feat(tui): vision-aware image paste with non-vision fallback

[wqymi]

Summary

Makes TUI image paste vision-aware, with an end-to-end guarantee that non-vision models never receive image bytes (which cause hallucination), plus a discovery path so the agent knows which vision models it can dispatch to.

Vision-aware paste + fallback

• Unify vision capability: removed the redundant hardcoded supportsImageInput special-casing in transform.ts; capabilities.input.image is now the single source of truth (models.dev already marks mimo-v2.5 / mimo-v2-omni as image-capable). Kept one explicit mimo-auto override (free-tier routing alias absent from models.dev).
• TUI paste routing: vision model → existing base64 attach; non-vision model → spill clipboard image to a temp file and insert an @file reference (matching autocomplete's file-part shape), with a toast. Same fallback when pasting an image-file path. PDF/SVG handling unchanged. New Clipboard.spillImage helper. i18n key across all 7 locales.
• Build-time notice: system.ts injects a <vision-capability> block for non-vision models only — tells them they can't see images and how to proceed.
• Read tool harness: a non-vision model reading an image gets a warning (no base64 attachment); vision models and PDF reads unchanged. Uses Effect.catchDefect so an unresolvable model degrades to non-vision rather than crashing.

Model discovery (so --model is usable)

• actor models verb: new read-only subcommand listing available models (actor models), optionally filtered to vision-capable (actor models --vision), count-capped. Mirrors the subagent_type discoverability fix.
• --model description now points at actor models for valid values.
• Non-vision hints name up to 3 real vision models (from Provider.list(), deterministic) + point to actor models --vision for the rest — in both the system prompt and the Read-tool warning. Zero configured vision models → suggests configuring one or using OCR. The discovery loop closes: read image → warning names a model → actor models --vision → actor run ... --model <real vision model>.

Test Plan

• bun typecheck clean (all 12 turbo tasks pass on pre-push)
• bun test test/tool/actor-models.test.ts test/tool/read-vision-harness.test.ts test/tool/read.test.ts test/cli/tui/clipboard-spill.test.ts → 44 pass, 0 fail
• Non-vision model + paste image → temp file + @file reference + toast; vision model → base64 attach unchanged
• Non-vision model + read image → warning naming a real vision model + actor models --vision; vision model → attachment
• Unresolvable/missing model → degrades to non-vision (no crash), covered by tests
• actor models lists all; actor models --vision filters to image-capable; --limit caps
• Non-vision system prompt contains <vision-capability> with real model names; vision model does not

Plans: docs/compose/plans/2026-07-01-tui-paste-image-vision-fallback.md, docs/compose/plans/2026-07-01-actor-models-discovery.md

[danhduykhan-code]

Summary

Makes TUI image paste vision-aware. When the active model supports vision, images attach as before. When it does not, images degrade to a file-path reference instead of ever putting base64 bytes into a non-vision model's context (which causes hallucination). The guarantee is enforced end-to-end: at the paste boundary, at system-prompt build time, and at the Read tool.

• Unify vision capability (A): removed the redundant hardcoded supportsImageInput special-casing in transform.ts; capabilities.input.image is now the single source of truth (models.dev already marks mimo-v2.5 / mimo-v2-omni as image-capable). Kept one explicit mimo-auto override (free-tier routing alias absent from models.dev).
• TUI paste routing (B): vision model → existing base64 attach; non-vision model → spill clipboard image to a temp file and insert an @file reference (matching autocomplete's file-part shape), with a toast. Same fallback when pasting an image-file path. PDF/SVG handling unchanged. New Clipboard.spillImage helper. i18n key added across all 7 locales.
• Build-time notice (C.1): system.ts injects a <vision-capability> block for non-vision models only, telling them they can't see images — dispatch a vision subagent (actor run ... --model <vision model>) with the image path, or use hexdump for raw binary needs.
• Read tool harness (C.2): a non-vision model reading an image gets a warning (no base64 attachment); vision models and PDF reads are unchanged. Uses Effect.catchDefect so an unresolvable model degrades to non-vision rather than crashing.

Test Plan

• bun typecheck clean (all 12 turbo tasks pass on pre-push)
• bun test test/cli/tui/clipboard-spill.test.ts test/tool/read-vision-harness.test.ts test/tool/read.test.ts → 41 pass, 0 fail
• Non-vision model + paste clipboard image → temp file written, @<file> inserted, toast shown
• Vision model + paste clipboard image → base64 attachment unchanged
• Non-vision model + read on an image → warning, no attachment; vision model → attachment present
• Unresolvable / missing model → degrades to non-vision (no crash), covered by tests
• Non-vision model system prompt contains <vision-capability>; vision model does not

Plan: docs/compose/plans/2026-07-01-tui-paste-image-vision-fallback.md