fix(context): preserve multimodal content during compression

[qWaitCrypto]

What does this PR do?

Fixes a ContextCompressor bug when conversation history contains structured or multimodal content (e.g. list-valued message content with text/image blocks).

Before this change, ContextCompressor assumed msg["content"] was always a string in several compression paths. That caused two problems:

1. Summary serialization emitted Python reprs for multimodal content instead of readable text, degrading summary quality
2. merge-summary-into-tail raised TypeError: can only concatenate str (not "list") to str when the first tail message had list-valued content

This patch adds a _coerce_content_to_text() helper that normalizes structured content to readable text where text is needed, while preserving multimodal list structure when merging a summary into the tail message.

Manually verified against real multimodal request shapes from OpenAI-style chat completions, OpenAI-style responses, and an Anthropic-compatible messages endpoint.

Related Issue

No existing issue — discovered during code review of the compression paths.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• Added _coerce_content_to_text() in agent/context_compressor.py to normalize structured / multimodal message content into readable text for compression logic.
• Updated compression paths in agent/context_compressor.py to use normalized text for summary serialization, tail token estimation, old tool-result pruning, and system-note appending.
• Fixed merge-summary-into-tail in agent/context_compressor.py so list-valued multimodal content keeps its original structure by prepending a new text block instead of flattening the whole message into a string.
• Added a compatibility fallback for input_image parts that use url instead of image_url.
• Added regression coverage in tests/agent/test_context_compressor.py for:
  • readable multimodal serialization (including input_image.url fallback)
  • pruning non-string tool content
  • merge-summary-into-tail with multimodal content while preserving structure

How to Test

python3 -m pytest tests/agent/test_context_compressor.py -q

41 passed, 41 warnings in 20.94s

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: Windows 11 (WSL2 Ubuntu)

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 or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

[teknium1]

Thanks for the thorough write-up and the tests! The bugs you identified are real — but the same fix landed on main independently before this PR could be merged.

Automated hermes-sweeper review — closing as implemented on main.

• The two helpers that address these crashes — _content_text_for_contains() and _append_text_to_content() — were added in commit 1e8254e59 (fix(agent): guard context compressor against structured message content).
• That commit fixes both the system-note injection path (agent/context_compressor.py ~line 1218) and the summary-merge-into-tail path (~line 1235) using the same safe-prepend/append approach you proposed.
• Regression tests for both crash paths were also added in the same commit to tests/agent/test_context_compressor.py.
• PR fix(agent): handle structured and multimodal content in ContextCompressor #14143, which you cross-referenced, was the salvaged version of this fix — it was closed but its commit is reachable from main.

If you find any remaining edge cases not covered by the current implementation, feel free to open a new focused issue or PR.