@miller-tech/uap 1.185.0 → 1.185.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.185.0",
+  "version": "1.185.1",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",
@@ -21,7 +21,7 @@
     "start": "node dist/bin/cli.js",
     "test": "vitest",
     "test:ci": "vitest run",
-    "test:enforcers": "UAP_PROXY_ENV_AUTOLOAD=0 python3 -m unittest tools.agents.tests.test_enforcement_self_protect tools.agents.tests.test_schema_diff_gate tools.agents.tests.test_stream_telemetry tools.agents.tests.test_project_telemetry_events tools.agents.tests.test_workdir_scope_enforcer tools.agents.tests.test_gate_evidence tools.agents.tests.test_expert_review_pr_scope tools.agents.tests.test_path_containment tools.agents.tests.test_path_normalizer_hardened tools.agents.tests.test_sandbox_tool_strip tools.agents.tests.test_proxy_env_loader tools.agents.tests.test_stream_required_tool tools.agents.tests.test_tool_call_wrapper_profiles tools.agents.tests.test_tool_convert_cache tools.agents.tests.test_doubling_break tools.agents.tests.test_error_loop_ignores_correctives tools.agents.tests.test_attractor_detection tools.agents.tests.test_client_disconnect tools.agents.tests.test_confidence_escalation tools.agents.tests.test_coordination_ban tools.agents.tests.test_coordination_early_ban tools.agents.tests.test_cycle_break_exploration tools.agents.tests.test_deferral_break tools.agents.tests.test_deliver_autoroute tools.agents.tests.test_delivery_enforcement_all_langs tools.agents.tests.test_delivery_enforcement_exemptions tools.agents.tests.test_delivery_enforcement_filepath tools.agents.tests.test_delivery_enforcement_web_and_bash tools.agents.tests.test_disconnect_watcher tools.agents.tests.test_empty_maxtokens_recovery tools.agents.tests.test_empty_tool_loop_break tools.agents.tests.test_enforcer_escape_hatches tools.agents.tests.test_error_loop_break tools.agents.tests.test_finalize_suppression tools.agents.tests.test_malformed_unclosed_think tools.agents.tests.test_mandate_beats_recon tools.agents.tests.test_mandate_deliver tools.agents.tests.test_overflow_truncate_count_tokens tools.agents.tests.test_passthrough_oauth tools.agents.tests.test_project_telemetry tools.agents.tests.test_proxy_auth_headers tools.agents.tests.test_prune_preserve_force_write tools.agents.tests.test_recon_deliver_gate tools.agents.tests.test_session_admission tools.agents.tests.test_stream_heartbeat tools.agents.tests.test_stuck_break_reattach tools.agents.tests.test_turn_count_breaker_periodic tools.agents.tests.test_upstream_chokepoint tools.agents.tests.test_vision_passthrough tools.agents.tests.test_worktree_required tools.agents.tests.test_enforcer_suite_coverage tools.agents.tests.test_validate_plan_gate tools.agents.tests.test_validate_plan_inside_project tools.agents.tests.test_anthropic_proxy_streaming tools.agents.tests.test_delivery_enforcement_worktree tools.agents.tests.test_output_token_ceilings tools.agents.tests.test_tool_narrowing_core",
+    "test:enforcers": "UAP_PROXY_ENV_AUTOLOAD=0 python3 -m unittest tools.agents.tests.test_enforcement_self_protect tools.agents.tests.test_schema_diff_gate tools.agents.tests.test_stream_telemetry tools.agents.tests.test_project_telemetry_events tools.agents.tests.test_workdir_scope_enforcer tools.agents.tests.test_gate_evidence tools.agents.tests.test_expert_review_pr_scope tools.agents.tests.test_path_containment tools.agents.tests.test_path_normalizer_hardened tools.agents.tests.test_sandbox_tool_strip tools.agents.tests.test_proxy_env_loader tools.agents.tests.test_stream_required_tool tools.agents.tests.test_tool_call_wrapper_profiles tools.agents.tests.test_tool_convert_cache tools.agents.tests.test_doubling_break tools.agents.tests.test_error_loop_ignores_correctives tools.agents.tests.test_attractor_detection tools.agents.tests.test_client_disconnect tools.agents.tests.test_confidence_escalation tools.agents.tests.test_coordination_ban tools.agents.tests.test_coordination_early_ban tools.agents.tests.test_cycle_break_exploration tools.agents.tests.test_deferral_break tools.agents.tests.test_deliver_autoroute tools.agents.tests.test_delivery_enforcement_all_langs tools.agents.tests.test_delivery_enforcement_exemptions tools.agents.tests.test_delivery_enforcement_filepath tools.agents.tests.test_delivery_enforcement_web_and_bash tools.agents.tests.test_disconnect_watcher tools.agents.tests.test_empty_maxtokens_recovery tools.agents.tests.test_empty_tool_loop_break tools.agents.tests.test_enforcer_escape_hatches tools.agents.tests.test_error_loop_break tools.agents.tests.test_finalize_suppression tools.agents.tests.test_malformed_unclosed_think tools.agents.tests.test_mandate_beats_recon tools.agents.tests.test_mandate_deliver tools.agents.tests.test_overflow_truncate_count_tokens tools.agents.tests.test_passthrough_oauth tools.agents.tests.test_project_telemetry tools.agents.tests.test_proxy_auth_headers tools.agents.tests.test_prune_preserve_force_write tools.agents.tests.test_recon_deliver_gate tools.agents.tests.test_session_admission tools.agents.tests.test_stream_heartbeat tools.agents.tests.test_stuck_break_reattach tools.agents.tests.test_turn_count_breaker_periodic tools.agents.tests.test_upstream_chokepoint tools.agents.tests.test_vision_passthrough tools.agents.tests.test_worktree_required tools.agents.tests.test_enforcer_suite_coverage tools.agents.tests.test_validate_plan_gate tools.agents.tests.test_validate_plan_inside_project tools.agents.tests.test_anthropic_proxy_streaming tools.agents.tests.test_delivery_enforcement_worktree tools.agents.tests.test_output_token_ceilings tools.agents.tests.test_tool_narrowing_core tools.agents.tests.test_models_context_window",
     "test:coverage": "vitest --coverage",
     "bench": "vitest --config vitest.bench.config.ts",
     "lint": "eslint src --ext .ts",

package/tools/agents/scripts/anthropic_proxy.py

@@ -12650,6 +12650,57 @@ def _parse_anthropic_sse_to_message(raw: bytes) -> dict | None:
     }
 
 
+ADVERTISED_MODEL_IDS = (
+    "claude-haiku-4-5-20251001",
+    "claude-sonnet-4-6",
+    "claude-sonnet-5-20250514",
+    "claude-fable-5",
+    "qwen36-35b-a3b-iq4xs",
+)
+
+# Keys OpenAI-compatible clients probe for a model's context window. There is no
+# standard, so emit the common spellings rather than betting on one: hermes reads
+# context_length / context_window / max_context_length / max_model_len / n_ctx
+# (agent/model_metadata.py:_CONTEXT_LENGTH_KEYS), LiteLLM and vLLM prefer
+# max_model_len, LM Studio uses max_context_length. They are all the same number.
+_CONTEXT_WINDOW_KEYS = (
+    "context_length",
+    "context_window",
+    "max_context_length",
+    "max_model_len",
+    "n_ctx",
+)
+
+
+def _model_entry(model_id: str) -> dict:
+    """One /v1/models row, carrying the context window when we know it.
+
+    Advertising this is not cosmetic. A client that cannot discover the window
+    cannot size its own history to it, so it grows unbounded and the FIRST thing
+    that notices is this proxy — which can then only prune blind, after the
+    prompt is already built.
+
+    Live, 2026-08-04: hermes has a context compressor and probes for exactly
+    these keys. We advertised bare {"id", "object"} rows, its model cache held no
+    entry for our model, so the compressor never engaged. It sent 470 messages /
+    219,957 tokens against a 130,048 window (169%), and the proxy CRITICAL PRUNEd
+    290 of them to fit — 61 such events in 18 hours. Raising the window from
+    86,784 to 130,048 had not helped, because the growth was never sized to the
+    window in the first place.
+
+    Only advertised for models this proxy serves LOCALLY. A model that
+    round-trips to api.anthropic.com has its own (much larger) window, and
+    stamping the local llama.cpp figure on it would make clients truncate
+    needlessly — a worse bug than the one being fixed, so when in doubt emit
+    nothing and leave the client on its own defaults.
+    """
+    entry = {"id": model_id, "object": "model"}
+    if PROXY_CONTEXT_WINDOW > 0 and not _should_passthrough_model(model_id):
+        for key in _CONTEXT_WINDOW_KEYS:
+            entry[key] = PROXY_CONTEXT_WINDOW
+    return entry
+
+
 @app.get("/v1/models")
 async def models():
     """Return available model list.
@@ -12666,15 +12717,7 @@ async def models():
     ANTHROPIC_PASSTHROUGH_MODELS=__local_only__ is set, all IDs (including
     the Claude ones below) are served by the local llama.cpp backend.
     """
-    return {
-        "data": [
-            {"id": "claude-haiku-4-5-20251001", "object": "model"},
-            {"id": "claude-sonnet-4-6", "object": "model"},
-            {"id": "claude-sonnet-5-20250514", "object": "model"},
-            {"id": "claude-fable-5", "object": "model"},
-            {"id": "qwen36-35b-a3b-iq4xs", "object": "model"},
-        ]
-    }
+    return {"data": [_model_entry(mid) for mid in ADVERTISED_MODEL_IDS]}
 
 
 @app.get("/health")

package/tools/agents/tests/test_models_context_window.py

@@ -0,0 +1,84 @@
+"""/v1/models must advertise the context window for locally-served models.
+
+Regression (hermes, 2026-08-04): the endpoint returned bare {"id","object"}
+rows. Hermes HAS a context compressor and probes for context_length /
+context_window / max_context_length / max_model_len / n_ctx, found none, and its
+model cache held no entry for our model — so the compressor never engaged. It
+sent 470 messages / 219,957 tokens against a 130,048 window (169%) and the proxy
+CRITICAL PRUNEd 290 of them; 61 such events in 18 hours. Raising the window from
+86,784 to 130,048 had not helped, because the growth was never sized to the
+window at all.
+
+A client that cannot discover the window cannot size its history to it.
+"""
+import importlib.util
+import os
+import unittest
+from pathlib import Path
+
+proxy_path = Path(__file__).resolve().parents[3] / "tools" / "agents" / "scripts" / "anthropic_proxy.py"
+
+
+def load_proxy(window="130048", passthrough=None):
+    """Import a fresh proxy module under the given env (constants bind at import)."""
+    os.environ["PROXY_CONTEXT_WINDOW"] = window
+    if passthrough is None:
+        os.environ.pop("ANTHROPIC_PASSTHROUGH_MODELS", None)
+    else:
+        os.environ["ANTHROPIC_PASSTHROUGH_MODELS"] = passthrough
+    spec = importlib.util.spec_from_file_location("anthropic_proxy_ctx", proxy_path)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+class ModelsAdvertiseContextWindowTest(unittest.TestCase):
+    def tearDown(self):
+        os.environ.pop("PROXY_CONTEXT_WINDOW", None)
+        os.environ.pop("ANTHROPIC_PASSTHROUGH_MODELS", None)
+
+    def test_local_model_carries_the_window(self):
+        ap = load_proxy()
+        e = ap._model_entry("qwen36-35b-a3b-iq4xs")
+        self.assertEqual(e["context_length"], 130048)
+
+    def test_every_probed_key_is_emitted(self):
+        # There is no standard key. Clients disagree, so emit the common
+        # spellings — missing the one a client happens to read is the same
+        # failure as advertising nothing.
+        ap = load_proxy()
+        e = ap._model_entry("qwen36-35b-a3b-iq4xs")
+        for key in ("context_length", "context_window", "max_context_length",
+                    "max_model_len", "n_ctx"):
+            self.assertEqual(e.get(key), 130048, key)
+
+    def test_passthrough_models_do_not_get_the_local_window(self):
+        # A model that round-trips to api.anthropic.com has a much larger window.
+        # Stamping the local llama.cpp figure on it would make clients truncate
+        # needlessly — worse than the bug being fixed.
+        ap = load_proxy(passthrough=None)  # default patterns: Claude passes through
+        for mid in ("claude-sonnet-4-6", "claude-haiku-4-5-20251001"):
+            self.assertNotIn("context_length", ap._model_entry(mid), mid)
+
+    def test_local_only_sentinel_means_every_id_is_local(self):
+        ap = load_proxy(passthrough="__local_only__")
+        for mid in ap.ADVERTISED_MODEL_IDS:
+            self.assertEqual(ap._model_entry(mid).get("context_length"), 130048, mid)
+
+    def test_unset_window_advertises_nothing(self):
+        # Better to say nothing than to assert a wrong number.
+        ap = load_proxy(window="0")
+        self.assertEqual(
+            ap._model_entry("qwen36-35b-a3b-iq4xs"), {"id": "qwen36-35b-a3b-iq4xs", "object": "model"}
+        )
+
+    def test_entry_always_keeps_the_openai_shape(self):
+        ap = load_proxy()
+        for mid in ap.ADVERTISED_MODEL_IDS:
+            e = ap._model_entry(mid)
+            self.assertEqual(e["id"], mid)
+            self.assertEqual(e["object"], "model")
+
+
+if __name__ == "__main__":
+    unittest.main()