@miller-tech/uap 1.42.5 → 1.43.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.42.5",
+  "version": "1.43.0",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/src/policies/enforcers/delivery_enforcement.py

@@ -6,8 +6,9 @@ non-trivial coding work goes through the `uap deliver` convergence loop (which
 drives a model to verified completion against the real gates) rather than
 ad-hoc hand edits.
 
-SAFETY — default mode is ADVISORY (always allows, logs a nudge), so installing
-this policy never breaks editing. Strict enforcement is opt-in:
+DEFAULT mode is BLOCK: substantive source edits must route through `uap deliver`
+(set UAP_ENFORCE_DELIVERY=advisory to relax to a logged nudge). Escape hatches
+(UAP_DELIVER_ACTIVE / UAP_DELIVER_BYPASS) are always honored:
 
   UAP_ENFORCE_DELIVERY=block   # direct source edits outside a deliver context
                                # are blocked (exit 2)
@@ -84,13 +85,13 @@ def main() -> None:
         "manual edit."
     )
 
-    mode = os.environ.get("UAP_ENFORCE_DELIVERY", "advisory").lower()
+    mode = os.environ.get("UAP_ENFORCE_DELIVERY", "block").lower()
     if mode == "block":
         emit(False, msg)
 
-    # Advisory (default): never blocks. Surface the nudge, then allow.
+    # Advisory (opt-out): never blocks. Surface the nudge, then allow.
     print(f"[delivery-enforcement advisory] {msg}", file=sys.stderr)
-    emit(True, "advisory: nudge logged (set UAP_ENFORCE_DELIVERY=block to enforce)")
+    emit(True, "advisory: nudge logged (block is the default; UAP_ENFORCE_DELIVERY=advisory relaxes)")
 
 
 if __name__ == "__main__":

package/src/policies/schemas/policies/delivery-enforcement.md

@@ -32,13 +32,13 @@ the expectation explicit and, when a team opts in, enforces it.
 
 Python enforcer `delivery_enforcement.py`.
 
-**Default mode is ADVISORY** — it always allows the edit and logs a one-line
-nudge toward `uap deliver`. Installing the policy therefore never breaks normal
-editing.
+**Default mode is BLOCK** — a direct source edit outside a deliver context is
+blocked (exit 2) until the work is routed through `uap deliver` (or
+`UAP_DELIVER_ACTIVE`/`UAP_DELIVER_BYPASS` is set).
 
-**Strict mode is opt-in** via `UAP_ENFORCE_DELIVERY=block`: a direct source
-edit outside a deliver context is then blocked (exit 2) until the work is routed
-through `uap deliver` or `UAP_DELIVER_BYPASS=1` is set.
+**Advisory mode is opt-out** via `UAP_ENFORCE_DELIVERY=advisory` — it then
+always allows the edit and logs a one-line nudge toward `uap deliver` instead
+of blocking.
 
 Exempt by construction: non-source files; `docs/`, `scripts/`, `policies/`,
 `src/policies/`, test files (deliver protects those itself); and tooling

package/tools/agents/opencode_uap_agent.py

@@ -37,26 +37,35 @@ logger = logging.getLogger(__name__)
 # --------------------------------------------------------------------------- #
 
 
-def _make_opencode_config(api_endpoint: str) -> dict:
+def _make_opencode_config(
+    api_endpoint: str, model_ref: str = "llama.cpp/qwen35-a3b-iq4xs"
+) -> dict:
+    # model_ref is the harbor `-m` value, e.g. "llama.cpp/qwen3.6-a3b-iq4xs".
+    # Derive the provider + model id so opencode registers EXACTLY the model
+    # harbor selects (avoids ProviderModelNotFoundError when the local model
+    # changes, e.g. Qwen3.5 -> Qwen3.6).
+    provider_id, sep, model_id = model_ref.partition("/")
+    if not sep:
+        provider_id, model_id = "llama.cpp", model_ref
     return {
         "$schema": "https://opencode.ai/config.json",
         "provider": {
-            "llama.cpp": {
+            provider_id: {
                 "npm": "@ai-sdk/openai-compatible",
-                "name": "llama-server (local Qwen3.5)",
+                "name": "llama-server (local)",
                 "options": {
                     "baseURL": api_endpoint,
-                    "apiKey": "sk-qwen35b",
+                    "apiKey": "sk-local",
                 },
                 "models": {
-                    "qwen35-a3b-iq4xs": {
-                        "name": "Qwen3.5 35B A3B (IQ4_XS)",
+                    model_id: {
+                        "name": model_id,
                         "limit": {"context": 262144, "output": 81920},
                     }
                 },
             }
         },
-        "model": "llama.cpp/qwen35-a3b-iq4xs",
+        "model": f"{provider_id}/{model_id}",
     }
 
 
@@ -998,7 +1007,10 @@ class OpenCodeBaseline(BaseInstalledAgent):
         if version:
             variables["version"] = version
         variables["opencode_config"] = json.dumps(
-            _make_opencode_config(self._api_endpoint), indent=2
+            _make_opencode_config(
+                self._api_endpoint, self.model_name or "llama.cpp/qwen35-a3b-iq4xs"
+            ),
+            indent=2,
         )
         variables["api_endpoint"] = self._api_endpoint
         return variables
@@ -1081,11 +1093,20 @@ class OpenCodeUAP(BaseInstalledAgent):
         version = self.version()
         if version:
             variables["version"] = version
-        # Layer 1: opencode.json points to proxy at localhost:11435
-        # The proxy forwards to the real LLM endpoint and injects tool_choice="required"
-        proxy_endpoint = "http://127.0.0.1:11435/v1"
+        # Modern local models (Qwen3.6+) emit native OpenAI tool calls; the
+        # deprecated GBNF-forcing tool-choice proxy (Layer 1) breaks them. Point
+        # opencode at the model endpoint directly by default; set
+        # UAP_BENCH_PROXY=1 to restore the qwen3.5-era proxy path (port 11435).
+        endpoint = (
+            "http://127.0.0.1:11435/v1"
+            if os.environ.get("UAP_BENCH_PROXY") == "1"
+            else self._api_endpoint
+        )
         variables["opencode_config"] = json.dumps(
-            _make_opencode_config(proxy_endpoint), indent=2
+            _make_opencode_config(
+                endpoint, self.model_name or "llama.cpp/qwen35-a3b-iq4xs"
+            ),
+            indent=2,
         )
         variables["api_endpoint"] = self._api_endpoint
         # NOTE: CLAUDE.md is now built dynamically per-task in create_run_agent_commands