LibreChat/packages/data-provider/src/bedrock.ts
Danny Avila 034b672d0c
🫧 feat: Claude Opus 4.7 Reasoning Visibility (#12701)
* 🫧 fix: Restore Claude Opus 4.7 Reasoning Visibility

Claude Opus 4.7 omits `thinking` content from Messages API responses by
default — empty thinking blocks still stream, but the `thinking` field is
blank unless the caller passes `display: "summarized"` in the adaptive
thinking config. This silenced the LibreChat "Thoughts" UI for Anthropic
(and Anthropic-on-Bedrock) adaptive models.

- Extend `ThinkingConfigAdaptive` in `packages/api/src/types/anthropic.ts`
  with an optional `display: 'summarized' | 'omitted'` field
- Emit `{ type: 'adaptive', display: 'summarized' }` from
  `configureReasoning` in `packages/api/src/endpoints/anthropic/helpers.ts`
- Emit `{ type: 'adaptive', display: 'summarized' }` from
  `bedrockInputParser` in `packages/data-provider/src/bedrock.ts` and
  update the local `ThinkingConfig` union
- Update existing adaptive-thinking assertions to include the new field
- Add dedicated tests asserting `display: 'summarized'` flows through
  both the Anthropic endpoint and the Bedrock parser

See https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7#thinking-content-omitted-by-default

* refactor: Gate `display: summarized` on Opus 4.7+

Narrow the reasoning-visibility opt-in to the models that actually omit
thinking content by default, instead of applying it to every adaptive
model. Pre-Opus-4.7 adaptive models (Opus 4.6, Sonnet 4.6) already return
summaries, so sending the field is unnecessary noise.

- Add `omitsThinkingByDefault(model)` in `packages/data-provider/src/bedrock.ts`
  that returns true only for Opus 4.7+ (including future majors like Opus 5+)
- Bedrock parser now only attaches `display: 'summarized'` when the helper
  matches, keeping the adaptive object unchanged for older models
- Anthropic endpoint `configureReasoning` uses the same helper so its emit
  path matches the Bedrock one
- Tests: replace the blanket `display: 'summarized'` assertions with
  model-specific ones (Opus 4.7 gets it, Opus 4.6 / Sonnet 4.6 do not),
  add a dedicated `omitsThinkingByDefault` suite covering naming variants
  and future versions

* feat: Configurable Thought Visibility for Anthropic Adaptive Models

Expose the Anthropic `thinking.display` API field as a user-facing
parameter so users can override the `auto` default (which stays as the
Opus-4.7+ opt-in added earlier in this PR). Also fixes the CI type error
by widening the adaptive thinking type assignment via a resolver helper
that returns a properly-typed object.

- Add `ThinkingDisplay` enum (`auto` | `summarized` | `omitted`) and
  matching zod schema in `packages/data-provider/src/schemas.ts`
- Add `thinkingDisplay` to `tConversationSchema`, `anthropicSettings`, and
  the pick lists for Bedrock input/parser + Anthropic agent params
- Add `resolveThinkingDisplay(model, explicit)` helper in
  `packages/data-provider/src/bedrock.ts` that returns the wire value or
  undefined (auto → model default, explicit → always honored)
- `bedrockInputParser` now reads `thinkingDisplay` from input and emits
  `display` only when the resolver returns a value; strips the field
  on non-adaptive-model branches so it does not leak
- `configureReasoning` in the Anthropic endpoint threads
  `thinkingDisplay` through, uses the resolver, and casts the adaptive
  config to `AnthropicClientOptions['thinking']` so the widened shape
  compiles against the stale installed SDK types
- Add UI slider for `thinkingDisplay` in `parameterSettings.ts` next to
  `effort`, with three-position `com_ui_auto` / `com_ui_summarized` /
  `com_ui_omitted` labels
- Add translation keys `com_endpoint_anthropic_thinking_display`,
  `com_endpoint_anthropic_thinking_display_desc`, `com_ui_summarized`,
  `com_ui_omitted`
- Add tests: `resolveThinkingDisplay` suite (5 cases covering auto /
  explicit / unknown input), parser round-trip tests for all three
  modes on Opus 4.6 and Opus 4.7, Anthropic endpoint tests for explicit
  summarized/omitted overrides

* fix: Drop `thinkingDisplay` When Adaptive Thinking Is Disabled

If a user turns adaptive thinking off but had previously selected a
`thinkingDisplay` value, the stale field was left in `additionalFields`
and ended up merged into the Bedrock request's
`additionalModelRequestFields`. That leaks a non-Bedrock key into the
payload and can round-trip back into `llmConfig`.

- Delete `additionalFields.thinkingDisplay` alongside `thinking` and
  `thinkingBudget` in the `thinking === false` branch of
  `bedrockInputParser`
- Add a regression test asserting `thinking`, `thinkingBudget`, and
  `thinkingDisplay` are all absent when adaptive thinking is disabled on
  an Opus 4.7 request

Reported by chatgpt-codex-connector on PR #12701.

* refactor: Consolidate `ThinkingDisplay` Types and Preserve Persisted Display

Address review findings on PR #12701:

- [Codex P2] `bedrockInputSchema.transform` now extracts
  `thinking.display` from persisted `additionalModelRequestFields` back
  into the top-level `thinkingDisplay` field so explicit `'omitted'`
  round-trips through storage instead of being silently reverted to
  `'summarized'` on the next parse.
- [Codex P2] `getLLMConfig` in the Anthropic endpoint now reads
  `.display` from a persisted `thinking` object (agents store the full
  Anthropic shape) and uses it as the fallback for `thinkingDisplay`
  when no top-level override is present.
- [Audit #2] Collapse the three parallel wire-value types into a single
  `ThinkingDisplayWireValue = Exclude<ThinkingDisplay, 'auto'>` exported
  from `schemas.ts`; remove the duplicate `ThinkingDisplay` alias in
  `packages/api/src/types/anthropic.ts` (which collided with the enum
  name) and the `ThinkingDisplayValue` alias in `bedrock.ts`.
- [Audit #3] Add `thinkingDisplay` to the `TEndpointOption` pick list
  next to `effort`.
- [Audit #4] Add a TODO comment next to the `as
  AnthropicClientOptions['thinking']` cast explaining the stale
  `@librechat/agents` SDK types that require it.
- Add tests: four round-trip cases asserting `bedrockInputSchema`
  recovers `display` from persisted AMRF (Opus 4.7 omitted, pre-4.7
  summarized, unknown-value ignore, explicit top-level wins), and two
  `getLLMConfig` cases asserting the Anthropic endpoint preserves and
  overrides persisted `thinking.display`.

* fix: Preserve Persisted `thinking.display` in bedrockInputParser

The parser constructed a fresh adaptive thinking config without looking at
any `display` already embedded in the incoming
`additionalModelRequestFields.thinking`. On round-trip through
`initializeBedrock`, a persisted user choice of `'omitted'` on Opus 4.7+
was silently reverted to `'summarized'` by the auto fallback.

- Extract `extractPersistedDisplay` helper and reuse it in both the
  schema transform (form-state round-trip) and the parser (wire-request
  round-trip)
- `bedrockInputParser` now feeds the persisted display as the resolver's
  explicit value when no top-level `thinkingDisplay` override is set
- Add regression tests: parser preserves `display: 'omitted'` for
  persisted Opus 4.7 AMRF, and top-level `thinkingDisplay` still wins
  over persisted AMRF display

Reported by chatgpt-codex-connector (P1) on PR #12701.
2026-04-16 21:56:52 -04:00

526 lines
17 KiB
TypeScript

import { z } from 'zod';
import * as s from './schemas';
const DEFAULT_ENABLED_MAX_TOKENS = 8192;
const DEFAULT_THINKING_BUDGET = 2000;
const bedrockReasoningConfigValues = new Set<string>(Object.values(s.BedrockReasoningConfig));
type ThinkingConfig =
| { type: 'enabled'; budget_tokens: number }
| { type: 'adaptive'; display?: s.ThinkingDisplayWireValue };
/**
* Resolves the final `thinking.display` value for an adaptive-thinking request.
*
* Starting with Claude Opus 4.7, the Messages API returns empty `thinking`
* blocks unless the request sets `thinking.display`. This helper encodes the
* three user-facing modes — `'auto'` (LibreChat decides), `'summarized'`, and
* `'omitted'` — into the wire value (or `undefined` when the field should be
* left off).
*
* See https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7#thinking-content-omitted-by-default
*/
/**
* Safely extracts a nested `thinking.display` string from a persisted
* `additionalModelRequestFields` object, returning `undefined` if the shape
* isn't what we expect.
*/
function extractPersistedDisplay(amrf: unknown): string | undefined {
if (typeof amrf !== 'object' || amrf === null) {
return undefined;
}
const thinking = (amrf as Record<string, unknown>).thinking;
if (typeof thinking !== 'object' || thinking === null) {
return undefined;
}
const display = (thinking as Record<string, unknown>).display;
return typeof display === 'string' ? display : undefined;
}
export function resolveThinkingDisplay(
model: string,
explicit?: s.ThinkingDisplay | string | null,
): s.ThinkingDisplayWireValue | undefined {
if (explicit === s.ThinkingDisplay.summarized) {
return s.ThinkingDisplay.summarized;
}
if (explicit === s.ThinkingDisplay.omitted) {
return s.ThinkingDisplay.omitted;
}
if (omitsThinkingByDefault(model)) {
return s.ThinkingDisplay.summarized;
}
return undefined;
}
type AnthropicReasoning = {
thinking?: ThinkingConfig | boolean;
thinkingBudget?: number;
};
type AnthropicInput = BedrockConverseInput & {
additionalModelRequestFields: BedrockConverseInput['additionalModelRequestFields'] &
AnthropicReasoning;
};
/** Extracts opus major/minor version from both naming formats */
function parseOpusVersion(model: string): { major: number; minor: number } | null {
const nameFirst = model.match(/claude-opus[-.]?(\d+)(?:[-.](\d+))?/);
if (nameFirst) {
return {
major: parseInt(nameFirst[1], 10),
minor: nameFirst[2] != null ? parseInt(nameFirst[2], 10) : 0,
};
}
const numFirst = model.match(/claude-(\d+)(?:[-.](\d+))?-opus/);
if (numFirst) {
return {
major: parseInt(numFirst[1], 10),
minor: numFirst[2] != null ? parseInt(numFirst[2], 10) : 0,
};
}
return null;
}
/** Extracts sonnet major/minor version from both naming formats.
* Uses single-digit minor capture to avoid matching date suffixes (e.g., -20250514). */
function parseSonnetVersion(model: string): { major: number; minor: number } | null {
const nameFirst = model.match(/claude-sonnet[-.]?(\d+)(?:[-.](\d)(?!\d))?/);
if (nameFirst) {
return {
major: parseInt(nameFirst[1], 10),
minor: nameFirst[2] != null ? parseInt(nameFirst[2], 10) : 0,
};
}
const numFirst = model.match(/claude-(\d+)(?:[-.](\d)(?!\d))?-sonnet/);
if (numFirst) {
return {
major: parseInt(numFirst[1], 10),
minor: numFirst[2] != null ? parseInt(numFirst[2], 10) : 0,
};
}
return null;
}
/** Checks if a model supports adaptive thinking (Opus 4.6+, Sonnet 4.6+) */
export function supportsAdaptiveThinking(model: string): boolean {
const opus = parseOpusVersion(model);
if (opus && (opus.major > 4 || (opus.major === 4 && opus.minor >= 6))) {
return true;
}
const sonnet = parseSonnetVersion(model);
if (sonnet != null && (sonnet.major > 4 || (sonnet.major === 4 && sonnet.minor >= 6))) {
return true;
}
return false;
}
/**
* Checks if a model omits `thinking` content from responses by default.
*
* Starting with Claude Opus 4.7, the Messages API returns empty `thinking`
* blocks unless the request explicitly opts in via `thinking.display =
* "summarized"`. This helper narrows the opt-in to Opus 4.7+ (and any future
* major Opus version) so older adaptive-thinking models are left untouched.
*
* See https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7#thinking-content-omitted-by-default
*/
export function omitsThinkingByDefault(model: string): boolean {
const opus = parseOpusVersion(model);
if (opus && (opus.major > 4 || (opus.major === 4 && opus.minor >= 7))) {
return true;
}
return false;
}
/** Checks if a model qualifies for the context-1m beta header (Sonnet 4+, Opus 4.6+, Opus 5+) */
export function supportsContext1m(model: string): boolean {
const sonnet = parseSonnetVersion(model);
if (sonnet != null && sonnet.major >= 4) {
return true;
}
const opus = parseOpusVersion(model);
if (opus && (opus.major > 4 || (opus.major === 4 && opus.minor >= 6))) {
return true;
}
return false;
}
/**
* Gets the appropriate anthropic_beta headers for Bedrock Anthropic models.
* Bedrock uses `anthropic_beta` (with underscore) in additionalModelRequestFields.
*
* @param model - The Bedrock model identifier (e.g., "anthropic.claude-sonnet-4-20250514-v1:0")
* @returns Array of beta header strings, or empty array if not applicable
*/
function getBedrockAnthropicBetaHeaders(model: string): string[] {
const betaHeaders: string[] = [];
const isClaudeThinkingModel =
model.includes('anthropic.claude-3-7-sonnet') ||
/anthropic\.claude-(?:[4-9](?:\.\d+)?(?:-\d+)?-(?:sonnet|opus|haiku)|(?:sonnet|opus|haiku)-[4-9])/.test(
model,
);
const isSonnet4PlusModel =
/anthropic\.claude-(?:sonnet-[4-9]|[4-9](?:\.\d+)?(?:-\d+)?-sonnet)/.test(model);
if (isClaudeThinkingModel) {
betaHeaders.push('output-128k-2025-02-19');
}
if (isSonnet4PlusModel || supportsAdaptiveThinking(model)) {
betaHeaders.push('context-1m-2025-08-07');
}
return betaHeaders;
}
export const bedrockInputSchema = s.tConversationSchema
.pick({
/* LibreChat params; optionType: 'conversation' */
modelLabel: true,
promptPrefix: true,
resendFiles: true,
iconURL: true,
greeting: true,
spec: true,
maxOutputTokens: true,
maxContextTokens: true,
artifacts: true,
/* Bedrock params; optionType: 'model' */
region: true,
system: true,
model: true,
maxTokens: true,
temperature: true,
topP: true,
stop: true,
thinking: true,
thinkingBudget: true,
effort: true,
thinkingDisplay: true,
reasoning_effort: true,
promptCache: true,
/* Catch-all fields */
topK: true,
additionalModelRequestFields: true,
})
.transform((obj) => {
if ((obj as AnthropicInput).additionalModelRequestFields?.thinking != null) {
const _obj = obj as AnthropicInput;
const thinking = _obj.additionalModelRequestFields.thinking;
obj.thinking = !!thinking;
obj.thinkingBudget =
typeof thinking === 'object' && 'budget_tokens' in thinking
? thinking.budget_tokens
: undefined;
if (obj.thinkingDisplay == null) {
const persistedDisplay = extractPersistedDisplay({ thinking });
if (
persistedDisplay === s.ThinkingDisplay.summarized ||
persistedDisplay === s.ThinkingDisplay.omitted
) {
obj.thinkingDisplay = persistedDisplay as s.ThinkingDisplay;
}
}
delete obj.additionalModelRequestFields;
}
return s.removeNullishValues(obj);
})
.catch(() => ({}));
export type BedrockConverseInput = z.infer<typeof bedrockInputSchema>;
export const bedrockInputParser = s.tConversationSchema
.pick({
/* LibreChat params; optionType: 'conversation' */
modelLabel: true,
promptPrefix: true,
resendFiles: true,
iconURL: true,
greeting: true,
spec: true,
artifacts: true,
maxOutputTokens: true,
maxContextTokens: true,
/* Bedrock params; optionType: 'model' */
region: true,
model: true,
maxTokens: true,
temperature: true,
topP: true,
stop: true,
thinking: true,
thinkingBudget: true,
effort: true,
thinkingDisplay: true,
reasoning_effort: true,
promptCache: true,
/* Catch-all fields */
topK: true,
additionalModelRequestFields: true,
})
.catchall(z.any())
.transform((data) => {
const knownKeys = [
'modelLabel',
'promptPrefix',
'resendFiles',
'iconURL',
'greeting',
'spec',
'maxOutputTokens',
'artifacts',
'additionalModelRequestFields',
'region',
'model',
'maxTokens',
'temperature',
'topP',
'stop',
'promptCache',
];
const additionalFields: Record<string, unknown> = {};
const typedData = data as Record<string, unknown>;
Object.entries(typedData).forEach(([key, value]) => {
if (!knownKeys.includes(key)) {
if (key === 'topK') {
additionalFields['top_k'] = value;
} else {
additionalFields[key] = value;
}
delete typedData[key];
}
});
/** Default thinking and thinkingBudget for 'anthropic.claude-3-7-sonnet' models, if not defined */
if (
typeof typedData.model === 'string' &&
(typedData.model.includes('anthropic.claude-3-7-sonnet') ||
/anthropic\.claude-(?:[4-9](?:\.\d+)?(?:-\d+)?-(?:sonnet|opus|haiku)|(?:sonnet|opus|haiku)-[4-9])/.test(
typedData.model,
))
) {
const isAdaptive = supportsAdaptiveThinking(typedData.model as string);
if (isAdaptive) {
const effort = additionalFields.effort;
if (effort && typeof effort === 'string' && effort !== '') {
additionalFields.output_config = { effort };
}
delete additionalFields.effort;
if (additionalFields.thinking === false) {
delete additionalFields.thinking;
delete additionalFields.thinkingBudget;
delete additionalFields.thinkingDisplay;
} else {
/**
* Persisted agent `model_parameters` round-trip back through this
* parser with the prior `thinking.display` embedded in
* `additionalModelRequestFields`. Surface it as the resolver's
* explicit value when no top-level `thinkingDisplay` is set so the
* prior user choice (e.g. 'omitted') survives instead of being
* clobbered by the Opus 4.7+ auto → 'summarized' fallback.
*/
const topLevelDisplay = additionalFields.thinkingDisplay as
| s.ThinkingDisplay
| string
| null
| undefined;
const persistedDisplay = extractPersistedDisplay(typedData.additionalModelRequestFields);
const thinkingConfig: ThinkingConfig = { type: 'adaptive' };
const display = resolveThinkingDisplay(
typedData.model as string,
topLevelDisplay ?? persistedDisplay,
);
if (display) {
thinkingConfig.display = display;
}
additionalFields.thinking = thinkingConfig;
delete additionalFields.thinkingBudget;
delete additionalFields.thinkingDisplay;
}
} else {
if (additionalFields.thinking === undefined) {
additionalFields.thinking = true;
} else if (additionalFields.thinking === false) {
delete additionalFields.thinking;
delete additionalFields.thinkingBudget;
}
if (additionalFields.thinking === true && additionalFields.thinkingBudget === undefined) {
additionalFields.thinkingBudget = DEFAULT_THINKING_BUDGET;
}
delete additionalFields.effort;
delete additionalFields.thinkingDisplay;
}
/** Anthropic uses 'effort' via output_config, not reasoning_config */
delete additionalFields.reasoning_effort;
if ((typedData.model as string).includes('anthropic.')) {
const betaHeaders = getBedrockAnthropicBetaHeaders(typedData.model as string);
if (betaHeaders.length > 0) {
additionalFields.anthropic_beta = betaHeaders;
}
}
} else {
delete additionalFields.thinking;
delete additionalFields.thinkingBudget;
delete additionalFields.effort;
delete additionalFields.thinkingDisplay;
delete additionalFields.output_config;
delete additionalFields.anthropic_beta;
const reasoningEffort = additionalFields.reasoning_effort;
delete additionalFields.reasoning_effort;
if (
typeof reasoningEffort === 'string' &&
bedrockReasoningConfigValues.has(reasoningEffort)
) {
additionalFields.reasoning_config = reasoningEffort;
}
}
const isAnthropicModel =
typeof typedData.model === 'string' && typedData.model.includes('anthropic.');
/** Strip stale fields from previously-persisted additionalModelRequestFields */
if (
typeof typedData.additionalModelRequestFields === 'object' &&
typedData.additionalModelRequestFields != null
) {
const amrf = typedData.additionalModelRequestFields as Record<string, unknown>;
if (!isAnthropicModel) {
delete amrf.anthropic_beta;
delete amrf.thinking;
delete amrf.thinkingBudget;
delete amrf.effort;
delete amrf.output_config;
delete amrf.reasoning_config;
} else {
delete amrf.reasoning_config;
delete amrf.reasoning_effort;
}
}
/** Default promptCache for claude and nova models, if not defined */
if (
typeof typedData.model === 'string' &&
(typedData.model.includes('claude') || typedData.model.includes('nova'))
) {
if (typedData.promptCache === undefined) {
typedData.promptCache = true;
}
} else if (typedData.promptCache === true) {
typedData.promptCache = undefined;
}
if (Object.keys(additionalFields).length > 0) {
typedData.additionalModelRequestFields = {
...((typedData.additionalModelRequestFields as Record<string, unknown> | undefined) || {}),
...additionalFields,
};
}
if (typedData.maxOutputTokens !== undefined) {
typedData.maxTokens = typedData.maxOutputTokens;
} else if (typedData.maxTokens !== undefined) {
typedData.maxOutputTokens = typedData.maxTokens;
}
return s.removeNullishValues(typedData) as BedrockConverseInput;
})
.catch(() => ({}));
/**
* Configures the "thinking" parameter based on given input and thinking options.
*
* @param data - The parsed Bedrock request options object
* @returns The object with thinking configured appropriately
*/
function configureThinking(data: AnthropicInput): AnthropicInput {
const updatedData = { ...data };
const thinking = updatedData.additionalModelRequestFields?.thinking;
if (thinking === true) {
updatedData.maxTokens =
updatedData.maxTokens ?? updatedData.maxOutputTokens ?? DEFAULT_ENABLED_MAX_TOKENS;
delete updatedData.maxOutputTokens;
const thinkingConfig: ThinkingConfig = {
type: 'enabled',
budget_tokens:
updatedData.additionalModelRequestFields?.thinkingBudget ?? DEFAULT_THINKING_BUDGET,
};
if (thinkingConfig.budget_tokens > updatedData.maxTokens) {
thinkingConfig.budget_tokens = Math.floor(updatedData.maxTokens * 0.9);
}
updatedData.additionalModelRequestFields!.thinking = thinkingConfig;
delete updatedData.additionalModelRequestFields!.thinkingBudget;
} else if (
typeof thinking === 'object' &&
thinking != null &&
(thinking as { type: string }).type === 'adaptive'
) {
if (updatedData.maxTokens == null && updatedData.maxOutputTokens != null) {
updatedData.maxTokens = updatedData.maxOutputTokens;
}
delete updatedData.maxOutputTokens;
delete updatedData.additionalModelRequestFields!.thinkingBudget;
}
return updatedData;
}
export const bedrockOutputParser = (data: Record<string, unknown>) => {
const knownKeys = [...Object.keys(s.tConversationSchema.shape), 'topK', 'top_k'];
let result: Record<string, unknown> = {};
// Extract known fields from the root level
Object.entries(data).forEach(([key, value]) => {
if (knownKeys.includes(key)) {
result[key] = value;
}
});
// Extract known fields from additionalModelRequestFields
if (
typeof data.additionalModelRequestFields === 'object' &&
data.additionalModelRequestFields !== null
) {
Object.entries(data.additionalModelRequestFields as Record<string, unknown>).forEach(
([key, value]) => {
if (knownKeys.includes(key)) {
if (key === 'top_k') {
result['topK'] = value;
} else if (key === 'thinking' || key === 'thinkingBudget') {
return;
} else {
result[key] = value;
}
}
},
);
}
// Handle maxTokens and maxOutputTokens
if (result.maxTokens !== undefined && result.maxOutputTokens === undefined) {
result.maxOutputTokens = result.maxTokens;
} else if (result.maxOutputTokens !== undefined && result.maxTokens === undefined) {
result.maxTokens = result.maxOutputTokens;
}
result = configureThinking(result as AnthropicInput);
const amrf = result.additionalModelRequestFields as Record<string, unknown> | undefined;
if (!amrf || Object.keys(amrf).length === 0) {
delete result.additionalModelRequestFields;
}
return result;
};