[Optimize]: Resolution budgets, crop/bbox coordinate tools, and image-workflow guidance for vision tools #2248

Summary

Upgrade BitFun's built-in vision tools with four capabilities borrowed from Qwen-MM-Plugins' design:

1. Resolution budget presets for analyze_image / view_image (small / normal / large)
2. Coordinate closed loop: new crop_image and draw_bbox tools using 0-1000 normalized coordinates
3. Small-image upscaling to a minimum pixel floor (capped at 2x linear) so tiny crops stay legible for VLMs
4. Built-in decision skill (image-workflow) teaching the agent when to view vs analyze vs crop, and how to pick a budget

Background

The current vision path is two fixed-limit tools plus a batch pre-analysis path:

• analyze_image sends the image + prompt to the configured image-understanding model. Every call pays full image tokens, and the only size control is the hard provider cap (1MB tool cap / provider dimension limits). There is no way for the agent to trade detail against cost.
• view_image attaches an image to the primary model context, also with no resolution control.
• optimize_image_with_size_limit never upscales, so a tiny crop (e.g. a 64x64 region) is sent as-is and OCR/detail suffers.
• The only coordinate handling is a prose coordinate_note warning the model that positions in the analysis are estimates. There is no tool to actually crop or annotate a static image, so the estimate cannot be turned into an action loop.
• Tool descriptions are one-liners; the model has no guidance on which tool to use when.

Qwen-MM-Plugins solves these with token-budget-driven resolution presets (256/1024/2048 visual tokens), a crop/draw_bbox pair that consumes grounding output in the same 0-1000 normalized coordinate system, an upscale floor, and declarative SKILL.md decision documents.

Proposed changes

1. Budget presets (optional parameter, legacy behavior unchanged)

• Add ImageBudget (small/normal/large) mapping to pixel targets via token budgets (256/1024/2048 x 32^2 = ~512^2 / ~1024^2 / ~1448^2), clamped by the provider ImageLimits ceiling.
• Add an optional budget parameter to analyze_image and view_image. When omitted, the existing path runs unchanged (no behavior change for existing callers/configs).
• Fix the resize_note wording which currently hard-codes "downscaled" and would misreport upscaling.

2. Coordinate closed loop: crop_image + draw_bbox

• crop_image(path, box[x1,y1,x2,y2 in 0-1000], output_path?): validates and clamps the box, saves the cropped region next to the source by default, and returns pixel-mapped coordinates. Attaches a preview when the primary model supports multimodal tool output; degrades to a text summary otherwise.
• draw_bbox(path, boxes[{bbox,label?}], output_path?): draws rectangles with an adaptive line width and a fixed palette. v1 draws boxes only (no text labels, no new dependencies).
• Both support remote workspace paths (read + write through workspace filesystem services).
• Closed loop: the model reports a region in normalized coordinates, crop_image cuts it, then analyze_image(crop_path, budget="large") re-inspects the region at high resolution.

3. Small-image upscaling

• In the budget path only: images below the minimum pixel floor (~512^2) are upscaled, capped at 2x linear so tiny images do not burn tokens pointlessly (larger regions should be re-cropped instead).

4. Built-in decision skill

• New builtin skill image-workflow (SKILL.md): tool-selection decision table, budget guidance (small preview / normal default / large detail; large on crops), coordinate discipline (normalized 0-1000, model coordinates are estimates, use reported width/height/was_resized to convert), and cost tips (image tokens dominate; crop before repeated full-image calls).

Non-goals (follow-ups)

• 32px patch-grid snapping for Qwen-style providers
• Text label rendering in draw_bbox (needs font rasterization)
• Budget support for the ImageAnalyzer batch pre-analysis path