ACP: expose contextTier as a session config option (parity with interactive /model picker)
- 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 feature or problem you'd like to solve
The interactive CLI lets you change the context window tier mid-session (via the `/model` context tier picker), but the ACP server does not expose `contextTier` as a session config option. ACP clients can only set the tier at spawn time with `--context `, and have no way to discover or change it once the session is running.
### Proposed solution
Expose the context tier over ACP with the same shape already used for `reasoning_effort`:
1. Include a `contextTier` entry in the `configOptions` array returned by `session/new` (and `session/load`), of `type: "select"`, with the eligible values for the current model (`default`, `long_context`) and the current value.
2. Accept `session/set_config_option` with `configId: "contextTier"` and apply it to the live session, exactly as the `/model` picker does interactively.
3. Only advertise the option for tiered-pricing models, mirroring the interactive behaviour ("context tier picker appears for eligible models").
This closes a parity gap rather than adding new functionality: the capability already exists in the runtime and is reachable from the interactive UI and from `settings.json`; it is just not reachable through ACP.
Issue #1994 ("Allow reasoning effort discovery and setting via ACP") is the direct precedent — `reasoning_effort` is now a first-class ACP config option, and `contextTier` is the remaining sibling setting that isn't.
### Example prompts or workflows
1. An ACP client (Zed, or any ACP editor integration) shows a "Context Size" picker next to the model and reasoning-effort pickers, populated from `session/new`, instead of having to hardcode the tier list per CLI version.
2. A user on Claude Opus hits a large refactor mid-conversation and switches from `default` to `long_context` without losing session state — today the client must kill and respawn `copilot --acp --context long_context` and resume, or the control is simply hidden.
3. A client drops back from `long_context` to `default` after the large task is done, to avoid the surcharge for the rest of the conversation.
4. A client reads the current tier from `session/load` when reconnecting to an existing session, so the UI reflects reality instead of guessing.
### Additional context
Reproduced against GitHub Copilot CLI **1.0.75**.
**The interactive CLI supports hot changes** — from `copilot help config`:
```
contextTier: context window tier for tiered-pricing models (e.g., "default" or "long_context").
- Can also be set with --context flag (overrides persisted setting)
- Can be changed with /model command (context tier picker appears for eligible models)
```
**ACP does not.** Minimal probe over `copilot --acp` (initialize → `session/new` → `session/set_config_option`):
```
configOptions ids: ["mode","model","reasoning_effort","allow_all"]
contextTier -> {"code":-32602,"message":"Unknown config option 'contextTier'."}
context -> {"code":-32602,"message":"Unknown config option 'context'."}
contextSize -> {"code":-32602,"message":"Unknown config option 'contextSize'."}
```
So `reasoning_effort` is exposed but `contextTier` is not, even though both are per-session model settings adjustable from the interactive `/model` flow.
**Current workaround and why it's poor.** The tier can only be pinned at spawn time via `--context `. Changing it for a running conversation requires terminating the agent process and respawning it with a different flag, then reloading the session — an intrusive lifecycle operation for what is a per-request setting, and one that ACP clients cannot make feel like the interactive picker.
贡献指南
调研方向
以现有的 reasoning_effort ACP 配置选项为先例开始,然后跟踪 session/new 和 session/load 的通告与处理,以及 session/set_config_option 的处理。将这些路径与交互式 /model 选择器和 --context 的行为进行比较。当符合条件的模型公开当前的 contextTier,并且客户端可以在实时会话期间更改它而无需重启时,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
- api, cli
- Issue 类型
- 功能
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 活跃度
- 活跃
- 描述清晰度
- 基本清楚
- 新手友好度
- 65/100