运行时
- 作者仓库星标 0
- 作者仓库 skills-registry
- 领域
- 通用
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @tomevault-io · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- macOS · Linux · Windows
- 运行环境要求
- Node.js
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 命令执行
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# Gemini CLI Runtime Contract
This skill defines how `polycli:polycli-provider-agent` (the subagent) interacts with the
polycli companion script for Gemini. Only invoked from within `polycli:polycli-provider-agent`.
## Primary helper
```bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/polycli-companion.bundle.mjs" rescue --provider gemini "<prompt>" --json
```
## Commands available to the agent
The provider-agent forwards whatever subcommand the caller supplied. Use `rescue` for multi-step agent work and `ask` for one-shot questions; the rest are slash-command-driven user flows.
| Command | Used by agent? | Purpose |
|---------|---------------|---------|
| `rescue` | Yes | Multi-step agent task (600s base, 1200s for deep-reasoning models — see Latency expectations) |
| `ask` | Yes | One-shot question (120s base, 240s for deep-reasoning models — see Latency expectations) |
| `setup` | No | User checks installation |
| `health` | No | User runs end-to-end probe |
| `review` | No | User triggers code review |
| `adversarial-review` | No | User triggers adversarial review |
| `status` | No | User checks job status |
| `result` | No | User fetches completed output |
| `cancel` | No | User cancels background job |
| `timing` | No | User inspects timing history |
Resumable thread state is exposed via the `--resume-last` flag, not a separate subcommand.
## Latency expectations
Only **deep-reasoning** Gemini variants (Pro / Thinking series) routinely spend 30s–several minutes silently reasoning before emitting visible text. Flash and other non-reasoning variants typically stream within seconds. This is upstream behavior, not a polycli stall.
To absorb the reasoning latency without inflating other models' budgets, the companion applies a **model-scoped** multiplier:
```js
PROVIDER_TIMEOUT_MULTIPLIERS = {
gemini: { "gemini-3.1-pro-preview": 2 },
opencode: { "kimi-for-coding/k2p6": 2 }, // also a code-reasoning model
}
```
(The opencode entry was added 2026-05-02 after a HumanEval/130 Tribonacci probe hit the 120s ask ceiling; the same pattern can be extended to any other reasoning model id.)
Resolution rules (per `resolveTimeoutMs` in `polycli-companion.mjs`):
- `--model gemini-3.1-pro-preview` → ×2 budget (the explicit reasoning case)
- No `--model`, cached upstream-default model is `gemini-3.1-pro-preview` → ×2 budget (the common case today)
- `--model <some-flash-or-other>` → ×1 (caller explicitly chose a non-reasoning model)
- Unknown / cache empty / non-gemini provider → ×1
Effective ceilings when the multiplier applies:
| kind | base | gemini reasoning |
|---|---|---|
| `ask` | 120s | 240s |
| `rescue` | 600s | 1200s |
| `review` | 300s | 600s |
| `adversarial-review` | 300s | 600s |
`health` keeps the universal 60s budget regardless of model — gemini's health prompt is sentinel-only and does not exercise reasoning.
If you cannot estimate how long a gemini prompt will take (e.g. open-ended diagnosis on a large diff), prefer `--background` and poll with `/polycli:status` / `/polycli:result` instead of relying on the foreground timeout. The background worker reuses the same `execution.timeout` (so gemini still gets the model-resolved budget), but the calling shell is freed immediately and the user can decide how long to wait.
When upstream releases a new reasoning-capable Gemini model id, add it to `PROVIDER_TIMEOUT_MULTIPLIERS.gemini`. Do not blanket-multiply the whole provider — that would over-budget Flash users.
## Routing controls
These are CLI flags, not prompt text:
- `--background` — run async, return job ID immediately
- `--write` — allow Gemini to modify files (maps to `--approval-mode auto_edit`)
- `--resume-last` — continue previous Gemini thread
- `--fresh` — start new thread (ignore previous)
- `--model <model>` — override the default model
- `--effort <low|medium|high>` — reasoning effort level
- `--prompt-file <path>` — read prompt from file instead of positional args
- `--json` — always use this for machine-readable output
## Safety rules
- **Preserve prompt text as-is.** Only reshape via `gemini-prompting` skill.
- **Never inspect the repo** from the agent. Claude does that.
- **Return stdout exactly.** No independent analysis.
- **Return nothing** if invocation fails (let Claude handle the error).
---
> Source: [bbingz/polycli](https://github.com/bbingz/polycli) — distributed by [TomeVault](https://tomevault.io).
<!-- tomevault:4.0:skill_md:2026-05-23 -->
章节与文件索引
按说明文件中的顺序排列。
代理可用的命令
延迟预期
路由控制
安全规则
文件/命令 -> polycli:polycli-provider-agent · rescue · ask · setup · health · review · adversarial-review · status
内容校验值 -> 6197fa86710a
当前原文的自动摘译,尚未人工审校。
代理可用的命令
提供者代理转发调用者提供的任何子命令。使用
rescue进行多步骤代理工作,使用ask进行一次性问题;其余的是斜线命令驱动的用户流程。命令·由特工使用? · 目的rescue· 是 · 多步骤智能体任务(基础 600 秒,深度推理模型 1200 秒 — 请参见延迟预期)ask· 是 · 一次性问题(基础 120 秒,深度推理模型 240 秒 — 请参见延迟预期)延迟预期
只有深度推理的双子座变体(Pro/Thinking 系列)通常会花 30 秒——几分钟的时间默默推理,然后发出可见的文本。 Flash 和其他非推理变体通常会在几秒钟内传输。这是上游行为,而不是 Polycli 停顿。为了吸收推理延迟而不增加其他模型的预算,配套应用了模型范围的乘数:(在 HumanEval/130 Tribonacci 探针达到 120 秒询问上限后,于 2026 年 5 月 2 日添加了开放代码条目;相同的模式可以扩展到任何其他推理模型 ID。)
路由控制
这些是 CLI 标志,而不是提示文本:
--background— 运行异步,立即返回作业 ID--write— 允许 Gemini 修改文件(映射到--approval-mode auto_edit)--resume-last— 继续上一个 Gemini 线程--fresh— 启动新线程(忽略上一个)安全规则
按原样保留提示文本。只能通过
gemini-prompting技能重塑。切勿检查代理的存储库。克劳德就是这么做的。准确返回标准输出。没有独立分析。如果调用失败则不返回任何内容(让 Claude 处理错误)。来源:bbingz/polycli — 由 TomeVault 分发。