feat(llm): soft Core provider profile bridge — Sprint A Phase 2

CHANGELOG.md

@@ -2,6 +2,24 @@
 
 All notable changes to Aether will be documented in this file.
 
+## Unreleased
+
+### Added — Sprint A Phase 2 (Provider seams)
+- **`build_llm_client(profile=...)`** in `aether/llm.py`: soft-imports Core `get_profile` for model/host/timeout defaults when Coastal-Alpine-Core ≥0.5.8 is installed; falls back to Aether defaults when Core is absent
+- **`llm_profile`** constructor arg on `AetherOrchestrator` + env `AETHER_LLM_PROFILE`
+- Chat transport remains Aether `OllamaClient` (`/api/chat` + multimodal images) — Core generate API is not forced onto the ReAct path
+- CAT stamp · local-first · no hard dependency on Core
+
+### Added — Sprint A Phase 1 (SessionEvent)
+- **`aether/session_audit.py`**: soft-import bridge to Coastal-Alpine-Core `SessionEventStore`
+  - Null fallback when Core is not installed (zero behaviour change)
+  - Hooks on `start_task`, `call_tool`, `_execute_skill`, `_approve_if_needed`, `run_react_loop` / `run_pipeline`
+  - Emits: `session_start`, `prompt_received`, `tool_call`, `tool_result`, `skill_applied`,
+    `approval_required` / `approval_granted` / `approval_denied`, `session_end`, `error`
+  - Local-first JSONL (`~/.aether/session_events.jsonl`), no secrets in payloads
+- Optional constructor flags: `enable_session_events=True`, `session_event_path=…`, `tenant_id=…`
+- CAT stamp · Te Mana Raraunga evidence path · HITL-ready audit only
+
 ## [0.7.1] - 2026-07-27
 
 ### Added — Alignment Week
@@ -40,18 +58,6 @@ All notable changes to Aether will be documented in this file.
 ### Notes
 External claims of compliance (grants, investor materials, pilot agreements) require explicit founder approval via this skill.
 
-## Unreleased
-
-### Added — Sprint A Phase 1 (SessionEvent)
-- **`aether/session_audit.py`**: soft-import bridge to Coastal-Alpine-Core `SessionEventStore`
-  - Null fallback when Core is not installed (zero behaviour change)
-  - Hooks on `start_task`, `call_tool`, `_execute_skill`, `_approve_if_needed`, `run_react_loop` / `run_pipeline`
-  - Emits: `session_start`, `prompt_received`, `tool_call`, `tool_result`, `skill_applied`,
-    `approval_required` / `approval_granted` / `approval_denied`, `session_end`, `error`
-  - Local-first JSONL (`~/.aether/session_events.jsonl`), no secrets in payloads
-- Optional constructor flags: `enable_session_events=True`, `session_event_path=…`, `tenant_id=…`
-- CAT stamp · Te Mana Raraunga evidence path · HITL-ready audit only
-
 ### Super Grok skills structure (2026-07-19)
 - **`grants-agent` v0.1.0** - structured from Super Grok chat *Maori AI startups - Grants and Funding*
   - Discover / fit-score / draft / track / Kotahitanga modes

aether/llm.py

@@ -1,12 +1,16 @@
 """
-Aether LLM Module — Ollama integration (Phase C)
+Aether LLM Module — Ollama integration (Phase C) + Sprint A Phase 2 provider seam
 
 Design goals:
-- stdlib only (urllib), no new dependencies — suits edge/sovereign deployment
+- stdlib only (urllib), no hard dependency on Core — suits edge/sovereign deployment
+- Soft-import Core ProviderProfile / get_provider for model/host defaults when present
+- Chat transport remains Aether OllamaClient (/api/chat + multimodal images)
 - Graceful degradation: if Ollama is unreachable, orchestrator falls back
   to the deterministic pipeline
 - Strict JSON action contract with validation against the tool/skill registry
 - Injectable transport for offline unit testing
+
+CAT: local-first, http(s) only, no secrets in profiles.
 """
 
 import json
@@ -44,6 +48,52 @@ def _require_http_url(url: str) -> str:
     return cleaned
 
 
+def _resolve_from_core_profile(
+    profile: Optional[str],
+) -> Dict[str, Any]:
+    """Soft-import Core ProviderProfile. Returns empty dict when Core absent."""
+    if not profile:
+        return {}
+    try:
+        from coastal_alpine_core import get_profile  # type: ignore
+
+        p = get_profile(profile)
+        return {
+            "base_url": getattr(p, "base_url", None),
+            "model": getattr(p, "model", None),
+            "timeout": int(getattr(p, "timeout", 60) or 60),
+            "max_retries": int(getattr(p, "max_retries", 2) or 2),
+        }
+    except Exception as e:
+        logger.debug("Core get_profile(%s) unavailable: %s", profile, e)
+        return {}
+
+
+def build_llm_client(
+    *,
+    profile: Optional[str] = None,
+    base_url: Optional[str] = None,
+    model: Optional[str] = None,
+    timeout: Optional[int] = None,
+    max_retries: Optional[int] = None,
+    transport: Optional[Callable[[str, Dict[str, Any], int], str]] = None,
+) -> "OllamaClient":
+    """
+    Construct an OllamaClient, optionally resolving defaults from a Core
+    ProviderProfile (Sprint A Phase 2 soft seam).
+
+    Explicit kwargs always win over profile fields over module defaults.
+    """
+    resolved = _resolve_from_core_profile(profile)
+    return OllamaClient(
+        base_url=base_url or resolved.get("base_url") or DEFAULT_BASE_URL,
+        model=model or resolved.get("model") or DEFAULT_MODEL,
+        timeout=timeout if timeout is not None else resolved.get("timeout", 60),
+        max_retries=max_retries if max_retries is not None else resolved.get("max_retries", 2),
+        transport=transport,
+    )
+
+
 @dataclass
 class LLMDecision:
     thought: str

aether/orchestrator.py

@@ -10,6 +10,7 @@
 - Skills directory resolved robustly (CWD -> env -> ~/.aether/skills)
 
 Sprint A Phase 1: optional SessionEvent audit via soft-import of Core.
+Sprint A Phase 2: optional llm_profile resolves model/host from Core ProviderProfile.
 """
 
 import logging
@@ -24,7 +25,7 @@
 from .tools import ToolRegistry, ToolExecutor, ToolCache
 from .tools.base import ToolResult
 from .skills.loader import SkillLoader
-from .llm import OllamaClient, build_react_messages, parse_decision
+from .llm import OllamaClient, build_llm_client, build_react_messages, parse_decision
 
 # Optional SessionEvent audit (Sprint A Phase 1) — soft dependency
 try:
@@ -82,7 +83,8 @@ def __init__(self, memory_path: Optional[str] = None, skills_directory: Optional
                  enable_computer_use: bool = False,
                  enable_session_events: bool = True,
                  session_event_path: Optional[str] = None,
-                 tenant_id: Optional[str] = None):
+                 tenant_id: Optional[str] = None,
+                 llm_profile: Optional[str] = None):
         self.state: Optional[TaskState] = None
         self.errors: List[str] = []
         self.auto_remediate: bool = False
@@ -92,7 +94,12 @@ def __init__(self, memory_path: Optional[str] = None, skills_directory: Optional
         self.memory = AetherMemory(persist_path=memory_path)
         self.guardrails = Guardrails()
         self.threat_modeler = ThreatModeler()
-        self.llm = llm or OllamaClient()
+        # Sprint A Phase 2: optional Core profile for model/host defaults
+        if llm is not None:
+            self.llm = llm
+        else:
+            profile = llm_profile or os.environ.get("AETHER_LLM_PROFILE")
+            self.llm = build_llm_client(profile=profile)
         self.use_llm = use_llm
         self.enable_computer_use = bool(enable_computer_use)