github / github/copilot-cli

Background task agents silently substitute requested model without warning

未关闭
#3,266 0 条评论 2 个 reaction 已指派 0 人 在 GitHub 查看
area:agents area:models
主要语言
Shell
星标
11.2k
派生
1.9k
平均合并
14 小时 16 分钟
30 天内合并 PR
6

描述

### Describe the bug

When launching a background agent via the task tool with an explicit model parameter, the agent silently substitutes
a different model and completes without any indication of the substitution — until you inspect the model: field in
the result metadata.

Environment: GitHub Copilot CLI v1.0.44, Windows

### Affected version

GitHub Copilot CLI 1.0.44

### Steps to reproduce the behavior

Steps to reproduce:

1. Launch a background agent with model: "claude-opus-4.7" or model: "gpt-5.5" or model: "gpt-5.3-codex", by prompting "do a tri-model review using claude opus 4.7, gpt 5.5 and gpt 5.3 codex" with claude sonnet as the active model.
2. Agent completes with status: idle
3. Read result — model: claude-sonnet-4.6 appears in the metadata, not the requested model

Actual behaviour:

Agent silently downgrades to claude-sonnet-4.6 and returns a result. No warning, no error, nothing in the output
indicating the substitution occurred. The substitution is only discoverable by reading the raw metadata line model:
in the agent result.

Why this matters:

Workflows that depend on specific model capabilities (e.g., multi-model review loops requiring claude-opus-4.7,
gpt-5.5, and gpt-5.3-codex for independent perspectives) are silently invalidated. The caller believes they got
three independent reviews when in fact two were the same model. gpt-5.3-codex appeared to work correctly; it was
claude-opus-4.7 and gpt-5.5 that substituted.

### Expected behavior

Either:

- Use the requested model, or
- Fail with a clear error: "Model 'claude-opus-4.7' is not available in this environment" so the caller can decide
how to proceed

### Additional context

_No response_

贡献指南

打开贡献指南

调研方向

首先,使用 claude-opus-4.7 或 gpt-5.5 等明确指定的模型复现后台任务,然后检查结果元数据,其中会显示被替换的模型。跟踪任务工具的模型选择,并使完成操作要么使用请求的模型,要么报告该模型不可用;验证不再发生静默替换。

由索引模型根据 Issue 内容生成。

评估

技术栈
shell
领域
ai, cli
Issue 类型
缺陷
难度
3/5
预计耗时
1-2 天
活跃度
冷清
描述清晰度
基本清楚
新手友好度
52/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。