gamai-plugins 0.2.2__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.
- gamai_plugins-0.2.2/LICENSE +21 -0
- gamai_plugins-0.2.2/PKG-INFO +220 -0
- gamai_plugins-0.2.2/README.md +202 -0
- gamai_plugins-0.2.2/gamai_plugins/__init__.py +2 -0
- gamai_plugins-0.2.2/gamai_plugins/cli.py +134 -0
- gamai_plugins-0.2.2/gamai_plugins/mcp.py +89 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/protean/plugin.json +64 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/protean/run.py +788 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/protean/seccomp.json +134 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/team/plugin.json +105 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/team/run.py +59 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/team/team_core.py +642 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/trust/plugin.json +76 -0
- gamai_plugins-0.2.2/gamai_plugins/plugins/trust/run.py +680 -0
- gamai_plugins-0.2.2/gamai_plugins.egg-info/PKG-INFO +220 -0
- gamai_plugins-0.2.2/gamai_plugins.egg-info/SOURCES.txt +25 -0
- gamai_plugins-0.2.2/gamai_plugins.egg-info/dependency_links.txt +1 -0
- gamai_plugins-0.2.2/gamai_plugins.egg-info/entry_points.txt +6 -0
- gamai_plugins-0.2.2/gamai_plugins.egg-info/requires.txt +3 -0
- gamai_plugins-0.2.2/gamai_plugins.egg-info/top_level.txt +1 -0
- gamai_plugins-0.2.2/pyproject.toml +40 -0
- gamai_plugins-0.2.2/setup.cfg +4 -0
- gamai_plugins-0.2.2/tests/test_cli.py +40 -0
- gamai_plugins-0.2.2/tests/test_mcp.py +23 -0
- gamai_plugins-0.2.2/tests/test_protean.py +74 -0
- gamai_plugins-0.2.2/tests/test_team.py +222 -0
- gamai_plugins-0.2.2/tests/test_trust.py +262 -0
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 gamAI
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,220 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: gamai-plugins
|
|
3
|
+
Version: 0.2.2
|
|
4
|
+
Summary: Portable trust, sandbox and team plugins for AI agents (MCP + CLI).
|
|
5
|
+
Author: gamAI
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Keywords: ai-agents,mcp,sandbox,trust,team
|
|
8
|
+
Classifier: Programming Language :: Python :: 3
|
|
9
|
+
Classifier: Operating System :: POSIX :: Linux
|
|
10
|
+
Classifier: Operating System :: Microsoft :: Windows
|
|
11
|
+
Classifier: Topic :: Security
|
|
12
|
+
Requires-Python: >=3.10
|
|
13
|
+
Description-Content-Type: text/markdown
|
|
14
|
+
License-File: LICENSE
|
|
15
|
+
Provides-Extra: crypto
|
|
16
|
+
Requires-Dist: cryptography>=42; extra == "crypto"
|
|
17
|
+
Dynamic: license-file
|
|
18
|
+
|
|
19
|
+
# gamAI plugins
|
|
20
|
+
|
|
21
|
+
Portable, dependency-light plugins for AI agents: **policy evaluation**,
|
|
22
|
+
**sandbox execution** and **team coordination**. The same code is exposed as an
|
|
23
|
+
MCP server, a CLI and a plain Python function, so it can run in Claude Code,
|
|
24
|
+
opencode, Codex, Cursor, LangChain or a custom agent loop.
|
|
25
|
+
|
|
26
|
+
> Status: `0.2.2` — small, explicit about its limits, and hardened against the
|
|
27
|
+
> failures found in an external-style audit. `protean` is a namespace sandbox,
|
|
28
|
+
> not a VM and not a hardened multi-tenant boundary.
|
|
29
|
+
|
|
30
|
+
## Plugins
|
|
31
|
+
|
|
32
|
+
| Plugin | What it does | Default |
|
|
33
|
+
|---|---|---|
|
|
34
|
+
| `trust` | Evaluates allow/deny policy with audit; verifies package integrity (never executes a manifest). | missing/invalid policy ⇒ deny |
|
|
35
|
+
| `protean` | Runs a command in a sandbox with a lease receipt, risk-routed substrates and JSONL evidence. | `low` = bubblewrap |
|
|
36
|
+
| `team` | Deterministic goal split, isolated clone, persistent state, at most one optional LLM call per task. | deterministic fallback |
|
|
37
|
+
|
|
38
|
+
## Install
|
|
39
|
+
|
|
40
|
+
```bash
|
|
41
|
+
pip install -e .
|
|
42
|
+
# optional Ed25519 verification for trust
|
|
43
|
+
pip install -e '.[crypto]'
|
|
44
|
+
```
|
|
45
|
+
|
|
46
|
+
Linux is recommended for `protean`:
|
|
47
|
+
|
|
48
|
+
```bash
|
|
49
|
+
sudo apt-get install bubblewrap # risk=low
|
|
50
|
+
# Docker is optional; risk=medium/high use docker runc
|
|
51
|
+
```
|
|
52
|
+
|
|
53
|
+
`risk=low` mounts a minimal read-only root (`/usr/bin`, `/usr/sbin`,
|
|
54
|
+
`/usr/lib`, `/usr/share`, `/bin`, `/sbin`, `/lib`, `/lib64`). `/etc` is a tmpfs
|
|
55
|
+
with an allowlist of non-secret files; `/etc/shadow`, `gshadow`, `sudoers`,
|
|
56
|
+
host SSH keys and private TLS keys are not exposed. `/root`, `/home`, `/opt`,
|
|
57
|
+
`/var` and `/run` are not mounted. `/tmp` is a 32 MiB tmpfs, and `prlimit`
|
|
58
|
+
caps address space and process count when available.
|
|
59
|
+
|
|
60
|
+
**The allowlist checks only the first executable token.** Shell syntax can run
|
|
61
|
+
other binaries present in the mounted root. The sandbox, not the allowlist, is
|
|
62
|
+
the security boundary. Available binaries depend on the substrate: the default
|
|
63
|
+
Docker image has `sh` and `python3`, but not `bash`, `git` or `node`; use a
|
|
64
|
+
custom `PROTEAN_IMAGE` if you need them.
|
|
65
|
+
|
|
66
|
+
## MCP
|
|
67
|
+
|
|
68
|
+
```json
|
|
69
|
+
{
|
|
70
|
+
"mcpServers": {
|
|
71
|
+
"gamai-plugins": {
|
|
72
|
+
"command": "gamai-plugins-mcp"
|
|
73
|
+
}
|
|
74
|
+
}
|
|
75
|
+
}
|
|
76
|
+
```
|
|
77
|
+
|
|
78
|
+
The server exposes six namespaced tools:
|
|
79
|
+
|
|
80
|
+
```text
|
|
81
|
+
trust_managed_policy
|
|
82
|
+
trust_marketplace_verify
|
|
83
|
+
protean_run
|
|
84
|
+
team_create
|
|
85
|
+
team_dispatch
|
|
86
|
+
team_status
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
## CLI
|
|
90
|
+
|
|
91
|
+
```bash
|
|
92
|
+
gamai-plugins list
|
|
93
|
+
gamai-plugins call trust managed_policy '{"action":"load"}'
|
|
94
|
+
gamai-plugins call protean run '{"cmd":"echo hi","risk":"low","timeout_s":5}'
|
|
95
|
+
gamai-plugins call team team_create '{"goal":"review this change","agents":2}'
|
|
96
|
+
gamai-plugins mcp
|
|
97
|
+
```
|
|
98
|
+
|
|
99
|
+
Or the per-plugin entry points:
|
|
100
|
+
|
|
101
|
+
```bash
|
|
102
|
+
gamai-trust managed_policy '{"action":"check","tool":"shell"}'
|
|
103
|
+
gamai-protean run '{"cmd":"uname -a","risk":"low"}'
|
|
104
|
+
gamai-team team_status '{"team_id":"team-..."}'
|
|
105
|
+
```
|
|
106
|
+
|
|
107
|
+
CLI exit code is non-zero when a tool returns `ok:false` or a trust decision is
|
|
108
|
+
`deny`, so shell/CI gates can use it.
|
|
109
|
+
|
|
110
|
+
## Python
|
|
111
|
+
|
|
112
|
+
```python
|
|
113
|
+
from gamai_plugins.cli import call
|
|
114
|
+
|
|
115
|
+
print(call("protean", "run", {"cmd": "echo hi", "risk": "low", "timeout_s": 5}))
|
|
116
|
+
```
|
|
117
|
+
|
|
118
|
+
## Policy and integrity semantics
|
|
119
|
+
|
|
120
|
+
`trust_managed_policy` is an **evaluator**, not an enforcement point. The caller
|
|
121
|
+
must enforce the returned `decision`. A valid policy without `default` is
|
|
122
|
+
`deny`; deny rules win. Paths are normalized and shell segments are evaluated
|
|
123
|
+
separately.
|
|
124
|
+
|
|
125
|
+
`trust_marketplace_verify` returns:
|
|
126
|
+
|
|
127
|
+
- `integrity_ok` — every expected file matches and there is no unexpected,
|
|
128
|
+
truncated or symlinked content;
|
|
129
|
+
- `trusted` — integrity is complete and the signature verifies with a **pinned**
|
|
130
|
+
key (`public_key` argument or `GAMAI_TRUST_PUBKEY`);
|
|
131
|
+
- `status` — `verified`, `self_signed`, `unsigned`, `failed` or `unverified`.
|
|
132
|
+
|
|
133
|
+
A self-signed manifest is not trusted. Without a manifest the result is
|
|
134
|
+
`unverified`, not a silent pass. See [docs/MANIFEST.md](docs/MANIFEST.md) for
|
|
135
|
+
the canonical signing form and `scripts/sign_manifest.py`.
|
|
136
|
+
|
|
137
|
+
## Team semantics
|
|
138
|
+
|
|
139
|
+
`team_create` creates a shared clone (not a source worktree) with
|
|
140
|
+
`core.hooksPath` pointing at an empty directory, so an untrusted repository
|
|
141
|
+
cannot execute hooks later. State is written before isolation, so a slow clone
|
|
142
|
+
does not leave a registered workspace without state.
|
|
143
|
+
|
|
144
|
+
`team` prompts and role hints are currently Bosnian. That is intentional, not a
|
|
145
|
+
translation layer; make it configurable if you need another language. Dispatch
|
|
146
|
+
is serialized per team: the LLM call happens while the team lock is held, so
|
|
147
|
+
concurrent dispatches wait rather than racing state.
|
|
148
|
+
|
|
149
|
+
## Configuration
|
|
150
|
+
|
|
151
|
+
| Variable | Used by | Meaning |
|
|
152
|
+
|---|---|---|
|
|
153
|
+
| `GAMAI_POLICY` | trust | policy JSON; default `~/.config/gamai/trust/policy.json` |
|
|
154
|
+
| `GAMAI_AUDIT` | trust | audit JSONL; default `~/.local/state/gamai/trust/audit.jsonl` |
|
|
155
|
+
| `GAMAI_AUDIT_DISABLE` | trust | `1` disables audit **only when the policy does not require audit** |
|
|
156
|
+
| `GAMAI_TRUST_PUBKEY` | trust | pinned Ed25519 key (PEM, 64 hex, or base64 raw 32 bytes) |
|
|
157
|
+
| `GAMAI_HOME` | trust | base config dir; default `~/.config/gamai` |
|
|
158
|
+
| `GAMAI_STATE` | trust | base state dir; default `~/.local/state/gamai` |
|
|
159
|
+
| `PROTEAN_HOME` | protean | data dir; default `~/.config/gamai/protean` |
|
|
160
|
+
| `PROTEAN_IMAGE` | protean | Docker image for medium/high; default `python:3.12-alpine` |
|
|
161
|
+
| `PROTEAN_HIGH_REMOTE` | protean | `1` allows high risk to use the configured remote gVisor |
|
|
162
|
+
| `PROTEAN_REMOTE_HOST` | protean | optional remote host for `critical` + gVisor |
|
|
163
|
+
| `PROTEAN_REMOTE_PORT`, `PROTEAN_REMOTE_KEY`, `PROTEAN_REMOTE_USER` | protean | SSH details for the remote substrate |
|
|
164
|
+
| `PROTEAN_KNOWN_HOSTS` | protean | SSH known_hosts; default `~/.ssh/known_hosts` |
|
|
165
|
+
| `PROTEAN_SSH_STRICT` | protean | `yes`, `accept-new` (default) or `no` |
|
|
166
|
+
| `GAMAI_TEAM_STATE_DIR` | team | state dir; default `~/.local/state/gamai/teams` |
|
|
167
|
+
| `GAMAI_TEAM_WORKTREES` | team | workspace dir; default `~/.cache/gamai/teams` |
|
|
168
|
+
| `GAMAI_TEAM_REPO` | team | default repo for isolation when `repo` is omitted |
|
|
169
|
+
| `GAMAI_TEAM_LLM_BUDGET` | team | max LLM calls per team; default 50 |
|
|
170
|
+
| `GAMAI_TEAM_LLM_TIMEOUT` | team | per-call timeout; default 40 s |
|
|
171
|
+
| `GAMAI_LLM_URL` | team | optional HTTPS endpoint for one LLM call per task; HTTP only for loopback |
|
|
172
|
+
| `GAMAI_LLM_TOKEN` | team | optional `X-Token` sent to that endpoint |
|
|
173
|
+
| `GAMAI_TEAM_AGENT` | team | default backend name; default `agent` |
|
|
174
|
+
|
|
175
|
+
No secret is stored in this repository. `protean` creates a per-install
|
|
176
|
+
`lease.key` at runtime with mode `600`.
|
|
177
|
+
|
|
178
|
+
## Safety model
|
|
179
|
+
|
|
180
|
+
- **Trust is fail-closed:** missing/invalid policy, invalid manifest or a
|
|
181
|
+
required audit failure yields deny/failure.
|
|
182
|
+
- **Manifest is data:** `trust` hashes and verifies packages; it never executes
|
|
183
|
+
a manifest.
|
|
184
|
+
- **Lease is a receipt, not authorization:** `protean` mints and verifies the
|
|
185
|
+
lease in the same process. It binds command/genome/TTL for evidence, but it
|
|
186
|
+
is not an external policy decision.
|
|
187
|
+
- **Substrate routing:** `low` uses bubblewrap; `medium`/`high` use Docker with
|
|
188
|
+
no network, dropped capabilities, read-only root and resource limits;
|
|
189
|
+
`critical` is denied unless a remote gVisor host is explicitly configured.
|
|
190
|
+
- **Bounded output:** `protean` captures at most 64 KiB per stream and returns
|
|
191
|
+
at most 8,000 characters; evidence runs are pruned to the newest 200.
|
|
192
|
+
- **Audit cannot be disabled when required:** `GAMAI_AUDIT_DISABLE=1` is
|
|
193
|
+
ignored when the policy says `audit: true`; audit failure becomes
|
|
194
|
+
`deny`/`audit_failed`.
|
|
195
|
+
- **Team is not an authorization boundary.** It coordinates work; the backend
|
|
196
|
+
decides what it executes. LLM redirects are not followed and public plain
|
|
197
|
+
HTTP is refused; the configured HTTPS origin still receives the token by
|
|
198
|
+
design.
|
|
199
|
+
- **Docker socket is root-equivalent.** Do not expose `protean` medium/high to
|
|
200
|
+
untrusted users.
|
|
201
|
+
- **Trust score is heuristic:** `trust_score` is a documented 0–100 heuristic,
|
|
202
|
+
not a cryptographic proof.
|
|
203
|
+
|
|
204
|
+
Read [SECURITY.md](SECURITY.md) and [THREAT_MODEL.md](THREAT_MODEL.md) before
|
|
205
|
+
using this in production.
|
|
206
|
+
|
|
207
|
+
## Development
|
|
208
|
+
|
|
209
|
+
```bash
|
|
210
|
+
python -m unittest discover -s tests -v
|
|
211
|
+
python -m gamai_plugins.mcp --selftest
|
|
212
|
+
python scripts/smoke_mcp.py
|
|
213
|
+
```
|
|
214
|
+
|
|
215
|
+
The tests do not need network access, Docker, or an LLM. `protean` low-risk
|
|
216
|
+
tests use bubblewrap when available; Windows runs the same suite without it.
|
|
217
|
+
|
|
218
|
+
## License
|
|
219
|
+
|
|
220
|
+
MIT.
|
|
@@ -0,0 +1,202 @@
|
|
|
1
|
+
# gamAI plugins
|
|
2
|
+
|
|
3
|
+
Portable, dependency-light plugins for AI agents: **policy evaluation**,
|
|
4
|
+
**sandbox execution** and **team coordination**. The same code is exposed as an
|
|
5
|
+
MCP server, a CLI and a plain Python function, so it can run in Claude Code,
|
|
6
|
+
opencode, Codex, Cursor, LangChain or a custom agent loop.
|
|
7
|
+
|
|
8
|
+
> Status: `0.2.2` — small, explicit about its limits, and hardened against the
|
|
9
|
+
> failures found in an external-style audit. `protean` is a namespace sandbox,
|
|
10
|
+
> not a VM and not a hardened multi-tenant boundary.
|
|
11
|
+
|
|
12
|
+
## Plugins
|
|
13
|
+
|
|
14
|
+
| Plugin | What it does | Default |
|
|
15
|
+
|---|---|---|
|
|
16
|
+
| `trust` | Evaluates allow/deny policy with audit; verifies package integrity (never executes a manifest). | missing/invalid policy ⇒ deny |
|
|
17
|
+
| `protean` | Runs a command in a sandbox with a lease receipt, risk-routed substrates and JSONL evidence. | `low` = bubblewrap |
|
|
18
|
+
| `team` | Deterministic goal split, isolated clone, persistent state, at most one optional LLM call per task. | deterministic fallback |
|
|
19
|
+
|
|
20
|
+
## Install
|
|
21
|
+
|
|
22
|
+
```bash
|
|
23
|
+
pip install -e .
|
|
24
|
+
# optional Ed25519 verification for trust
|
|
25
|
+
pip install -e '.[crypto]'
|
|
26
|
+
```
|
|
27
|
+
|
|
28
|
+
Linux is recommended for `protean`:
|
|
29
|
+
|
|
30
|
+
```bash
|
|
31
|
+
sudo apt-get install bubblewrap # risk=low
|
|
32
|
+
# Docker is optional; risk=medium/high use docker runc
|
|
33
|
+
```
|
|
34
|
+
|
|
35
|
+
`risk=low` mounts a minimal read-only root (`/usr/bin`, `/usr/sbin`,
|
|
36
|
+
`/usr/lib`, `/usr/share`, `/bin`, `/sbin`, `/lib`, `/lib64`). `/etc` is a tmpfs
|
|
37
|
+
with an allowlist of non-secret files; `/etc/shadow`, `gshadow`, `sudoers`,
|
|
38
|
+
host SSH keys and private TLS keys are not exposed. `/root`, `/home`, `/opt`,
|
|
39
|
+
`/var` and `/run` are not mounted. `/tmp` is a 32 MiB tmpfs, and `prlimit`
|
|
40
|
+
caps address space and process count when available.
|
|
41
|
+
|
|
42
|
+
**The allowlist checks only the first executable token.** Shell syntax can run
|
|
43
|
+
other binaries present in the mounted root. The sandbox, not the allowlist, is
|
|
44
|
+
the security boundary. Available binaries depend on the substrate: the default
|
|
45
|
+
Docker image has `sh` and `python3`, but not `bash`, `git` or `node`; use a
|
|
46
|
+
custom `PROTEAN_IMAGE` if you need them.
|
|
47
|
+
|
|
48
|
+
## MCP
|
|
49
|
+
|
|
50
|
+
```json
|
|
51
|
+
{
|
|
52
|
+
"mcpServers": {
|
|
53
|
+
"gamai-plugins": {
|
|
54
|
+
"command": "gamai-plugins-mcp"
|
|
55
|
+
}
|
|
56
|
+
}
|
|
57
|
+
}
|
|
58
|
+
```
|
|
59
|
+
|
|
60
|
+
The server exposes six namespaced tools:
|
|
61
|
+
|
|
62
|
+
```text
|
|
63
|
+
trust_managed_policy
|
|
64
|
+
trust_marketplace_verify
|
|
65
|
+
protean_run
|
|
66
|
+
team_create
|
|
67
|
+
team_dispatch
|
|
68
|
+
team_status
|
|
69
|
+
```
|
|
70
|
+
|
|
71
|
+
## CLI
|
|
72
|
+
|
|
73
|
+
```bash
|
|
74
|
+
gamai-plugins list
|
|
75
|
+
gamai-plugins call trust managed_policy '{"action":"load"}'
|
|
76
|
+
gamai-plugins call protean run '{"cmd":"echo hi","risk":"low","timeout_s":5}'
|
|
77
|
+
gamai-plugins call team team_create '{"goal":"review this change","agents":2}'
|
|
78
|
+
gamai-plugins mcp
|
|
79
|
+
```
|
|
80
|
+
|
|
81
|
+
Or the per-plugin entry points:
|
|
82
|
+
|
|
83
|
+
```bash
|
|
84
|
+
gamai-trust managed_policy '{"action":"check","tool":"shell"}'
|
|
85
|
+
gamai-protean run '{"cmd":"uname -a","risk":"low"}'
|
|
86
|
+
gamai-team team_status '{"team_id":"team-..."}'
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
CLI exit code is non-zero when a tool returns `ok:false` or a trust decision is
|
|
90
|
+
`deny`, so shell/CI gates can use it.
|
|
91
|
+
|
|
92
|
+
## Python
|
|
93
|
+
|
|
94
|
+
```python
|
|
95
|
+
from gamai_plugins.cli import call
|
|
96
|
+
|
|
97
|
+
print(call("protean", "run", {"cmd": "echo hi", "risk": "low", "timeout_s": 5}))
|
|
98
|
+
```
|
|
99
|
+
|
|
100
|
+
## Policy and integrity semantics
|
|
101
|
+
|
|
102
|
+
`trust_managed_policy` is an **evaluator**, not an enforcement point. The caller
|
|
103
|
+
must enforce the returned `decision`. A valid policy without `default` is
|
|
104
|
+
`deny`; deny rules win. Paths are normalized and shell segments are evaluated
|
|
105
|
+
separately.
|
|
106
|
+
|
|
107
|
+
`trust_marketplace_verify` returns:
|
|
108
|
+
|
|
109
|
+
- `integrity_ok` — every expected file matches and there is no unexpected,
|
|
110
|
+
truncated or symlinked content;
|
|
111
|
+
- `trusted` — integrity is complete and the signature verifies with a **pinned**
|
|
112
|
+
key (`public_key` argument or `GAMAI_TRUST_PUBKEY`);
|
|
113
|
+
- `status` — `verified`, `self_signed`, `unsigned`, `failed` or `unverified`.
|
|
114
|
+
|
|
115
|
+
A self-signed manifest is not trusted. Without a manifest the result is
|
|
116
|
+
`unverified`, not a silent pass. See [docs/MANIFEST.md](docs/MANIFEST.md) for
|
|
117
|
+
the canonical signing form and `scripts/sign_manifest.py`.
|
|
118
|
+
|
|
119
|
+
## Team semantics
|
|
120
|
+
|
|
121
|
+
`team_create` creates a shared clone (not a source worktree) with
|
|
122
|
+
`core.hooksPath` pointing at an empty directory, so an untrusted repository
|
|
123
|
+
cannot execute hooks later. State is written before isolation, so a slow clone
|
|
124
|
+
does not leave a registered workspace without state.
|
|
125
|
+
|
|
126
|
+
`team` prompts and role hints are currently Bosnian. That is intentional, not a
|
|
127
|
+
translation layer; make it configurable if you need another language. Dispatch
|
|
128
|
+
is serialized per team: the LLM call happens while the team lock is held, so
|
|
129
|
+
concurrent dispatches wait rather than racing state.
|
|
130
|
+
|
|
131
|
+
## Configuration
|
|
132
|
+
|
|
133
|
+
| Variable | Used by | Meaning |
|
|
134
|
+
|---|---|---|
|
|
135
|
+
| `GAMAI_POLICY` | trust | policy JSON; default `~/.config/gamai/trust/policy.json` |
|
|
136
|
+
| `GAMAI_AUDIT` | trust | audit JSONL; default `~/.local/state/gamai/trust/audit.jsonl` |
|
|
137
|
+
| `GAMAI_AUDIT_DISABLE` | trust | `1` disables audit **only when the policy does not require audit** |
|
|
138
|
+
| `GAMAI_TRUST_PUBKEY` | trust | pinned Ed25519 key (PEM, 64 hex, or base64 raw 32 bytes) |
|
|
139
|
+
| `GAMAI_HOME` | trust | base config dir; default `~/.config/gamai` |
|
|
140
|
+
| `GAMAI_STATE` | trust | base state dir; default `~/.local/state/gamai` |
|
|
141
|
+
| `PROTEAN_HOME` | protean | data dir; default `~/.config/gamai/protean` |
|
|
142
|
+
| `PROTEAN_IMAGE` | protean | Docker image for medium/high; default `python:3.12-alpine` |
|
|
143
|
+
| `PROTEAN_HIGH_REMOTE` | protean | `1` allows high risk to use the configured remote gVisor |
|
|
144
|
+
| `PROTEAN_REMOTE_HOST` | protean | optional remote host for `critical` + gVisor |
|
|
145
|
+
| `PROTEAN_REMOTE_PORT`, `PROTEAN_REMOTE_KEY`, `PROTEAN_REMOTE_USER` | protean | SSH details for the remote substrate |
|
|
146
|
+
| `PROTEAN_KNOWN_HOSTS` | protean | SSH known_hosts; default `~/.ssh/known_hosts` |
|
|
147
|
+
| `PROTEAN_SSH_STRICT` | protean | `yes`, `accept-new` (default) or `no` |
|
|
148
|
+
| `GAMAI_TEAM_STATE_DIR` | team | state dir; default `~/.local/state/gamai/teams` |
|
|
149
|
+
| `GAMAI_TEAM_WORKTREES` | team | workspace dir; default `~/.cache/gamai/teams` |
|
|
150
|
+
| `GAMAI_TEAM_REPO` | team | default repo for isolation when `repo` is omitted |
|
|
151
|
+
| `GAMAI_TEAM_LLM_BUDGET` | team | max LLM calls per team; default 50 |
|
|
152
|
+
| `GAMAI_TEAM_LLM_TIMEOUT` | team | per-call timeout; default 40 s |
|
|
153
|
+
| `GAMAI_LLM_URL` | team | optional HTTPS endpoint for one LLM call per task; HTTP only for loopback |
|
|
154
|
+
| `GAMAI_LLM_TOKEN` | team | optional `X-Token` sent to that endpoint |
|
|
155
|
+
| `GAMAI_TEAM_AGENT` | team | default backend name; default `agent` |
|
|
156
|
+
|
|
157
|
+
No secret is stored in this repository. `protean` creates a per-install
|
|
158
|
+
`lease.key` at runtime with mode `600`.
|
|
159
|
+
|
|
160
|
+
## Safety model
|
|
161
|
+
|
|
162
|
+
- **Trust is fail-closed:** missing/invalid policy, invalid manifest or a
|
|
163
|
+
required audit failure yields deny/failure.
|
|
164
|
+
- **Manifest is data:** `trust` hashes and verifies packages; it never executes
|
|
165
|
+
a manifest.
|
|
166
|
+
- **Lease is a receipt, not authorization:** `protean` mints and verifies the
|
|
167
|
+
lease in the same process. It binds command/genome/TTL for evidence, but it
|
|
168
|
+
is not an external policy decision.
|
|
169
|
+
- **Substrate routing:** `low` uses bubblewrap; `medium`/`high` use Docker with
|
|
170
|
+
no network, dropped capabilities, read-only root and resource limits;
|
|
171
|
+
`critical` is denied unless a remote gVisor host is explicitly configured.
|
|
172
|
+
- **Bounded output:** `protean` captures at most 64 KiB per stream and returns
|
|
173
|
+
at most 8,000 characters; evidence runs are pruned to the newest 200.
|
|
174
|
+
- **Audit cannot be disabled when required:** `GAMAI_AUDIT_DISABLE=1` is
|
|
175
|
+
ignored when the policy says `audit: true`; audit failure becomes
|
|
176
|
+
`deny`/`audit_failed`.
|
|
177
|
+
- **Team is not an authorization boundary.** It coordinates work; the backend
|
|
178
|
+
decides what it executes. LLM redirects are not followed and public plain
|
|
179
|
+
HTTP is refused; the configured HTTPS origin still receives the token by
|
|
180
|
+
design.
|
|
181
|
+
- **Docker socket is root-equivalent.** Do not expose `protean` medium/high to
|
|
182
|
+
untrusted users.
|
|
183
|
+
- **Trust score is heuristic:** `trust_score` is a documented 0–100 heuristic,
|
|
184
|
+
not a cryptographic proof.
|
|
185
|
+
|
|
186
|
+
Read [SECURITY.md](SECURITY.md) and [THREAT_MODEL.md](THREAT_MODEL.md) before
|
|
187
|
+
using this in production.
|
|
188
|
+
|
|
189
|
+
## Development
|
|
190
|
+
|
|
191
|
+
```bash
|
|
192
|
+
python -m unittest discover -s tests -v
|
|
193
|
+
python -m gamai_plugins.mcp --selftest
|
|
194
|
+
python scripts/smoke_mcp.py
|
|
195
|
+
```
|
|
196
|
+
|
|
197
|
+
The tests do not need network access, Docker, or an LLM. `protean` low-risk
|
|
198
|
+
tests use bubblewrap when available; Windows runs the same suite without it.
|
|
199
|
+
|
|
200
|
+
## License
|
|
201
|
+
|
|
202
|
+
MIT.
|
|
@@ -0,0 +1,134 @@
|
|
|
1
|
+
#!/usr/bin/env python3
|
|
2
|
+
"""One CLI for all portable GAMA plugins."""
|
|
3
|
+
import json
|
|
4
|
+
import subprocess
|
|
5
|
+
import sys
|
|
6
|
+
from pathlib import Path
|
|
7
|
+
|
|
8
|
+
ROOT = Path(__file__).resolve().parent / "plugins"
|
|
9
|
+
PLUGINS = ("trust", "protean", "team")
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
def manifest(plugin):
|
|
13
|
+
with open(ROOT / plugin / "plugin.json", encoding="utf-8") as fh:
|
|
14
|
+
return json.load(fh)
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
def _public_name(plugin, tool_name):
|
|
18
|
+
return tool_name if tool_name.startswith(plugin + "_") else "%s_%s" % (plugin, tool_name)
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
def tools():
|
|
22
|
+
out = []
|
|
23
|
+
for plugin in PLUGINS:
|
|
24
|
+
m = manifest(plugin)
|
|
25
|
+
for tool in m.get("tools", []):
|
|
26
|
+
out.append({
|
|
27
|
+
"name": _public_name(plugin, tool["name"]),
|
|
28
|
+
"plugin": plugin,
|
|
29
|
+
"tool": tool["name"],
|
|
30
|
+
"description": tool.get("description", ""),
|
|
31
|
+
"inputSchema": tool.get("parameters", {"type": "object"}),
|
|
32
|
+
"timeout_s": tool.get("timeout_s", 90),
|
|
33
|
+
})
|
|
34
|
+
return out
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
def call(plugin, tool, args, timeout=None):
|
|
38
|
+
if plugin not in PLUGINS:
|
|
39
|
+
return {"ok": False, "err": "unknown plugin %r" % plugin}
|
|
40
|
+
path = ROOT / plugin / "run.py"
|
|
41
|
+
if not path.is_file():
|
|
42
|
+
return {"ok": False, "err": "plugin entrypoint missing"}
|
|
43
|
+
if timeout is None:
|
|
44
|
+
timeout = 90
|
|
45
|
+
for entry in tools():
|
|
46
|
+
if entry["plugin"] == plugin and entry["tool"] == tool:
|
|
47
|
+
timeout = int(entry.get("timeout_s") or 90)
|
|
48
|
+
break
|
|
49
|
+
try:
|
|
50
|
+
p = subprocess.run(
|
|
51
|
+
[sys.executable, str(path), tool],
|
|
52
|
+
input=json.dumps(args, ensure_ascii=False),
|
|
53
|
+
text=True, encoding="utf-8", errors="replace",
|
|
54
|
+
capture_output=True, timeout=timeout,
|
|
55
|
+
cwd=str(ROOT / plugin),
|
|
56
|
+
)
|
|
57
|
+
except subprocess.TimeoutExpired:
|
|
58
|
+
return {"ok": False, "err": "timeout after %ss" % timeout}
|
|
59
|
+
raw = (p.stdout or "").strip().splitlines()
|
|
60
|
+
for line in reversed(raw):
|
|
61
|
+
try:
|
|
62
|
+
return json.loads(line)
|
|
63
|
+
except Exception:
|
|
64
|
+
continue
|
|
65
|
+
return {"ok": False, "err": (p.stderr or p.stdout or "empty output").strip()[:500]}
|
|
66
|
+
|
|
67
|
+
|
|
68
|
+
def _print(obj):
|
|
69
|
+
print(json.dumps(obj, ensure_ascii=False, indent=2))
|
|
70
|
+
|
|
71
|
+
|
|
72
|
+
def main(argv=None):
|
|
73
|
+
argv = list(sys.argv[1:] if argv is None else argv)
|
|
74
|
+
if not argv or argv[0] in ("-h", "--help", "help"):
|
|
75
|
+
print("usage: gamai-plugins list | call <plugin> <tool> [json] | mcp")
|
|
76
|
+
return 0
|
|
77
|
+
cmd = argv[0]
|
|
78
|
+
if cmd == "list":
|
|
79
|
+
_print({"ok": True, "tools": tools()})
|
|
80
|
+
return 0
|
|
81
|
+
if cmd == "call":
|
|
82
|
+
if len(argv) < 3:
|
|
83
|
+
_print({"ok": False, "err": "usage: call <plugin> <tool> [json]"})
|
|
84
|
+
return 2
|
|
85
|
+
plugin, tool = argv[1], argv[2]
|
|
86
|
+
try:
|
|
87
|
+
args = json.loads(argv[3]) if len(argv) > 3 else {}
|
|
88
|
+
except Exception as exc:
|
|
89
|
+
_print({"ok": False, "err": "invalid json: %s" % exc})
|
|
90
|
+
return 2
|
|
91
|
+
if not isinstance(args, dict):
|
|
92
|
+
_print({"ok": False, "err": "arguments must be a JSON object"})
|
|
93
|
+
return 2
|
|
94
|
+
res = call(plugin, tool, args)
|
|
95
|
+
_print(res)
|
|
96
|
+
return 0 if res.get("ok") and res.get("allowed") is not False else 1
|
|
97
|
+
if cmd == "mcp":
|
|
98
|
+
from .mcp import main as mcp_main
|
|
99
|
+
return mcp_main()
|
|
100
|
+
_print({"ok": False, "err": "unknown command %r" % cmd})
|
|
101
|
+
return 2
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
def _plugin_main(plugin, argv=None):
|
|
105
|
+
argv = list(sys.argv[1:] if argv is None else argv)
|
|
106
|
+
if not argv:
|
|
107
|
+
_print({"ok": False, "err": "usage: gamai-%s <tool> [json]" % plugin})
|
|
108
|
+
return 2
|
|
109
|
+
args = {}
|
|
110
|
+
if len(argv) > 1:
|
|
111
|
+
try:
|
|
112
|
+
args = json.loads(argv[1])
|
|
113
|
+
except Exception as exc:
|
|
114
|
+
_print({"ok": False, "err": "invalid json: %s" % exc})
|
|
115
|
+
return 2
|
|
116
|
+
res = call(plugin, argv[0], args)
|
|
117
|
+
_print(res)
|
|
118
|
+
return 0 if res.get("ok") and res.get("allowed") is not False else 1
|
|
119
|
+
|
|
120
|
+
|
|
121
|
+
def trust_main(argv=None):
|
|
122
|
+
return _plugin_main("trust", argv)
|
|
123
|
+
|
|
124
|
+
|
|
125
|
+
def protean_main(argv=None):
|
|
126
|
+
return _plugin_main("protean", argv)
|
|
127
|
+
|
|
128
|
+
|
|
129
|
+
def team_main(argv=None):
|
|
130
|
+
return _plugin_main("team", argv)
|
|
131
|
+
|
|
132
|
+
|
|
133
|
+
if __name__ == "__main__":
|
|
134
|
+
raise SystemExit(main())
|
|
@@ -0,0 +1,89 @@
|
|
|
1
|
+
#!/usr/bin/env python3
|
|
2
|
+
"""Minimal stdio MCP server for the portable GAMA plugins (stdlib only)."""
|
|
3
|
+
import json
|
|
4
|
+
import sys
|
|
5
|
+
|
|
6
|
+
from . import __version__
|
|
7
|
+
from .cli import call, tools
|
|
8
|
+
|
|
9
|
+
PROTOCOL = "2024-11-05"
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
def _tool_list():
|
|
13
|
+
return [{
|
|
14
|
+
"name": t["name"],
|
|
15
|
+
"description": "[%s] %s" % (t["plugin"], t["description"]),
|
|
16
|
+
"inputSchema": t["inputSchema"],
|
|
17
|
+
} for t in tools()]
|
|
18
|
+
|
|
19
|
+
|
|
20
|
+
def _find(name):
|
|
21
|
+
for t in tools():
|
|
22
|
+
if t["name"] == name:
|
|
23
|
+
return t
|
|
24
|
+
return None
|
|
25
|
+
|
|
26
|
+
|
|
27
|
+
def handle(req):
|
|
28
|
+
if not isinstance(req, dict):
|
|
29
|
+
return None
|
|
30
|
+
rid = req.get("id")
|
|
31
|
+
method = req.get("method")
|
|
32
|
+
if method == "initialize":
|
|
33
|
+
return {"jsonrpc": "2.0", "id": rid, "result": {
|
|
34
|
+
"protocolVersion": PROTOCOL,
|
|
35
|
+
"capabilities": {"tools": {}},
|
|
36
|
+
"serverInfo": {"name": "gamai-plugins", "version": __version__},
|
|
37
|
+
}}
|
|
38
|
+
if "id" not in req:
|
|
39
|
+
return None
|
|
40
|
+
if method == "notifications/initialized":
|
|
41
|
+
return None
|
|
42
|
+
if method == "ping":
|
|
43
|
+
return {"jsonrpc": "2.0", "id": rid, "result": {}}
|
|
44
|
+
if method == "tools/list":
|
|
45
|
+
return {"jsonrpc": "2.0", "id": rid, "result": {"tools": _tool_list()}}
|
|
46
|
+
if method == "tools/call":
|
|
47
|
+
params = req.get("params") or {}
|
|
48
|
+
name = params.get("name") or ""
|
|
49
|
+
args = params.get("arguments") or {}
|
|
50
|
+
t = _find(name)
|
|
51
|
+
if t is None:
|
|
52
|
+
return {"jsonrpc": "2.0", "id": rid, "error": {"code": -32601, "message": "unknown tool %r" % name}}
|
|
53
|
+
res = call(t["plugin"], t["tool"], args, t.get("timeout_s"))
|
|
54
|
+
text = json.dumps(res, ensure_ascii=False, indent=2)
|
|
55
|
+
return {"jsonrpc": "2.0", "id": rid, "result": {
|
|
56
|
+
"content": [{"type": "text", "text": text}],
|
|
57
|
+
"isError": bool(res.get("ok") is False or res.get("allowed") is False),
|
|
58
|
+
}}
|
|
59
|
+
return {"jsonrpc": "2.0", "id": rid, "error": {"code": -32601, "message": "unknown method %r" % method}}
|
|
60
|
+
|
|
61
|
+
|
|
62
|
+
def main(argv=None):
|
|
63
|
+
argv = list(sys.argv[1:] if argv is None else argv)
|
|
64
|
+
for stream in (sys.stdin, sys.stdout):
|
|
65
|
+
try:
|
|
66
|
+
stream.reconfigure(encoding="utf-8", errors="replace")
|
|
67
|
+
except Exception:
|
|
68
|
+
pass
|
|
69
|
+
if "--selftest" in argv:
|
|
70
|
+
print(json.dumps({"ok": True, "tools": len(tools()), "names": [t["name"] for t in tools()]}, indent=2))
|
|
71
|
+
return 0
|
|
72
|
+
for line in sys.stdin:
|
|
73
|
+
line = line.strip()
|
|
74
|
+
if not line:
|
|
75
|
+
continue
|
|
76
|
+
try:
|
|
77
|
+
req = json.loads(line)
|
|
78
|
+
except Exception:
|
|
79
|
+
continue
|
|
80
|
+
requests = req if isinstance(req, list) else [req]
|
|
81
|
+
for item in requests:
|
|
82
|
+
out = handle(item)
|
|
83
|
+
if out is not None:
|
|
84
|
+
print(json.dumps(out, ensure_ascii=False), flush=True)
|
|
85
|
+
return 0
|
|
86
|
+
|
|
87
|
+
|
|
88
|
+
if __name__ == "__main__":
|
|
89
|
+
raise SystemExit(main())
|