steerable-agent-runtime 0.6.8__tar.gz → 0.6.10__tar.gz

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: steerable-agent-runtime
-Version: 0.6.8
+Version: 0.6.10
 Summary: Steerable agent runtime: LLM, tool, storage, and transport adapters.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "steerable-agent-runtime"
-version = "0.6.8"
+version = "0.6.10"
 description = "Steerable agent runtime: LLM, tool, storage, and transport adapters."
 readme = "README.md"
 requires-python = ">=3.10"

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/src/steerable_agent_runtime/loop.py

@@ -877,6 +877,13 @@ class CoreLoop:
             else None
         )
         wrap_up = False
+        # Terminal that triggered the current wrap-up when a budget wall
+        # (maxRounds) offered narration. The wrap-up summary is an interim
+        # report, not task completion: the no-tools terminal below restores
+        # this decision so the host's auto-continue sees budget_exhausted and
+        # resumes with a fresh budget instead of mistaking the summary for a
+        # finished task.
+        wrap_up_terminal: CompletionDecision | None = None
         wrap_up_tool_rounds_used = 0
         idle_cut_count = 0
 
@@ -1051,7 +1058,10 @@ class CoreLoop:
                 )
                 yield LoopEvent(
                     "budget_exhausted",
-                    {"kind": "rounds", "rounds": self._config.max_rounds},
+                    # "budget" not "kind": the sidecar wraps loop events as
+                    # {"kind": <event kind>, **data} — a data["kind"] would
+                    # clobber the envelope kind hosts dispatch on.
+                    {"budget": "rounds", "rounds": self._config.max_rounds},
                 )
                 narrate = await self._offer_narration(
                     "", round_index, had_tool_calls=True, ctx=ctx,
@@ -1060,6 +1070,7 @@ class CoreLoop:
                 if narrate is not None:
                     prompt, reason = narrate
                     completion_redos += 1
+                    wrap_up_terminal = decision
                     mark_wrap_up()
                     manager.append_fragment(NarrationRequest(prompt))
                     yield LoopEvent(
@@ -1362,7 +1373,7 @@ class CoreLoop:
                             if exhausted:
                                 yield LoopEvent(
                                     "budget_exhausted",
-                                    {"kind": "tokens", "used": budget_state.tokens_used},
+                                    {"budget": "tokens", "used": budget_state.tokens_used},
                                 )
                                 decision = CompletionDecision(
                                     status="budget_exhausted", reason="token budget exceeded"
@@ -1586,7 +1597,13 @@ class CoreLoop:
 
             # ── decide: no tool calls → terminal (unless a hook vetoes) ──
             if not tool_calls:
-                if content.strip():
+                if wrap_up_terminal is not None:
+                    # Budget-wall wrap-up: the narration summary is an interim
+                    # report, not a finished task. Restore the wall's terminal
+                    # (budget_exhausted) so the host auto-continues with a
+                    # fresh budget instead of stopping on a half-done task.
+                    decision = wrap_up_terminal
+                elif content.strip():
                     decision = CompletionDecision(
                         status="completed",
                         reason="assistant produced final response with no pending tools",

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/src/steerable_agent_runtime/model_resolve.py

@@ -19,7 +19,7 @@ silent defaults are exactly how the drift stayed invisible.
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import Literal
+from typing import Any, Literal
 
 from .model_catalog import MODEL_ENTRIES, PROVIDER_ENTRIES
 
@@ -173,6 +173,82 @@ def provider_endpoint(provider: str) -> ProviderEndpoint | None:
     return ProviderEndpoint(provider=provider, api_base_url=api, env_vars=env)
 
 
+#: Sidecar factory kinds the host settings picker maps onto. Catalog ids
+#: not in this table (and whose models are not all Anthropic-format) speak
+#: OpenAI-compatible chat completions.
+_WIRE_KIND_BY_PROVIDER: dict[str, str] = {
+    "anthropic": "anthropic",
+    "google": "google",
+    "google-vertex": "google",
+    "google-vertex-anthropic": "anthropic",
+    "xai": "openai-responses",
+}
+
+#: models.dev omits ``api`` for first-party SDKs. Host pickers still need a
+#: fillable URL so the user is not dropped into a blank field for OpenAI /
+#: Anthropic / Gemini / Groq / xAI.
+_SDK_DEFAULT_BASE_URLS: dict[str, str] = {
+    "openai": "https://api.openai.com/v1",
+    "anthropic": "https://api.anthropic.com",
+    "google": "https://generativelanguage.googleapis.com",
+    "groq": "https://api.groq.com/openai/v1",
+    "xai": "https://api.x.ai/v1",
+}
+
+
+def _wire_kind(provider_id: str, api: str | None, formats: set[str]) -> str:
+    mapped = _WIRE_KIND_BY_PROVIDER.get(provider_id)
+    if mapped:
+        return mapped
+    if api and "/anthropic" in api.lower():
+        return "anthropic"
+    if formats and formats <= {"anthropic"}:
+        return "anthropic"
+    return "openai_compat"
+
+
+def describe_catalog_providers() -> list[dict[str, Any]]:
+    """Host settings pickers: one row per catalogued serving provider.
+
+    Same service pattern as ``compat.describe`` / ``presets.describe``: the
+    bundled catalog is the source of truth for default URLs and chat-capable
+    model ids. ``wireKind`` is the sidecar factory kind. ``apiBaseUrl`` is
+    None when neither the catalog nor the first-party SDK default knows an
+    endpoint — the host then collects one by hand. Embedding / TTS / video
+    ids (``tool_format == "none"``) are omitted so the model list is what a
+    chat agent can actually call.
+    """
+    models_by_provider: dict[str, list[str]] = {}
+    formats_by_provider: dict[str, set[str]] = {}
+    for key, row in MODEL_ENTRIES.items():
+        provider_id, _, model_id = key.partition("/")
+        if not provider_id or not model_id:
+            continue
+        _context, _modalities, tool_format, _reasoning = row
+        if tool_format == "none":
+            continue
+        models_by_provider.setdefault(provider_id, []).append(model_id)
+        formats_by_provider.setdefault(provider_id, set()).add(tool_format)
+
+    described: list[dict[str, Any]] = []
+    for provider_id in sorted(PROVIDER_ENTRIES):
+        api, env_vars = PROVIDER_ENTRIES[provider_id]
+        models = models_by_provider.get(provider_id, [])
+        if not models:
+            continue
+        formats = formats_by_provider.get(provider_id, set())
+        described.append(
+            {
+                "id": provider_id,
+                "apiBaseUrl": api or _SDK_DEFAULT_BASE_URLS.get(provider_id),
+                "envVars": list(env_vars),
+                "wireKind": _wire_kind(provider_id, api, formats),
+                "models": sorted(models),
+            }
+        )
+    return described
+
+
 def catalog_provider_for_base_url(base_url: str | None) -> str | None:
     """The catalog provider serving ``base_url``, matched by URL host.
 

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/src/steerable_agent_runtime.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: steerable-agent-runtime
-Version: 0.6.8
+Version: 0.6.10
 Summary: Steerable agent runtime: LLM, tool, storage, and transport adapters.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/tests/test_antihallucination.py

@@ -17,6 +17,7 @@ from steerable_agent_runtime import (
     AntiHallucinationConfig,
     AntiHallucinationHooks,
     CoreLoop,
+    LoopConfig,
     LoopEvent,
     RouterToolExecutor,
     ToolRouter,
@@ -521,6 +522,42 @@ async def test_narration_round_on_empty_terminal() -> None:
     )
 
 
+@pytest.mark.asyncio
+async def test_narration_at_max_rounds_wall_stays_budget_exhausted() -> None:
+    """The maxRounds wall interrupts work; the narration summary it triggers
+    is an interim report, not task completion. The terminal must keep
+    budget_exhausted so the host's auto-continue resumes with a fresh budget
+    (a completed status would silently stop a half-done task)."""
+    provider = make_provider(
+        [
+            {"content": "", "tool_calls": [tc("get_data", call_id="c1")]},
+            {"content": "", "tool_calls": [tc("get_data", {"source": "a"}, "c2")]},
+            {"content": "", "tool_calls": [tc("get_data", {"source": "b"}, "c3")]},
+            # narration wrap-up round (tools withheld): interim summary
+            {"content": "阶段性总结：已取数 3 次。"},
+        ]
+    )
+    hooks = AntiHallucinationHooks(provider, AntiHallucinationConfig())
+    loop = CoreLoop(
+        provider,
+        RouterToolExecutor(make_router_with_tool()),
+        LoopConfig(max_rounds=3),
+        hooks=hooks,
+    )
+    events = await collect(loop.run([LLMMessage.text_of("user", "取数")], tools=TOOLS))
+
+    final = final_completion(events)
+    assert final["status"] == "budget_exhausted"
+    assert "maxRounds" in final["reason"]
+    # The interim summary still rides the completion and the transcript.
+    assert final["textLength"] == len("阶段性总结：已取数 3 次。")
+    narrations = [
+        e for e in events
+        if e.kind == "hook_action" and e.data.get("action") == "narrate"
+    ]
+    assert len(narrations) == 1
+
+
 @pytest.mark.asyncio
 async def test_narration_second_chance_on_empty_wrap_up() -> None:
     """Real-data replay finding: the wrap-up round itself can come back empty

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/tests/test_gateway_catalog.py

@@ -179,6 +179,12 @@ async def test_fetch_caches_within_ttl(fake_httpx) -> None:
     assert second.entries == first.entries
 
 
+async def test_fetch_ttl_zero_bypasses_cache(fake_httpx) -> None:
+    await fetch_gateway_models("http://gw.test/v1")
+    await fetch_gateway_models("http://gw.test/v1", ttl_sec=0)
+    assert fake_httpx.calls == 2
+
+
 async def test_fetch_serves_stale_on_refresh_failure(fake_httpx) -> None:
     await fetch_gateway_models("http://gw.test/v1")
     fake_httpx.payload = ConnectionError("gateway down")

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/tests/test_loop.py

@@ -410,7 +410,7 @@ async def test_token_budget_exhausted() -> None:
     events = await collect(loop.run([LLMMessage.text_of("user", "hi")]))
     decision = final_completion(events)
     assert decision["status"] == "budget_exhausted"
-    assert any(e.kind == "budget_exhausted" and e.data["kind"] == "tokens" for e in events)
+    assert any(e.kind == "budget_exhausted" and e.data["budget"] == "tokens" for e in events)
 
 
 @pytest.mark.asyncio

{steerable_agent_runtime-0.6.8 → steerable_agent_runtime-0.6.10}/tests/test_model_resolve.py

@@ -84,3 +84,44 @@ def test_provider_endpoint_none_for_first_party_default() -> None:
 
 def test_provider_endpoint_unknown_provider() -> None:
     assert provider_endpoint("no-such-provider") is None
+
+
+def test_describe_catalog_providers_covers_featured_vendors() -> None:
+    from steerable_agent_runtime.model_resolve import describe_catalog_providers
+
+    described = describe_catalog_providers()
+    by_id = {row["id"]: row for row in described}
+    assert "deepseek" in by_id
+    assert by_id["deepseek"]["apiBaseUrl"] == "https://api.deepseek.com"
+    assert by_id["deepseek"]["wireKind"] == "openai_compat"
+    assert "deepseek-v4-flash" in by_id["deepseek"]["models"]
+
+    assert by_id["anthropic"]["wireKind"] == "anthropic"
+    assert by_id["anthropic"]["apiBaseUrl"] == "https://api.anthropic.com"
+    assert any(m.startswith("claude-") for m in by_id["anthropic"]["models"])
+
+    assert by_id["google"]["wireKind"] == "google"
+    assert by_id["google"]["apiBaseUrl"] == "https://generativelanguage.googleapis.com"
+    assert "gemini-embedding-001" not in by_id["google"]["models"]
+
+    assert by_id["xai"]["wireKind"] == "openai-responses"
+    assert by_id["minimax"]["wireKind"] == "anthropic"
+    assert by_id["minimax"]["apiBaseUrl"] and "/anthropic" in by_id["minimax"]["apiBaseUrl"]
+
+    # First-party openai has no catalog api field; the describe row still
+    # fills the SDK default so a host picker can auto-set the URL.
+    assert by_id["openai"]["apiBaseUrl"] == "https://api.openai.com/v1"
+
+
+def test_describe_catalog_providers_skips_providers_without_chat_models() -> None:
+    from steerable_agent_runtime.model_resolve import describe_catalog_providers
+
+    described = describe_catalog_providers()
+    for row in described:
+        assert row["models"], row["id"]
+        assert row["wireKind"] in {
+            "openai_compat",
+            "anthropic",
+            "google",
+            "openai-responses",
+        }