feat: force adaptive thinking via KIMI_MODEL_ADAPTIVE_THINKING#232
feat: force adaptive thinking via KIMI_MODEL_ADAPTIVE_THINKING#232RealKai42 wants to merge 2 commits into
Conversation
Add the KIMI_MODEL_ADAPTIVE_THINKING env var and a matching
adaptive_thinking model-alias field that force adaptive thinking
(thinking: { type: 'adaptive' }) on or off, overriding the Anthropic
model-name version inference.
This lets custom-named staff endpoints that back an adaptive-capable
model opt in even when the model name does not encode a parseable Claude
version (which would otherwise fall back to budget-based thinking).
The kosong AnthropicChatProvider resolves the adaptive decision once in
withThinking() as `_adaptiveThinking ?? supportsAdaptiveThinking(model)`
and threads it through clampEffort/supportsEffortParam so forced adaptive
also unlocks the max effort level.
🦋 Changeset detectedLatest commit: 983399c The changes in this PR will be included in the next version bump. This PR includes changesets to release 4 packages
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
commit: |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 298c22518a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // Explicitly declare adaptive-thinking support, overriding the kosong | ||
| // model-name version inference. Needed for custom-named Anthropic endpoints | ||
| // whose model name does not encode a parseable Claude version. | ||
| adaptiveThinking: z.boolean().optional(), |
There was a problem hiding this comment.
For the documented custom-named Anthropic case, setting only adaptive_thinking = true leaves capabilities empty. The TUI model picker uses ModelAlias.capabilities directly (apps/kimi-code/src/tui/components/dialogs/model-selector.ts:71-75) and treats aliases without thinking as unsupported, so switching to such an alias forces thinking=false before the request ever reaches the provider; the runtime capability resolver also still relies on declared/detected capabilities (packages/agent-core/src/session/provider-manager.ts:200-205). Please have the new flag imply the thinking capability, or require capabilities = ["thinking"] alongside it in the config docs, so the advertised one-field opt-in actually enables thinking for custom model names.
Useful? React with 👍 / 👎.
| } | ||
|
|
||
| const effectiveEffort = clampEffort(effort, this._model); | ||
| const effectiveEffort = clampEffort(effort, this._model, adaptive); |
There was a problem hiding this comment.
When adaptiveThinking=true is used for the documented custom-named endpoint case, xhigh still goes through clampEffort with the opaque model name and is downgraded to high because isOpus47(model) cannot recognize the backing model. That makes KIMI_MODEL_ADAPTIVE_THINKING=true fail to fully override name-based inference for custom Claude Opus 4.7/4.8 endpoints that support xhigh, even though max is preserved; the override needs a way to keep xhigh for known-capable custom endpoints or document that it cannot.
Useful? React with 👍 / 👎.
A model alias that sets adaptive_thinking=true (or env KIMI_MODEL_ADAPTIVE_THINKING=true) without listing a thinking capability previously resolved to thinking=false for custom-named endpoints not in the capability catalog. The model picker then classified the alias as "thinking unsupported" and forced thinking off (persisting default_thinking=false). Treat a forced adaptive flag as advertising the thinking capability in resolveModelCapabilities, so the config.toml one-field opt-in agrees with the KIMI_MODEL_* env path (which already defaults capabilities to include thinking).
1 participant
Summary
Adds a way to explicitly force adaptive thinking (
thinking: { type: 'adaptive' }) on Anthropic-compatible endpoints, overriding the model-name version inference.Today the adaptive-vs-budget decision in kosong's
AnthropicChatProvider.withThinking()is gated solely bysupportsAdaptiveThinking(model), which parses the model name (FAMILY_VERSION_RE, requires Claude ≥ 4.6). Custom-named staff endpoints (e.g.coding-model-okapi-0527-vibe) fail that regex and always fall back to budget-based thinking — even when the backing model supports adaptive. There was no way to opt in.This adds two aligned knobs:
KIMI_MODEL_ADAPTIVE_THINKING=true(part of theKIMI_MODEL_*channel)adaptive_thinkingon a[models.<alias>]entry (works for hand-writtenconfig.tomlaliases too)Both flow down to a new
adaptiveThinking?: booleanoption on the kosong provider.withThinking()now resolves the decision once as_adaptiveThinking ?? supportsAdaptiveThinking(model)and threads it throughclampEffort/supportsEffortParam, so forcing adaptive also unlocks themaxeffort level. Unset = current behavior (infer from name).Layers touched
packages/kosong/src/providers/anthropic.ts— new option + resolved-once adaptive decisionpackages/agent-core/src/config/schema.ts—adaptiveThinkingonModelAliasSchemapackages/agent-core/src/config/env-model.ts— parseKIMI_MODEL_ADAPTIVE_THINKING(genericparseBooleanVar, fail-fast)packages/agent-core/src/session/provider-manager.ts— thread alias field into the anthropic provider configenv-vars.md+config-files.md) + changesetRisk
Forcing adaptive on an endpoint whose backing model does not support it makes the API reject the request. The docs call this out; leave it unset unless you know the backing model supports adaptive thinking.
Test Plan
TDD throughout (failing test first, then implementation):
adaptiveThinking=trueforces adaptive on an unversioned model name;maxnot clamped; no-flag stays budget-based (regression guard); explicitfalseforces budget on a 4.6 nameKIMI_MODEL_ADAPTIVE_THINKINGmaps onto the alias (true/false/omitted) and rejects invalid input; aliasadaptiveThinkingforwarded to the anthropic provider config and omitted when unset