fix(models): preserve AvailableModels capability metadata

[SenZh]

Summary

Fixes three related metadata/runtime issues discovered while using Cursor models in OpenCode on slower networks.

1. Model discovery metadata (AvailableModels path)

• Increase AvailableModels RPC timeout (5s -> 30s) and config discovery timeout (35s) to avoid silent fallback to GetUsableModels
• Normalize reasoning_effort, supportsImages, and supportsThinking from Cursor API responses
• Infer context window from variant params, contextTokenLimitForMaxMode, and tooltip markdown (including variant-level tooltip, e.g. Gemini 1M)
• Take Math.max context across grouped variants
• Emit per-model image/reasoning flags in provider and config descriptors instead of hardcoding all models as image-capable with 200k context
• Set attachment: true on config model descriptors so OpenCode does not strip image inputs before they reach the proxy

2. Session title generation (NameAgent path)

• Increase NameAgent RPC timeout (5s -> 15s); live calls often need ~6-7s due to h2-bridge subprocess + network RTT
• Add timeout/failure logging instead of silently returning empty titles
• Broaden title-generation prompt detection for OpenCode 1.18 ("Generate a brief title", etc.)

Root cause

Models: On slow links, the default 5s RPC budget caused AvailableModels to time out and silently degrade to GetUsableModels, flattening capability metadata (notably context window to 200k). The normalizer also ignored Gemini-style reasoning_effort, never mapped API image/thinking flags, missed variant-level tooltip context, and config descriptors lacked attachment: true — causing all models to appear text-only in the UI.

Titles: OpenCode sends agent=title requests through the plugin proxy. The plugin routes these to Cursor's NameAgent RPC (not the user's chat model). The default 5s client timeout was shorter than typical end-to-end latency, so title generation silently no-op'd and sessions stayed at New session - <timestamp>.

Test plan

• npm run build
• npm run test (smoke suite — regression cases for reasoning_effort, supportsImages, variant tooltip context, contextTokenLimitForMaxMode, config attachment, grouped variant Math.max)
• Live verification: Gemini 3.8 Flash shows image input + 1M context; session titles auto-generate after first message
• Maintainer: verify mixed context sizes (200k/272k/300k/1M) on your network after merge

[SenZh]

Hi maintainers — thanks for maintaining this plugin!

This PR bundles three fixes found during real-world use with Cursor OAuth on a slower network. All changes are covered by the smoke test suite and verified locally against live Cursor API responses.

What was broken:

1. Model list — After ~5s, AvailableModels timed out -> silent fallback to GetUsableModels -> every model showed 200k context and incorrect capabilities. Gemini 3.8 Flash showed text-only despite API returning supportsImages: true.
2. Image input — Even when metadata was correct, config descriptors missing attachment: true caused OpenCode to strip images before the proxy.
3. Session titles — NameAgent RPC (used for agent=title) hit the 5s default timeout (~6-7s actual). Titles stayed as New session - <timestamp>. Other providers work because they use standard chat/completions, not this RPC path.

Note on CI: The repo only has a manual Release workflow — no PR checks run automatically. I ran npm run build and npm run test locally; all pass.

Happy to split into separate PRs or adjust anything. Let me know if you'd like a live repro script or logs.

[thaoran]

+1 — hit exactly this on 2.5.1 (OpenCode 1.18.9, macOS, OpenChamber desktop).

Independently traced the same root cause before finding this PR: the live AvailableModels call takes 11–18s here (measured repeatedly, ~296KB payload), so the hardcoded 5s timeoutMs in cursor-rpc always kills the bridge and silently degrades to GetUsableModels — every model ends up at DEFAULT_CONTEXT_WINDOW = 200k with flat legacy-slug IDs and no 1m/300k/272k tiers.

Verified the fix direction locally: bumping the RPC default from 5s → 30s makes a fresh opencode models register the full grouped catalog (104 models), and opencode debug config shows correct limits (claude-sonnet-5 → 300k, gpt-5.6-sol → 272k, -1m tiers → 1M).

Looking forward to this being merged — the tooltip-based context inference for models whose variants don't carry a context param (grok/gemini) is a nice touch beyond the timeout fix. I had filed this as an issue earlier today; removing it as a duplicate of this PR.