Models with long IDs in recent crash /models menu with BUTTON_DATA_INVALID (Telegram 64-byte callback_data limit)

[n3r066]

Symptom

Pressing /models (or any path that shows model-selection-menu) returns:

Планировщик: 🔴 Не удалось получить список моделей

Log:

[ERROR] [ModelHandler] Error showing model menu: GrammyError:
  Call to 'sendMessage' failed! (400: Bad Request: BUTTON_DATA_INVALID)
    at async showModelSelectionMenu
      (.../bot/menus/model-selection-menu.js:56:9)

Cause

model-selection-menu.js builds inline-keyboard callback_data as:

keyboard.text(labelWithCheck, `model:${model.providerID}:${model.modelID}`).row();

Telegram Bot API caps callback_data at 64 bytes. So any model whose
providerID + : + modelID exceeds ~53 chars blows the request.

In our case the offending entry lives in ~/.local/state/opencode/model.json
→ recent:

{
  "providerID": "fireworks",
  "modelID": "accounts/hubabuba3227-1hvtqlh/deployments/kpwpvuky"
}

model:fireworks:accounts/hubabuba3227-1hvtqlh/deployments/kpwpvuky
= 66 bytes → BUTTON_DATA_INVALID.

This is purely a function of the arbitrary model ID; it happens with any model
whose path is long (Fireworks accounts deployments, custom fine-tunes, etc.).
The model itself is otherwise valid; only the inline-menu rendering breaks.

Repro

1. Configure an OpenCode provider whose modelID is longer than ~53 chars on
   the joined model:<providerID>:<modelID> string.
2. Trigger that model so it lands in ~/.local/state/opencode/model.json →
   recent.
3. In Telegram, send /models.
4. Telegram responds with 400 BUTTON_DATA_INVALID; the bot catches it and
   replies with the localised model.menu.error text.

Suggested fixes (any one)

1. Shorten/tile the callback: store an opaque short id (e.g. base32 hash or
   incrementing index) in callback_data, look up the real
   {providerID,modelID} from an in-memory map keyed by that id.
2. Skip / rewrite override labels so the joined payload fits 64 bytes; surface
   an upsell in the menu text when truncation occurs.
3. Add a model-id length validator at startup that warns or refuses to render
   models whose callback_data would exceed 64 bytes.

Option 1 is the most robust; option 3 is the cheapest patch.

Environment

• Bot version: v0.22.1 (npm @grinev/opencode-telegram-bot@latest)
• OpenCode CLI/opencode serve: 1.17.13
• Telegram Bot API: latest (2026-07-04)
• Confirmed: end-to-end Telegram polling and OpenCode serve are healthy; only
  the inline-keyboard payload fails.