fix(bedrock): decode base64 data URL to bytes for Converse image source

[liuhao1024]

What does this PR do?

_convert_content_to_converse() in agent/bedrock_adapter.py passes the base64-encoded ASCII string from a data URL directly as ImageBlock.source.bytes. Bedrock's Converse API expects raw bytes — boto3 base64-encodes them at the wire layer. The result is double-encoding, and Bedrock rejects every image with:

ValidationException: Failed to sanitize image: Invalid or unsupported image format

Root Cause

Line 482 of agent/bedrock_adapter.py:

"source": {"bytes": data},   # BUG: 'data' is still base64-encoded string

data is the portion after the comma in data:image/png;base64,iVBORw0KGgo... — a base64 ASCII string. It needs to be decoded to raw bytes first.

Related Issue

Fixes #33317

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• See commit messages for detailed changes

How to Test

1. Run pytest tests/ -q — all tests should pass
2. Verify the specific scenario described above is resolved

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 26.4.1

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture and workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A

Code Intelligence

• Analyzed: agent/bedrock_adapter.py:_convert_content_to_converse (called from convert_messages_to_converse, used by Bedrock transport for all multimodal messages)
• Blast radius: LOW — single function, single code path (data URL images only), no impact on remote URL or text content
• Related patterns: other adapters (anthropic, openai) handle data URLs differently — anthropic uses base64.b64decode() (correct), openai passes string (correct for its API). Bedrock is the only adapter that needs raw bytes.

Fixes #33317

[alt-glitch]

Duplicate fix — PR #28085 already addresses the same bedrock image double-encoding bug (base64 string passed as source.bytes instead of decoded raw bytes). See also issue #33317 which was triaged as a duplicate of #28085.

[JiaDe-Wu]

Solid work. We hit the same double-encoding issue and fixed it in #34742. The root cause is identical — base64 string passed to source.bytes instead of decoded raw bytes. boto3 re-encodes at the wire layer. Your fix and ours are the same two-line change. Whichever lands first should close the other.

[liuhao1024]

@alt-glitch @JiaDe-Wu Thanks for the context. I see #28085 and #34742 both address the same base64 double-encoding issue. Since all three PRs (#28085, #34742, #33536) are still open, I'll wait for the maintainer to decide which approach to merge. If either #28085 or #34742 merges first, I'm happy to close this PR as a duplicate.

[teknium1]

Thanks for the focused Bedrock regression fix. The premise is confirmed on current main: agent/bedrock_adapter.py:525 extracts the data-URL payload and agent/bedrock_adapter.py:534 currently forwards that base64 string as image.source.bytes.

Problems

• The proposed decode at agent/bedrock_adapter.py:535 is unguarded and non-strict. The converter accepts any data: URL, so malformed payloads may either be silently converted to junk bytes or raise while building the request. The related open implementation in #28085 uses base64.b64decode(..., validate=True), catches decode errors, and skips empty/malformed image blocks.

Suggested changes

• Use strict decoding, catch binascii.Error/ValueError, and preserve adjacent text when discarding an invalid image block.
• Add a malformed-data-URL regression case alongside the existing conversion test in tests/agent/test_bedrock_adapter.py:332.

This is an automated hermes-sweeper review.

[teknium1]

Please use strict decoding and handle malformed payloads here. This branch accepts every data: URL, while plain b64decode() may accept junk characters or raise for invalid padding. Use validate=True, catch decode errors, and skip empty/malformed image blocks so an invalid attachment does not fail conversion of the whole message.