Wire AWS Bedrock into research-run routing (main-process)

docs/CHANGELOG.md

@@ -7,6 +7,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+
+- **AWS Bedrock is now routable for research runs (main-process).** Registered
+  the Amazon Nova models (`us.amazon.nova-micro-v1:0`, `-lite-`, `-pro-`) with
+  real prices, and wired the model→provider router and the standard-pipeline
+  dispatch so a stage model set to a Nova id (e.g. `AI_REASONING_MODEL`) runs
+  through the Bedrock `converse` API and prices correctly through the cost gate.
+  Bedrock routing is refused inside a supervised MCP worker — AWS credentials
+  are deliberately never inherited by workers, and a regression test pins that
+  boundary. (Azure Foundry routing is tracked as a follow-up: its model id is a
+  user deployment name, which needs a small deployment-pricing mechanism.)
+
 ## [1.37.2] - 2026-07-22
 
 ### Added

examples/deploy/aws-bedrock/README.md

@@ -52,13 +52,21 @@ primr keys test bedrock     # free, auth-only: "authenticated; N foundation mode
 ```
 
 This validates the deployment (credentials, region, and visible foundation
-models). Note that full research-pipeline routing through Bedrock is **not yet
-wired**: the model-name→provider router has no Bedrock branch, so a Bedrock
-model id set via `AI_REASONING_MODEL` (e.g. `us.amazon.nova-lite-v1:0`) is not
-recognized as a Bedrock model — it falls through to the first-party xAI path and
-the run cannot even be priced. To route utility-tier inference through a custom
-endpoint today, use the OpenAI-compatible gateway seam (`LOCAL_LLM_BASE_URL` /
-`LOCAL_LLM_API_KEY`).
+models). To route a pipeline stage through Bedrock, set the stage model env var
+to a registered Nova model id (the `us.` cross-region inference-profile form):
+
+```bash
+AI_REASONING_MODEL=us.amazon.nova-pro-v1:0   # reasoning/analysis stage
+AI_REPORT_MODEL=us.amazon.nova-pro-v1:0      # section-writing stage
+AI_FAST_MODEL=us.amazon.nova-lite-v1:0       # scraping/utility stage
+```
+
+The registered ids — `us.amazon.nova-micro-v1:0`, `us.amazon.nova-lite-v1:0`,
+`us.amazon.nova-pro-v1:0` — carry real prices, so the mandatory cost estimate
+prices them correctly, and routing dispatches them to the Bedrock `converse`
+API. **Main-process only:** Bedrock routing is refused inside a supervised MCP
+worker, which deliberately does not carry AWS credentials. For other
+Bedrock-hosted models, use their inference-profile id via the same env vars.
 
 ## 4. Clean up
 

src/primr/ai/llm.py

@@ -199,7 +199,13 @@ def llm(
     # than fall through to the Gemini code path below (Gemini API rejects
     # unknown model names with 404). The xAI branch above stays separate
     # because grok_llm carries xAI-specific session-token bookkeeping.
-    if config is not None and config.provider in ("openai", "anthropic", "ollama"):
+    if config is not None and config.provider in (
+        "openai",
+        "anthropic",
+        "ollama",
+        "bedrock",
+        "foundry",
+    ):
         from primr.ai.routing import get_provider_for_model
 
         log_chat_interaction(prompt, f"Model: {model_name} ({config.provider} dispatch)")

src/primr/ai/routing.py

@@ -369,6 +369,19 @@ def get_provider_for_model(model_name: str) -> Provider:
         return _get_anthropic_provider()
     if provider_name == "ollama":
         return _get_ollama_provider()
+    if provider_name in ("bedrock", "foundry"):
+        # Bedrock/Foundry credentials live only in the main process; a
+        # supervised worker deliberately never inherits AWS/Azure secrets
+        # (config/env.py). Refuse rather than fall through to a broken call.
+        if os.environ.get("PRIMR_SUPERVISED_WORKER") == "1":
+            raise ValueError(
+                f"Model {model_name!r} uses the {provider_name!r} provider, which "
+                "is main-process only. Supervised workers do not carry AWS/Azure "
+                "credentials by design; run this route in the main process."
+            )
+        from primr.ai.providers.registry import get_registered_provider_for_model
+
+        return get_registered_provider_for_model(model_name)
 
     raise ValueError(
         f"Model {model_name!r} has provider {provider_name!r} which has no "

src/primr/ai/stage_routing.py

@@ -617,6 +617,14 @@ def _provider_configured(config: ModelConfig) -> bool:
         return bool(os.getenv("OPENAI_API_KEY"))
     if provider == "anthropic":
         return bool(os.getenv("ANTHROPIC_API_KEY"))
+    if provider == "bedrock":
+        return bool(
+            os.getenv("AWS_BEARER_TOKEN_BEDROCK")
+            or os.getenv("AWS_ACCESS_KEY_ID")
+            or os.getenv("AWS_PROFILE")
+        )
+    if provider == "foundry":
+        return bool(os.getenv("AZURE_OPENAI_API_KEY"))
     return provider == "ollama"
 
 

src/primr/config/model_registry.py

@@ -902,6 +902,54 @@ class ModelRegistry:
         supports_multimodal=False,
     )
 
+    # =========================================================================
+    # AMAZON BEDROCK - Amazon Nova family via the boto3 converse API.
+    # Model ids are cross-region inference-profile ids (the `us.` prefix is
+    # required for on-demand invocation). Prices verified July 2026 against the
+    # AWS Bedrock pricing page. Cached-input (Bedrock prompt caching, ~75% off)
+    # is left unset so the estimate gate assumes no cache discount (conservative).
+    # Bedrock routing is MAIN-PROCESS ONLY — see providers/registry.py; the
+    # supervised-worker credential boundary (config/env.py) is not crossed.
+    # =========================================================================
+    BEDROCK_NOVA_MICRO = ModelConfig(
+        name="us.amazon.nova-micro-v1:0",
+        display_name="Amazon Nova Micro (Bedrock)",
+        provider="bedrock",
+        cost_per_1m_input_tokens=0.035,
+        cost_per_1m_output_tokens=0.14,
+        max_input_tokens=128_000,
+        max_output_tokens=5_120,
+        supports_thinking=False,
+        supports_tools=True,
+        supports_multimodal=False,
+    )
+
+    BEDROCK_NOVA_LITE = ModelConfig(
+        name="us.amazon.nova-lite-v1:0",
+        display_name="Amazon Nova Lite (Bedrock)",
+        provider="bedrock",
+        cost_per_1m_input_tokens=0.06,
+        cost_per_1m_output_tokens=0.24,
+        max_input_tokens=300_000,
+        max_output_tokens=5_120,
+        supports_thinking=False,
+        supports_tools=True,
+        supports_multimodal=True,
+    )
+
+    BEDROCK_NOVA_PRO = ModelConfig(
+        name="us.amazon.nova-pro-v1:0",
+        display_name="Amazon Nova Pro (Bedrock)",
+        provider="bedrock",
+        cost_per_1m_input_tokens=0.80,
+        cost_per_1m_output_tokens=3.20,
+        max_input_tokens=300_000,
+        max_output_tokens=5_120,
+        supports_thinking=False,
+        supports_tools=True,
+        supports_multimodal=True,
+    )
+
     # =========================================================================
     # DEEP RESEARCH AGENT - Autonomous research producing 12+ page reports
     # This is a SEPARATE API (Interactions API), not generate_content

src/primr/config/models.py

@@ -281,6 +281,10 @@ class PrimrModels:
         ModelRegistry.OLLAMA_LLAMA4_SCOUT.name: ModelRegistry.OLLAMA_LLAMA4_SCOUT,
         ModelRegistry.OLLAMA_GLM_4_6.name: ModelRegistry.OLLAMA_GLM_4_6,
         ModelRegistry.OLLAMA_PHI4_14B.name: ModelRegistry.OLLAMA_PHI4_14B,
+        # Amazon Bedrock (Nova family, main-process routing only)
+        ModelRegistry.BEDROCK_NOVA_MICRO.name: ModelRegistry.BEDROCK_NOVA_MICRO,
+        ModelRegistry.BEDROCK_NOVA_LITE.name: ModelRegistry.BEDROCK_NOVA_LITE,
+        ModelRegistry.BEDROCK_NOVA_PRO.name: ModelRegistry.BEDROCK_NOVA_PRO,
     }
 
     @classmethod

tests/test_ai/test_bedrock_routing.py

@@ -0,0 +1,67 @@
+"""Amazon Bedrock is routable in the main process only, and never crosses the
+supervised-worker credential boundary.
+
+Bedrock (Nova) models are registered with real prices so the mandatory cost
+gate can price them. Routing reaches ``BedrockProvider`` through the registry
+builder. Crucially, a supervised worker must refuse Bedrock routing (AWS
+secrets are deliberately stripped from worker environments), and the env
+allowlist must never be widened to let them through.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from primr.config.env import is_supervised_worker_env_allowed
+from primr.config.models import PrimrModels
+
+BEDROCK_MODELS = [
+    "us.amazon.nova-micro-v1:0",
+    "us.amazon.nova-lite-v1:0",
+    "us.amazon.nova-pro-v1:0",
+]
+
+
+@pytest.mark.parametrize("model", BEDROCK_MODELS)
+def test_bedrock_model_is_registered_and_priced(model: str) -> None:
+    config = PrimrModels.get_model_config(model)
+    assert config is not None, f"{model} must be in ALL_MODELS"
+    assert config.provider == "bedrock"
+    assert config.cost_per_1m_input_tokens > 0
+    assert config.cost_per_1m_output_tokens > 0
+    # The mandatory pre-run estimate must not KeyError on a Bedrock model.
+    assert PrimrModels.calculate_cost(model, 100_000, 20_000) > 0
+
+
+def test_bedrock_routes_to_registry_provider_in_main_process(monkeypatch) -> None:
+    monkeypatch.delenv("PRIMR_SUPERVISED_WORKER", raising=False)
+    sentinel = object()
+    import primr.ai.providers.registry as registry
+
+    monkeypatch.setattr(registry, "get_registered_provider_for_model", lambda name: sentinel)
+    from primr.ai.routing import get_provider_for_model
+
+    assert get_provider_for_model("us.amazon.nova-micro-v1:0") is sentinel
+
+
+def test_bedrock_routing_refused_in_supervised_worker(monkeypatch) -> None:
+    monkeypatch.setenv("PRIMR_SUPERVISED_WORKER", "1")
+    from primr.ai.routing import get_provider_for_model
+
+    with pytest.raises(ValueError, match="main-process only"):
+        get_provider_for_model("us.amazon.nova-micro-v1:0")
+
+
+@pytest.mark.parametrize(
+    "secret",
+    [
+        "AWS_BEARER_TOKEN_BEDROCK",
+        "AWS_SECRET_ACCESS_KEY",
+        "AWS_ACCESS_KEY_ID",
+        "AZURE_OPENAI_API_KEY",
+    ],
+)
+def test_cloud_infra_secrets_never_enter_a_worker(secret: str) -> None:
+    # Regression pin: nobody may "fix" Bedrock/Foundry worker routing by
+    # widening the allowlist. These stay stripped from supervised workers.
+    assert is_supervised_worker_env_allowed(secret) is False