fix(bedrock): streaming fallback to Converse API + image base64 decode + bearer token routing

[JiaDe-Wu]

Three fixes for the Bedrock Claude path that share a common theme: the AnthropicBedrock SDK doesn't handle all Bedrock scenarios correctly, and the native Converse API path is the reliable fallback.

1. Streaming fallback (new)

When the AnthropicBedrock SDK's stream accumulator receives a Bedrock error event (throttling, overload, 5xx) before message_start, it raises RuntimeError: Unexpected event order and discards the actual error payload. After 3 retries the session dies.

Fix: detect this specific RuntimeError in the conversation loop and auto-switch api_mode from anthropic_messages to bedrock_converse for the rest of the session. The Converse API handles these error events gracefully via boto3's native error handling.

2. Image base64 decode (fixes #33317)

_convert_content_to_converse() passed the base64 string directly as source.bytes. boto3 re-encodes at the wire layer, resulting in double-encoding. Bedrock rejects with "Failed to sanitize image".

Fix: base64.b64decode(data) before passing to source.bytes.

3. Bearer token routing (fixes #28156 part 1)

Users with AWS_BEARER_TOKEN_BEDROCK were routed through the AnthropicBedrock SDK (which only supports SigV4 signing), causing RuntimeError: could not resolve credentials from session.

Fix: check for bearer token env var in the dual-path routing logic. If present, always use bedrock_converse regardless of model.

Files

• agent/conversation_loop.py -- streaming fallback handler
• agent/bedrock_adapter.py -- base64 decode fix
• hermes_cli/runtime_provider.py -- bearer token routing
• tests/agent/test_bedrock_adapter.py -- 3 new tests

Testing

121 bedrock_adapter tests passing (Python 3.11 + 3.14).

Ref: #33317, #28156, #14737

[alt-glitch]

Significant overlap with existing open PRs:

• Image base64 fix: duplicate of fix(bedrock): support bearer-token auth + raw image bytes for aux vision #28085 and fix(bedrock): decode base64 data URL to bytes for Converse image source #33536 (both open, same two-line fix)
• Bearer token routing: overlaps with fix(bedrock): route Claude through Converse API when bearer token auth #24507 and fix(bedrock): correct api_mode detection for bearer-token auth #26531 (both open, same Converse API routing)
• Streaming fallback: related to fix(anthropic_stream): surface real error payload instead of cryptic "Unexpected event order" #14737 (Anthropic SDK error surface)

This PR bundles all three fixes together. Consider whether to merge this or the individual PRs.

[JiaDe-Wu]

Good callout. The individual PRs (#28085, #33536, #24507, #26531) are all 50-200 commits stale and would need rebasing. This PR bundles the fixes onto current main with tests, so it can merge as-is without conflict resolution. Happy to close if maintainers prefer to rebase the originals instead.

[RamsAI-bot]

For anyone tracking bearer-token Bedrock support: the two PRs that look closest to a complete fix are #24507 (main-loop Converse routing + AWS_BEARER_TOKEN_BEDROCK → AWS_BEARER_TOKEN promotion) and #28085 (auxiliary dual-path Converse client). They were verified to compose cleanly (cherry-pick, 0 conflicts) and together cover both the main loop and auxiliary tasks (title generation / compression; vision still has a separate gap).

Relative to this PR: the routing here omits the token promotion that #24507 carries (which matters for cron jobs / subagents that don't inherit the interactive session's env), and the base64 image-decode fix is also included in #28085. The streaming-overload → Converse fallback in this PR is independently useful for SigV4 hosts.

[RamsAI-bot]

Update — the aux vision gap noted in my comment above is now resolved by #28085's second commit (acd5de2da), verified working against real bearer-token Bedrock. So #24507 + #28085 (in full) now cover the main loop and all aux tasks — title generation, compression, and vision — under AWS_BEARER_TOKEN_BEDROCK.

[teknium1]

Thanks for consolidating the Bedrock reports. The image and bearer-routing premises are both still present on current main, but this bundle needs a focused salvage rather than a direct application.

Problems

• agent/bedrock_adapter.py:91-101 creates boto3 clients without promoting AWS_BEARER_TOKEN_BEDROCK to boto3's AWS_BEARER_TOKEN. This PR routes bearer users to Converse but does not provide that token to the new client path; related PR #24507 contains the missing promotion.
• The new fallback changes agent.api_mode permanently from a string-matched RuntimeError. Current agent/agent_init.py:895-915 initializes _bedrock_guardrail_config only when Converse is selected at startup, so this transition would lose configured guardrails.
• No test exercises the new conversation-loop fallback; the added tests cover only conversion and provider resolution.
• base64.b64decode(data) is non-strict and the fallback still sends arbitrary bytes for malformed input.

Suggested changes

• Add guarded bearer-token promotion in the boto3 client factory with regression coverage.
• Use a tested Converse-transition helper that preserves Bedrock runtime state, including guardrails, and restrict fallback to a confirmed retryable stream failure.
• Strictly validate data URLs and skip malformed image blocks; add a live-path fallback regression test.

Automated hermes-sweeper review.

[teknium1]

base64.b64decode() defaults to non-strict parsing, so malformed payloads can silently decode to junk; the exception fallback then still sends arbitrary bytes to Bedrock. Use validate=True and omit an invalid image block rather than forwarding it.