emrg: read tool returns vision-format image blocks for local images (rant 2026-08-24T14:36:01)

[argszero]

Summary

Rant 2026-08-24T14:36:01: users with a vision-capable model (e.g. deepseek-v4-flash-vision-exp) had no way to let the model see a local image — read on a binary file returned UnicodeDecodeError → "binary file?" error, and bash base64 output is text, not a vision image block. Only clipboard paste (TUI /image) could reach vision format.

Changes

• emrg/tools/read_tool.py: detect image extensions (.png/.jpg/.jpeg/.gif/.webp, case-insensitive) → return a structured JSON ref {"type": "image", "path": ..., "mime": ...} instead of a binary error; MAX_IMAGE_SIZE (256KB) cap with a resize hint (sips/ImageMagick) for oversized images; tool description updated to advertise image support.
• emrg/server/daemon.py: new _tool_content_for_llm() helper wired into all 3 tool-message sites (main tool loop + 2 continue paths). It converts the read-tool image ref into an OpenAI vision content block (data-URL base64) for vision-capable models, degrades to a text placeholder for non-vision models, and passes non-image results through unchanged.
• llm.py: no change needed — tool messages already pass content through as-is, and OpenAI/DeepSeek accept content arrays for tool messages.
• tests: 5 read-tool tests (image ref, mime mapping, size cap, uppercase ext, description) + 4 daemon tests (vision conversion, non-vision degradation, passthrough, missing image). Agent.md count synced (1040 → 1049).

Acceptance

• read /path/to/img.png returns a vision block instead of a binary error
• vision-capable models can answer based on image content
• non-vision models degrade to a text placeholder without error
• existing text reads pass regression

[argszero]

✅ LGTM — cycle 144335+1. Reviewed against rant 2026-08-24T14:36:01:

• read_tool.py returns a structured image ref ({"type":"image",path,mime}) for .png/.jpg/.jpeg/.gif/.webp instead of erroring on binary read; 256KB cap with a concrete resize hint; extension matching is case-insensitive.
• daemon._tool_content_for_llm() correctly degrades: vision model → OpenAI [text, image_url] data-URL block; non-vision model → text placeholder; non-image results passthrough unchanged; missing file → graceful '[Image unavailable]' warning instead of a crash.
• All 3 tool-message construction sites (main loop + 2 resume paths) are wired through the helper.
• Tests cover vision/non-vision/passthrough/missing-image + read_tool mime/size/uppercase/description; Agent.md count synced 1049. CI green (test 1m21s + test-windows 59s, run 32698965107).

[argszero]

✅ LGTM (2/3) — cycle 2026-08-24T15:03, independent re-verification. Head cd2b995 unchanged, MERGEABLE/CLEAN, CI double-green (test 1m21s + test-windows 59s, run 32698965107). Change set matches the rant 2026-08-24T14:36:01 acceptance items (vision-format image refs from ReadTool; daemon-side conversion with non-vision degradation; 3 tool-message sites wired; 9 new tests; Agent.md count synced).

[argszero]

✅ LGTM (3/3) — cycle 2026-08-24T15:10, final independent verification. Head cd2b995 unchanged, MERGEABLE/CLEAN, CI double-green (test 1m21s + test-windows 59s, run 32698965107). Three consecutive approvals from three distinct cycles (145700, 150307, 151015), no requested changes in between. Change set verified against rant 2026-08-24T14:36:01: ReadTool returns vision-format image refs for local images, daemon._tool_content_for_llm converts to OpenAI vision content blocks with non-vision degradation, all 3 tool-message sites wired, 9 new tests, Agent.md count synced. Merging.