@armadra/agent 0.7.6 → 0.8.0

package/CHANGELOG.md

@@ -5,6 +5,19 @@ English · [简体中文](CHANGELOG.zh-CN.md)
 > This file is in English starting with 0.6.0. Release notes for 0.1 through 0.5.1 are in Chinese in
 > [CHANGELOG.zh-CN.md](CHANGELOG.zh-CN.md). New entries go into both files.
 
+## 0.8.0 (2026-10-10)
+
+### External agents
+
+- **New `pi` driver**: `task(agent="pi")` drives `pi --mode rpc` (verified with pi 1.1.0). pi has no approval channel of its own, so ama loads a minimal extension for that run (`-e`, written to a per-session temp dir and deleted on close; pi's config is untouched) that routes every non-read-only tool call to ama: rejected in `plan` / `allowlist` (which also only enable `read,grep,find,ls`), `edit` / `write` allowed from `auto-edit` up, everything else asks you. Dialogs from your own pi extensions are cancelled, never answered. Two-turn resume, steer, interrupt, USD usage and context occupancy are reported. The community `pi-acp` adapter is not listed: it never sends permission requests (#198).
+- **Cursor CLI listed as unverified**: `task(agent="cursor")` runs `cursor-agent acp` per Cursor's ACP docs (modes `agent` / `plan` / `ask`); not tested locally because it is not installed (#198).
+- **Codex one-shot mode works outside git repos**: `codex exec` / `exec resume` now pass `--skip-git-repo-check` (ama only starts external agents in directories it already trusts); previously they failed immediately in a non-git directory (#198).
+- **ACP adapters no longer stay in a looser mode**: explicit mode maps for claude-agent-acp (`auto-edit` → `acceptEdits`, `full-auto` → `auto`, so a user default of `bypassPermissions` is replaced), codex-acp (`plan` / `default` → `read-only`, `auto-edit` → `workspace-write`, `auto` / `full-auto` → `agent`; `plan` used to refuse to start) and Copilot (URL mode ids `…#plan` / `…#agent`). Modes that open every permission are never chosen (#198).
+- **`model` reaches ACP agents**: `agents.<id>.model` and the `task` `model` argument are now applied through the agent's `model` session config option (matched by value, name, or the model part of `provider/model`); when nothing matches you get a notice and the agent's default model (#198).
+- **Usage numbers corrected**: Codex (app-server and `exec`) and Copilot report input tokens including cache hits, and Copilot's ACP usage is a session total; ama now subtracts the cached part and takes per-turn differences, so totals no longer count cache reads twice. Claude stream-json now reports context occupancy and window like the other drivers (#198).
+- **Verified ranges updated** after a real-CLI audit on 2026-10-10: claude-agent-acp 0.89, codex-acp 2.2, Copilot 1.0.95, OpenCode 1.18, pi 1.1; capability matrix in docs/agents.md, numbers in docs/benchmarks/external-agents-2026-10.md (#198).
+- **`model_unavailable` for ChatGPT models the account cannot use**: a codex-backend 400 "model is not supported when using Codex with a ChatGPT account" now fails with `model_unavailable` and tells you to refresh the list with `ama models discover chatgpt` (the backend withdrew `gpt-6-astra`, which an older discovery cache still listed) (#198).
+
 ## 0.7.6 (2026-10-10)
 
 ### Model efficiency

package/CHANGELOG.zh-CN.md

@@ -4,6 +4,19 @@
 
 > 从 0.6.0 起 [CHANGELOG.md](CHANGELOG.md) 为英文，本文件保留中文记录（0.1–0.5.1 的完整历史在此）。新条目两份都要加。
 
+## 0.8.0（2026-10-10）
+
+### 外部 Agent
+
+- **新增 `pi` 驱动**：`task(agent="pi")` 驱动 `pi --mode rpc`（pi 1.1.0 实测）。pi 自己没有审批通道，ama 以 `-e` 给这一次运行加载一个最小扩展（写在每个会话自己的临时目录、关会话即删，不碰 pi 的配置），把只读工具以外的调用交给 ama：`plan` / `allowlist` 直接拒绝（并只开 `read,grep,find,ls`），`auto-edit` 及更宽放行 `edit` / `write`，其余问你。你自己的 pi 扩展弹出的对话框一律取消、不代答。支持两轮续聊、steer、中断、美元用量与上下文占用。社区适配器 `pi-acp` 从不发权限请求，不收录（#198）。
+- **收录 Cursor CLI（未验证）**：`task(agent="cursor")` 按官方 ACP 文档运行 `cursor-agent acp`（模式 `agent` / `plan` / `ask`）；本机未安装，未实测（#198）。
+- **Codex 一次性模式在非 git 目录可用**：`codex exec` / `exec resume` 带上 `--skip-git-repo-check`（ama 只在自己已信任的目录里起外部 Agent）；此前在非 git 目录直接失败（#198）。
+- **ACP 适配器不再停在更宽的模式**：claude-agent-acp（`auto-edit` → `acceptEdits`、`full-auto` → `auto`，用户缺省为 `bypassPermissions` 时也会改回）、codex-acp（`plan` / `default` → `read-only`、`auto-edit` → `workspace-write`、`auto` / `full-auto` → `agent`；此前 `plan` 拒绝启动）、Copilot（URL 形式的模式 id `…#plan` / `…#agent`）写入显式映射，从不选放开全部权限的模式（#198）。
+- **`model` 交给 ACP Agent**：`agents.<id>.model` 与 `task` 的 `model` 参数经 Agent 的 `model` 会话配置项设置（按值、名称或 `provider/model` 的模型部分匹配）；没有匹配时提示并用它的缺省模型（#198）。
+- **用量口径修正**：Codex（app-server 与 `exec`）和 Copilot 报的 input 含缓存命中，Copilot 的 ACP usage 还是会话累计；ama 现在扣掉缓存部分、按回合取差值，总量不再把缓存读算两次。Claude stream-json 也像其它驱动一样上报上下文占用与窗口（#198）。
+- **已验证区间更新**：2026-10-10 用真实 CLI 实测，claude-agent-acp 0.89、codex-acp 2.2、Copilot 1.0.95、OpenCode 1.18、pi 1.1；能力矩阵见 docs/agents.md，数字见 docs/benchmarks/external-agents-2026-10.md（#198）。
+- **ChatGPT 账户不可用的模型报 `model_unavailable`**：codex 后端以 400「model is not supported when using Codex with a ChatGPT account」拒绝时，报 `model_unavailable` 并提示用 `ama models discover chatgpt` 刷新模型表（后端已撤下 `gpt-6-astra`，旧的发现缓存里还有它）（#198）。
+
 ## 0.7.6（2026-10-10）
 
 ### 模型调用效率

package/dist/ai/apis/chatgpt-backend.d.ts

@@ -10,7 +10,9 @@
  * - 出错恢复（每种至多一次）：401 → 强制刷新重试；SIWC 400 `subscription_sharing_unsupported_capability` → 删去
  *   `error.param` 重试；codex 400 `Instructions are not valid` → 本会话切 developer-message 重试；
  * - 错误映射（文案带码前缀，宿主按码判断）：429 配额 → `quota_exceeded`（不重试）；失效 → `auth_expired`；
- *   SIWC 403 `subscription_sharing_user_not_eligible` → `not_eligible`（i18n 文案列出原因）；503 留给会话层现有重试；
+ *   SIWC 403 `subscription_sharing_user_not_eligible` → `not_eligible`（i18n 文案列出原因）；codex 400「model is not
+ *   supported when using Codex with a ChatGPT account」→ `model_unavailable`（发现缓存过期时常见，提示刷新）；
+ *   503 留给会话层现有重试；
  * - 渠道跟随登录方式由会话层做（auth/chatgpt/follow.ts）；到这里仍不符说明用户显式写了 `@渠道`，报
  *   `chatgpt_flavor_mismatch`（i18n）；
  * - 用量：`cost = 0`、`billing: "subscription"`。

package/dist/ai/apis/chatgpt-backend.js

@@ -10,7 +10,9 @@
  * - 出错恢复（每种至多一次）：401 → 强制刷新重试；SIWC 400 `subscription_sharing_unsupported_capability` → 删去
  *   `error.param` 重试；codex 400 `Instructions are not valid` → 本会话切 developer-message 重试；
  * - 错误映射（文案带码前缀，宿主按码判断）：429 配额 → `quota_exceeded`（不重试）；失效 → `auth_expired`；
- *   SIWC 403 `subscription_sharing_user_not_eligible` → `not_eligible`（i18n 文案列出原因）；503 留给会话层现有重试；
+ *   SIWC 403 `subscription_sharing_user_not_eligible` → `not_eligible`（i18n 文案列出原因）；codex 400「model is not
+ *   supported when using Codex with a ChatGPT account」→ `model_unavailable`（发现缓存过期时常见，提示刷新）；
+ *   503 留给会话层现有重试；
  * - 渠道跟随登录方式由会话层做（auth/chatgpt/follow.ts）；到这里仍不符说明用户显式写了 `@渠道`，报
  *   `chatgpt_flavor_mismatch`（i18n）；
  * - 用量：`cost = 0`、`billing: "subscription"`。
@@ -220,6 +222,13 @@ export function recoverChatGptError(error, model, compat, options, apiKey, tried
             return { kind: "retry" };
         }
     }
+    if (backend === "codex" &&
+        error.status === 400 &&
+        /not supported when using codex with a chatgpt account/i.test(error.body))
+        return {
+            kind: "fail",
+            error: new Error(`model_unavailable: ${msg().auth.request.modelUnavailable(`${provider}/${model.id}`)}`),
+        };
     if (backend === "codex" &&
         error.status === 400 &&
         /instructions are not valid/i.test(error.body) &&