fix(catalog): shorten no-thinking gateway prefix to no-think/

[Rahulsharma0810]

Problem

The current no-thinking gateway prefix claude-3-omniroute-no-thinking/ is 34 characters long. When combined with a provider prefix and model name, the full ID routinely exceeds the visible width of terminal-based AI clients.

In OpenCode and OpenClaw (Claude-based terminal clients), the model picker column is fixed-width. With the old prefix, every single no-thinking variant is truncated — users can see claude-3-omniroute-no-thinking/cc/claude-o but not the actual model name at the end. This makes the entries effectively unusable for manual selection.

Screenshot from OpenCode model picker (searching for "thinking"):

All 35 visible entries show only claude-3-omniroute-no-thinking/<provider>/claude-o — the model suffix is cut off in every row. The provider prefix alone consumes the entire column width.

Solution

Rename the prefix from claude-3-omniroute-no-thinking/ → no-think/ (9 characters, a 74% reduction).

After the change:

no-think/cc/claude-opus-4-6     OmniRoute   ← fully visible
no-think/gh/claude-haiku-4.5    OmniRoute   ← fully visible
no-think/opencode-zen/claude-s  OmniRoute   ← fully visible

Second fix: canonical prefix consistency (commit 2263686)

When MODELS_CATALOG_PREFIX_MODE=canonical (from PR #4427), the rest of the catalog emits
canonical provider prefixes (e.g. claude/claude-opus-4-6). But appendNoThinkingVariants
was wrapping whatever alias-prefixed id it found in finalModels unchanged, producing
no-think/cc/claude-opus-4-6 even though canonical mode was active.

Fix: added normalizeProviderPrefix(qualifiedId, aliasToCanonical) helper inside
noThinkingAlias.ts. appendNoThinkingVariants now accepts an optional aliasToCanonical?
map; the catalog passes aliasToProviderId only in canonical mode so no-think ids stay
consistent with the active prefix mode.

Mode	Example no-think id
dual / alias (default)	no-think/cc/claude-opus-4-6
canonical	no-think/claude/claude-opus-4-6

What changed

• open-sse/utils/noThinkingAlias.ts:
  • NO_THINKING_PREFIX renamed to "no-think/"
  • normalizeProviderPrefix() helper added (private)
  • appendNoThinkingVariants() accepts optional aliasToCanonical? param
• src/app/api/v1/models/catalog.ts: passes aliasToProviderId when prefixMode === "canonical"
• Comments in src/shared/constants/modelSpecs.ts and src/sse/handlers/chat.ts updated
• tests/unit/no-thinking-alias.test.ts: all string literals updated; 2 new canonical tests

What is unchanged

• All detection, stripping, routing, and catalog logic is unchanged — only the constant string value and the new optional normalization
• applyNoThinkingAlias(), isNoThinkingAlias(), stripNoThinkingAlias(), shouldExposeNoThinkingAlias() all work identically
• Default behavior (dual/alias modes) is byte-identical to before

Migration note

Clients that saved a model like claude-3-omniroute-no-thinking/cc/claude-opus-4-6 in their config will need to update to no-think/cc/claude-opus-4-6. Inbound requests using the old prefix will not be recognized after this change. If backward compat is required, a temporary dual-recognition shim can be added to isNoThinkingAlias().

Tests

tests 14 | pass 14 | fail 0

New tests:

✔ appendNoThinkingVariants normalizes alias prefix to canonical when aliasToCanonical map is provided
✔ appendNoThinkingVariants keeps alias prefix when no map is provided

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[diegosouzapw]

Obrigado, @Rahulsharma0810! 🙌 A ideia (prefixo curto no-think/ para não truncar no model-picker de clientes de terminal) é ótima. Reconstruí a PR sobre o tip da release/v3.8.34 mantendo apenas o rename do prefixo — a branch original carregava, da base stale, dois reverts que precisei descartar: a remoção de preParsedBody/resolveChatRequestBody no handleChat (reabria o OOM #4380) e a reversão do maxOutputTokens em modelSpecs (desfazia o #4584). Testes verdes. Mergeando para a release.