@namzu/sdk 62.4.0 → 62.6.0

package/src/pricing/catalogue.generated.ts

@@ -33,6 +33,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 10,
 					outputCostPer1M: 50,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 1, writeCostPer1M: 12.5 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -41,6 +42,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 10,
 					outputCostPer1M: 50,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.25, writeCostPer1M: 12.5 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -49,6 +51,8 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 8,
 					outputCostPer1M: 40,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.2, writeCostPer1M: 10 },
+					fastMultiplier: 2,
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -57,6 +61,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 1,
 					outputCostPer1M: 5,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.1, writeCostPer1M: 1.25 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -65,6 +70,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 10,
 					outputCostPer1M: 50,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 1, writeCostPer1M: 12.5 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -73,6 +79,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 10,
 					outputCostPer1M: 50,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.25, writeCostPer1M: 12.5 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -81,6 +88,8 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 8,
 					outputCostPer1M: 40,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.2, writeCostPer1M: 10 },
+					fastMultiplier: 2,
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -89,6 +98,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 15,
 					outputCostPer1M: 75,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 1.5, writeCostPer1M: 18.75 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -97,6 +107,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 15,
 					outputCostPer1M: 75,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 1.5, writeCostPer1M: 18.75 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -105,6 +116,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 5,
 					outputCostPer1M: 25,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.5, writeCostPer1M: 6.25 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -113,6 +125,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 5,
 					outputCostPer1M: 25,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.5, writeCostPer1M: 6.25 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -121,6 +134,8 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 5,
 					outputCostPer1M: 25,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.5, writeCostPer1M: 6.25 },
+					fastMultiplier: 2,
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -129,6 +144,8 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 5,
 					outputCostPer1M: 25,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.5, writeCostPer1M: 6.25 },
+					fastMultiplier: 2,
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -137,6 +154,8 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 4,
 					outputCostPer1M: 20,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.2, writeCostPer1M: 5 },
+					fastMultiplier: 2,
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -145,6 +164,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 3,
 					outputCostPer1M: 15,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.3, writeCostPer1M: 3.75 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -153,6 +173,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 3,
 					outputCostPer1M: 15,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.3, writeCostPer1M: 3.75 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -161,6 +182,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 3,
 					outputCostPer1M: 15,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.3, writeCostPer1M: 3.75 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -169,6 +191,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 2,
 					outputCostPer1M: 10,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.2, writeCostPer1M: 2.5 },
+					webSearchCostPer1000: 10,
 				},
 			],
 			[
@@ -177,6 +200,7 @@ export const VENDOR_RATES: readonly VendorRates[] = [
 					inputCostPer1M: 2,
 					outputCostPer1M: 10,
 					cache: { promptIncludesCacheReads: false, readCostPer1M: 0.1, writeCostPer1M: 2.5 },
+					webSearchCostPer1000: 10,
 				},
 			],
 		]),

package/src/pricing/rates.source.json

@@ -19,6 +19,17 @@
 		"inference. Its turns are priced at zero, which is KNOWN-free and therefore",
 		"distinct from unknown. Such a vendor needs no model rows.",
 		"",
+		"fastMultiplier (optional) is the vendor's published multiple of every rate for",
+		"a request served in fast mode (verified 2026-11-10 against the fast-mode page:",
+		"2x on the full context window, stacking with cache pricing). A model without it",
+		"has no published fast price, so a fast request on it is reported as unpriced.",
+		"",
+		"webSearchPer1000 (optional, per vendor) is the price in USD of 1,000 hosted web",
+		"searches, billed per request on top of tokens (verified 2026-11-10 against the",
+		"pricing page's web search section: $10 per 1,000, a failed search is not billed).",
+		"It is copied onto every model row of that vendor. Absent means searches are not",
+		"part of the total.",
+		"",
 		"Matching is EXACT on a normalised id (lowercased, one trailing -YYYYMMDD or",
 		"@YYYYMMDD snapshot suffix removed). Deliberately not prefix or substring:",
 		"within one vendor here, ids one character apart differ in price by 4x, so a",
@@ -59,6 +70,7 @@
 			"verified": "2026-11-10",
 			"promptIncludesCacheReads": false,
 			"reportsCacheWrites": true,
+			"webSearchPer1000": 10,
 			"note": "The last two rows exist because the driver's own offline catalogue offers those models and this table did not price them — a lookup-key gap that reads as `cost unknown` and is invisible to the generator's own check, which only proves the module matches this file. `providers/anthropic` asserts the two lists agree. Vendor list prices, every row checked against its pricing page on the `verified` date. Cache read is the vendor's published 0.1x of the input rate, except on the six models it prices differently: 0.05x on claude-opus-5-5 and claude-sonnet-5-5 and 0.025x on claude-fable-5-1, claude-mythos-5-1, claude-fable-5-2 and claude-mythos-5-2, entered at the published rate rather than derived. claude-haiku-5-5 has no row on purpose: the vendor prices it by prompt length ($0.10 in / $0.50 out up to 100,000 prompt tokens, $0.50 / $2.50 above, cache counted in the length), this table holds one rate per model, and either single rate would be wrong for the other half of requests, so its cost reports unknown. Cache write is its published 1.25x for the five-minute entry, which is the default TTL and the one the driver's `cacheControl: { type: 'auto' }` asks for. claude-sonnet-5 launched at an introductory $2/$10 with $3/$15 scheduled from 2026-09-01, and this table recorded the scheduled rate; the vendor has since made $2/$10 the standard price and cancelled the increase. The table has no time dimension, so a future promotional window is still recorded at the list rate and over-reported rather than mis-modelled.",
 			"models": [
 				{
@@ -66,14 +78,16 @@
 					"inputPer1M": 8,
 					"outputPer1M": 40,
 					"cacheReadPer1M": 0.2,
-					"cacheWritePer1M": 10
+					"cacheWritePer1M": 10,
+					"fastMultiplier": 2
 				},
 				{
 					"id": "claude-mythos-5-2",
 					"inputPer1M": 8,
 					"outputPer1M": 40,
 					"cacheReadPer1M": 0.2,
-					"cacheWritePer1M": 10
+					"cacheWritePer1M": 10,
+					"fastMultiplier": 2
 				},
 				{
 					"id": "claude-fable-5-1",
@@ -94,7 +108,8 @@
 					"inputPer1M": 4,
 					"outputPer1M": 20,
 					"cacheReadPer1M": 0.2,
-					"cacheWritePer1M": 5
+					"cacheWritePer1M": 5,
+					"fastMultiplier": 2
 				},
 				{
 					"id": "claude-sonnet-5-5",
@@ -122,14 +137,16 @@
 					"inputPer1M": 5,
 					"outputPer1M": 25,
 					"cacheReadPer1M": 0.5,
-					"cacheWritePer1M": 6.25
+					"cacheWritePer1M": 6.25,
+					"fastMultiplier": 2
 				},
 				{
 					"id": "claude-opus-4-8",
 					"inputPer1M": 5,
 					"outputPer1M": 25,
 					"cacheReadPer1M": 0.5,
-					"cacheWritePer1M": 6.25
+					"cacheWritePer1M": 6.25,
+					"fastMultiplier": 2
 				},
 				{
 					"id": "claude-opus-4-7",

package/src/provider/fallback.ts

@@ -22,10 +22,12 @@
  */
 
 import { GENAI } from '../constants/telemetry/index.js'
+import { DEFAULT_MAX_REQUEST_RICH_CONTENT_BYTES } from '../runtime/query/request-rich-content.js'
 import { classifyProviderError, isAbortError } from '../types/provider/errors.js'
 import type {
 	ChatCompletionParams,
 	LLMProvider,
+	ProviderCapabilities,
 	ReasoningEffort,
 	StreamChunk,
 	ThinkingConfig,
@@ -513,7 +515,7 @@ export function withProviderFallback(
 			return first.provider.name
 		},
 		get capabilities() {
-			return first.provider.capabilities
+			return chainCapabilities(members, first.provider.capabilities)
 		},
 		chatStream,
 		...(first.provider.listModels
@@ -604,3 +606,33 @@ export function withProviderFallback(
 			: {}),
 	} as LLMProvider
 }
+
+/**
+ * The head's declarations, with the request rich-content budget narrowed to
+ * what EVERY member can take: a request is built once and may be served by any
+ * of them, so a head with a large budget must not send a fallback more than it
+ * accepts. A member that declares nothing counts as the kernel default. When
+ * no member declares one the head's capabilities are returned untouched.
+ */
+function chainCapabilities(
+	members: readonly { readonly provider: Pick<LLMProvider, 'capabilities'> }[],
+	head: ProviderCapabilities | undefined,
+): ProviderCapabilities | undefined {
+	if (!members.some((m) => m.provider.capabilities?.maxRequestRichContentBytes !== undefined)) {
+		return head
+	}
+	const budget = Math.min(
+		...members.map(
+			(m) =>
+				m.provider.capabilities?.maxRequestRichContentBytes ??
+				DEFAULT_MAX_REQUEST_RICH_CONTENT_BYTES,
+		),
+	)
+	return {
+		supportsTools: true,
+		supportsStreaming: true,
+		supportsFunctionCalling: true,
+		...head,
+		maxRequestRichContentBytes: budget,
+	}
+}

package/src/public-runtime.ts

@@ -478,7 +478,7 @@ export type {
 // did nothing.
 export { SESSION_MEMORY_TAG, createMemoryPromoter } from './turn/index.js'
 export type { MemoryPromoterOptions, SessionMemoryCategory } from './turn/index.js'
-export { createMemoryRecallStep } from './turn/memory-recall.js'
+export { createMemoryRecallStep, recalledMemoryIds } from './turn/memory-recall.js'
 export type {
 	MemoryRecallOptions,
 	MemorySelection,

package/src/runtime/query/auto-review/classifier.ts

@@ -38,6 +38,7 @@ import {
 } from '../../../utils/backoff.js'
 import { SCOPE_ATTRIBUTE } from '../../../utils/log/types.js'
 import { type Logger, resolveLogger } from '../../../utils/logger.js'
+import { chatStreamWithSamplingRepair } from '../iteration/side-call.js'
 import {
 	STAGE_ONE_INSTRUCTION,
 	STAGE_TWO_INSTRUCTION,
@@ -209,7 +210,7 @@ export function createAutoReviewClassifier(
 				// ignores it would otherwise hold the call open past its deadline, and
 				// a turn cancelled while it waits would wait with it.
 				const collected = collectChatCompletion(
-					options.provider.chatStream({
+					chatStreamWithSamplingRepair(options.provider, {
 						model: options.model,
 						messages,
 						maxTokens: stage === 1 ? 64 : 8_192,

package/src/runtime/query/index.ts

@@ -393,6 +393,29 @@ export interface QueryParams {
 	 */
 	refusalFallback?: 'auto' | 'off'
 
+	/**
+	 * Recorded-context keys (see `RecordedContextItem.key`) to leave out of the
+	 * request when the provider refuses it. On a refusal, before any model
+	 * fallback, the loop re-runs the step once with every entry filed under one
+	 * of these keys left out of the request. The history itself is not changed,
+	 * so nothing is lost for a later turn. The refused request is announced by a
+	 * `model_refusal` event with `outcome: 'withheld'`, and the steps that
+	 * follow, in this turn, see the keys in `PrepareStepContext.withheldKeys` so
+	 * a stage can stop adding what was just set aside. When the entries are
+	 * recalled memories the event carries their ids. A refusal of the retry
+	 * falls through to `refusalFallback` unchanged. Nothing happens when the
+	 * request carried no entry under a listed key. Absent means none.
+	 */
+	withholdOnRefusal?: readonly string[]
+
+	/**
+	 * Recorded-context keys to leave out of every request of this turn from the
+	 * first one, for a host that remembers an earlier `outcome: 'withheld'`
+	 * refusal and does not want the next turn refused again. History is not
+	 * changed. Seen by stages as `PrepareStepContext.withheldKeys`.
+	 */
+	withheldKeys?: readonly string[]
+
 	/**
 	 * Model-visible size cap for a single tool result. Over-budget output is
 	 * spilled to the turn directory and replaced with a head+tail preview
@@ -1918,6 +1941,10 @@ export async function* query(params: QueryParams): AsyncGenerator<SessionEvent,
 				? { maxOutputContinuations: params.maxOutputContinuations }
 				: {}),
 			...(params.refusalFallback !== undefined ? { refusalFallback: params.refusalFallback } : {}),
+			...(params.withholdOnRefusal && params.withholdOnRefusal.length > 0
+				? { withholdOnRefusal: [...params.withholdOnRefusal] }
+				: {}),
+			withheldKeys: new Set(params.withheldKeys ?? []),
 			...(params.inboundMessages ? { inboundMessages: params.inboundMessages } : {}),
 			...(params.reminders ? { reminders: params.reminders } : {}),
 			...(params.durableInbound ? { durableInbound: params.durableInbound } : {}),

package/src/runtime/query/iteration/index.ts

@@ -30,6 +30,7 @@ import {
 import { getTracer } from '../../../telemetry/runtime-accessors.js'
 import { STRUCTURED_OUTPUT_TOOL_NAME } from '../../../tools/builtins/structuredOutput.js'
 import type { ToolManager } from '../../../toolsets/manager.js'
+import { RECALLED_MEMORY_ENTRY_KEY, recalledMemoryIds } from '../../../turn/memory-recall.js'
 import type { CostInfo, TokenUsage } from '../../../types/common/index.js'
 import { NamzuError } from '../../../types/errors/index.js'
 import type { MessageId, SessionId, TurnId } from '../../../types/ids/index.js'
@@ -40,6 +41,8 @@ import {
 	createRuntimeContextMessage,
 	createSystemMessage,
 	createToolMessage,
+	turnContextEntries,
+	turnContextKey,
 } from '../../../types/message/index.js'
 import { classifyProviderError, readOutputOverflow } from '../../../types/provider/errors.js'
 import type { ChatCompletionResponse, ReasoningEffort } from '../../../types/provider/index.js'
@@ -69,10 +72,10 @@ import {
 	snapshotRequestContext,
 } from '../request-context.js'
 import {
-	DEFAULT_MAX_REQUEST_RICH_CONTENT_BYTES,
 	type RequestImageIdentity,
 	markProviderRejectedImage,
 	projectRequestRichContent,
+	resolveMaxRequestRichContentBytes,
 } from '../request-rich-content.js'
 import { withoutVoidAssistantTurns } from '../request-void-turns.js'
 import { formatSteeringNote, isOperatorUserMessage } from '../steering.js'
@@ -255,6 +258,20 @@ export class IterationOrchestrator {
 	}
 
 	/** Whether a request for `model` may carry images: the driver's catalogue and this run's own evidence agree. */
+	/**
+	 * The request rich-content budget, resolved each time it is read: the
+	 * caller's `turnConfig` override if one was given, else what the provider
+	 * serving now declares. Never written back into `turnConfig`, which would
+	 * turn the first provider's declaration into an override that a resumed
+	 * turn, a nested agent or a later provider then inherits.
+	 */
+	private richContentBudget(): number {
+		return resolveMaxRequestRichContentBytes(
+			this.ctx.turnConfig.maxRequestRichContentBytes,
+			this.ctx.provider.capabilities?.maxRequestRichContentBytes,
+		)
+	}
+
 	private modelReadsImages(model: string): boolean {
 		const serving = this.servingModel(model)
 		return (
@@ -332,6 +349,8 @@ export class IterationOrchestrator {
 	private refusalFallbackUsed = false
 	/** The request has already been re-run on the same model once this turn. */
 	private refusalRetryUsed = false
+	/** The request has already been re-run once without its withheld recorded entries. */
+	private refusalWithholdUsed = false
 	/**
 	 * `turn_end` (and, for a child, `subagent_stop`) fired when the model
 	 * finished and no hook asked for more. `finalizeTurn` then does not fire
@@ -898,7 +917,16 @@ export class IterationOrchestrator {
 							turnId: recorder.turnId,
 							memories: recalled,
 						})
-					const baseMessages = recorder.messages
+					const withheld = this.ctx.withheldKeys
+					// A withheld key is left out of the request only: the history keeps
+					// its entries, so a later turn or a restore still has them.
+					const baseMessages =
+						withheld && withheld.size > 0
+							? recorder.messages.filter((message) => {
+									const key = turnContextKey(message)
+									return key === undefined || !withheld.has(key)
+								})
+							: recorder.messages
 
 					// Step guidance is appended to the REQUEST, never pushed onto
 					// the turn's history: it applies to this step only, and pushing
@@ -967,8 +995,7 @@ export class IterationOrchestrator {
 					if (policyNotice) requestHistory.push(stepContextMessage(policyNotice))
 					const projected = projectRequestRichContent(
 						withoutVoidAssistantTurns(this.projectObservations(requestHistory)),
-						this.ctx.turnConfig.maxRequestRichContentBytes ??
-							DEFAULT_MAX_REQUEST_RICH_CONTENT_BYTES,
+						this.richContentBudget(),
 						{ imagesReadable: this.modelReadsImages(stepModel) },
 					)
 					// Request-only data (the context contributions, notes about the
@@ -1197,6 +1224,7 @@ export class IterationOrchestrator {
 							cacheControl: { type: 'auto' },
 							...(turnConfig.thinking ? { thinking: turnConfig.thinking } : {}),
 							...(stepEffort ? { effort: stepEffort } : {}),
+							...(turnConfig.speed ? { speed: turnConfig.speed } : {}),
 							...(!forceFinalize && turnConfig.webSearch
 								? { webSearch: turnConfig.webSearch }
 								: {}),
@@ -1355,6 +1383,9 @@ export class IterationOrchestrator {
 									},
 								}
 							: {}),
+						...(response.usage.cacheMiss !== undefined
+							? { cacheMiss: response.usage.cacheMiss }
+							: {}),
 						...contextFigures,
 					})
 
@@ -3180,6 +3211,42 @@ export class IterationOrchestrator {
 			return 'settled'
 		}
 
+		const withholdKeys = this.ctx.withholdOnRefusal
+		if (
+			!input.forceFinalize &&
+			!this.refusalWithholdUsed &&
+			withholdKeys &&
+			this.ctx.withheldKeys
+		) {
+			const carried = withholdKeys.filter(
+				(key) =>
+					!this.ctx.withheldKeys?.has(key) &&
+					recorder.messages.some((message) => turnContextKey(message) === key),
+			)
+			if (carried.length > 0) {
+				this.refusalWithholdUsed = true
+				for (const key of carried) this.ctx.withheldKeys.add(key)
+				const memoryIds = carried.includes(RECALLED_MEMORY_ENTRY_KEY)
+					? recalledMemoryIds(turnContextEntries(recorder.messages, RECALLED_MEMORY_ENTRY_KEY))
+					: []
+				await this.ctx.emitEvent({
+					type: 'model_refusal',
+					turnId: recorder.turnId,
+					iteration: iterationNum,
+					model: servedBy.model,
+					providerId: servedBy.providerId,
+					outcome: 'withheld',
+					withheldKeys: carried,
+					...(memoryIds.length > 0 ? { withheldMemoryIds: memoryIds } : {}),
+					...(refusal.category ? { category: refusal.category } : {}),
+					...(refusal.explanation ? { explanation: refusal.explanation } : {}),
+				})
+				await closeIteration()
+				yield* this.ctx.drainPending()
+				return 'retry'
+			}
+		}
+
 		if (!input.forceFinalize && !this.refusalFallbackUsed && this.ctx.refusalFallback !== 'off') {
 			const next = await this.refusalFallbackModel(input.stepModel, refusal)
 			if (next !== undefined) {
@@ -3369,7 +3436,7 @@ export class IterationOrchestrator {
 			]
 			const finalMessages = projectRequestRichContent(
 				withoutVoidAssistantTurns(this.projectObservations(finalHistory)),
-				this.ctx.turnConfig.maxRequestRichContentBytes ?? DEFAULT_MAX_REQUEST_RICH_CONTENT_BYTES,
+				this.richContentBudget(),
 				{ imagesReadable: this.modelReadsImages(model) },
 			)
 			// The closing directive is the last thing the model reads: after
@@ -3411,6 +3478,7 @@ export class IterationOrchestrator {
 				// is exactly the shape a field goes missing from — so it is tested
 				// separately rather than assumed to have been kept in step.
 				...(this.ctx.turnConfig.effort ? { effort: this.ctx.turnConfig.effort } : {}),
+				...(this.ctx.turnConfig.speed ? { speed: this.ctx.turnConfig.speed } : {}),
 				// Cancellable too: a Stop during the closing summary must not
 				// stream to completion.
 				signal: this.ctx.abortController.signal,

package/src/runtime/query/iteration/phases/context.ts

@@ -226,6 +226,10 @@ export interface IterationContext {
 	readonly maxOutputContinuations?: number
 	/** See `QueryParams.refusalFallback`. */
 	readonly refusalFallback?: 'auto' | 'off'
+	/** See `QueryParams.withholdOnRefusal`. */
+	readonly withholdOnRefusal?: readonly string[]
+	/** Recorded-context keys left out of the request; grows when a refusal is retried without them. */
+	readonly withheldKeys?: Set<string>
 
 	/**
 	 * What the driver said this model's context window is, resolved once.

package/src/runtime/query/iteration/request-repair.ts

@@ -39,7 +39,13 @@ import { isProviderRejectedImageError } from './provider-rejected-image.js'
  * compaction summaries, do not pass through it.
  */
 
-export type RejectedParameter = 'effort' | 'thinking' | 'temperature' | 'topP' | 'responseFormat'
+export type RejectedParameter =
+	| 'effort'
+	| 'thinking'
+	| 'temperature'
+	| 'topP'
+	| 'topK'
+	| 'responseFormat'
 
 export interface MediaRemoval {
 	readonly count: number
@@ -110,6 +116,7 @@ const PARAMETER_NAMES: Record<RejectedParameter, string> = {
 	thinking: 'thinking',
 	temperature: 'temperature',
 	topP: 'top[_ ]?p',
+	topK: 'top[_ ]?k',
 	responseFormat:
 		'(?:response_format|output_config(?:\\.format)?|json_schema|json_object|structured\\s+output)',
 }
@@ -124,6 +131,10 @@ const PARAMETER_PATTERNS = Object.fromEntries(
 	]),
 ) as Record<RejectedParameter, { named: RegExp; refused: RegExp }>
 
+const SAMPLING = ['temperature', 'topP', 'topK'] as const satisfies readonly RejectedParameter[]
+const SAMPLING_GROUP =
+	/\btemperature\b.{0,80}\btop[_ ]?[pk]\b|\btop[_ ]?[pk]\b.{0,80}\btemperature\b/i
+
 const BOTH_SAMPLING =
 	/temperature.{0,60}top[_ ]?p.{0,80}(?:cannot|can['’]t|may\s+not|should\s+not).{0,20}(?:both|together)/i
 
@@ -156,6 +167,7 @@ function rejectedParameters(
 		thinking: params.thinking !== undefined,
 		temperature: params.temperature !== undefined,
 		topP: params.topP !== undefined,
+		topK: params.topK !== undefined,
 		responseFormat: params.responseFormat !== undefined,
 	} satisfies Record<RejectedParameter, boolean>
 	const found: RejectedParameter[] = []
@@ -165,6 +177,12 @@ function rejectedParameters(
 		if (structured ? pattern.named.test(text) : pattern.refused.test(text)) found.push(name)
 	}
 	if (found.length === 0 && set.topP && BOTH_SAMPLING.test(text)) found.push('topP')
+	// The vendor words a refusal of sampling controls as a group ("Omit
+	// temperature, top_p, and top_k"), so naming one of them in a refusal that
+	// names the others is enough for whichever of them was sent.
+	if (found.length === 0 && SAMPLING_GROUP.test(text)) {
+		for (const name of SAMPLING) if (set[name]) found.push(name)
+	}
 	return found
 }
 
@@ -726,6 +744,7 @@ function withoutParameters(
 	if (names.has('thinking')) next.thinking = undefined
 	if (names.has('temperature')) next.temperature = undefined
 	if (names.has('topP')) next.topP = undefined
+	if (names.has('topK')) next.topK = undefined
 	if (names.has('responseFormat')) next.responseFormat = undefined
 	return next
 }
@@ -741,3 +760,25 @@ function isHandledElsewhere(error: unknown, params: ChatCompletionParams): boole
 	}
 	return isImageInputRefusal(error)
 }
+
+/**
+ * Which of `temperature`, `top_p` and `top_k` a provider refused on this
+ * request, from the same reading of its rejection the main turn uses. Empty
+ * for any other failure, so a caller can strip exactly these and nothing else.
+ */
+export function rejectedSamplingParameters(
+	error: unknown,
+	params: ChatCompletionParams,
+): RejectedParameter[] {
+	const rejection = classifyRequestRejection(error, params)
+	if (rejection?.kind !== 'parameter') return []
+	return rejection.parameters.filter((name) => (SAMPLING as readonly string[]).includes(name))
+}
+
+/** The request without the named sampling parameters. */
+export function withoutSamplingParameters(
+	params: ChatCompletionParams,
+	names: ReadonlySet<RejectedParameter>,
+): ChatCompletionParams {
+	return withoutParameters(params, names)
+}

package/src/runtime/query/iteration/side-call.ts

@@ -0,0 +1,60 @@
+import type {
+	ChatCompletionParams,
+	LLMProvider,
+	StreamChunk,
+} from '../../../types/provider/index.js'
+import {
+	type RejectedParameter,
+	rejectedSamplingParameters,
+	withoutSamplingParameters,
+} from './request-repair.js'
+
+/**
+ * What each model was found to refuse, for the life of the provider object.
+ * Weakly keyed so a dropped provider takes its entry with it.
+ */
+const refused = new WeakMap<object, Map<string, Set<RejectedParameter>>>()
+
+/**
+ * A model call made outside the turn loop (a compaction summary, its verifier,
+ * the auto-review classifier) that asks for a deterministic answer with
+ * `temperature: 0`. Current models refuse sampling parameters outright, and
+ * the turn loop already recovers from that; this gives side calls the same
+ * recovery instead of letting the call fail and the feature with it.
+ *
+ * The request goes out as asked. If the provider refuses a sampling parameter
+ * BEFORE any output arrived, it is resent without exactly those parameters,
+ * and the finding is kept for the model so later calls omit them up front.
+ * Nothing is kept unless the resend was accepted, and any other failure is
+ * thrown unchanged.
+ */
+export async function* chatStreamWithSamplingRepair(
+	provider: LLMProvider,
+	params: ChatCompletionParams,
+): AsyncGenerator<StreamChunk> {
+	const perModel = refused.get(provider) ?? new Map<string, Set<RejectedParameter>>()
+	refused.set(provider, perModel)
+	const known = perModel.get(params.model)
+	let request = known ? withoutSamplingParameters(params, known) : params
+	let stripped: Set<RejectedParameter> | undefined
+	for (;;) {
+		let yielded = false
+		try {
+			for await (const chunk of provider.chatStream(request)) {
+				yielded = true
+				yield chunk
+			}
+			if (stripped) {
+				const merged = new Set([...(perModel.get(params.model) ?? []), ...stripped])
+				perModel.set(params.model, merged)
+			}
+			return
+		} catch (error) {
+			if (yielded || stripped) throw error
+			const names = rejectedSamplingParameters(error, request)
+			if (names.length === 0) throw error
+			stripped = new Set(names)
+			request = withoutSamplingParameters(request, stripped)
+		}
+	}
+}

package/src/runtime/query/iteration/step-shaping.ts

@@ -162,6 +162,9 @@ export function stepContext(
 			),
 		},
 		steps: steps(),
+		...(ctx.withheldKeys && ctx.withheldKeys.size > 0
+			? { withheldKeys: [...ctx.withheldKeys] }
+			: {}),
 		prepared,
 	}
 }