feat: add Aforge benchmark runtime

.env.example

@@ -128,7 +128,8 @@
 # only provider credential, else claude_code. An invalid value is logged as
 # a warning and ignored. Leave this UNSET on an OpenRouter-only deployment —
 # auto-select already picks open_code and the deepseek-v4-flash-0731 default.
-# SWE_DEFAULT_RUNTIME=claude_code  # or: open_code, codex
+# SWE_DEFAULT_RUNTIME=claude_code  # or: aforge, open_code, codex
+# AFORGE_BIN=/absolute/path/to/aforge  # optional; otherwise aforge must be on PATH
 
 # Default model when callers don't pass `models` in the request config.
 # Applies to all 16 agent roles for whichever runtime is active. Caller

README.md

@@ -279,6 +279,28 @@ python -m pip install -e ".[dev]"
 
 ### 3. Run
 
+#### Benchmark with unreleased aforge-v2
+
+Build aforge-v2 once, point AgentField at it, and select the runtime. No SWE-AF
+code or prompt changes are required:
+
+```bash
+cd /path/to/aforge-v2
+go build -o /absolute/path/to/bin/aforge ./cmd/aforge
+
+export OPENROUTER_API_KEY=sk-or-v1-...
+export AFORGE_BIN=/absolute/path/to/bin/aforge
+export SWE_DEFAULT_RUNTIME=aforge
+export SWE_DEFAULT_MODEL=openrouter/deepseek/deepseek-v4-flash-0731
+python -m swe_af
+```
+
+The draft pins the exact AgentField harness commit used for these benchmarks.
+Replace that source pin with the released SDK version before merging. The
+existing Docker image does not bundle the unreleased aforge-v2 binary; add it
+to the container or run the node on the host until Aforge publishes an
+installable artifact.
+
 ```bash
 af                 # starts AgentField control plane on :8080
 python -m swe_af   # registers node id "swe-planner"
@@ -820,7 +842,7 @@ Pass `config` to `build` or `execute`. Full schema: [`swe_af/execution/schemas.p
 
 | Key                       | Default         | Description                                           |
 | ------------------------- | --------------- | ----------------------------------------------------- |
-| `runtime`                 | `"claude_code"` | Model runtime: `"claude_code"`, `"open_code"`, or `"codex"`. The default also honors the `SWE_DEFAULT_RUNTIME` env var when no `runtime` is passed in `config` — set it on the deployment so callers don't need to plumb a config through. |
+| `runtime`                 | `"claude_code"` | Model runtime: `"aforge"`, `"claude_code"`, `"open_code"`, or `"codex"`. The default also honors the `SWE_DEFAULT_RUNTIME` env var when no `runtime` is passed in `config` — set it on the deployment so callers don't need to plumb a config through. |
 | `models`                  | `null`          | Flat role-model map (`default` + role keys below). Without a caller-supplied value, the `SWE_DEFAULT_MODEL` env var is used as the default for all roles — set it on the deployment to pin a model without code changes. Caller `models.default` or per-role keys still win. |
 | `max_coding_iterations`   | `5`             | Inner-loop retry budget                               |
 | `max_advisor_invocations` | `2`             | Middle-loop advisor budget                            |

agentfield-package.yaml

@@ -53,7 +53,7 @@ user_environment:
       type: secret
       scope: global
     - name: SWE_DEFAULT_RUNTIME
-      description: Coding runtime for every role (claude_code | open_code | codex)
+      description: Coding runtime for every role (aforge | claude_code | open_code | codex)
     - name: SWE_DEFAULT_MODEL
       description: Override the model id for every role (e.g. openrouter/deepseek/deepseek-v4-flash-0731)
     - name: AGENTFIELD_SERVER

pyproject.toml

@@ -6,7 +6,7 @@ requires-python = ">=3.12"
 dependencies = [
     # >=0.1.96 ships ReasonerFailed, which build() raises so an empty build
     # reports `failed` (not `succeeded`) with its result preserved (#82 Gap 2).
-    "agentfield>=0.1.113",
+    "agentfield @ git+https://github.com/Agent-Field/agentfield.git@60b0aa76df604575429af60fb0ccffc9aaca7745#subdirectory=sdk/python",
     "pydantic>=2.0",
     # Compatibility pin: newer SDK builds have surfaced
     # "Unknown message type: rate_limit_event" during streaming.

requirements-docker.txt

@@ -2,7 +2,7 @@
 #
 # Same runtime dependencies as requirements.txt.
 
-agentfield>=0.1.111
+agentfield @ git+https://github.com/Agent-Field/agentfield.git@60b0aa76df604575429af60fb0ccffc9aaca7745#subdirectory=sdk/python
 pydantic>=2.0
 claude-agent-sdk==0.1.20
 hax-sdk>=0.2.4

requirements.txt

@@ -2,7 +2,7 @@
 #
 # Install: python -m pip install -r requirements.txt
 
-agentfield>=0.1.113
+agentfield @ git+https://github.com/Agent-Field/agentfield.git@60b0aa76df604575429af60fb0ccffc9aaca7745#subdirectory=sdk/python
 pydantic>=2.0
 claude-agent-sdk==0.1.20
 hax-sdk>=0.2.4

swe_af/execution/schemas.py

@@ -616,6 +616,9 @@ class QASynthesisResult(BaseModel):
 _OPENROUTER_AUTO_DEFAULT_MODEL = "openrouter/deepseek/deepseek-v4-flash-0731"
 
 _RUNTIME_BASE_MODELS: dict[str, dict[str, str]] = {
+    "aforge": {
+        **{field: _OPENROUTER_AUTO_DEFAULT_MODEL for field in ALL_MODEL_FIELDS},
+    },
     "claude_code": {
         **{field: "sonnet" for field in ALL_MODEL_FIELDS},
         "qa_synthesizer_model": "haiku",
@@ -656,7 +659,7 @@ def _codex_default_model() -> str:
     return _CODEX_CHATGPT_MODEL if _codex_uses_chatgpt_auth() else _CODEX_API_KEY_MODEL
 
 
-def _runtime_to_provider(runtime: str) -> Literal["claude", "opencode", "codex"]:
+def _runtime_to_provider(runtime: str) -> Literal["aforge", "claude", "opencode", "codex"]:
     return runtime_to_harness_provider(runtime)  # type: ignore[return-value]
 
 
@@ -677,7 +680,7 @@ def _openrouter_only_env() -> bool:
     return bool(os.getenv("OPENROUTER_API_KEY", "").strip())
 
 
-def _default_runtime() -> Literal["claude_code", "open_code", "codex"]:
+def _default_runtime() -> Literal["aforge", "claude_code", "open_code", "codex"]:
     """Default runtime, honoring the ``SWE_DEFAULT_RUNTIME`` env var.
 
     Lets the deployer pick the runtime without every caller having to pass
@@ -930,7 +933,7 @@ class BuildConfig(BaseModel):
 
     model_config = ConfigDict(extra="forbid")
 
-    runtime: Literal["claude_code", "open_code", "codex"] = Field(default_factory=_default_runtime)
+    runtime: Literal["aforge", "claude_code", "open_code", "codex"] = Field(default_factory=_default_runtime)
     models: dict[str, str] | None = None
 
     max_review_iterations: int = 2
@@ -1049,7 +1052,7 @@ def model_post_init(self, __context: Any) -> None:
         _validate_flat_models(self.models)
 
     @property
-    def ai_provider(self) -> Literal["claude", "opencode", "codex"]:
+    def ai_provider(self) -> Literal["aforge", "claude", "opencode", "codex"]:
         return _runtime_to_provider(self.runtime)
 
     @property
@@ -1239,7 +1242,7 @@ class ExecutionConfig(BaseModel):
 
     model_config = ConfigDict(extra="forbid")
 
-    runtime: Literal["claude_code", "open_code", "codex"] = Field(default_factory=_default_runtime)
+    runtime: Literal["aforge", "claude_code", "open_code", "codex"] = Field(default_factory=_default_runtime)
     models: dict[str, str] | None = None
     _resolved_models: dict[str, str] = PrivateAttr(default_factory=dict)
 
@@ -1291,7 +1294,7 @@ def _model_for(self, field_name: str) -> str:
         return self._resolved_models[field_name]
 
     @property
-    def ai_provider(self) -> Literal["claude", "opencode", "codex"]:
+    def ai_provider(self) -> Literal["aforge", "claude", "opencode", "codex"]:
         return _runtime_to_provider(self.runtime)
 
     @property

swe_af/fast/app.py

@@ -58,6 +58,8 @@ def _runtime_to_provider(runtime: str) -> str:
         return "claude"
     if runtime == "codex":
         return "codex"
+    if runtime == "aforge":
+        return "aforge"
     return "opencode"
 
 

swe_af/fast/schemas.py

@@ -20,6 +20,7 @@
 _OPEN_CODE_DEFAULT = "openrouter/deepseek/deepseek-v4-flash-0731"
 
 _RUNTIME_DEFAULTS: dict[str, str] = {
+    "aforge": _OPEN_CODE_DEFAULT,
     "claude_code": _CLAUDE_CODE_DEFAULT,
     "open_code": _OPEN_CODE_DEFAULT,
     # codex is resolved dynamically (auth-mode dependent); see _runtime_default().
@@ -126,7 +127,7 @@ class FastBuildConfig(BaseModel):
 
     model_config = ConfigDict(extra="forbid")
 
-    runtime: Literal["claude_code", "open_code", "codex"] = Field(default_factory=_default_fast_runtime)
+    runtime: Literal["aforge", "claude_code", "open_code", "codex"] = Field(default_factory=_default_fast_runtime)
     models: dict[str, str] | None = None
     max_tasks: int = 10
     task_timeout_seconds: int = 300

swe_af/issue/schemas.py

@@ -106,7 +106,7 @@ class IssueBuildConfig(BaseModel):
 
     model_config = ConfigDict(extra="forbid")
 
-    runtime: Literal["claude_code", "open_code", "codex"] = Field(
+    runtime: Literal["aforge", "claude_code", "open_code", "codex"] = Field(
         default_factory=_default_runtime
     )
     models: dict[str, str] | None = None

swe_af/runtime/providers.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-RUNTIME_VALUES = ("claude_code", "open_code", "codex")
+RUNTIME_VALUES = ("aforge", "claude_code", "open_code", "codex")
 
 
 def normalize_runtime_provider(runtime: str) -> str:
@@ -12,6 +12,8 @@ def normalize_runtime_provider(runtime: str) -> str:
         return "claude_code"
     if value in {"open_code", "opencode"}:
         return "open_code"
+    if value in {"aforge", "aforge_v2", "aforge-v2"}:
+        return "aforge"
     if value == "codex":
         return "codex"
     raise ValueError(f"Unsupported runtime provider: {runtime}")
@@ -24,6 +26,8 @@ def runtime_to_harness_provider(runtime: str) -> str:
         return "claude"
     if normalized == "open_code":
         return "opencode"
+    if normalized == "aforge":
+        return "aforge"
     return "codex"
 
 
@@ -34,4 +38,6 @@ def runtime_to_harness_adapter(runtime: str) -> str:
         return "claude-code"
     if normalized == "open_code":
         return "opencode"
+    if normalized == "aforge":
+        return "aforge"
     return "codex"

tests/test_model_config.py

@@ -89,6 +89,11 @@ def test_open_code_defaults(self) -> None:
         for field in ALL_MODEL_FIELDS:
             self.assertEqual(resolved[field], "openrouter/deepseek/deepseek-v4-flash-0731")
 
+    def test_aforge_defaults(self) -> None:
+        resolved = resolve_runtime_models(runtime="aforge", models=None)
+        for field in ALL_MODEL_FIELDS:
+            self.assertEqual(resolved[field], "openrouter/deepseek/deepseek-v4-flash-0731")
+
     def test_models_default_applies_to_all(self) -> None:
         resolved = resolve_runtime_models(
             runtime="claude_code",
@@ -127,6 +132,14 @@ def test_open_code_runtime_provider(self) -> None:
             resolved = cfg.resolved_models()
         self.assertEqual(resolved["coder_model"], "openrouter/deepseek/deepseek-v4-flash-0731")
 
+    def test_aforge_runtime_provider(self) -> None:
+        cfg = BuildConfig(runtime="aforge")
+        self.assertEqual(cfg.ai_provider, "aforge")
+        self.assertEqual(
+            cfg.resolved_models()["coder_model"],
+            "openrouter/deepseek/deepseek-v4-flash-0731",
+        )
+
 
 class TestOpenRouterAutoSelection(unittest.TestCase):
     """When only an OpenRouter key is present (no explicit runtime), SWE-AF

tests/test_runtime_provider_routing.py

@@ -7,6 +7,11 @@ def test_runtime_to_harness_adapter_supports_codex() -> None:
     assert runtime_to_harness_adapter("codex") == "codex"
 
 
+def test_runtime_to_harness_adapter_supports_aforge_aliases() -> None:
+    assert runtime_to_harness_adapter("aforge") == "aforge"
+    assert runtime_to_harness_adapter("aforge-v2") == "aforge"
+
+
 def test_execution_agents_source_uses_shared_runtime_adapter() -> None:
     import inspect
     import swe_af.reasoners.execution_agents as execution_agents