fix(bedrock): fall back to non-streaming InvokeModel when IAM denies InvokeModelWithResponseStream

[teknium1]

Summary

AWS Bedrock now works with InvokeModel-only IAM policies: when IAM denies bedrock:InvokeModelWithResponseStream, the agent falls back to the non-streaming converse() path instead of looping on AccessDenied.

Root cause: the agent loop hard-prefers streaming (conversation_loop.py:977), and an AWS AccessDeniedException for the streaming action never matched the existing "stream not supported" auto-fallback — so least-privilege IAM users could not use Bedrock at all.

Changes

• agent/bedrock_adapter.py: new is_streaming_access_denied_error() (ClientError code check + wrapped-SDK message match); call_converse_stream() falls back to converse() on denial.
• agent/chat_completion_helpers.py: bedrock_converse streaming branch retries inline via converse() and sets _disable_streaming so subsequent turns skip the doomed stream attempt; the streaming retry block also recognizes the denial for the AnthropicBedrock SDK path (cheap message pre-check before importing bedrock_adapter, avoiding its lazy boto3 install for unrelated providers).
• Both paths print a one-line notice naming the IAM action that restores streaming.

Validation

	Before	After
InvokeModel-only IAM	AccessDenied retry loop, unusable	inline fallback to converse(), session continues non-streaming
Full IAM	streaming	streaming (unchanged)

Targeted tests: tests/agent/test_bedrock_adapter.py + tests/run_agent/test_streaming.py — 162/162 pass, including 7 new tests (detector positives/negatives, inline fallback, ThrottlingException still propagates).

Infographic

[github-actions]

🔎 Lint report: hermes/hermes-fb8b4d98 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10715 on HEAD, 10712 on base (🆕 +3)

🆕 New issues (1):

Rule	Count
invalid-assignment	1

First entries

agent/bedrock_adapter.py:237: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to `<class 'ClientError'>`

✅ Fixed issues: none

Unchanged: 5606 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[liuhao1024]

Verified ✅ — thorough IAM fallback with good coverage across all three call sites.

What was checked:

• Traced the 3 fallback paths: call_converse_stream() in bedrock_adapter.py, _bedrock_call() in chat_completion_helpers.py, and the AnthropicBedrock SDK _call() path. All three correctly detect InvokeModelWithResponseStream denial and fall back to the non-streaming equivalent.
• Verified is_streaming_access_denied_error() detection is appropriately scoped: requires invokemodelwithresponsestream in message and an auth-related error code (AccessDeniedException / UnauthorizedException) or "not authorized" phrasing. A ValidationException mentioning the action name won't trigger the fallback (test confirms).
• Checked the lazy-import guard in _call(): invokemodelwithresponsestream in the lowered error message is checked first, avoiding a boto3 import for unrelated providers' errors. Smart optimization.
• The _disable_streaming = True flag persists for the session — once denied, subsequent turns skip streaming immediately instead of re-failing.
• Confirmed test coverage: detection unit tests (positive, negative, wrapped SDK form), integration test with mock client, error propagation test ensuring non-IAM errors still raise.