models: allow model variants to override limit (context/output) #56582

Summary

Model variants can only carry settings, headers, and body — they cannot override a model's limit (context/output tokens). This makes it impossible to offer multiple context-window tiers (e.g. 256K vs 1M) for the same upstream model without duplicating the whole model entry. Some providers expose this as a first-class control (e.g. the CodeBuddy VS Code extension has a per-model "context length" selector), so an equivalent knob in OpenCode would be useful.

Environment

• opencode version: 2.0.32
• OS: Linux 7.0.0-38-generic (Ubuntu, x86_64)
• Terminal: xterm-256color
• Shell: /bin/bash
• Install/channel: latest (release)
• Active plugins: none found in config

Reproduction

1. Define a custom openai-compatible model in opencode.jsonc with "limit": { "context": 1000000, "output": 131072 }.
2. Add variants to the same model entry.
3. Try to give one variant a smaller context budget (e.g. context: 262144) — the variant schema accepts only settings, headers, and body, so limit cannot be varied.

Expected Behavior

Variant-level limit override (merge order provider → model → variant), so a single catalog entry can expose context tiers:

codebuddy/glm-5.3-flash#1m
codebuddy/glm-5.3-flash#256k

Alternatively, a native per-session context-length selector would achieve the same.

Actual Behavior

limit is static per model entry. The only workaround is duplicating the entry with an alias modelID:

"glm-5.3-flash": {
  "name": "T-GLM-5.3-Flash（快速·1M）",
  "limit": { "context": 1000000, "output": 131072 }
},
"glm-5.3-flash-256k": {
  "modelID": "glm-5.3-flash",
  "name": "T-GLM-5.3-Flash（快速·256K）",
  "limit": { "context": 262144, "output": 131072 }
}

This works, but pollutes the model list with one entry per context tier.

Additional Context

• Docs (https://opencode.ai/v2/docs/models/) state: "Each variant can contain settings, headers, and body."
• limit.context is used by OpenCode for context accounting / compaction timing, so a smaller tier genuinely changes behavior, not just display.
• The motivating setup is a custom openai-compatible provider where the backend serves the same model at different context budgets (CodeBuddy's client picks per request; OpenCode currently can't).