agent-intern 0.30.2__tar.gz → 0.30.3__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (21) hide show
  1. {agent_intern-0.30.2 → agent_intern-0.30.3}/PKG-INFO +28 -15
  2. {agent_intern-0.30.2 → agent_intern-0.30.3}/README.md +27 -14
  3. {agent_intern-0.30.2 → agent_intern-0.30.3}/agent_intern.egg-info/PKG-INFO +28 -15
  4. {agent_intern-0.30.2 → agent_intern-0.30.3}/codex_bridge.py +1 -1
  5. {agent_intern-0.30.2 → agent_intern-0.30.3}/copilot_bridge.py +1 -1
  6. {agent_intern-0.30.2 → agent_intern-0.30.3}/cursor_bridge.py +75 -7
  7. {agent_intern-0.30.2 → agent_intern-0.30.3}/grok_bridge.py +1 -1
  8. {agent_intern-0.30.2 → agent_intern-0.30.3}/opencode_bridge.py +14 -4
  9. {agent_intern-0.30.2 → agent_intern-0.30.3}/proc_tree.py +94 -1
  10. {agent_intern-0.30.2 → agent_intern-0.30.3}/pyproject.toml +1 -1
  11. {agent_intern-0.30.2 → agent_intern-0.30.3}/server.py +4 -4
  12. {agent_intern-0.30.2 → agent_intern-0.30.3}/swarm.py +2 -2
  13. {agent_intern-0.30.2 → agent_intern-0.30.3}/swarm_watch.py +2 -1
  14. {agent_intern-0.30.2 → agent_intern-0.30.3}/LICENSE +0 -0
  15. {agent_intern-0.30.2 → agent_intern-0.30.3}/agent_intern.egg-info/SOURCES.txt +0 -0
  16. {agent_intern-0.30.2 → agent_intern-0.30.3}/agent_intern.egg-info/dependency_links.txt +0 -0
  17. {agent_intern-0.30.2 → agent_intern-0.30.3}/agent_intern.egg-info/entry_points.txt +0 -0
  18. {agent_intern-0.30.2 → agent_intern-0.30.3}/agent_intern.egg-info/requires.txt +0 -0
  19. {agent_intern-0.30.2 → agent_intern-0.30.3}/agent_intern.egg-info/top_level.txt +0 -0
  20. {agent_intern-0.30.2 → agent_intern-0.30.3}/kimi_bridge.py +0 -0
  21. {agent_intern-0.30.2 → agent_intern-0.30.3}/setup.cfg +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: agent-intern
3
- Version: 0.30.2
3
+ Version: 0.30.3
4
4
  Summary: MCP server bridging the Antigravity (agy), OpenAI Codex, GitHub Copilot, Cursor, opencode, Grok Build, and Kimi Code CLIs so Claude Code can drive them as sub-agents
5
5
  Author: Sinan Tüfekçi
6
6
  License: MIT
@@ -64,12 +64,17 @@ file path.
64
64
 
65
65
  ## Quick start
66
66
 
67
- **1. Register the server** (needs [uv](https://docs.astral.sh/uv/)):
67
+ **1. Install it** (needs [uv](https://docs.astral.sh/uv/)). The plugin is the recommended way, because
68
+ it adds three slash commands on top of the server:
68
69
 
69
70
  ```bash
70
- claude mcp add -s user agent-intern -- uvx agent-intern
71
+ claude plugin marketplace add SinanTufekci/agent-intern
72
+ claude plugin install agent-intern@agent-intern
71
73
  ```
72
74
 
75
+ Or register just the MCP server: `claude mcp add -s user agent-intern -- uvx agent-intern`. Pick
76
+ one or the other, since doing both gives Claude every tool twice.
77
+
73
78
  **2. Install at least one backend CLI and sign in once.** Any one of them works on its own:
74
79
 
75
80
  | If you have… | Install | Sign in |
@@ -83,11 +88,20 @@ claude mcp add -s user agent-intern -- uvx agent-intern
83
88
  **3. Restart Claude Code and just ask.** The server ships its own routing guide as MCP instructions,
84
89
  so Claude knows which tool fits — you don't have to name them:
85
90
 
86
- > *"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."*
87
- > *"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."*
88
- > *"Summarise each of the six files in src/handlers in parallel with a swarm."*
91
+ - *"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."*
92
+ - *"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."*
93
+ - *"Summarise each of the six files in src/handlers in parallel with a swarm."*
94
+
95
+ With the plugin you also get three slash commands:
89
96
 
90
- Any `*_status` tool (say, *"run antigravity_status"*) checks a backend's setup without spending quota.
97
+ | Command | What it does |
98
+ |---|---|
99
+ | `/agent-intern:second-opinion [--council]` | Sends your diff to a reviewer from another model family, or to 2–3 of them in parallel. Claude then checks each finding against the code and marks it agree, disagree or unsure. |
100
+ | `/agent-intern:image <what to draw>` | Gemini draws it, the file is saved into your project, and Claude looks at the result. |
101
+ | `/agent-intern:doctor` | Shows which backends are installed and signed in, with the next step for any that aren't. Spends no quota. |
102
+
103
+ Without the plugin, any `*_status` tool (say, *"run antigravity_status"*) checks a backend without
104
+ spending quota.
91
105
 
92
106
  > [!TIP]
93
107
  > uvx pins the version it first caches, so nothing updates behind your back. Every `*_status` call
@@ -96,14 +110,12 @@ Any `*_status` tool (say, *"run antigravity_status"*) checks a backend's setup w
96
110
 
97
111
  ## What it's for
98
112
 
99
- | | |
100
- |---|---|
101
- | 🎨 **Images, inside Claude Code** | `antigravity_image` has Gemini draw it and returns the saved file — no extra API key, no extra tool. |
102
- | 🧠 **A second opinion** | A different model family reviews what Claude just wrote. Their blind spots rarely overlap. |
103
- | 🐝 **Parallel fan-out** | `agent_swarm` runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers). |
104
- | 💸 **Cheaper grunt work** | Bulk renames, boilerplate and first-pass ports burn *their* quota instead of Claude's tokens. |
105
- | 🆓 **No subscription? Still works** | opencode's free models answer with zero credentials — slow, but free. |
106
- | 🔌 **Zero new auth** | Piggybacks the CLI logins you already have. The bridge manages no keys of its own. |
113
+ - 🎨 **Images, inside Claude Code.** `antigravity_image` has Gemini draw it and returns the saved file — no extra API key, no extra tool.
114
+ - 🧠 **A second opinion.** A different model family reviews what Claude just wrote. Their blind spots rarely overlap.
115
+ - 🐝 **Parallel fan-out.** `agent_swarm` runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers).
116
+ - 💸 **Cheaper grunt work.** Bulk renames, boilerplate and first-pass ports burn *their* quota instead of Claude's tokens.
117
+ - 🆓 **No subscription? Still works.** opencode's free models answer with zero credentials — slow, but free.
118
+ - 🔌 **Zero new auth.** Piggybacks the CLI logins you already have. The bridge manages no keys of its own.
107
119
 
108
120
  ### Watch it work
109
121
 
@@ -203,6 +215,7 @@ the same thread. No private APIs, no token handling: it only bridges what the CL
203
215
 
204
216
  The CLIs behind this bridge update themselves, so most breakage is upstream drift rather than a bug
205
217
  in the bridge. A report that includes the relevant `*_status` output usually pins it down in one go.
218
+ [Contributing guide](https://github.com/SinanTufekci/agent-intern/blob/main/CONTRIBUTING.md) ·
206
219
  [Open an issue](https://github.com/SinanTufekci/agent-intern/issues/new/choose) ·
207
220
  [Start a discussion](https://github.com/SinanTufekci/agent-intern/discussions) · Developed on Windows —
208
221
  confirmations from macOS and Linux are very welcome.
@@ -31,12 +31,17 @@ file path.
31
31
 
32
32
  ## Quick start
33
33
 
34
- **1. Register the server** (needs [uv](https://docs.astral.sh/uv/)):
34
+ **1. Install it** (needs [uv](https://docs.astral.sh/uv/)). The plugin is the recommended way, because
35
+ it adds three slash commands on top of the server:
35
36
 
36
37
  ```bash
37
- claude mcp add -s user agent-intern -- uvx agent-intern
38
+ claude plugin marketplace add SinanTufekci/agent-intern
39
+ claude plugin install agent-intern@agent-intern
38
40
  ```
39
41
 
42
+ Or register just the MCP server: `claude mcp add -s user agent-intern -- uvx agent-intern`. Pick
43
+ one or the other, since doing both gives Claude every tool twice.
44
+
40
45
  **2. Install at least one backend CLI and sign in once.** Any one of them works on its own:
41
46
 
42
47
  | If you have… | Install | Sign in |
@@ -50,11 +55,20 @@ claude mcp add -s user agent-intern -- uvx agent-intern
50
55
  **3. Restart Claude Code and just ask.** The server ships its own routing guide as MCP instructions,
51
56
  so Claude knows which tool fits — you don't have to name them:
52
57
 
53
- > *"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."*
54
- > *"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."*
55
- > *"Summarise each of the six files in src/handlers in parallel with a swarm."*
58
+ - *"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."*
59
+ - *"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."*
60
+ - *"Summarise each of the six files in src/handlers in parallel with a swarm."*
61
+
62
+ With the plugin you also get three slash commands:
56
63
 
57
- Any `*_status` tool (say, *"run antigravity_status"*) checks a backend's setup without spending quota.
64
+ | Command | What it does |
65
+ |---|---|
66
+ | `/agent-intern:second-opinion [--council]` | Sends your diff to a reviewer from another model family, or to 2–3 of them in parallel. Claude then checks each finding against the code and marks it agree, disagree or unsure. |
67
+ | `/agent-intern:image <what to draw>` | Gemini draws it, the file is saved into your project, and Claude looks at the result. |
68
+ | `/agent-intern:doctor` | Shows which backends are installed and signed in, with the next step for any that aren't. Spends no quota. |
69
+
70
+ Without the plugin, any `*_status` tool (say, *"run antigravity_status"*) checks a backend without
71
+ spending quota.
58
72
 
59
73
  > [!TIP]
60
74
  > uvx pins the version it first caches, so nothing updates behind your back. Every `*_status` call
@@ -63,14 +77,12 @@ Any `*_status` tool (say, *"run antigravity_status"*) checks a backend's setup w
63
77
 
64
78
  ## What it's for
65
79
 
66
- | | |
67
- |---|---|
68
- | 🎨 **Images, inside Claude Code** | `antigravity_image` has Gemini draw it and returns the saved file — no extra API key, no extra tool. |
69
- | 🧠 **A second opinion** | A different model family reviews what Claude just wrote. Their blind spots rarely overlap. |
70
- | 🐝 **Parallel fan-out** | `agent_swarm` runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers). |
71
- | 💸 **Cheaper grunt work** | Bulk renames, boilerplate and first-pass ports burn *their* quota instead of Claude's tokens. |
72
- | 🆓 **No subscription? Still works** | opencode's free models answer with zero credentials — slow, but free. |
73
- | 🔌 **Zero new auth** | Piggybacks the CLI logins you already have. The bridge manages no keys of its own. |
80
+ - 🎨 **Images, inside Claude Code.** `antigravity_image` has Gemini draw it and returns the saved file — no extra API key, no extra tool.
81
+ - 🧠 **A second opinion.** A different model family reviews what Claude just wrote. Their blind spots rarely overlap.
82
+ - 🐝 **Parallel fan-out.** `agent_swarm` runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers).
83
+ - 💸 **Cheaper grunt work.** Bulk renames, boilerplate and first-pass ports burn *their* quota instead of Claude's tokens.
84
+ - 🆓 **No subscription? Still works.** opencode's free models answer with zero credentials — slow, but free.
85
+ - 🔌 **Zero new auth.** Piggybacks the CLI logins you already have. The bridge manages no keys of its own.
74
86
 
75
87
  ### Watch it work
76
88
 
@@ -170,6 +182,7 @@ the same thread. No private APIs, no token handling: it only bridges what the CL
170
182
 
171
183
  The CLIs behind this bridge update themselves, so most breakage is upstream drift rather than a bug
172
184
  in the bridge. A report that includes the relevant `*_status` output usually pins it down in one go.
185
+ [Contributing guide](https://github.com/SinanTufekci/agent-intern/blob/main/CONTRIBUTING.md) ·
173
186
  [Open an issue](https://github.com/SinanTufekci/agent-intern/issues/new/choose) ·
174
187
  [Start a discussion](https://github.com/SinanTufekci/agent-intern/discussions) · Developed on Windows —
175
188
  confirmations from macOS and Linux are very welcome.
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: agent-intern
3
- Version: 0.30.2
3
+ Version: 0.30.3
4
4
  Summary: MCP server bridging the Antigravity (agy), OpenAI Codex, GitHub Copilot, Cursor, opencode, Grok Build, and Kimi Code CLIs so Claude Code can drive them as sub-agents
5
5
  Author: Sinan Tüfekçi
6
6
  License: MIT
@@ -64,12 +64,17 @@ file path.
64
64
 
65
65
  ## Quick start
66
66
 
67
- **1. Register the server** (needs [uv](https://docs.astral.sh/uv/)):
67
+ **1. Install it** (needs [uv](https://docs.astral.sh/uv/)). The plugin is the recommended way, because
68
+ it adds three slash commands on top of the server:
68
69
 
69
70
  ```bash
70
- claude mcp add -s user agent-intern -- uvx agent-intern
71
+ claude plugin marketplace add SinanTufekci/agent-intern
72
+ claude plugin install agent-intern@agent-intern
71
73
  ```
72
74
 
75
+ Or register just the MCP server: `claude mcp add -s user agent-intern -- uvx agent-intern`. Pick
76
+ one or the other, since doing both gives Claude every tool twice.
77
+
73
78
  **2. Install at least one backend CLI and sign in once.** Any one of them works on its own:
74
79
 
75
80
  | If you have… | Install | Sign in |
@@ -83,11 +88,20 @@ claude mcp add -s user agent-intern -- uvx agent-intern
83
88
  **3. Restart Claude Code and just ask.** The server ships its own routing guide as MCP instructions,
84
89
  so Claude knows which tool fits — you don't have to name them:
85
90
 
86
- > *"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."*
87
- > *"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."*
88
- > *"Summarise each of the six files in src/handlers in parallel with a swarm."*
91
+ - *"Ask Gemini to draw a pixel-art rocket for the README header and save it under assets/."*
92
+ - *"Have Copilot review the diff you just wrote — read-only — and tell me where it disagrees with you."*
93
+ - *"Summarise each of the six files in src/handlers in parallel with a swarm."*
94
+
95
+ With the plugin you also get three slash commands:
89
96
 
90
- Any `*_status` tool (say, *"run antigravity_status"*) checks a backend's setup without spending quota.
97
+ | Command | What it does |
98
+ |---|---|
99
+ | `/agent-intern:second-opinion [--council]` | Sends your diff to a reviewer from another model family, or to 2–3 of them in parallel. Claude then checks each finding against the code and marks it agree, disagree or unsure. |
100
+ | `/agent-intern:image <what to draw>` | Gemini draws it, the file is saved into your project, and Claude looks at the result. |
101
+ | `/agent-intern:doctor` | Shows which backends are installed and signed in, with the next step for any that aren't. Spends no quota. |
102
+
103
+ Without the plugin, any `*_status` tool (say, *"run antigravity_status"*) checks a backend without
104
+ spending quota.
91
105
 
92
106
  > [!TIP]
93
107
  > uvx pins the version it first caches, so nothing updates behind your back. Every `*_status` call
@@ -96,14 +110,12 @@ Any `*_status` tool (say, *"run antigravity_status"*) checks a backend's setup w
96
110
 
97
111
  ## What it's for
98
112
 
99
- | | |
100
- |---|---|
101
- | 🎨 **Images, inside Claude Code** | `antigravity_image` has Gemini draw it and returns the saved file — no extra API key, no extra tool. |
102
- | 🧠 **A second opinion** | A different model family reviews what Claude just wrote. Their blind spots rarely overlap. |
103
- | 🐝 **Parallel fan-out** | `agent_swarm` runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers). |
104
- | 💸 **Cheaper grunt work** | Bulk renames, boilerplate and first-pass ports burn *their* quota instead of Claude's tokens. |
105
- | 🆓 **No subscription? Still works** | opencode's free models answer with zero credentials — slow, but free. |
106
- | 🔌 **Zero new auth** | Piggybacks the CLI logins you already have. The bridge manages no keys of its own. |
113
+ - 🎨 **Images, inside Claude Code.** `antigravity_image` has Gemini draw it and returns the saved file — no extra API key, no extra tool.
114
+ - 🧠 **A second opinion.** A different model family reviews what Claude just wrote. Their blind spots rarely overlap.
115
+ - 🐝 **Parallel fan-out.** `agent_swarm` runs N tasks at once and can mix backends in a single call (~2.8× at 3 Gemini workers).
116
+ - 💸 **Cheaper grunt work.** Bulk renames, boilerplate and first-pass ports burn *their* quota instead of Claude's tokens.
117
+ - 🆓 **No subscription? Still works.** opencode's free models answer with zero credentials — slow, but free.
118
+ - 🔌 **Zero new auth.** Piggybacks the CLI logins you already have. The bridge manages no keys of its own.
107
119
 
108
120
  ### Watch it work
109
121
 
@@ -203,6 +215,7 @@ the same thread. No private APIs, no token handling: it only bridges what the CL
203
215
 
204
216
  The CLIs behind this bridge update themselves, so most breakage is upstream drift rather than a bug
205
217
  in the bridge. A report that includes the relevant `*_status` output usually pins it down in one go.
218
+ [Contributing guide](https://github.com/SinanTufekci/agent-intern/blob/main/CONTRIBUTING.md) ·
206
219
  [Open an issue](https://github.com/SinanTufekci/agent-intern/issues/new/choose) ·
207
220
  [Start a discussion](https://github.com/SinanTufekci/agent-intern/discussions) · Developed on Windows —
208
221
  confirmations from macOS and Linux are very welcome.
@@ -497,7 +497,7 @@ def run_codex_streaming(
497
497
  before = set() if continue_conv else _rollout_names()
498
498
  args = build_args(prompt, workspace, sandbox, model, resume_session, out_path, json_stream=True)
499
499
  try:
500
- proc = subprocess.Popen(
500
+ proc = proc_tree.popen(
501
501
  args,
502
502
  cwd=workspace,
503
503
  stdin=subprocess.DEVNULL,
@@ -437,7 +437,7 @@ def run_copilot_streaming(
437
437
  args = build_args(prompt, workspace, sandbox, model, session_id, json_stream=True)
438
438
 
439
439
  state: dict = {"answer": "", "delta": ""}
440
- proc = subprocess.Popen(
440
+ proc = proc_tree.popen(
441
441
  args,
442
442
  cwd=workspace,
443
443
  stdin=subprocess.DEVNULL,
@@ -99,6 +99,55 @@ def _resolve_bin() -> str:
99
99
 
100
100
  CURSOR_BIN = _resolve_bin()
101
101
 
102
+ # A cursor-agent release directory under the install's versions/: YYYY.MM.DD, an
103
+ # optional -HH-MM-SS, then -<commit>. Copied from cursor-agent.ps1's own filter.
104
+ _VERSION_DIR_RE = re.compile(r"^(\d{4})\.(\d{1,2})\.(\d{1,2})(-\d{2}-\d{2}-\d{2})?-[a-f0-9]+$")
105
+
106
+
107
+ def _launch_prefix(bin_path: str, windows: Optional[bool] = None) -> list[str]:
108
+ """argv prefix that runs cursor-agent, bypassing cmd.exe when it is the Windows shim.
109
+
110
+ The installer's `cursor-agent.CMD` hands its arguments to cmd.exe (`%*`, with
111
+ delayed expansion on), which re-parses them — so a prompt containing `"` and `&`
112
+ runs a shell command on the host, whatever the sandbox says. See proc_tree's
113
+ batch-shim note. The shim only forwards to `cursor-agent.ps1`, which runs
114
+ `node.exe index.js` from the install dir if present, else from the newest
115
+ `versions/<YYYY.MM.DD>-<commit>/`. Doing that selection here and launching node
116
+ directly is the same program with cmd.exe and PowerShell taken out. Verified
117
+ live: `--version` and `status` answer identically both ways. Anything we can't
118
+ resolve falls back to the shim, where proc_tree.check_args refuses a dangerous
119
+ prompt instead of running it. `windows` defaults to the real platform; tests pass
120
+ it explicitly, since faking os.name breaks pathlib off Windows.
121
+ """
122
+ if windows is None:
123
+ windows = os.name == "nt"
124
+ if not windows or not proc_tree.is_batch_file(bin_path):
125
+ return [bin_path]
126
+ base = Path(bin_path).parent
127
+ flat = (base / "node.exe", base / "index.js")
128
+ if all(f.is_file() for f in flat):
129
+ return [str(f) for f in flat]
130
+ best: Optional[tuple[int, str]] = None
131
+ versions = base / "versions"
132
+ for d in versions.iterdir() if versions.is_dir() else ():
133
+ m = _VERSION_DIR_RE.match(d.name)
134
+ if not m or not d.is_dir():
135
+ continue
136
+ key = (int(f"{m.group(1)}{int(m.group(2)):02d}{int(m.group(3)):02d}"), d.name)
137
+ if best is None or key > best:
138
+ best = key
139
+ if best is not None:
140
+ node, entry = versions / best[1] / "node.exe", versions / best[1] / "index.js"
141
+ if node.is_file() and entry.is_file():
142
+ return [str(node), str(entry)]
143
+ return [bin_path]
144
+
145
+
146
+ def _launch() -> list[str]:
147
+ """The argv prefix for every cursor-agent call (see _launch_prefix)."""
148
+ return _launch_prefix(CURSOR_BIN)
149
+
150
+
102
151
  # cursor's state home. Chats live under ~/.cursor/chats/<workspace-hash>/<chat-id>/
103
152
  # (each chat dir has meta.json + store.db). No documented override, so this is
104
153
  # fixed to the home dir; we only READ these files (for the restart-proof continue
@@ -142,10 +191,29 @@ def _spawn_kwargs() -> dict:
142
191
  shim doesn't flash a console; POSIX starts a new session.
143
192
  """
144
193
  if os.name == "nt":
145
- return {"creationflags": getattr(subprocess, "CREATE_NO_WINDOW", 0x08000000)}
194
+ return {
195
+ "creationflags": getattr(subprocess, "CREATE_NO_WINDOW", 0x08000000),
196
+ "env": _env(),
197
+ }
146
198
  return {"start_new_session": True}
147
199
 
148
200
 
201
+ def _env() -> dict:
202
+ """Process env carrying what cursor-agent.ps1 would have set.
203
+
204
+ Needed now that _launch_prefix runs node directly instead of the ps1: the
205
+ launcher records how it was invoked and points node's compile cache at a fixed
206
+ dir (without it every call recompiles). Both are setdefault, so a value the user
207
+ exported still wins.
208
+ """
209
+ env = dict(os.environ)
210
+ env.setdefault("CURSOR_INVOKED_AS", os.path.basename(CURSOR_BIN))
211
+ local = env.get("LOCALAPPDATA")
212
+ if local:
213
+ env.setdefault("NODE_COMPILE_CACHE", os.path.join(local, "cursor-compile-cache"))
214
+ return env
215
+
216
+
149
217
  def normalize_workspace(ws: Optional[str]) -> str:
150
218
  """Absolute path for `ws`, or the server's cwd when omitted."""
151
219
  return os.path.abspath(ws) if ws else os.getcwd()
@@ -192,7 +260,7 @@ def create_chat(workspace: str) -> str:
192
260
  good measure and let the subsequent ask create the workspace's chat dir.
193
261
  """
194
262
  proc = proc_tree.run_captured(
195
- [CURSOR_BIN, "create-chat"],
263
+ [*_launch(), "create-chat"],
196
264
  cwd=workspace,
197
265
  stdin=subprocess.DEVNULL,
198
266
  timeout=30,
@@ -319,7 +387,7 @@ def list_models() -> list[str]:
319
387
  return _MODELS_CACHE
320
388
  try:
321
389
  proc = proc_tree.run_captured(
322
- [CURSOR_BIN, "models"],
390
+ [*_launch(), "models"],
323
391
  stdin=subprocess.DEVNULL,
324
392
  timeout=20,
325
393
  **_TEXT,
@@ -393,7 +461,7 @@ def build_args(
393
461
  last.
394
462
  """
395
463
  args = [
396
- CURSOR_BIN,
464
+ *_launch(),
397
465
  "-p",
398
466
  "--output-format",
399
467
  "stream-json" if json_stream else "text",
@@ -503,7 +571,7 @@ def run_cursor_streaming(
503
571
  args = build_args(prompt, workspace, sandbox, model, chat_id, json_stream=True)
504
572
 
505
573
  state: dict = {"answer": "", "assistant": ""}
506
- proc = subprocess.Popen(
574
+ proc = proc_tree.popen(
507
575
  args,
508
576
  cwd=workspace,
509
577
  stdin=subprocess.DEVNULL,
@@ -582,7 +650,7 @@ def cursor_version() -> Optional[str]:
582
650
  """`cursor-agent --version` first line, or None if cursor can't be run."""
583
651
  try:
584
652
  proc = proc_tree.run_captured(
585
- [CURSOR_BIN, "--version"],
653
+ [*_launch(), "--version"],
586
654
  stdin=subprocess.DEVNULL,
587
655
  timeout=15,
588
656
  **_TEXT,
@@ -604,7 +672,7 @@ def auth_status() -> tuple[bool, str]:
604
672
  """
605
673
  try:
606
674
  proc = proc_tree.run_captured(
607
- [CURSOR_BIN, "status"],
675
+ [*_launch(), "status"],
608
676
  stdin=subprocess.DEVNULL,
609
677
  timeout=20,
610
678
  **_TEXT,
@@ -520,7 +520,7 @@ def run_grok_streaming(
520
520
  args = build_args(prompt, workspace, sandbox, model, resume_id, use_continue, json_stream=True)
521
521
 
522
522
  state: dict = {}
523
- proc = subprocess.Popen(
523
+ proc = proc_tree.popen(
524
524
  args,
525
525
  cwd=workspace,
526
526
  stdin=subprocess.DEVNULL,
@@ -102,10 +102,20 @@ OPENCODE_BIN_ENV = os.environ.get("OPENCODE_BIN", "opencode")
102
102
 
103
103
 
104
104
  def _resolve_bin() -> str:
105
- """Full path to the opencode executable (see OPENCODE_BIN_ENV note)."""
105
+ """Full path to the opencode executable (see OPENCODE_BIN_ENV note).
106
+
107
+ On Windows the npm install is an `opencode.cmd` shim, and a shim hands its
108
+ arguments to cmd.exe — which re-parses them, so a prompt containing `"` and `&`
109
+ would run a shell command on the host (see proc_tree's batch-shim note). The npm
110
+ shim only forwards to `node_modules/opencode-ai/bin/opencode.exe`, a native
111
+ binary, so launch that directly: same program, no cmd.exe. If the shim can't be
112
+ resolved, proc_tree.check_args refuses a dangerous prompt instead of running it.
113
+ """
106
114
  if os.path.sep in OPENCODE_BIN_ENV or os.path.isfile(OPENCODE_BIN_ENV):
107
- return OPENCODE_BIN_ENV
108
- return shutil.which(OPENCODE_BIN_ENV) or OPENCODE_BIN_ENV
115
+ found = OPENCODE_BIN_ENV
116
+ else:
117
+ found = shutil.which(OPENCODE_BIN_ENV) or OPENCODE_BIN_ENV
118
+ return proc_tree.npm_shim_target(found) or found
109
119
 
110
120
 
111
121
  OPENCODE_BIN = _resolve_bin()
@@ -543,7 +553,7 @@ def _run_impl(
543
553
  args = build_args(prompt, workspace, sandbox, model, resume_id, use_continue, agent)
544
554
 
545
555
  state: dict = {}
546
- proc = subprocess.Popen(
556
+ proc = proc_tree.popen(
547
557
  args,
548
558
  cwd=workspace,
549
559
  stdin=subprocess.DEVNULL,
@@ -32,6 +32,7 @@ platform and its worse failure mode, completely unaddressed.
32
32
  from __future__ import annotations
33
33
 
34
34
  import os
35
+ import re
35
36
  import signal
36
37
  import subprocess
37
38
  from typing import Optional
@@ -93,6 +94,98 @@ def kill_tree(proc: subprocess.Popen) -> None:
93
94
  pass
94
95
 
95
96
 
97
+ # ----------------------------------------------------------------- Windows batch shims
98
+ #
99
+ # A `.cmd`/`.bat` file cannot be executed directly: `CreateProcess` runs it as
100
+ # `cmd.exe /c "<file> <args>"`, so every argument is re-parsed by cmd.exe — with
101
+ # `%var%` expansion, `!var!` expansion when the shim enables delayed expansion, and
102
+ # `&`/`|`/`<`/`>` as command separators. Python's `list2cmdline` quotes for the
103
+ # MSVCRT parser, not for cmd.exe, and cmd.exe does not honour its backslash-escaped
104
+ # quotes. So an argument such as `x" & calc & "` closes the quote early and the rest
105
+ # RUNS AS A COMMAND — verified live against a shim. The prompt is exactly such an
106
+ # argument, and it can carry text Claude read from an untrusted file or web page.
107
+ # That escapes every `sandbox` setting: the command runs in the shim's cmd.exe, on
108
+ # the host, before the agent (and its sandbox) ever starts.
109
+ #
110
+ # Correct escaping for cmd.exe is possible only for some inputs (Rust's fix for the
111
+ # same bug, CVE-2024-24576, refuses the rest). Two defences instead:
112
+ # 1. each bridge resolves its shim to the real executable where it can
113
+ # (`npm_shim_target` below; cursor and muse have their own launch paths), so
114
+ # cmd.exe is never involved;
115
+ # 2. `check_args` refuses to hand any metacharacter to a shim that is still left,
116
+ # turning a would-be injection into a clear error. Every spawn goes through
117
+ # `run_captured` / `popen` below, which call it — a test guards that.
118
+ BATCH_SUFFIXES = (".cmd", ".bat")
119
+ CMD_METACHARS = frozenset('"%!^&|<>\r\n')
120
+
121
+
122
+ def is_batch_file(executable: str) -> bool:
123
+ """True when `executable` is a .cmd/.bat file, which Windows runs via cmd.exe."""
124
+ return str(executable).lower().endswith(BATCH_SUFFIXES)
125
+
126
+
127
+ def check_args(args: list[str], windows: Optional[bool] = None) -> None:
128
+ """Refuse to pass cmd.exe metacharacters to a batch-file shim. Never mutates.
129
+
130
+ A no-op off Windows and for real executables. Raises ValueError naming the
131
+ offending characters, so the caller fails loudly instead of running whatever the
132
+ argument smuggled in. See the note above for why escaping is not attempted.
133
+ `windows` defaults to the real platform; tests pass it rather than faking os.name.
134
+ """
135
+ if windows is None:
136
+ windows = os.name == "nt"
137
+ if not windows or not args or not is_batch_file(args[0]):
138
+ return
139
+ for arg in args[1:]:
140
+ bad = sorted(set(str(arg)) & CMD_METACHARS)
141
+ if bad:
142
+ shown = ", ".join(repr(c) for c in bad)
143
+ raise ValueError(
144
+ f"refusing to run {args[0]!r}: it is a batch-file shim, which Windows runs "
145
+ f"through cmd.exe, and an argument contains characters cmd.exe would "
146
+ f"interpret ({shown}) — that is how a prompt injects a shell command. "
147
+ f"Point the backend's *_BIN env var at the real executable instead."
148
+ )
149
+
150
+
151
+ def npm_shim_target(shim: str) -> Optional[str]:
152
+ """The native .exe an npm cmd-shim launches, or None if it isn't one.
153
+
154
+ npm's generated shims end in a line like
155
+ "%dp0%\\node_modules\\opencode-ai\\bin\\opencode.exe" %*
156
+ where `%dp0%` is the shim's own directory. When the target is a real `.exe`,
157
+ launching it directly behaves identically minus cmd.exe — so the prompt never
158
+ meets cmd's parser. Shims that run `node script.js` are left alone (None): there
159
+ the correct target depends on npm's node-resolution logic, and guessing wrong is
160
+ worse than the loud refusal `check_args` gives.
161
+ """
162
+ if not is_batch_file(shim):
163
+ return None
164
+ try:
165
+ with open(shim, encoding="utf-8", errors="replace") as f:
166
+ text = f.read()
167
+ except OSError:
168
+ return None
169
+ base = os.path.dirname(os.path.abspath(shim))
170
+ for quoted in reversed(re.findall(r'"([^"]+\.exe)"', text, flags=re.IGNORECASE)):
171
+ low = quoted.lower()
172
+ for token in ("%dp0%", "%~dp0"):
173
+ if low.startswith(token):
174
+ # The shim spells its path with backslashes; "/" joins correctly on
175
+ # every OS, so the resolution is testable off Windows too.
176
+ rel = quoted[len(token) :].replace("\\", "/").lstrip("/")
177
+ candidate = os.path.normpath(os.path.join(base, rel))
178
+ if os.path.isfile(candidate):
179
+ return candidate
180
+ return None
181
+
182
+
183
+ def popen(args: list[str], **popen_kwargs) -> subprocess.Popen:
184
+ """`subprocess.Popen` behind the batch-shim check. Use this, never Popen directly."""
185
+ check_args(args)
186
+ return subprocess.Popen(args, **popen_kwargs)
187
+
188
+
96
189
  def run_captured(
97
190
  args: list[str],
98
191
  *,
@@ -112,7 +205,7 @@ def run_captured(
112
205
  pipes open. Reading the pipes only *after* the tree is dead is what makes the
113
206
  second read terminate; that ordering is the fix, not a detail of it.
114
207
  """
115
- proc = subprocess.Popen(
208
+ proc = popen(
116
209
  args,
117
210
  stdout=subprocess.PIPE,
118
211
  stderr=subprocess.PIPE,
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "agent-intern"
3
- version = "0.30.2"
3
+ version = "0.30.3"
4
4
  description = "MCP server bridging the Antigravity (agy), OpenAI Codex, GitHub Copilot, Cursor, opencode, Grok Build, and Kimi Code CLIs so Claude Code can drive them as sub-agents"
5
5
  readme = "README.md"
6
6
  requires-python = ">=3.10"
@@ -564,7 +564,7 @@ mcp = FastMCP("agent-intern", instructions=SERVER_INSTRUCTIONS)
564
564
  # installed package metadata, which goes stale on editable installs). Keep in
565
565
  # sync with pyproject.toml's version. Compared at startup against the latest
566
566
  # tag on GitHub so a long-lived clone learns when to `git pull`.
567
- __version__ = "0.30.2"
567
+ __version__ = "0.30.3"
568
568
 
569
569
  # Logs go to stderr (stdout is the MCP protocol channel). Quiet by default;
570
570
  # set AGY_BRIDGE_DEBUG=1 for per-call diagnostics. See _configure_logging.
@@ -3035,7 +3035,7 @@ def _open_watch_window(url: str, rid: str = _MAIN) -> None:
3035
3035
  return
3036
3036
  for exe in _chromium_app_browsers():
3037
3037
  try:
3038
- subprocess.Popen(
3038
+ proc_tree.popen(
3039
3039
  [exe, f"--app={url}", f"--window-size={_WATCH_WINDOW_SIZE}"],
3040
3040
  stdin=subprocess.DEVNULL,
3041
3041
  stdout=subprocess.DEVNULL,
@@ -3106,7 +3106,7 @@ def _run_agy_watched(
3106
3106
  except Exception: # noqa: BLE001 - the viewer is best-effort, never fatal
3107
3107
  pass
3108
3108
 
3109
- proc = subprocess.Popen(
3109
+ proc = proc_tree.popen(
3110
3110
  args,
3111
3111
  cwd=workspace,
3112
3112
  stdin=subprocess.DEVNULL,
@@ -3247,7 +3247,7 @@ def _run_agy_image_watched(
3247
3247
  except Exception: # noqa: BLE001 - viewer is best-effort
3248
3248
  pass
3249
3249
 
3250
- proc = subprocess.Popen(
3250
+ proc = proc_tree.popen(
3251
3251
  args,
3252
3252
  cwd=workspace,
3253
3253
  stdin=subprocess.DEVNULL,
@@ -502,7 +502,7 @@ def _run_text_worker_watched(
502
502
  if model:
503
503
  args += ["--model", model]
504
504
  args += ["-p", prompt]
505
- proc = subprocess.Popen(
505
+ proc = proc_tree.popen(
506
506
  args,
507
507
  cwd=workspace,
508
508
  stdin=subprocess.DEVNULL,
@@ -738,7 +738,7 @@ def _run_image_worker_watched(index, prompt, target, workspace, timeout_s) -> Wo
738
738
  os.makedirs(workspace, exist_ok=True)
739
739
  wrapped = server._wrap_image_prompt(prompt, target)
740
740
  args = server._agy_base_args(timeout_s) + ["-p", wrapped]
741
- proc = subprocess.Popen(
741
+ proc = proc_tree.popen(
742
742
  args,
743
743
  cwd=workspace,
744
744
  stdin=subprocess.DEVNULL,
@@ -24,6 +24,7 @@ import webbrowser
24
24
  from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
25
25
  from typing import Optional
26
26
 
27
+ import proc_tree
27
28
  import server
28
29
  from server import (
29
30
  _chromium_app_browsers,
@@ -209,7 +210,7 @@ def _launch(url: str, w: int, h: int, x: Optional[int] = None, y: Optional[int]
209
210
  ]
210
211
  for exe in _chromium_app_browsers():
211
212
  try:
212
- subprocess.Popen(
213
+ proc_tree.popen(
213
214
  [exe, *flags],
214
215
  stdin=subprocess.DEVNULL,
215
216
  stdout=subprocess.DEVNULL,
File without changes
File without changes