@namzu/sdk 52.0.0 → 52.1.0

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

@@ -0,0 +1,743 @@
+import { createHash } from 'node:crypto'
+
+import { repairToolMessageHistory, toolHistoryRepairChanged } from '../../../compaction/dangling.js'
+import { isProviderRequestError } from '../../../provider/errors.js'
+import { isImageInputRefusal } from '../../../provider/image-input-refusal.js'
+import type {
+	AssistantMessage,
+	Message,
+	MessageAttachment,
+	ToolResultBlock,
+} from '../../../types/message/index.js'
+import type {
+	ChatCompletionParams,
+	LLMProvider,
+	StreamChunk,
+} from '../../../types/provider/index.js'
+import { findSingleRequestImage } from '../request-rich-content.js'
+import { isProviderRejectedImageError } from './provider-rejected-image.js'
+
+/**
+ * Repair-and-retry for a request a provider rejects as invalid (a 400 that is
+ * not a context overflow), so a conversation does not dead-end on it.
+ *
+ * Mirrors the reference terminal's table of silent 400 recoveries: pair tool
+ * calls with results before a send, strip rejected reasoning, drop a
+ * parameter the model refuses, and take rejected media out of the OUTGOING
+ * request. Stored history is never edited.
+ *
+ * **Nothing is remembered on a guess.** Every repair is a guess about what the
+ * provider objected to. What it teaches (a parameter to leave out, reasoning to
+ * strip, media to omit, a media limit) is kept for the run only once a resend
+ * WITHOUT the offending part was accepted. A guess that did not fix the
+ * request is forgotten and the original rejection is thrown, so a wrong guess
+ * never changes a later request.
+ *
+ * It decorates one model call of the agent loop (a turn, or the closing
+ * summary), so a rejection is only acted on when nothing has been accepted yet:
+ * resending then cannot duplicate output. Other callers of a provider, such as
+ * compaction summaries, do not pass through it.
+ */
+
+export type RejectedParameter = 'effort' | 'thinking' | 'temperature' | 'topP' | 'responseFormat'
+
+export interface MediaRemoval {
+	readonly count: number
+	readonly kind: 'image' | 'document' | 'media'
+	/** Some of it was attached to the person's latest message. */
+	readonly current: boolean
+	/** It was dropped to fit a count the provider accepts, not because a file was faulty. */
+	readonly limit: boolean
+}
+
+export interface RequestRepairHooks {
+	/** Debug-level record of a silent repair. Never shown to the person. */
+	readonly debug?: (message: string) => void
+	/** Media was taken out of a request that then went through. Called once per removal. */
+	readonly onMediaRemoved?: (removal: MediaRemoval) => void | Promise<void>
+	/**
+	 * The member that serves the next request. A fallback chain moves between
+	 * members, and what a member refused is a fact about that member, not about
+	 * the head of the chain.
+	 */
+	readonly servedBy?: () => { readonly providerId: string; readonly model?: string } | undefined
+}
+
+type Rejection =
+	| { readonly kind: 'tool-pairing' }
+	| { readonly kind: 'reasoning' }
+	| { readonly kind: 'parameter'; readonly parameters: readonly RejectedParameter[] }
+	| {
+			readonly kind: 'media'
+			readonly generic: boolean
+			readonly messageIndex?: number
+			readonly contentIndex?: number
+	  }
+
+/** Most repair resends for one original request, across every send it makes. */
+export const MAX_REPAIRS_PER_REQUEST = 5
+const MAX_STRIPS_PER_INDEX = 3
+/** The reference drops media in these steps for an unprocessable request. */
+const GENERIC_MEDIA_LIMITS: readonly number[] = [510, 255, 127]
+
+const PARAMETER_CODES: ReadonlySet<string> = new Set([
+	'unsupported_parameter',
+	'unsupported_value',
+	'unknown_parameter',
+])
+
+const REASONING_CODES: ReadonlySet<string> = new Set([
+	'invalid_encrypted_content',
+	'invalid_reasoning',
+	'invalid_signature',
+])
+
+const TOOL_PAIRING =
+	/\btool_use\b.{0,120}\btool_result\b|\btool_result\b.{0,120}\btool_use\b|unexpected\s+`?tool_use_id`?|tool_use\s+ids?\s+must\s+be\s+unique|\btool_call_id\b|role\s+'tool'\s+must\s+be\s+a\s+response|'tool_calls'\s+must\s+be\s+followed|Expected\s+toolResult\s+blocks|function\s+call\s+turn\s+comes\s+immediately|number\s+of\s+function\s+response\s+parts/i
+
+const REASONING =
+	/\bthought[_ ]?signature\b|`?signature`?\s+in\s+`?(?:thinking|redacted_thinking)|\b(?:thinking|redacted_thinking)\s+blocks?\b.{0,160}\b(?:cannot|can['’]t|must|invalid|modified|expected)|\bexpected\s+`?(?:thinking|redacted_thinking)`?|\bencrypted[_ ]content\b|\breasoning\s+item\b|\bitem\s+'rs_[^']*'|\bof\s+type\s+'reasoning'/i
+
+/** A path into the tool list: wording about a tool schema is never about a request parameter or media. */
+const TOOL_SCHEMA_PATH =
+	/\b(?:tools?|functions?)[.[]\s*\d+|input_schema|\.properties\b|json\s+schema/i
+
+const REFUSAL =
+	"(?:unsupported|not\\s+supported|does\\s?n[o']t\\s+support|do\\s+not\\s+support|unrecognized|unknown\\s+(?:request\\s+)?(?:parameter|argument|field)|not\\s+allowed|deprecated|only\\s+the\\s+default)"
+
+const PARAMETER_NAMES: Record<RejectedParameter, string> = {
+	effort: '(?:reasoning[_. -]?effort|effort)',
+	thinking: 'thinking',
+	temperature: 'temperature',
+	topP: 'top[_ ]?p',
+	responseFormat:
+		'(?:response_format|output_config(?:\\.format)?|json_schema|json_object|structured\\s+output)',
+}
+
+const PARAMETER_PATTERNS = Object.fromEntries(
+	(Object.entries(PARAMETER_NAMES) as [RejectedParameter, string][]).map(([name, token]) => [
+		name,
+		{
+			named: new RegExp(`\\b${token}\\b`, 'i'),
+			refused: new RegExp(`${REFUSAL}.{0,80}\\b${token}\\b|\\b${token}\\b.{0,80}${REFUSAL}`, 'i'),
+		},
+	]),
+) as Record<RejectedParameter, { named: RegExp; refused: RegExp }>
+
+const BOTH_SAMPLING =
+	/temperature.{0,60}top[_ ]?p.{0,80}(?:cannot|can['’]t|may\s+not|should\s+not).{0,20}(?:both|together)/i
+
+const IMAGE_FAULT =
+	/\b(?:images?|pictures?|photos?)\b.{0,100}\b(?:too\s+(?:large|big)|exceeds?|larger\s+than|maximum\s+(?:size|dimensions?)|dimensions?|(?:could|can)\s?not\s+be\s+(?:processed|decoded|loaded|read|fetched|downloaded)|failed\s+to\s+(?:process|decode|load|read|download)|unsupported|not\s+supported|corrupt\w*|malformed|truncated|invalid)\b|\b(?:(?:could|can)\s?not|failed\s+to|unable\s+to)\s+(?:process|decode|load|read|download|fetch)\b.{0,60}\b(?:images?|pictures?)\b|\bunsupported\s+(?:image|media)\b|\binvalid\s+image\b/i
+const DOCUMENT_FAULT =
+	/\bdocument_block\b|\b(?:documents?|pdfs?)\b.{0,80}\b(?:too\s+(?:large|big)|exceeds?\s+(?:the\s+)?(?:maximum|size|page)|page\s+limit|password|corrupt\w*|(?:could|can)\s?not\s+be\s+(?:processed|read|decoded)|failed\s+to\s+(?:process|read|decode))\b|\bpassword[- ]protected\s+pdf\b|\bpdf\b.{0,40}\bpages?\b.{0,20}\b(?:limit|maximum|exceeds?)\b/i
+const GENERIC_MEDIA =
+	/\bunprocessable\b.{0,100}\b(?:images?|media|documents?)\b|\b(?:images?|media|documents?)\b.{0,100}\bunprocessable\b|\btoo\s+many\s+(?:images?|documents?|media|attachments?)\b|\bmedia_budget\b|\bimage_block\b/i
+const NAMES_MEDIA = /\b(?:images?|pictures?|photos?|documents?|pdfs?|media|attachments?)\b/i
+
+function wireIndexes(text: string): { messageIndex?: number; contentIndex?: number } {
+	const both = /\bmessages?[.[]\s*(\d+)\s*\]?[.[]\s*content[.[]\s*(\d+)/i.exec(text)
+	if (both?.[1] !== undefined && both[2] !== undefined) {
+		return { messageIndex: Number(both[1]), contentIndex: Number(both[2]) }
+	}
+	const only = /\bmessages?[.[]\s*(\d+)\b/i.exec(text)
+	return only?.[1] !== undefined ? { messageIndex: Number(only[1]) } : {}
+}
+
+function rejectedParameters(
+	text: string,
+	code: string | undefined,
+	params: ChatCompletionParams,
+): RejectedParameter[] {
+	if (TOOL_SCHEMA_PATH.test(text)) return []
+	const structured = code !== undefined && PARAMETER_CODES.has(code)
+	const set = {
+		effort: params.effort !== undefined,
+		thinking: params.thinking !== undefined,
+		temperature: params.temperature !== undefined,
+		topP: params.topP !== undefined,
+		responseFormat: params.responseFormat !== undefined,
+	} satisfies Record<RejectedParameter, boolean>
+	const found: RejectedParameter[] = []
+	for (const name of Object.keys(set) as RejectedParameter[]) {
+		if (!set[name]) continue
+		const pattern = PARAMETER_PATTERNS[name]
+		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')
+	return found
+}
+
+function requestHasMedia(params: ChatCompletionParams): boolean {
+	return listMedia(params.messages).length > 0
+}
+
+/**
+ * What kind of invalid request this is, from the provider's structured code
+ * where it sent one and from its wording only as a fallback.
+ */
+export function classifyRequestRejection(
+	error: unknown,
+	params: ChatCompletionParams,
+): Rejection | null {
+	if (!isProviderRequestError(error) || error.kind !== 'bad_request') return null
+	if (
+		error.status !== undefined &&
+		error.status !== 400 &&
+		error.status !== 413 &&
+		error.status !== 422
+	) {
+		return null
+	}
+	const text = error.detail ?? error.message
+	const code = error.providerCode
+
+	if (code !== undefined && REASONING_CODES.has(code)) return { kind: 'reasoning' }
+	if (REASONING.test(text)) return { kind: 'reasoning' }
+	if (TOOL_PAIRING.test(text)) return { kind: 'tool-pairing' }
+
+	const parameters = rejectedParameters(text, code, params)
+	if (parameters.length > 0) return { kind: 'parameter', parameters }
+
+	if (TOOL_SCHEMA_PATH.test(text)) return null
+	if (code === 'invalid_image') return { kind: 'media', generic: false, ...wireIndexes(text) }
+	if (GENERIC_MEDIA.test(text)) return { kind: 'media', generic: true }
+	// A bare 413 or 422 says nothing about media: it counts only when the
+	// request carries some and the provider's words name it.
+	if ((error.status === 413 || error.status === 422) && NAMES_MEDIA.test(text)) {
+		return requestHasMedia(params) ? { kind: 'media', generic: true } : null
+	}
+	if (IMAGE_FAULT.test(text) || DOCUMENT_FAULT.test(text)) {
+		return { kind: 'media', generic: false, ...wireIndexes(text) }
+	}
+	return null
+}
+
+function isAccepted(chunk: StreamChunk): boolean {
+	if (chunk.retry !== undefined || chunk.fallback !== undefined || chunk.error !== undefined) {
+		return false
+	}
+	return Boolean(
+		chunk.delta.content ||
+			chunk.delta.toolCalls?.length ||
+			chunk.delta.toolCallEnd ||
+			chunk.delta.reasoning ||
+			chunk.delta.citation ||
+			chunk.finishReason !== undefined ||
+			chunk.usage !== undefined ||
+			chunk.replayState !== undefined ||
+			chunk.textParts !== undefined,
+	)
+}
+
+const mediaFingerprint = (data: string, mediaType: string): string =>
+	createHash('sha256').update(mediaType).update('\0').update(data).digest('hex')
+
+const FAULT_NOTE =
+	'the provider could not process it. The file may be unsupported or corrupt; do not retry reading it. If you need to inspect it, use a shell command instead.'
+const LIMIT_NOTE =
+	'this conversation holds more files than the provider accepts at once, so the oldest are left out of the request. They are still in the conversation.'
+const removedNote = (kind: 'image' | 'document', limit: boolean): string =>
+	`[${kind} removed: ${limit ? LIMIT_NOTE : FAULT_NOTE}]`
+
+interface MediaSlot {
+	readonly messageIndex: number
+	readonly itemIndex: number
+	readonly source: 'user' | 'tool'
+	readonly kind: 'image' | 'document'
+	readonly fingerprint: string
+}
+
+function listMedia(messages: readonly Message[]): MediaSlot[] {
+	const slots: MediaSlot[] = []
+	for (const [messageIndex, message] of messages.entries()) {
+		if (message.role === 'user') {
+			for (const [itemIndex, attachment] of (message.attachments ?? []).entries()) {
+				if (attachment.type === 'stored') continue
+				slots.push({
+					messageIndex,
+					itemIndex,
+					source: 'user',
+					kind: attachment.type === 'document' ? 'document' : 'image',
+					fingerprint: mediaFingerprint(attachment.data, attachment.mediaType),
+				})
+			}
+		} else if (message.role === 'tool' && Array.isArray(message.content)) {
+			for (const [itemIndex, block] of message.content.entries()) {
+				if (block.type !== 'image' && block.type !== 'document') continue
+				slots.push({
+					messageIndex,
+					itemIndex,
+					source: 'tool',
+					kind: block.type,
+					fingerprint: mediaFingerprint(block.data, block.mediaType),
+				})
+			}
+		}
+	}
+	return slots
+}
+
+/** Replace the given media with a short text note, in a copy. */
+function removeMedia(
+	messages: readonly Message[],
+	slots: readonly MediaSlot[],
+	limit: boolean,
+): Message[] {
+	if (slots.length === 0) return messages as Message[]
+	const byMessage = new Map<number, Map<number, MediaSlot>>()
+	for (const slot of slots) {
+		const items = byMessage.get(slot.messageIndex) ?? new Map<number, MediaSlot>()
+		items.set(slot.itemIndex, slot)
+		byMessage.set(slot.messageIndex, items)
+	}
+	return messages.map((message, messageIndex) => {
+		const items = byMessage.get(messageIndex)
+		if (!items) return message
+		if (message.role === 'user') {
+			const attachments: readonly MessageAttachment[] = message.attachments ?? []
+			const kept = attachments.filter((_a, i) => !items.has(i))
+			const notes = [...items.values()]
+				.sort((a, b) => a.itemIndex - b.itemIndex)
+				.map((slot) => removedNote(slot.kind, limit))
+				.join('\n\n')
+			const { attachments: _drop, ...rest } = message
+			return {
+				...rest,
+				content: message.content.length > 0 ? `${message.content}\n\n${notes}` : notes,
+				...(kept.length > 0 ? { attachments: kept } : {}),
+			}
+		}
+		if (message.role === 'tool' && Array.isArray(message.content)) {
+			const content: ToolResultBlock[] = message.content.map((block, i) => {
+				const slot = items.get(i)
+				return slot ? { type: 'text', text: removedNote(slot.kind, limit) } : block
+			})
+			return { ...message, content }
+		}
+		return message
+	})
+}
+
+function stripReasoning(messages: readonly Message[]): Message[] {
+	let changed = false
+	const out = messages.map((message) => {
+		if (message.role !== 'assistant') return message
+		const assistant = message as AssistantMessage
+		const hasState =
+			assistant.source !== undefined &&
+			(assistant.source as { replayState?: unknown }).replayState !== undefined
+		if (assistant.reasoning === undefined && !hasState) return message
+		changed = true
+		const { reasoning: _reasoning, source, ...rest } = assistant
+		if (source === undefined) return rest as Message
+		const { replayState: _state, ...route } = source as typeof source & {
+			replayState?: unknown
+		}
+		return { ...rest, source: route } as Message
+	})
+	return changed ? out : (messages as Message[])
+}
+
+/**
+ * The media a provider's `messages.N.content.M` points at, mapped back through
+ * that driver's own serialisation. Exact for the two wires whose layout is
+ * known; anything else returns undefined and the caller does not guess.
+ *
+ * - the content-block wire: system turns are left out; a run of consecutive tool results
+ *   is ONE user entry whose blocks are those results; every other message,
+ *   including a user turn that follows tool results, is its own entry; a user
+ *   entry's blocks are its text (when not empty) and then each attachment.
+ * - the chat-completions wire: one entry per message, system included; only user attachments
+ *   travel as media (tool results are sent as text); blocks are the text (when
+ *   not empty) and then each attachment.
+ */
+function mediaAtWireIndex(
+	messages: readonly Message[],
+	slots: readonly MediaSlot[],
+	family: string,
+	wireIndex: number,
+	contentIndex: number | undefined,
+): MediaSlot[] | undefined {
+	if (family !== 'anthropic' && family !== 'openai') return undefined
+	const entries: number[][] = []
+	let inToolRun = false
+	for (const [index, message] of messages.entries()) {
+		if (family === 'anthropic') {
+			if (message.role === 'system') continue
+			if (message.role === 'tool') {
+				if (!inToolRun) entries.push([])
+				inToolRun = true
+				entries[entries.length - 1]?.push(index)
+				continue
+			}
+			inToolRun = false
+		}
+		entries.push([index])
+	}
+	const entry = entries[wireIndex]
+	if (entry === undefined) return undefined
+	const mine = slots.filter((slot) => entry.includes(slot.messageIndex))
+	if (mine.length === 0) return undefined
+	if (contentIndex === undefined) return mine
+	const first = entry[0]
+	const message = first === undefined ? undefined : messages[first]
+	if (message?.role === 'user' && !(family === 'openai' && mine.some((s) => s.source === 'tool'))) {
+		const itemIndex = contentIndex - (message.content.length > 0 ? 1 : 0)
+		const hit = mine.filter((slot) => slot.itemIndex === itemIndex)
+		return hit.length > 0 ? hit : mine
+	}
+	if (family === 'anthropic' && message?.role === 'tool') {
+		const owner = entry[contentIndex]
+		const hit = owner === undefined ? [] : mine.filter((slot) => slot.messageIndex === owner)
+		return hit.length > 0 ? hit : mine
+	}
+	return mine
+}
+
+interface Learned {
+	readonly latched: Map<string, Set<RejectedParameter>>
+	readonly reasoningOff: Set<string>
+	readonly rejectedMedia: Set<string>
+	mediaLimit: number | undefined
+}
+
+const emptyLearned = (): Learned => ({
+	latched: new Map(),
+	reasoningOff: new Set(),
+	rejectedMedia: new Set(),
+	mediaLimit: undefined,
+})
+
+/** The budget and the unconfirmed guesses of one original request, shared by every send it makes. */
+export interface RepairSession {
+	readonly spent: Set<string>
+	readonly stripsByIndex: Map<number, number>
+	pending: Learned
+	repairs: number
+	removal: MediaRemoval | undefined
+}
+
+export class RequestRepair {
+	private readonly learned: Learned = emptyLearned()
+
+	constructor(private readonly hooks: RequestRepairHooks = {}) {}
+
+	/** A fresh budget for one original request. */
+	session(): RepairSession {
+		return {
+			spent: new Set(),
+			stripsByIndex: new Map(),
+			pending: emptyLearned(),
+			repairs: 0,
+			removal: undefined,
+		}
+	}
+
+	private served(
+		provider: Pick<LLMProvider, 'id'>,
+		params: ChatCompletionParams,
+	): { providerId: string; model: string } {
+		const member = this.hooks.servedBy?.()
+		return {
+			providerId: member?.providerId ?? params.providerRoute?.providerId ?? provider.id,
+			model: member?.model ?? params.providerRoute?.model ?? params.model,
+		}
+	}
+
+	private routeKey(provider: Pick<LLMProvider, 'id'>, params: ChatCompletionParams): string {
+		const { providerId, model } = this.served(provider, params)
+		return `${providerId}|${model}`
+	}
+
+	/** Everything already confirmed, then the tool pairing the provider requires. */
+	private prepare(
+		provider: Pick<LLMProvider, 'id'>,
+		params: ChatCompletionParams,
+	): ChatCompletionParams {
+		const key = this.routeKey(provider, params)
+		let next: ChatCompletionParams = params
+
+		const latched = this.learned.latched.get(key)
+		if (latched) next = withoutParameters(next, latched)
+
+		let messages = next.messages
+		if (this.learned.reasoningOff.has(key)) messages = stripReasoning(messages)
+		messages = this.applyMediaMemory(messages)
+		messages = this.pair(messages)
+		return messages === next.messages ? next : { ...next, messages }
+	}
+
+	private applyMediaMemory(messages: Message[]): Message[] {
+		const { rejectedMedia, mediaLimit } = this.learned
+		if (rejectedMedia.size === 0 && mediaLimit === undefined) return messages
+		const slots = listMedia(messages)
+		const faulty = slots.filter((slot) => rejectedMedia.has(slot.fingerprint))
+		let out = removeMedia(messages, faulty, false)
+		if (mediaLimit !== undefined) {
+			const rest = listMedia(out)
+			if (rest.length > mediaLimit) {
+				out = removeMedia(out, rest.slice(0, rest.length - mediaLimit), true)
+			}
+		}
+		return out
+	}
+
+	/** The shared tool-history repair, on the outgoing copy. Duplicate call ids fail closed there, so the request goes out as it is. */
+	private pair(messages: Message[]): Message[] {
+		try {
+			const repaired = repairToolMessageHistory(messages)
+			if (!toolHistoryRepairChanged(repaired.report)) return messages
+			this.hooks.debug?.(`Repaired tool pairing before sending: ${JSON.stringify(repaired.report)}`)
+			return repaired.messages
+		} catch (error) {
+			this.hooks.debug?.(
+				`Tool pairing could not be repaired before sending: ${error instanceof Error ? error.message : String(error)}`,
+			)
+			return messages
+		}
+	}
+
+	/** What a resend that was accepted has proved. */
+	private commit(session: RepairSession): void {
+		const { pending } = session
+		for (const [key, names] of pending.latched) {
+			const set = this.learned.latched.get(key) ?? new Set<RejectedParameter>()
+			for (const name of names) set.add(name)
+			this.learned.latched.set(key, set)
+		}
+		for (const key of pending.reasoningOff) this.learned.reasoningOff.add(key)
+		for (const fingerprint of pending.rejectedMedia) this.learned.rejectedMedia.add(fingerprint)
+		if (pending.mediaLimit !== undefined) {
+			this.learned.mediaLimit = Math.min(
+				this.learned.mediaLimit ?? pending.mediaLimit,
+				pending.mediaLimit,
+			)
+		}
+		session.pending = emptyLearned()
+	}
+
+	/** The request to send after `rejection`, or null when nothing more can be tried. */
+	private next(
+		provider: Pick<LLMProvider, 'id'>,
+		params: ChatCompletionParams,
+		rejection: Rejection,
+		session: RepairSession,
+	): ChatCompletionParams | null {
+		const { spent, pending } = session
+		const key = this.routeKey(provider, params)
+		switch (rejection.kind) {
+			case 'tool-pairing': {
+				if (spent.has('tool-pairing')) return null
+				spent.add('tool-pairing')
+				const repaired = repairOnce(params.messages)
+				if (repaired === null) return null
+				this.hooks.debug?.(`Retrying after a tool-history rejection (${key})`)
+				return { ...params, messages: repaired }
+			}
+			case 'reasoning': {
+				if (spent.has('reasoning')) return null
+				spent.add('reasoning')
+				const messages = stripReasoning(params.messages)
+				if (messages === params.messages) return null
+				pending.reasoningOff.add(key)
+				let next: ChatCompletionParams = { ...params, messages }
+				// With thinking on, a tool turn that carries no thinking block is
+				// itself refused ("Expected `thinking` … but found `tool_use`"), so
+				// the two go together: neither is asked for in this resend.
+				if (params.thinking !== undefined && params.thinking.type !== 'disabled') {
+					next = withoutParameters(next, new Set<RejectedParameter>(['thinking']))
+					const names = pending.latched.get(key) ?? new Set<RejectedParameter>()
+					names.add('thinking')
+					pending.latched.set(key, names)
+				}
+				this.hooks.debug?.(`Retrying without reasoning blocks (${key})`)
+				return next
+			}
+			case 'parameter': {
+				const fresh = rejection.parameters.filter((name) => !spent.has(`param:${name}`))
+				if (fresh.length === 0) return null
+				for (const name of fresh) spent.add(`param:${name}`)
+				const names = pending.latched.get(key) ?? new Set<RejectedParameter>()
+				for (const name of fresh) names.add(name)
+				pending.latched.set(key, names)
+				this.hooks.debug?.(
+					`Retrying without parameters the model refused (${key}): ${fresh.join(', ')}`,
+				)
+				return withoutParameters(params, new Set(fresh))
+			}
+			case 'media':
+				return this.nextWithoutMedia(provider, params, rejection, session)
+		}
+	}
+
+	private nextWithoutMedia(
+		provider: Pick<LLMProvider, 'id'>,
+		params: ChatCompletionParams,
+		rejection: Extract<Rejection, { kind: 'media' }>,
+		session: RepairSession,
+	): ChatCompletionParams | null {
+		const { spent, stripsByIndex, pending } = session
+		const slots = listMedia(params.messages)
+		if (slots.length === 0) return null
+
+		let victims: MediaSlot[] = []
+		let limited = false
+		if (!rejection.generic) {
+			if (rejection.messageIndex !== undefined) {
+				const used = stripsByIndex.get(rejection.messageIndex) ?? 0
+				if (used < MAX_STRIPS_PER_INDEX) {
+					stripsByIndex.set(rejection.messageIndex, used + 1)
+					victims =
+						mediaAtWireIndex(
+							params.messages,
+							slots,
+							this.served(provider, params).providerId,
+							rejection.messageIndex,
+							rejection.contentIndex,
+						) ?? []
+				}
+			}
+			if (victims.length === 0) {
+				// No exact index: the newest message that carries media.
+				const newest = slots[slots.length - 1]
+				if (newest === undefined) return null
+				const count = (stripsByIndex.get(-1) ?? 0) + 1
+				if (count > MAX_STRIPS_PER_INDEX) return null
+				stripsByIndex.set(-1, count)
+				victims = slots.filter((slot) => slot.messageIndex === newest.messageIndex)
+			}
+		} else {
+			// The reference's steps, then halving for smaller requests, down to
+			// none: each step keeps fewer than the request held, newest first.
+			const limit =
+				GENERIC_MEDIA_LIMITS.find((candidate) => candidate < slots.length) ??
+				Math.floor(slots.length / 2)
+			spent.add(`limit:${limit}`)
+			pending.mediaLimit = Math.min(pending.mediaLimit ?? limit, limit)
+			victims = slots.slice(0, slots.length - limit)
+			limited = true
+		}
+
+		// A count limit teaches a limit, not that these files are faulty.
+		if (!limited) for (const slot of victims) pending.rejectedMedia.add(slot.fingerprint)
+		this.hooks.debug?.(`Retrying without ${victims.length} media the provider rejected`)
+
+		let lastUser = -1
+		for (const [index, message] of params.messages.entries()) {
+			if (message.role === 'user') lastUser = index
+		}
+		const kinds = new Set(victims.map((slot) => slot.kind))
+		const removal: MediaRemoval = {
+			count: victims.length,
+			kind: kinds.size === 1 ? ([...kinds][0] as 'image' | 'document') : 'media',
+			current: victims.some((slot) => slot.source === 'user' && slot.messageIndex === lastUser),
+			limit: limited,
+		}
+		const before = session.removal
+		session.removal = before
+			? {
+					count: before.count + removal.count,
+					kind: before.kind === removal.kind ? before.kind : 'media',
+					current: before.current || removal.current,
+					limit: before.limit && removal.limit,
+				}
+			: removal
+		return { ...params, messages: removeMedia(params.messages, victims, limited) }
+	}
+
+	/**
+	 * Send `params`, repairing and resending a request the provider rejects as
+	 * invalid until it is accepted or nothing more can be changed. The original
+	 * rejection is thrown when nothing applies, and what the guesses taught is
+	 * forgotten unless a resend was accepted.
+	 */
+	async *stream(
+		provider: LLMProvider,
+		params: ChatCompletionParams,
+		session: RepairSession = this.session(),
+	): AsyncIterable<StreamChunk> {
+		let current = this.prepare(provider, params)
+		let firstError: unknown
+
+		for (;;) {
+			let produced = false
+			let reported = false
+			try {
+				for await (const chunk of provider.chatStream(current)) {
+					if (isAccepted(chunk)) {
+						produced = true
+						this.commit(session)
+						if (session.removal !== undefined && !reported) {
+							reported = true
+							const removal = session.removal
+							session.removal = undefined
+							await this.hooks.onMediaRemoved?.(removal)
+						}
+					}
+					yield chunk
+				}
+				this.commit(session)
+				if (session.removal !== undefined && !reported) {
+					const removal = session.removal
+					session.removal = undefined
+					await this.hooks.onMediaRemoved?.(removal)
+				}
+				return
+			} catch (error) {
+				if (current.signal?.aborted) throw current.signal.reason
+				if (produced) throw error
+				firstError ??= error
+				const forget = (): never => {
+					session.pending = emptyLearned()
+					session.removal = undefined
+					throw firstError
+				}
+				if (session.repairs >= MAX_REPAIRS_PER_REQUEST) return forget()
+				if (isHandledElsewhere(error, current)) return forget()
+				const rejection = classifyRequestRejection(error, current)
+				if (rejection === null) return forget()
+				const step = this.next(provider, current, rejection, session)
+				if (step === null) return forget()
+				session.repairs++
+				current = step
+			}
+		}
+	}
+}
+
+/** Re-run the shared repair on the outgoing copy; null when it changes nothing or cannot. */
+function repairOnce(messages: Message[]): Message[] | null {
+	try {
+		const repaired = repairToolMessageHistory(messages)
+		return toolHistoryRepairChanged(repaired.report) ? repaired.messages : null
+	} catch {
+		return null
+	}
+}
+
+function withoutParameters(
+	params: ChatCompletionParams,
+	names: ReadonlySet<RejectedParameter>,
+): ChatCompletionParams {
+	if (names.size === 0) return params
+	const next = { ...params }
+	if (names.has('effort')) next.effort = undefined
+	if (names.has('thinking')) next.thinking = undefined
+	if (names.has('temperature')) next.temperature = undefined
+	if (names.has('topP')) next.topP = undefined
+	if (names.has('responseFormat')) next.responseFormat = undefined
+	return next
+}
+
+/**
+ * Rejections another recovery already owns: a model that reads text only, and
+ * a single distinct image the provider calls invalid. Those keep their own
+ * commit points in `provider-rejected-image.ts`, which wraps this one.
+ */
+function isHandledElsewhere(error: unknown, params: ChatCompletionParams): boolean {
+	if (isProviderRejectedImageError(error) && findSingleRequestImage(params.messages) !== null) {
+		return true
+	}
+	return isImageInputRefusal(error)
+}