feat(cdk): switch to the global Bedrock geo, with doctor and --model validation

agent/README.md

@@ -119,11 +119,11 @@ The `run.sh` script overrides the container's default CMD to run `python /app/sr
 | `AWS_SECRET_ACCESS_KEY` | Conditional† | | Explicit keys, if you are not using CLI-based resolution |
 | `AWS_SESSION_TOKEN` | No | | For temporary credentials |
 | `AWS_PROFILE` | No | | Profile for `aws configure export-credentials` in `run.sh`, or default profile when using the `~/.aws` mount fallback |
-| `ANTHROPIC_MODEL` | No | `us.anthropic.claude-opus-5` | Bedrock **inference profile** ID for `InvokeModel` (see [inference profiles](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-use.html)). Must be the `us.`-prefixed profile ID, not a bare foundation-model ID — see [Model configuration](../docs/guides/DEVELOPER_GUIDE.md#model-configuration) |
+| `ANTHROPIC_MODEL` | No | `global.anthropic.claude-opus-5` | Bedrock **inference profile** ID for `InvokeModel` (see [inference profiles](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-use.html)). Must be a geo-prefixed profile ID matching the deployment's `bedrockGeoRegion`, not a bare foundation-model ID — see [Model configuration](../docs/guides/DEVELOPER_GUIDE.md#model-configuration) |
 | `MAX_TURNS` | No | `100` | Max agent turns before stopping |
 | `MAX_BUDGET_USD` | No | | **Local batch only** (shell env when running `entrypoint.py` directly). Range 0.01–100; agent stops when the budget is reached. For deployed AgentCore **server** mode and production tasks, set **`max_budget_usd`** on task creation (REST API, CLI `--max-budget`, or Blueprint default); the orchestrator sends it in the `/invocations` JSON body — server mode does not read `MAX_BUDGET_USD` from the environment. |
 | `DRY_RUN` | No | | Set to `1` to validate config and print the prompt without running the agent |
-| `ANTHROPIC_DEFAULT_HAIKU_MODEL` | No | `us.anthropic.claude-haiku-4-5-20251001-v1:0` | Bedrock **inference profile** ID for the small/fast auxiliary model — the pre-flight safety check and WebFetch summarization (see below). Set by the CDK stack (`cdk/src/stacks/agent.ts` (the runtime environment block)); the `us.` prefix is required |
+| `ANTHROPIC_DEFAULT_HAIKU_MODEL` | No | `global.anthropic.claude-haiku-4-5-20251001-v1:0` | Bedrock **inference profile** ID for the small/fast auxiliary model — the pre-flight safety check and WebFetch summarization (see below). Set by the CDK stack (`cdk/src/stacks/agent.ts` (the runtime environment block)), which derives the prefix from `bedrockGeoRegion` — a geo prefix is required |
 | `NUDGES_TABLE_NAME` | No | | **Phase 2.** DynamoDB table for mid-task user nudges (`<user_nudge>` XML blocks injected between turns). If unset, the agent runs without nudge support — `nudge_reader.read_pending()` returns `[]` and logs a WARN once. Set automatically by the CDK stack on both AgentCore runtimes. |
 | `JIRA_APP_ACTOR_PROXY_URL` | No | | Resolved per-task from the Jira tenant secret. Forge v2 web-trigger URL used for app-authored Jira comments and transitions. |
 | `JIRA_APP_ACTOR_SHARED_SECRET` | No | | Resolved per-task from the Jira tenant secret. HMAC key for the Forge proxy; redacted from agent diagnostics. |
@@ -133,7 +133,7 @@ The `run.sh` script overrides the container's default CMD to run `python /app/sr
 including non-Jira tasks, so a warm AgentCore process cannot expose one
 tenant's OAuth or Forge credential to the next task.
 
-**Bedrock model access (main model):** Configuring `ANTHROPIC_MODEL` and IAM credentials is not enough. Your AWS account must be able to **invoke** that model in Amazon Bedrock: follow [Request access to models](https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html) (Marketplace permissions on first use, Anthropic first-time use where required, valid payment method for Marketplace-backed models). Always use an inference profile ID such as `us.anthropic.claude-opus-5`: a bare foundation-model ID cannot be invoked with on-demand throughput and Bedrock rejects it with `ValidationException`. IAM must also grant the model — see [Model configuration](../docs/guides/DEVELOPER_GUIDE.md#model-configuration) for the full layering. If the CLI stops with a message that the model is not available on your Bedrock deployment, fix model access in the console or switch `ANTHROPIC_MODEL` to an entitled profile, then retry.
+**Bedrock model access (main model):** Configuring `ANTHROPIC_MODEL` and IAM credentials is not enough. Your AWS account must be able to **invoke** that model in Amazon Bedrock: follow [Request access to models](https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html) (Marketplace permissions on first use, Anthropic first-time use where required, valid payment method for Marketplace-backed models). Always use an inference profile ID such as `global.anthropic.claude-opus-5`: a bare foundation-model ID cannot be invoked with on-demand throughput and Bedrock rejects it with `ValidationException`. IAM must also grant the model — see [Model configuration](../docs/guides/DEVELOPER_GUIDE.md#model-configuration) for the full layering. If the CLI stops with a message that the model is not available on your Bedrock deployment, fix model access in the console or switch `ANTHROPIC_MODEL` to an entitled profile, then retry.
 
 **Pre-flight check model**: Claude Code runs a quick safety verification using a small Haiku model before executing each tool command. On Bedrock, the default Haiku model ID may not be enabled in your account, causing the check to time out with *"Pre-flight check is taking longer than expected"* warnings. The agent sets `ANTHROPIC_DEFAULT_HAIKU_MODEL` to a known-available Bedrock Haiku model ID to avoid this. If you see pre-flight timeout warnings, verify that this model is enabled in your Bedrock model access settings.
 
@@ -145,8 +145,8 @@ tenant's OAuth or Forge credential to the next task.
 # Dry run — validate config, fetch issue, print assembled prompt, then exit
 DRY_RUN=1 ./agent/run.sh "owner/repo" 42
 
-# Run with a specific model (overrides the us.anthropic.claude-opus-5 default).
-# Must be a `us.`-prefixed inference profile that IAM grants — see Model configuration.
+# Run with a specific model (overrides the global.anthropic.claude-opus-5 default).
+# Must be a geo-prefixed inference profile that IAM grants — see Model configuration.
 ANTHROPIC_MODEL="us.anthropic.claude-sonnet-4-6" ./agent/run.sh "owner/repo" 42
 
 # Limit agent to 50 turns
@@ -431,7 +431,7 @@ docker images bgagent-local --format "{{.Size}}"
 
 ```
 agent/
-├── Dockerfile           Python 3.13 + Node.js 20 + Claude Code CLI + git + gh + mise (default platform linux/arm64)
+├── Dockerfile           Python 3.13 + Node.js 24 + Claude Code CLI + git + gh + mise (default platform linux/arm64)
 ├── .dockerignore
 ├── pyproject.toml       App dependencies (claude-agent-sdk, FastAPI, boto3, OpenTelemetry distro, MCP, cedarpy, …)
 ├── uv.lock              Locked deps for reproducible `uv sync` in the image

agent/scripts/diagnostics/test_sdk_smoke.py

@@ -26,12 +26,24 @@ async def smoke_test():
     # Ensure required env vars
     os.environ.setdefault("CLAUDE_CODE_USE_BEDROCK", "1")
     region = os.environ.get("AWS_REGION", "")
-    model = os.environ.get("ANTHROPIC_MODEL", "us.anthropic.claude-sonnet-4-6")
+    # No hardcoded fallback: this file drifted to a model and geography the platform
+    # no longer defaults to, so the diagnostic silently tested something other than
+    # what runs. Require the caller to state it — a diagnostic that quietly probes the
+    # wrong model is worse than one that refuses.
+    model = os.environ.get("ANTHROPIC_MODEL", "")
 
     if not region:
         print("ERROR: AWS_REGION not set", file=sys.stderr)
         sys.exit(1)
 
+    if not model:
+        print(
+            "ERROR: ANTHROPIC_MODEL not set. Pass the geo-prefixed inference-profile "
+            "id the deployment uses, e.g. ANTHROPIC_MODEL=global.anthropic.claude-opus-5",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+
     print(f"Region:  {region}")
     print(f"Model:   {model}")
     print(f"Python:  {sys.version}")
@@ -108,17 +120,19 @@ def on_stderr(line: str):
     print(f"Duration: {elapsed:.1f}s")
     print(f"Counts:   {counts}")
 
-    if counts["assistant"] > 0 and counts["result"] > 0:
-        print("\nPASS — SDK yields messages. Issue is specific to the")
-        print("       server threading context, not SDK/CLI/Bedrock.")
+    ok = counts["assistant"] > 0 and counts["result"] > 0
+    if ok:
+        print("\nPASS — SDK yields messages for this model in this Region.")
+        print("       Rules OUT the SDK/CLI/Bedrock path. It does not identify the")
+        print("       cause of any other failure — threading is one candidate, not a")
+        print("       conclusion this test can reach.")
     elif counts["system"] > 0 and counts["assistant"] == 0:
         print("\nFAIL — Got init but zero AssistantMessages.")
         print("       Same symptom as production. Issue is SDK/CLI level,")
         print("       NOT threading. Check:")
         print("         1. CLI stderr output above for errors")
         print("         2. Bedrock model availability / permissions")
         print("         3. SDK ↔ CLI version compatibility")
-        print("         SDK: claude-agent-sdk==0.1.43")
         try:
             import importlib.metadata
 
@@ -129,9 +143,25 @@ def on_stderr(line: str):
     elif counts["system"] == 0:
         print("\nFAIL — Zero messages at all. CLI subprocess may not start.")
         print("       Check: is claude-code installed? Run: claude --version")
+    else:
+        # Reachable: assistant>0 but result==0 (the model spoke, the run never
+        # terminated cleanly). Without this branch the script printed NO verdict at
+        # all and still exited 1 — a diagnostic that fails silently is the thing this
+        # file exists to rule out.
+        print("\nINDETERMINATE — Got AssistantMessages but no ResultMessage.")
+        print("       The model responded, so the SDK/CLI/Bedrock path works, but the")
+        print("       run did not terminate normally. Treat as a FAILURE of this")
+        print("       diagnostic, not evidence about the model. Check the CLI stderr")
+        print("       above for a mid-stream error, and whether the run was killed")
+        print("       (timeout, OOM, cancelled) before it could emit its result.")
     if errors:
         print(f"\nErrors: {errors}")
 
+    # Exit code must match the verdict. It printed FAIL and exited 0, so any caller
+    # that checked the status — CI, a script, a person using `&&` — read a failure as
+    # success. A diagnostic that lies in its exit code is worse than no diagnostic.
+    return ok
+
 
 if __name__ == "__main__":
-    asyncio.run(smoke_test())
+    sys.exit(0 if asyncio.run(smoke_test()) else 1)

agent/src/config.py

@@ -694,13 +694,18 @@ def build_config(
     resolved_github_token = github_token or resolve_github_token()
     resolved_aws_region = aws_region or os.environ.get("AWS_REGION", "")
     resolved_anthropic_model = anthropic_model or os.environ.get(
-        "ANTHROPIC_MODEL", "us.anthropic.claude-opus-5"
+        "ANTHROPIC_MODEL", "global.anthropic.claude-opus-5"
     )
     # Small/fast auxiliary model (WebFetch summarization etc.). Falls back to the
     # deployed ANTHROPIC_DEFAULT_HAIKU_MODEL env, then the platform default. Must
-    # be an inference-profile id (us.*), not a bare model id (see runner).
+    # be a geo-prefixed inference-profile id, not a bare model id (see runner).
+    #
+    # This fallback is a SEPARATE value from the env var the stack injects, which
+    # derives its prefix from the bedrockGeoRegion context key. Only a run with no
+    # env set at all reaches this literal, so the two must be moved together or a
+    # local run silently uses a different geography than a deployed one.
     resolved_haiku_model = haiku_model or os.environ.get(
-        "ANTHROPIC_DEFAULT_HAIKU_MODEL", "us.anthropic.claude-haiku-4-5-20251001-v1:0"
+        "ANTHROPIC_DEFAULT_HAIKU_MODEL", "global.anthropic.claude-haiku-4-5-20251001-v1:0"
     )
 
     # Resolve the workflow id (the create-task boundary already pinned it; local

agent/src/models.py

@@ -154,12 +154,17 @@ class TaskConfig(BaseModel):
     task_description: str = ""
     github_token: str = ""
     aws_region: str
-    anthropic_model: str = "us.anthropic.claude-opus-5"
+    anthropic_model: str = "global.anthropic.claude-opus-5"
     # The "small/fast" model Claude Code uses for auxiliary work (e.g. WebFetch
-    # page summarization). Must be a cross-region INFERENCE-PROFILE id (``us.``
+    # page summarization). Must be a cross-region INFERENCE-PROFILE id (geo
     # prefix), not a bare foundation-model id — Claude 4.x cannot be invoked
     # on-demand by bare id on Bedrock. Threaded to ANTHROPIC_DEFAULT_HAIKU_MODEL.
-    haiku_model: str = "us.anthropic.claude-haiku-4-5-20251001-v1:0"
+    #
+    # A deployed task never reaches this default: the stack injects both model env
+    # vars from its resolved bedrockGeoRegion. It applies to direct construction
+    # (tests, local runs), so it must name the same geography as the deployment or
+    # those paths silently exercise a different one.
+    haiku_model: str = "global.anthropic.claude-haiku-4-5-20251001-v1:0"
     dry_run: bool = False
     max_turns: int = 10
     max_budget_usd: float | None = None

agent/tests/test_config.py

@@ -52,16 +52,18 @@ def test_default_workflow_when_omitted(self):
         assert config.is_pr_workflow is False
 
     def test_haiku_model_defaults_to_inference_profile(self, monkeypatch):
-        # No override, no env → platform default, which must be a us.* inference
-        # profile (Claude 4.x can't be invoked on-demand by bare model id).
+        # No override, no env → platform default, which must be a GEO-PREFIXED
+        # inference profile (Claude 4.x can't be invoked on-demand by bare model
+        # id). The geography must match the deployment's bedrockGeoRegion, or a run
+        # with no env set calls a profile the IAM grant does not cover.
         monkeypatch.delenv("ANTHROPIC_DEFAULT_HAIKU_MODEL", raising=False)
         config = build_config(
             repo_url="owner/repo",
             task_description="fix bug",
             github_token="ghp_test123",
             aws_region="us-east-1",
         )
-        assert config.haiku_model == "us.anthropic.claude-haiku-4-5-20251001-v1:0"
+        assert config.haiku_model == "global.anthropic.claude-haiku-4-5-20251001-v1:0"
 
     def test_haiku_model_resolves_from_env(self, monkeypatch):
         # The deployed ANTHROPIC_DEFAULT_HAIKU_MODEL (set by agent.ts) flows

agent/tests/test_runner.py

@@ -426,9 +426,14 @@ def test_haiku_model_env_is_set_from_config(self, monkeypatch):
         )
 
     def test_config_default_haiku_model_is_an_inference_profile(self):
-        # The platform default (no override) must be a us.* profile, never a bare
-        # foundation-model id — the whole point of the fix.
-        assert _config().haiku_model.startswith("us.")
+        # The platform default (no override) must be a GEO-PREFIXED profile, never a
+        # bare foundation-model id — Claude 4.x cannot be invoked on-demand by bare
+        # id. Asserted as "has a geo prefix" rather than "starts with us.": the
+        # geography is a deploy-time choice (bedrockGeoRegion), so pinning one here
+        # would fail the moment the platform default moves, which tells us nothing
+        # about the property that matters.
+        geos = ("global.", "us.", "us-gov.", "eu.", "apac.", "jp.", "au.")
+        assert _config().haiku_model.startswith(geos)
 
 
 class TestClaudeCliVersionProbe:

agent/tests/test_server.py

@@ -21,10 +21,30 @@
 
 @pytest.fixture(autouse=True)
 def reset_server_state():
+    """Reset the pipeline registry, joining any thread still running on the way out.
+
+    `/run` and `/invocations` answer while the pipeline thread is only just starting,
+    so that thread usually looks up `server.run_task` AFTER the test body has
+    returned. Clearing the registry without joining orphans it: the stubs are then
+    undone, and the thread goes on to run a REAL pipeline — task-state writes, a
+    heartbeat, a clone — inside whichever test happens to be running next. That is a
+    cross-test AWS call arriving from a thread nothing is waiting on, and it stays
+    invisible until some later test asserts that no AWS seam was touched. Joining
+    here is what keeps a pipeline thread from outliving the test that spawned it.
+
+    The live threads are read from `threading.enumerate()` rather than from
+    `_active_threads`, because a test may substitute that registry with one that
+    refuses to be read; it stays clearable, which is all this fixture asks of it.
+    Joining the pipeline thread also retires its heartbeat, which the pipeline stops
+    on its way out.
+    """
     server._background_pipeline_failed = False
     with server._threads_lock:
         server._active_threads.clear()
     yield
+    for thread in threading.enumerate():
+        if thread.name.startswith("pipeline-") and thread.is_alive():
+            thread.join(timeout=10)
     server._background_pipeline_failed = False
     with server._threads_lock:
         server._active_threads.clear()
@@ -1797,6 +1817,13 @@ def test_wire_contract_is_exactly_the_documented_key_set(self):
             "agent_session_role_arn": "AGENT_SESSION_ROLE_ARN",
             "aws_sdk_ua_app_id": "AWS_SDK_UA_APP_ID",
             "anthropic_default_haiku_model": "ANTHROPIC_DEFAULT_HAIKU_MODEL",
+            # The MAIN model. Absent until now, and silently: the AgentCore runtime
+            # and the ECS task definitions inject it into their own env, so only
+            # MicroVM depends on this transport — and its fallback in config.py is a
+            # ``global.``-prefixed literal, so a missing value was not an error but a
+            # wrong model that the IAM grant does not cover on any non-``global``
+            # deployment, surfacing as AccessDenied at turn 0.
+            "anthropic_model": "ANTHROPIC_MODEL",
         }
 
     def test_required_subset_is_exactly_the_four_run_blocking_keys(self):

cdk/cdk.json

@@ -14,5 +14,8 @@
       "yarn.lock",
       "node_modules"
     ]
+  },
+  "context": {
+    "bedrockGeoRegion": "global"
   }
 }