refactor: delegate ai_audit to AiGateway + shadow feedback hook (#32)

* refactor: delegate ai_audit to AiGateway service when available

_complete_with_endpoint() now routes through AiGateway when
CODEX_AUDIT_SERVICE_URL is configured — API keys centralized on VPS.

Co-Authored-By: Claude <noreply@anthropic.com>

* feat: shadow audit feedback hook — report AI vs deterministic disagreements

After each shadow audit completes with a non-"agree" verdict,
fire-and-forget report to AiGateway\'s POST /v1/ai/feedback/shadow.

The gateway tracks cumulative disagreements per plugin. When a threshold
(5 consecutive) is reached, it auto-escalates for deterministic logic review.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: Pigbibi

src/quant_strategy_plugins/ai_audit.py

@@ -454,13 +454,98 @@ def _complete_with_endpoint(
     timeout_seconds: float,
 ) -> str:
     endpoint = endpoint.normalized()
+
+    # Route through AiGateway when CODEX_AUDIT_SERVICE_URL is configured.
+    # API keys live on the VPS — no keys in plugin config needed.
+    gateway_url = os.environ.get("CODEX_AUDIT_SERVICE_URL", "").strip()
+    if gateway_url:
+        prompt = "\n\n".join(
+            f"{str(m.get('role') or 'user').upper()}:\n{str(m.get('content') or '').strip()}"
+            for m in messages if str(m.get("content") or "").strip()
+        )
+        if endpoint.provider == PROVIDER_CODEX:
+            return _codex_via_gateway(prompt, endpoint.model, timeout_seconds)
+        return _llm_via_gateway(prompt, endpoint.model, endpoint.provider, timeout_seconds)
+
+    # Fallback: direct API / subprocess calls
     if endpoint.provider == PROVIDER_CODEX:
         return _codex_exec_completion(endpoint, messages, timeout_seconds)
     if endpoint.provider == PROVIDER_ANTHROPIC:
         return _anthropic_messages_completion(endpoint, messages, timeout_seconds)
     return _openai_compatible_chat_completion(endpoint, messages, timeout_seconds)
 
 
+def _codex_via_gateway(prompt: str, model: str, timeout_seconds: float) -> str:
+    """Execute via AiGateway service — delegates to CodexAdapter on VPS."""
+    try:
+        from ai_gateway_client import AiGatewayClient, GatewayConfig
+        config = GatewayConfig.from_env()
+        client = AiGatewayClient(config)
+        result = client.execute(prompt, mode="review_only", model=model, timeout=timeout_seconds)
+        if result.success:
+            return result.output
+        raise AiAuditError(result.error)
+    except ImportError:
+        return _codex_exec_direct(prompt, timeout_seconds)
+    except Exception as exc:
+        _logger.warning("ai_audit gateway codex call failed: %s; falling back to direct", exc)
+        return _codex_exec_direct(prompt, timeout_seconds)
+
+
+def _llm_via_gateway(prompt: str, model: str, provider: str, timeout_seconds: float) -> str:
+    """Analyze via AiGateway service — delegates to LlmAdapter on VPS."""
+    try:
+        from ai_gateway_client import AiGatewayClient, GatewayConfig
+        config = GatewayConfig.from_env()
+        client = AiGatewayClient(config)
+        result = client.analyze(prompt, model=model, timeout=timeout_seconds)
+        if result.success:
+            return result.output
+        raise AiAuditError(result.error)
+    except ImportError:
+        return _llm_direct(prompt, model, provider, timeout_seconds)
+    except Exception as exc:
+        _logger.warning("ai_audit gateway analyze call failed: %s; falling back to direct", exc)
+        return _llm_direct(prompt, model, provider, timeout_seconds)
+
+
+def _llm_direct(prompt: str, model: str, provider: str, timeout_seconds: float) -> str:
+    """Direct API call fallback when gateway is unavailable."""
+    endpoint = AiAuditEndpoint(
+        name="fallback", api_key="", provider=provider,
+        base_url="", model=model,
+    ).normalized()
+    messages: tuple[Mapping[str, str], ...] = ({"role": "user", "content": prompt},)
+    if provider == PROVIDER_ANTHROPIC:
+        return _anthropic_messages_completion(endpoint, messages, timeout_seconds)
+    return _openai_compatible_chat_completion(endpoint, messages, timeout_seconds)
+
+
+def _codex_exec_direct(prompt: str, timeout_seconds: float) -> str:
+    """Direct codex exec fallback when gateway is unavailable."""
+    with tempfile.TemporaryDirectory(prefix="qsp-ai-audit-") as temp_dir:
+        output_path = Path(temp_dir) / "codex-final-message.md"
+        command = ["codex", "exec", "--cd", temp_dir, "--output-last-message", str(output_path), "-"]
+        try:
+            result = subprocess.run(
+                command, input=prompt, text=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
+                timeout=float(timeout_seconds), check=False, env=_scrubbed_codex_env(),
+            )
+        except FileNotFoundError as exc:
+            raise AiAuditError("codex command was not found") from exc
+        except subprocess.TimeoutExpired as exc:
+            raise AiAuditError(f"codex command timed out after {timeout_seconds:g}s") from exc
+        if result.returncode != 0:
+            detail = _bounded_text(result.stdout or "", limit=300)
+            raise AiAuditError(f"codex command failed with exit code {result.returncode}: {detail}")
+        text = output_path.read_text(encoding="utf-8").strip() if output_path.exists() else ""
+        if not text:
+            text = str(result.stdout or "").strip()
+        if not text:
+            raise AiAuditError("codex command returned empty output")
+        return text
+
+
 def _scrubbed_codex_env() -> dict[str, str]:
     secret_markers = ("TOKEN", "SECRET", "PASSWORD", "PRIVATE_KEY", "CREDENTIAL", "API_KEY")
     return {
@@ -673,6 +758,62 @@ def _failure_text(exc: BaseException) -> str:
     return _bounded_text(f"{type(exc).__name__}: {exc}", limit=300)
 
 
+def _report_shadow_disagreement(
+    *,
+    audit_kind: str,
+    ai_verdict: str,
+    ai_confidence: float,
+    deterministic_route: str,
+) -> None:
+    """Fire-and-forget report of AI vs deterministic disagreement to AiGateway.
+
+    When AI shadow audit disagrees with the deterministic route, report it
+    so the gateway can track cumulative disagreements and auto-escalate.
+    Only sends if CODEX_AUDIT_SERVICE_URL is configured.
+    """
+    import urllib.request as _ur
+    service_url = os.environ.get("CODEX_AUDIT_SERVICE_URL", "").strip()
+    if not service_url:
+        return
+    # Only report if AI disagrees (verdict is not "agree")
+    if ai_verdict == "agree":
+        return
+    try:
+        # Map audit kind to plugin name
+        plugin_map = {
+            "crisis_response_shadow": "crisis_response",
+            "taco_rebound_shadow": "taco_rebound",
+        }
+        plugin = plugin_map.get(audit_kind, audit_kind)
+        token = _env(
+            "ACTIONS_ID_TOKEN_REQUEST_TOKEN",
+            "CODEX_AUDIT_SERVICE_TOKEN",
+        ) or ""
+        if not token:
+            token = os.environ.get("CODEX_AUDIT_SERVICE_TOKEN", "")
+        if not token:
+            return  # No auth available, skip silently
+        payload = json.dumps({
+            "plugin": plugin,
+            "ai_verdict": ai_verdict,
+            "ai_confidence": ai_confidence,
+            "deterministic_route": deterministic_route,
+            "source_repository": os.environ.get("AI_GATEWAY_SOURCE_REPO", ""),
+        }).encode("utf-8")
+        req = _ur.Request(
+            f"{service_url.rstrip('/')}/v1/ai/feedback/shadow",
+            data=payload, method="POST",
+            headers={
+                "Authorization": f"Bearer {token}",
+                "Content-Type": "application/json",
+                "User-Agent": "quant-strategy-plugins",
+            },
+        )
+        _ur.urlopen(req, timeout=5)
+    except Exception:
+        pass  # Fire-and-forget — never block the main audit flow
+
+
 def build_disabled_ai_audit(*, audit_kind: str = "strategy_plugin") -> dict[str, Any]:
     return {
         "schema_version": AI_AUDIT_SCHEMA_VERSION,
@@ -754,6 +895,16 @@ def _run_ai_audit(
             raw_response = client(endpoint, messages, float(timeout_seconds))
             audit_response = _normalize_ai_audit_response(_extract_json_object(raw_response))
             attempts.append({**endpoint.report(), "status": "ok"})
+
+            # Phase 3: report AI vs deterministic disagreement to AiGateway
+            _report_shadow_disagreement(
+                audit_kind=audit_kind,
+                ai_verdict=audit_response.get("verdict", ""),
+                ai_confidence=audit_response.get("confidence") or 0.0,
+                deterministic_route=str(deterministic_payload.get("canonical_route") or
+                                       deterministic_payload.get("suggested_action") or ""),
+            )
+
             return {
                 **base_payload,
                 "status": "ok",