;
The ACP path (acpSession.ts) calls invocation.execute() directly and records the untruncated llmContent as the function response;
Auto chat compression can't catch it either: it checks the token count from the previous response, which excludes the new tool output, so a single huge output fires before compression runs.

Observed in practice

One test command produced 1.2M characters, jumping context from 98k to 530k tokens in a single turn; compression only ran on the next turn;
Another produced 2.7M characters, exceeding the 1,048,576 token limit and failing the turn.

Agent developers wiring gemini-cli into ACP clients with auto-approval (yolo) should avoid commands with large outputs until this is fixed.

Original post →

gemini-cli ACP mode skips tool output truncation, letting one command blow past the 1M token limit

More from coding & agent