feat(reviewer): support configuring a dedicated lightweight reviewer model (#615)

coworker/agent.py

@@ -244,6 +244,8 @@ def build_engine(
     # are user-global, preserving the "a repo can't enable this" invariant.
     auto_approve: Optional[bool] = None,
     auto_approve_shadow: Optional[bool] = None,
+    # Dedicated reviewer model (Issue #615). None ⇒ config.toml or session model.
+    reviewer_model: Optional[str] = None,
     # Persona-carried skill folders (OPE-58): the bundle's skills/ dir joins the loader so
     # its skills are readable by load_skill, not just listed by the filter.
     extra_skill_dirs: Optional[list[str | Path]] = None,
@@ -597,9 +599,14 @@ def _approval_extras(tool_name: str, _arguments: dict) -> dict:
     if live_on or shadow_on:
         from .reviewer import Reviewer
 
+        effective_reviewer_model = (
+            reviewer_model
+            or getattr(config, "reviewer_model", None)
+            or model
+        )
         engine.reviewer = Reviewer(
             provider=provider,
-            model=model,
+            model=effective_reviewer_model,
             known_world=engine.session_facts.world.render(),
         )
         # Shadow evaluation (Part 6 step 3): with only the shadow flag on, the reviewer is

coworker/config.py

@@ -52,6 +52,9 @@ class Config:
     # gates (zero false-allows; ≥30% fewer prompts) get measured on real sessions. Costs
     # one model call per card while on. Off by default; user-global only.
     auto_approve_shadow: bool = False
+    # Dedicated reviewer model for auto-approve mode (spec §1.5 / Issue #615).
+    # If None or empty, the reviewer inherits the session's model. User-global only.
+    reviewer_model: Optional[str] = None
     host: str = "127.0.0.1"
     port: int = 8765
     # Web search provider: "duckduckgo" (keyless default) | "tavily" | "brave" (need a key).
@@ -84,6 +87,7 @@ class Config:
     "allowed_domains",
     "auto_approve",
     "auto_approve_shadow",
+    "reviewer_model",
     "host",
     "port",
     "web_search_provider",
@@ -104,6 +108,7 @@ class Config:
     "allowed_domains",
     "auto_approve",
     "auto_approve_shadow",
+    "reviewer_model",
 }
 _WORKSPACE_FIELDS = _FIELDS - _GLOBAL_ONLY_FIELDS
 
@@ -139,9 +144,16 @@ def load_config(
 
     g = Path(global_path) if global_path is not None else global_config_path()
     if g.is_file():
-        for key, value in _read(g).items():
+        data = _read(g)
+        for key, value in data.items():
             if key in _FIELDS:
                 setattr(cfg, key, value)
+        if "reviewer" in data and isinstance(data["reviewer"], dict):
+            rm = data["reviewer"].get("model")
+            if isinstance(rm, str) and rm.strip():
+                cfg.reviewer_model = rm.strip()
+    if cfg.reviewer_model is not None and not str(cfg.reviewer_model).strip():
+        cfg.reviewer_model = None
     if workspace:
         w = Path(workspace).expanduser() / ".coworker" / "config.toml"
         if w.is_file():

coworker/server/app.py

@@ -1944,6 +1944,11 @@ def settings_set_auto_approve_shadow(body: dict) -> dict[str, Any]:
         # every approval card while the human still decides. Independent of the live flag.
         return manager.set_auto_approve_shadow((body or {}).get("auto_approve_shadow", False))
 
+    @app.post("/v1/settings/reviewer-model")
+    def settings_set_reviewer_model(body: dict) -> dict[str, Any]:
+        # Dedicated reviewer model for Auto-Approve (Issue #615).
+        return manager.set_reviewer_model((body or {}).get("reviewer_model"))
+
     @app.post("/v1/settings/pdf")
     def settings_set_pdf(body: dict) -> dict[str, Any]:
         # Token savings (owner ask, 2026-07-17): fallback mode for models without native

coworker/server/manager.py

@@ -724,6 +724,7 @@ def get_engine(
             # the next session build without a config.toml edit.
             auto_approve=self.auto_approve(),
             auto_approve_shadow=self.auto_approve_shadow(),
+            reviewer_model=self.reviewer_model(),
         )
         # An automation run rebuilt here (manual "Run now" over WS, durable resume) still
         # carries its task's standing allowances — the rules live on the task record.
@@ -3446,6 +3447,7 @@ def _selectable(m: str) -> bool:
             # Settings toggles and gate the composer's Auto-Approve mode entry.
             "auto_approve": self.auto_approve(),
             "auto_approve_shadow": self.auto_approve_shadow(),
+            "reviewer_model": self.reviewer_model(),
             "scratch_base": self._prefs.get("scratch_base")
             or self.DEFAULT_SCRATCH_BASE,
             # Real on-disk secrets location, so the UI shows the OS-native path instead of a
@@ -3551,6 +3553,25 @@ def set_auto_approve_shadow(self, on: Any) -> dict[str, Any]:
             "auto_approve_shadow": self.auto_approve_shadow(),
         }
 
+    def reviewer_model(self) -> Optional[str]:
+        from ..config import load_config
+
+        if "reviewer_model" in self._prefs and self._prefs["reviewer_model"]:
+            return str(self._prefs["reviewer_model"]).strip() or None
+        return load_config().reviewer_model
+
+    def set_reviewer_model(self, model: Any) -> dict[str, Any]:
+        val = str(model).strip() if model else ""
+        if val:
+            self._prefs["reviewer_model"] = val
+        else:
+            self._prefs.pop("reviewer_model", None)
+        self._save_prefs()
+        return {
+            "ok": True,
+            "reviewer_model": self.reviewer_model(),
+        }
+
     # -- PDF attachments / token savings (owner ask, 2026-07-17) ----------------
     DEFAULT_PDF_MAX_PAGES = 20
     DEFAULT_PDF_MAX_MB = 10

tests/test_auto_approve_settings.py

@@ -76,10 +76,73 @@ def test_build_engine_override_beats_config(tmp_path, monkeypatch):
     # config has it off by default; override to on → a reviewer is attached.
     engine = build_engine(agent=chat_agent(), auto_approve=True, auto_approve_shadow=False)
     assert engine.reviewer is not None
-    engine2 = build_engine(agent=chat_agent(), auto_approve=False, auto_approve_shadow=False)
+    engine2 = build_engine(
+        agent=chat_agent(), auto_approve=False, auto_approve_shadow=False
+    )
     assert engine2.reviewer is None
 
 
+def test_reviewer_model_endpoint_roundtrip(client):
+    s = client.get("/v1/settings").json()
+    assert s.get("reviewer_model") is None
+
+    # Set dedicated reviewer model
+    r = client.post(
+        "/v1/settings/reviewer-model",
+        json={"reviewer_model": "anthropic:claude-3-5-haiku"},
+    ).json()
+    assert r["ok"] and r["reviewer_model"] == "anthropic:claude-3-5-haiku"
+    assert (
+        client.get("/v1/settings").json()["reviewer_model"]
+        == "anthropic:claude-3-5-haiku"
+    )
+
+    # Clear dedicated model
+    r_clear = client.post(
+        "/v1/settings/reviewer-model", json={"reviewer_model": ""}
+    ).json()
+    assert r_clear["ok"] and r_clear["reviewer_model"] is None
+    assert client.get("/v1/settings").json()["reviewer_model"] is None
+
+
+def test_reviewer_model_falls_back_to_config(tmp_path, monkeypatch):
+    state = tmp_path / "state"
+    state.mkdir(parents=True)
+    (state / "config.toml").write_text('[reviewer]\nmodel = "ollama:llama3.2:3b"\n')
+    monkeypatch.setenv("COWORKER_STATE_DIR", str(state))
+    mgr = SessionManager(data_dir=tmp_path / "data")
+    assert mgr.reviewer_model() == "ollama:llama3.2:3b"
+
+    # Prefs override takes precedence
+    mgr.set_reviewer_model("anthropic:claude-3-5-haiku")
+    assert mgr.reviewer_model() == "anthropic:claude-3-5-haiku"
+
+
+def test_build_engine_reviewer_model(tmp_path, monkeypatch):
+    monkeypatch.setenv("COWORKER_STATE_DIR", str(tmp_path / "state"))
+    from coworker.agent import build_engine
+    from coworker.agents.chat import chat_agent
+
+    # When reviewer_model is supplied, engine.reviewer uses it
+    engine = build_engine(
+        agent=chat_agent(),
+        auto_approve=True,
+        model="gpt-5.6-sol",
+        reviewer_model="anthropic:claude-3-5-haiku",
+    )
+    assert engine.reviewer is not None
+    assert engine.reviewer.model == "anthropic:claude-3-5-haiku"
+
+    # When reviewer_model is omitted, engine.reviewer inherits session model
+    engine_default = build_engine(
+        agent=chat_agent(),
+        auto_approve=True,
+        model="gpt-5.6-sol",
+    )
+    assert engine_default.reviewer is not None
+    assert engine_default.reviewer.model == "gpt-5.6-sol"
+
+
 # -- metering (§1.7): durable reviewer stats from the audit store ------------------
 
 

tests/test_config.py

@@ -142,3 +142,26 @@ def test_cloud_endpoints_default_to_production():
     cfg = Config()
     assert cfg.cloud_base_url == "https://api.openworker.com"
     assert cfg.cloud_relay_ws_url.startswith("wss://")
+
+
+def test_reviewer_model_config(tmp_path):
+    g1 = tmp_path / "g1.toml"
+    g1.write_text('[reviewer]\nmodel = "anthropic:claude-3-5-haiku"\n')
+    cfg1 = load_config(global_path=g1)
+    assert cfg1.reviewer_model == "anthropic:claude-3-5-haiku"
+
+    g2 = tmp_path / "g2.toml"
+    g2.write_text('reviewer_model = "ollama:llama3.2:3b"\n')
+    cfg2 = load_config(global_path=g2)
+    assert cfg2.reviewer_model == "ollama:llama3.2:3b"
+
+    # Workspace config cannot override reviewer_model (user-global security invariant)
+    ws = tmp_path / "ws"
+    (ws / ".coworker").mkdir(parents=True)
+    (ws / ".coworker" / "config.toml").write_text(
+        '[reviewer]\nmodel = "untrusted:rogue-model"\n'
+        'reviewer_model = "untrusted:rogue"\n'
+    )
+    cfg3 = load_config(ws, global_path=g1)
+    assert cfg3.reviewer_model == "anthropic:claude-3-5-haiku"
+