fix(compressor): strip _multimodal tool-result images after compress()

[srojk34]

Problem

browser_vision and computer_use return screenshots as a _multimodal envelope:

{"_multimodal": True, "content": [...base64...], "text_summary": "..."}

Each screenshot can be several megabytes. _strip_historical_media only stripped user-attached images, so _multimodal tool results accumulated across turns and inflated every subsequent API request payload.

Providers enforce limits on byte size, not token count. Sessions with a handful of browser_vision calls produced HTTP 413 errors that survived every compression pass even when the token estimate was well within the model's context limit (issue #47339: 101 → 13 messages, still 413).

Fix

Add _strip_all_multimodal_tool_results(), called inside compress() after _sanitize_tool_pairs() and before _strip_historical_media().

It replaces every _multimodal envelope with its text_summary string:

{"_multimodal": True, "content": [...3MB base64...], "text_summary": "login page with a form"}
→ "[screenshot removed] login page with a form"

Why it's safe to strip all of them (including the most recent): compress() only runs between turns, never mid-turn. By the time compression triggers, the model always has an assistant response for every tool result in the conversation — the base64 has already been processed and retaining it serves no purpose.

Tests

6 new tests in TestStripAllMultimodalToolResults:

• single envelope stripped (no anchor exception)
• all envelopes stripped across multiple calls
• text_summary used as replacement text
• fallback placeholder when text_summary absent
• non-_multimodal tool results unchanged
• user/assistant messages unchanged

All 102 tests pass.

Fixes #47339.

[teknium1]

Thanks for the clear root-cause analysis and focused regression coverage. This is already implemented on current main; this is an automated hermes-sweeper review.

• Commit 122e5bc0373e8df65db751d5d820445aec275a95 (fix(agent): retry 413 after stripping vision payloads (#47339)) shipped in v2026.7.1.
• agent/conversation_loop.py:3422 strips retained vision payloads and retries after compression cannot reduce a 413 request further.
• run_agent.py:5194 removes image_url/input_image parts from tool messages while retaining their textual context.
• tests/run_agent/test_413_compression.py:230 covers the retained browser_vision payload case and verifies the retry omits the base64 data while preserving the screenshot text.
• The linked PR fix(agent): evict images before compression on 413 recovery (#47339) #47441 was closed as part of the same 413 Request Entity Too Large: context compression doesn't evict image/vision payloads — fails even at ~44k tokens on Copilot #47339 consolidation cluster; the shipping commit records that it was cherry-picked from fix(agent): retry 413 after stripping vision payloads (#47339) #47397.