bug: 'web'/'search' tool-category aliases silently resolve to no bound tool in headless --agent dispatch
- 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
## Summary
When an agent's frontmatter declares tools using the category aliases `web` or
`search`, those aliases silently resolve to **no bound tool at all** when the
agent is dispatched headlessly via `copilot --agent -p ... --allow-all-tools`.
There is no error or warning — the agent simply has no web-fetch/web-search
capability, which is easy to miss since nothing fails loudly.
Other category aliases in the same position work correctly: `read` → `view`,
`edit` → `edit`, `shell` → `powershell`, `agent` → `task/read_agent/list_agents`.
Only `web` and `search` are affected.
## Environment
- GitHub Copilot CLI 1.0.68 (Windows)
- Invocation: `copilot --agent -p "" --allow-all-tools`
## Repro
Agent frontmatter:
```yaml
---
name: tool-probe
tools:
- read
- web
- search
---
```
Dispatch:
```
copilot --agent tool-probe -p "List every tool you have available to you right now, one per line, then stop." --allow-all-tools
```
**Observed bound tools:** view, skill, sql — no web/search tool at all.
Changing the frontmatter to the literal tool names instead of the category aliases:
```yaml
tools:
- read
- web_fetch
- web_search
```
**Observed bound tools:** view, web_fetch, web_search, skill, sql — correct.
## Impact
Any agent relying on the documented web/search category aliases for web
access has zero web tooling in headless dispatch, with no error surfaced
anywhere. We found this affected 5 of our own agent specs (including a
Chief-of-Staff-style orchestrator whose own web lookups were silently broken)
before switching to literal tool names as a workaround.
## Suggested fix
Category-alias resolution for web/search should map to web_fetch/web_search
in headless --agent dispatch mode the same way shell correctly maps to
powershell and read maps to view.
贡献指南
调研方向
首先使用文档化的 web/search 复现步骤,跟踪 headless --agent 分发期间 frontmatter 工具的类别别名解析。将这些别名与 read 和 shell 进行比较,然后验证 web_fetch/web_search 的字面行为。当 web 和 search 在 headless 分发中绑定预期工具,并且静默失败由回归检查覆盖时,即表示完成。
由索引模型根据 Issue 内容生成。
评估
- 技术栈
- shell
- 领域
- 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
- 68/100