feat(cli-agent): support 7 more CLI agents headless

AGENTS.md

@@ -2,7 +2,7 @@
 
 ## Project Overview
 
-`aic` is an AI-powered git commit message generator (Rust CLI, edition 2024). It produces hunk-level atomic Conventional Commits: nothing staged → the LLM plans unstaged hunks into logical Batches → one commit per Batch; something staged → one commit. Two LLM backends: 12 API providers via `rig` (openai, anthropic, gemini, deepseek, groq, xai, mistral, openrouter, perplexity, together, ollama, openai-compatible) or external CLI agents (claude/codex/pi/opencode) run headless. Subcommands: `resolve` (merge-conflict resolution), `setup`, `use`, `list`, `update` (zipsign-verified self-update), `completion`.
+`aic` is an AI-powered git commit message generator (Rust CLI, edition 2024). It produces hunk-level atomic Conventional Commits: nothing staged → the LLM plans unstaged hunks into logical Batches → one commit per Batch; something staged → one commit. Two LLM backends: 12 API providers via `rig` (openai, anthropic, gemini, deepseek, groq, xai, mistral, openrouter, perplexity, together, ollama, openai-compatible) or external CLI agents (claude, codex, pi, opencode, omp, gemini, cursor, windsurf, copilot, trae, qwen) run headless. Subcommands: `resolve` (merge-conflict resolution), `setup`, `use`, `list`, `update` (zipsign-verified self-update), `completion`.
 
 ## Architecture & Data Flow
 

README.md

@@ -53,7 +53,7 @@ aic
 # → 3 atomic commits from your working tree
 ```
 
-> 💡 **No API key?** Skip the provider setup — use [Claude Code, Codex, pi, or opencode](#-no-api-key-use-your-ai-agent) instead. aic reuses their auth.
+> 💡 **No API key?** Skip the provider setup — use [Claude Code, Codex, pi, opencode, and 7 more CLI agents](#-no-api-key-use-your-ai-agent) instead. aic reuses their auth.
 
 > **Windows (PowerShell):** `irm https://github.com/CaicoLeung/aic/releases/latest/download/aic-installer.ps1 | iex`
 
@@ -63,6 +63,8 @@ aic
 
 Already have [Claude Code](https://docs.anthropic.com/claude/docs/claude-code), [OpenAI Codex](https://github.com/openai/codex), [pi](https://pi.dev), or [opencode](https://opencode.ai) installed and authenticated? aic can drive it in **headless mode** — no API key needed.
 
+Also supported: **oh-my-pi** (`omp`), **Gemini** (`gemini`), **Cursor** (`cursor-agent`), **Windsurf** (`devin` — Windsurf was renamed Devin Desktop), **GitHub Copilot** (`copilot`), **Trae** (`traecli`), and **Qwen Code** (`qwen`).
+
 ```sh
 aic setup    # → select "CLI agent" → pick your tool → done
 ```
@@ -75,20 +77,20 @@ command = "claude"
 args = ["-p", "{prompt}", "--output-format", "stream-json", "--include-partial-messages"]
 ```
 
-aic sends one prompt and reads the answer — it never runs the agent in tool-use mode. Each preset pins itself to read-only or text-only, so an injected instruction can't touch your working tree. Both backends' fields can coexist in the config; `backend_kind` selects the active one.
+aic sends one prompt and reads the answer — it never runs the agent in tool-use mode. Presets pin explicit flags where needed (codex's read-only sandbox, pi's `--no-tools`); the rest rely on headless print mode, where approval-gated tools can't run without a TTY — so an injected instruction can't touch your working tree. Both backends' fields can coexist in the config; `backend_kind` selects the active one.
 
-See [CLI-agent presets](#cli-agent-presets) for Codex, pi, and opencode.
+See [CLI-agent presets](#cli-agent-presets) for the full preset list.
 
 ---
 
 ## Features
 
 - **Hunk-level splitting** — one file, many concerns? Splits per-hunk into atomic commits, fully non-interactive
-- **Two backends** — API provider (12+ supported) or CLI agent (Claude Code, Codex, pi, opencode — no API key)
+- **Two backends** — API provider (12+ supported) or CLI agent (11 presets: claude, codex, pi, opencode, omp, gemini, cursor, windsurf, copilot, trae, qwen — no API key)
 - **Merge conflict resolution** — `aic resolve` proposes per-file resolutions you review, then finalizes the merge
 - **Live reasoning** — watch the model think as it decides the split
 - **Conventional Commits** — messages follow the [v1.0.0 spec](https://www.conventionalcommits.org/)
-- **Interactive setup** — `aic setup` is menu-driven; `aic use` switches between saved provider profiles and CLI agents (claude, codex, pi, opencode)
+- **Interactive setup** — `aic setup` is menu-driven; `aic use` switches between saved provider profiles and CLI agents (claude, codex, pi, opencode, omp, gemini, cursor, windsurf, copilot, trae, qwen)
 
 ## Installation
 
@@ -107,7 +109,7 @@ Shell completions: `aic completion` (bash, fish, zsh, nushell).
 | `aic` | Commit staged files. If nothing is staged, auto-split all unstaged changes into hunk-level atomic commits. |
 | `aic resolve` | Resolve git merge conflicts via the LLM. Review each file, then finalize. |
 | `aic setup` | Menu-driven config: API provider, CLI agent, or pre-commit confirmation. |
-| `aic use <name>` | Switch to a provider already configured via `aic setup`, or to a CLI agent (claude, codex, pi, opencode). |
+| `aic use <name>` | Switch to a provider already configured via `aic setup`, or to a CLI agent (claude, codex, pi, opencode, omp, gemini, cursor, windsurf, copilot, trae, qwen). |
 | `aic list` | Show resolved config and where each value comes from. |
 | `aic update` | Update to the latest release. |
 | `aic completion` | Install shell completions. |
@@ -169,7 +171,7 @@ OpenRouter and the OpenAI-compatible provider have no default model — set `mod
 
 ### CLI-agent presets
 
-Each preset ships a dedicated decoder for its CLI's stdout envelope, so aic can stream reasoning where the CLI exposes it and cleanly extract the answer.
+Presets with a decodable stdout envelope (claude, codex, pi, opencode, omp) get a decoder — omp reuses pi's, being a pi fork — so aic streams reasoning where the CLI exposes it and cleanly extracts the answer; the rest use plain print mode — stdout IS the answer.
 
 ```toml
 # OpenAI Codex — exec --json, read-only sandbox
@@ -192,6 +194,18 @@ command = "opencode"
 args = ["run", "--format", "json", "{prompt}"]
 ```
 
+Additional presets — all single-shot print mode (`-p`), answer on stdout:
+
+| preset | command | notes |
+|--------|---------|-------|
+| `omp` | `omp --mode json {prompt}` | pi fork; pi-shaped NDJSON, reasoning feed |
+| `gemini` | `gemini -p {prompt}` | shadows the `gemini` provider name in `aic use` — the Google API stays reachable via `aic use google` |
+| `cursor` | `cursor-agent -p {prompt}` | no `--trust` → runs untrusted (writes disabled) |
+| `windsurf` | `devin -p {prompt}` | Windsurf renamed to Devin Desktop; auth via `devin auth login` |
+| `copilot` | `copilot -p {prompt}` | tool use needs interactive approval headless mode can't give |
+| `trae` | `traecli -p {prompt}` | non-read tools gated behind a permission prompt |
+| `qwen` | `qwen -p {prompt}` | Qwen Code, gemini-cli lineage |
+
 The CLI must already be installed and logged in — aic does not install or authenticate it.
 
 See [ADR 0010](docs/adr/0010-cli-agent-backend.md) for the backend design and [ADR 0011](docs/adr/0011-explicit-backend-discriminator.md) for the `backend_kind` discriminator.

README.zh-CN.md

@@ -53,7 +53,7 @@ aic
 # → 从你的工作区生成 3 个原子提交
 ```
 
-> 💡 **没有 API key？** 跳过 provider 配置 —— 直接使用 [Claude Code、Codex、pi 或 opencode](#-没有-api-key复用你的-ai-agent)，aic 会复用它们的登录认证。
+> 💡 **没有 API key？** 跳过 provider 配置 —— 直接使用 [Claude Code、Codex、pi、opencode 以及另外 7 个 CLI agent](#-没有-api-key复用你的-ai-agent)，aic 会复用它们的登录认证。
 
 > **Windows (PowerShell):** `irm https://github.com/CaicoLeung/aic/releases/latest/download/aic-installer.ps1 | iex`
 
@@ -63,6 +63,8 @@ aic
 
 已经安装并登录了 [Claude Code](https://docs.anthropic.com/claude/docs/claude-code)、[OpenAI Codex](https://github.com/openai/codex)、[pi](https://pi.dev) 或 [opencode](https://opencode.ai)？aic 可以在 **headless 模式** 下驱动它 —— 无需 API key。
 
+同样支持：**oh-my-pi**（`omp`）、**Gemini**（`gemini`）、**Cursor**（`cursor-agent`）、**Windsurf**（`devin` —— Windsurf 已更名为 Devin Desktop）、**GitHub Copilot**（`copilot`）、**Trae**（`traecli`）和 **Qwen Code**（`qwen`）。
+
 ```sh
 aic setup    # → 选择 "CLI agent" → 选你的工具 → 完成
 ```
@@ -75,20 +77,20 @@ command = "claude"
 args = ["-p", "{prompt}", "--output-format", "stream-json", "--include-partial-messages"]
 ```
 
-aic 只发送一条 prompt 并读取回答 —— 绝不在 tool-use 模式下运行 agent。每个预设都锁定为只读或纯文本，因此注入的指令无法触碰你的工作区。两个 backend 的字段可以共存于配置文件中；`backend_kind` 决定哪个生效。
+aic 只发送一条 prompt 并读取回答 —— 绝不在 tool-use 模式下运行 agent。需要时预设会显式锁定权限（codex 的只读沙箱、pi 的 `--no-tools`）；其余预设依赖无头 print 模式 —— 没有 TTY，需审批的工具无法运行 —— 因此注入的指令无法触碰你的工作区。两个 backend 的字段可以共存于配置文件中；`backend_kind` 决定哪个生效。
 
-其他预设（Codex、pi、opencode）见 [CLI-agent 预设](#cli-agent-预设)。
+其他预设见 [CLI-agent 预设](#cli-agent-预设)。
 
 ---
 
 ## 功能特性
 
 - **Hunk 级别拆分** —— 一个文件、多种 concern？按 hunk 拆成多个原子提交，完全非交互
-- **两种 backend** —— API provider（支持 12+ 家）或 CLI agent（Claude Code、Codex、pi、opencode —— 无需 API key）
+- **两种 backend** —— API provider（支持 12+ 家）或 CLI agent（11 个预设：claude、codex、pi、opencode、omp、gemini、cursor、windsurf、copilot、trae、qwen —— 无需 API key）
 - **Merge 冲突解决** —— `aic resolve` 逐文件给出方案供你审核，然后完成 merge
 - **实时推理** —— 观看模型思考拆分方案的全过程
 - **Conventional Commits** —— message 遵循 [v1.0.0 规范](https://www.conventionalcommits.org/)
-- **交互式配置** —— `aic setup` 菜单驱动；`aic use` 在已保存的 provider 与 CLI agent（claude、codex、pi、opencode）之间切换
+- **交互式配置** —— `aic setup` 菜单驱动；`aic use` 在已保存的 provider 与 CLI agent（claude、codex、pi、opencode、omp、gemini、cursor、windsurf、copilot、trae、qwen）之间切换
 
 ## 安装
 
@@ -107,7 +109,7 @@ Shell 补全：`aic completion`（bash、fish、zsh、nushell）。
 | `aic` | 提交已 stage 的文件。若无 stage 内容，自动将所有未暂存改动拆分为 hunk 级别的原子提交。 |
 | `aic resolve` | 通过 LLM 解决 git merge 冲突。逐文件审核后完成 merge。 |
 | `aic setup` | 菜单驱动配置：API provider、CLI agent、或提交前确认。 |
-| `aic use <name>` | 切换到已通过 `aic setup` 配置过的 provider，或切换到 CLI agent（claude、codex、pi、opencode）。 |
+| `aic use <name>` | 切换到已通过 `aic setup` 配置过的 provider，或切换到 CLI agent（claude、codex、pi、opencode、omp、gemini、cursor、windsurf、copilot、trae、qwen）。 |
 | `aic list` | 展示已 resolve 的 config 及每个值的来源。 |
 | `aic update` | 更新到最新版本。 |
 | `aic completion` | 安装 shell 补全。 |
@@ -169,7 +171,7 @@ OpenRouter 和 OpenAI-compatible provider 没有默认 model —— 在 config 
 
 ### CLI-agent 预设
 
-每个预配有专用的解码器来处理对应 CLI 的 stdout 格式，因此 aic 能在 CLI 支持的情况下流式输出推理过程，并干净地提取回答。
+带可解码 stdout 封装的预设（claude、codex、pi、opencode、omp）配有解码器 —— omp 作为 pi 的复刻直接复用 pi 的解码器 —— aic 能在 CLI 支持的情况下流式输出推理过程并干净地提取回答；其余预设为纯打印模式 —— stdout 即回答。
 
 ```toml
 # OpenAI Codex — exec --json，只读沙箱
@@ -192,6 +194,18 @@ command = "opencode"
 args = ["run", "--format", "json", "{prompt}"]
 ```
 
+其余预设 —— 均为单次打印模式（`-p`），回答输出到 stdout：
+
+| 预设 | 命令 | 说明 |
+|------|------|------|
+| `omp` | `omp --mode json {prompt}` | pi 分支；pi 同构 NDJSON，带推理流 |
+| `gemini` | `gemini -p {prompt}` | 在 `aic use` 中遮蔽 `gemini` provider 名 —— Google API 仍可通过 `aic use google` 使用 |
+| `cursor` | `cursor-agent -p {prompt}` | 不带 `--trust` → 以未信任模式运行（禁用写入） |
+| `windsurf` | `devin -p {prompt}` | Windsurf 已更名 Devin Desktop；`devin auth login` 登录 |
+| `copilot` | `copilot -p {prompt}` | 工具调用需要 headless 无法给出的交互式审批 |
+| `trae` | `traecli -p {prompt}` | 非只读工具被权限提示门控 |
+| `qwen` | `qwen -p {prompt}` | Qwen Code，gemini-cli 血统 |
+
 该 CLI 必须已安装并登录 —— aic 不负责安装或认证。
 
 详见 [ADR 0010](docs/adr/0010-cli-agent-backend.md)（backend 设计）和 [ADR 0011](docs/adr/0011-explicit-backend-discriminator.md)（`backend_kind` 判别字段）。

docs/adr/0010-cli-agent-backend.md

@@ -47,12 +47,13 @@ placeholder), `timeout_secs`. All optional, so existing configs are unchanged 
 ### Generic command template, not one adapter per CLI
 
 The backend is a single `CliSpec { command, args, timeout_secs }` carrying an
-argv template with a literal `{prompt}` placeholder. Three presets
-(`cli_preset("claude" | "codex" | "pi")`) are offered as snippets by `aic setup`
-and the docs; any other CLI (Opencode, Copilot, OhMyPi, future ones) is covered
-by a custom `command`/`args` with **zero new code**. This deliberately rejects
-the "one hardcoded adapter per CLI" alternative — six adapters doing the same
-text-in/text-out job is a maintenance treadmill.
+argv template with a literal `{prompt}` placeholder. Presets are data, not
+adapters: the preset names live in one registry ([`PRESETS`], resolved by
+`cli_preset`), and adding one is a single match arm plus an `Encoding` choice —
+omp needed zero new decoder code because it reuses pi's. A CLI without a
+preset is still covered by a custom `command`/`args` with **zero new code**.
+This deliberately rejects the "one hardcoded adapter per CLI" alternative —
+many adapters doing the same text-in/text-out job is a maintenance treadmill.
 
 ### Dispatch: an enum, not `Box<dyn>`
 
@@ -76,23 +77,36 @@ single prompt and reads stdout. No tool loop is ever allowed: an agent that can
 run tools could act on prompt-injected instructions against the working tree.
 Print mode removes that entire class of risk.
 
-**Least-permission presets.** The promise above is enforced by the invocation
-itself, not by trusting each CLI's default — every preset pins itself to a
-text-only / read-only stance, because defaults differ and one (pi) is unsafe:
+**Least-permission presets.** The promise above is enforced by the invocation,
+not by tool-use mode: every preset runs single-shot print mode, where no TTY
+exists to answer an approval prompt. On top of that shared guarantee, presets
+pin explicit flags where the CLI's defaults need it — defaults differ and one
+(pi) is unsafe even in print mode:
 
 | Preset | Pinned flags | Why |
 | --- | --- | --- |
 | `claude` | `-p` + `--output-format stream-json --include-partial-messages` | `-p` is print mode (cannot prompt). The `stream-json` flags surface claude's `thinking_delta` as a live reasoning stream — without them, plain `-p` returns only the final answer and the batch-plan reasoning window stays empty. claude has no reliable `--no-tools` flag (`--allowedTools` is variadic and greedily consumes the prompt), so print mode's conservative default (it cannot prompt → no privileged tool auto-executes; `--dangerously-skip-permissions` stays opt-in) is the lever. The NDJSON envelope is decoded centrally, so the typed paths still receive plain JSON text. |
 | `codex` | `exec -s read-only` | `exec` runs non-interactively; the sandbox is pinned to `read-only` so model-generated shell commands cannot write, even if a global config widens the default. `--dangerously-bypass-approvals-and-sandbox` is opt-in. |
 | `pi` | `--no-tools -p` | **Required.** pi enables `read/bash/edit/write` tools by default; in print mode on a *trusted* project it can auto-run them (it cannot prompt) — effectively yolo. `--no-tools` disables all tools so print mode is genuinely text-only. |
 
-Custom `command`/`args` backends are the user's responsibility to harden. The
-`aic setup` wizard offers **only the four presets** (claude / codex / pi /
-opencode) — each ships a dedicated decoder for its CLI's stdout envelope, so
-aic can stream reasoning where the CLI exposes it and cleanly extract the
-answer. A hand-edited custom `command`/`args` still runs, but in plain-text mode
-with no reasoning feed and no envelope decoding; it is the config-edit escape
-hatch for a CLI without a preset, not a wizard option.
+Custom `command`/`args` backends are the user's responsibility to harden.
+
+The remaining presets pin **no** flags — deliberately. They rely on headless
+print mode itself: with no TTY, approval-gated tools cannot run (gemini,
+copilot, trae, qwen); cursor runs untrusted (writes disabled) because
+`--trust` is omitted; opencode and omp expose no equivalent of pi's
+`--no-tools`, so like claude their text-only stance rests on print mode's
+conservative default. This weakening of "pin flags everywhere" is accepted
+here on purpose: per-CLI permission flags drift between CLI versions, while
+the headless default is shared by all eleven and cannot regress per CLI.
+
+The `aic setup` wizard offers **only the presets** ([`PRESETS`]). The five
+with a decodable stdout envelope (claude, codex, pi, opencode — and omp, which
+reuses pi's decoder) get reasoning streamed or extracted where the CLI exposes
+it; the rest are plain print mode (stdout IS the answer). A hand-edited custom
+`command`/`args` still runs, but in plain-text mode with no reasoning feed and
+no envelope decoding; it is the config-edit escape hatch for a CLI without a
+preset, not a wizard option.
 
 ### Preset auto-migration