Bound tool image requests and clarify oversized payload errors

.github/workflows/windows-nsis-upgrade-regression.yml

@@ -395,6 +395,7 @@ jobs:
             ${{ runner.temp }}/nsis-packaged-proof/first-interaction/first-send.log
             ${{ runner.temp }}/security-first-send-trace.zip
             ${{ runner.temp }}/security-first-send/logs/desktop.log*
+            ${{ runner.temp }}/security-first-send/logs/gateway.log*
             ${{ runner.temp }}/security-first-send/first-send-main-console.jsonl
             candidate/audit-candidate.json
           if-no-files-found: warn

docs/configuration.md

@@ -271,6 +271,25 @@ OpenSquilla also carries provider registry entries for additional
 OpenAI-compatible or self-hosted backends. Use `opensquilla providers list` on
 your install to see the current catalog.
 
+Outgoing model requests have a separate JSON body limit of 24 MiB
+(25,165,824 bytes) by default. Set
+`OPENSQUILLA_PROVIDER_REQUEST_MAX_BODY_BYTES` in the Gateway environment to
+change the limit, or set it to `0` to disable this guard, then restart the
+Gateway. This counts the UTF-8 bytes actually sent to the provider, including
+base64 image data. The model's token budget measures a different limit: an
+image can fit its estimated token allowance while making the HTTP body too
+large. When the byte limit is exceeded, OpenSquilla attempts safe request-text
+compaction and otherwise reports a budget error before sending the request.
+An upstream service may enforce a lower body limit.
+
+Tool-produced images are prepared before model delivery, with each image
+limited to 2048 pixels on its longest edge and 1.5 MiB of base64 data. If an
+tool-image batch is larger than four images or 6 MiB, the model request
+includes at most four recent usable images and explicitly states how many were
+omitted. This applies to the first request after a tool returns images and to
+later history replay. The request-only reduction leaves the stored session
+unchanged; omitted images must be loaded again when their details are needed.
+
 Read: [`providers-and-models.md`](providers-and-models.md)
 
 ## Video Generation

opensquilla-webui/src/locales/de.json

@@ -4077,6 +4077,8 @@
       "credits": "Das Guthaben des Modelldienstkontos reicht nicht aus.",
       "modelUnavailable": "Das gewählte Modell ist nicht verfügbar.",
       "contextLimit": "Der Inhalt überschreitet die Modellkapazität. Kürzen Sie die Eingabe oder wählen Sie ein Modell mit größerem Kontextfenster.",
+      "requestTooLarge": "Der Modelldienst hat diese Anfrage wegen ihrer Größe abgelehnt. Entfernen Sie große Anhänge oder beginnen Sie einen neuen Chat mit einer kurzen Zusammenfassung.",
+      "outboundRequestTooLarge": "OpenSquilla hat die Anfrage gestoppt, weil der zu sendende Inhalt die Größenbegrenzung überschreitet. Entfernen Sie große Anhänge oder beginnen Sie einen neuen Chat mit einer kurzen Zusammenfassung.",
       "unsupported": "Dieses Modell unterstützt die Anfrage nicht.",
       "imageUnsupported": "Dieses Modell unterstützt keine Bilder.",
       "ensembleImageUnsupported": "Dieser Modus unterstützt keine Bilder.",

opensquilla-webui/src/locales/en.json

@@ -4180,6 +4180,8 @@
       "credits": "The model service account has insufficient credits.",
       "modelUnavailable": "The selected model is unavailable.",
       "contextLimit": "The content exceeds this model's capacity. Shorten the input or choose a larger-context model.",
+      "requestTooLarge": "The model service rejected this request as too large. Reduce large attachments or start a new chat with a short summary before trying again.",
+      "outboundRequestTooLarge": "OpenSquilla stopped this request because the content to send exceeds its size limit. Reduce large attachments or start a new chat with a short summary before trying again.",
       "unsupported": "This model does not support this request.",
       "imageUnsupported": "This model does not support images.",
       "ensembleImageUnsupported": "This mode does not support images.",

opensquilla-webui/src/locales/es.json

@@ -4077,6 +4077,8 @@
       "credits": "La cuenta del servicio del modelo no tiene saldo suficiente.",
       "modelUnavailable": "El modelo seleccionado no está disponible.",
       "contextLimit": "El contenido supera la capacidad del modelo. Acorta la entrada o elige un modelo con más contexto.",
+      "requestTooLarge": "El servicio del modelo rechazó esta solicitud por su tamaño. Reduce los archivos adjuntos grandes o inicia un chat nuevo con un resumen breve antes de volver a intentarlo.",
+      "outboundRequestTooLarge": "OpenSquilla detuvo esta solicitud porque el contenido que se iba a enviar supera su límite de tamaño. Reduce los archivos adjuntos grandes o inicia un chat nuevo con un resumen breve.",
       "unsupported": "Este modelo no admite esta solicitud.",
       "imageUnsupported": "Este modelo no admite imágenes.",
       "ensembleImageUnsupported": "Este modo no admite imágenes.",

opensquilla-webui/src/locales/fr.json

@@ -4077,6 +4077,8 @@
       "credits": "Le solde du compte du service de modèle est insuffisant.",
       "modelUnavailable": "Le modèle sélectionné est indisponible.",
       "contextLimit": "Le contenu dépasse la capacité du modèle. Raccourcissez l’entrée ou choisissez un modèle avec un contexte plus grand.",
+      "requestTooLarge": "Le service du modèle a rejeté cette requête car elle est trop volumineuse. Réduisez les pièces jointes volumineuses ou démarrez une nouvelle conversation avec un bref résumé avant de réessayer.",
+      "outboundRequestTooLarge": "OpenSquilla a arrêté cette requête car le contenu à envoyer dépasse sa limite de taille. Réduisez les pièces jointes volumineuses ou démarrez une nouvelle conversation avec un bref résumé.",
       "unsupported": "Ce modèle ne prend pas en charge cette demande.",
       "imageUnsupported": "Ce modèle ne prend pas en charge les images.",
       "ensembleImageUnsupported": "Ce mode ne prend pas en charge les images.",

opensquilla-webui/src/locales/ja.json

@@ -4077,6 +4077,8 @@
       "credits": "モデルサービスのアカウント残高が不足しています。",
       "modelUnavailable": "選択したモデルは利用できません。",
       "contextLimit": "内容がモデルの容量を超えています。入力を短くするか、コンテキストの大きいモデルを選んでください。",
+      "requestTooLarge": "モデルサービスがリクエストのサイズを理由に拒否しました。大きな添付ファイルを減らすか、新しいチャットで短い要約を伝えてから再試行してください。",
+      "outboundRequestTooLarge": "送信する内容がサイズ制限を超えるため、OpenSquilla がリクエストを停止しました。大きな添付ファイルを減らすか、新しいチャットで短い要約を伝えてください。",
       "unsupported": "このモデルはこのリクエストに対応していません。",
       "imageUnsupported": "このモデルは画像に対応していません。",
       "ensembleImageUnsupported": "このモードは画像に対応していません。",

opensquilla-webui/src/locales/zh-Hans.json

@@ -4179,6 +4179,8 @@
       "credits": "模型服务账户余额不足",
       "modelUnavailable": "当前模型不可用",
       "contextLimit": "内容超出当前模型容量，请缩短输入或选择更大上下文的模型",
+      "requestTooLarge": "模型服务拒绝了过大的请求。请减少大附件，或在新对话中提供简短摘要后再试。",
+      "outboundRequestTooLarge": "OpenSquilla 已拦截超出发送大小限制的请求。请减少大附件，或在新对话中提供简短摘要后再试。",
       "unsupported": "当前模型不支持此请求",
       "imageUnsupported": "当前模型不支持图片",
       "ensembleImageUnsupported": "当前模式不支持图片",

opensquilla-webui/src/utils/chat/chatErrorPresentation.test.ts

@@ -14,6 +14,8 @@ describe('Chat user error presentation', () => {
     ['provider_overloaded', 'busy'],
     ['auth_invalid', 'credentials', 'open-provider-settings'],
     ['context_overflow', 'contextLimit', 'choose-model'],
+    ['payload_too_large', 'requestTooLarge'],
+    ['wire_payload_too_large', 'outboundRequestTooLarge'],
     ['unsupported_feature', 'unsupported', 'choose-model'],
     ['insufficient_credits', 'credits', 'open-provider-settings'],
     ['model_not_found', 'modelUnavailable', 'choose-model'],
@@ -35,6 +37,8 @@ describe('Chat user error presentation', () => {
   it.each([
     ['no_provider', 'noProvider', 'open-model-settings'],
     ['provider_request_too_large', 'contextLimit', 'choose-model'],
+    ['provider_payload_too_large', 'requestTooLarge', undefined],
+    ['provider_wire_payload_too_large', 'outboundRequestTooLarge', undefined],
     ['provider_request_budget_exhausted', 'contextLimit', 'choose-model'],
     ['current_turn_context_exhausted', 'contextLimit', 'choose-model'],
     ['attachment_capacity_too_large', 'contextLimit', 'choose-model'],
@@ -129,7 +133,7 @@ describe('Chat user error presentation', () => {
     expect(chatErrorPresentation({ terminalStatus: 'timeout', code: 'provider_retry_after_deadline' })).toEqual({ messageKey: 'chat.errorMessage.timeout' })
     expect(chatErrorPresentation({ failureKind: 'insufficient_credits' })).toEqual({ messageKey: 'chat.errorMessage.credits', action: 'open-provider-settings' })
     expect(chatErrorPresentation({ failureKind: 'usage_limit' })).toEqual({ messageKey: 'chat.errorMessage.usageLimit' })
-    for (const code of ['400', '401', '403', '404', '429', '503']) expect(chatErrorPresentation({ code })).toEqual({ messageKey: 'chat.errorMessage.unknown' })
+    for (const code of ['400', '401', '403', '404', '413', '429', '503']) expect(chatErrorPresentation({ code })).toEqual({ messageKey: 'chat.errorMessage.unknown' })
   })
 
   it('never uses a general retry hint to offer whole-turn replay', () => {

opensquilla-webui/src/utils/chat/chatErrorPresentation.ts

@@ -69,6 +69,8 @@ const PROVIDER_PRESENTATIONS = new Map<string, ChatErrorPresentation>([
   ['provider_overloaded', presentation('busy')],
   ['auth_invalid', presentation('credentials', 'open-provider-settings')],
   ['context_overflow', presentation('contextLimit', 'choose-model')],
+  ['payload_too_large', presentation('requestTooLarge')],
+  ['wire_payload_too_large', presentation('outboundRequestTooLarge')],
   ['unsupported_feature', presentation('unsupported', 'choose-model')],
   ['insufficient_credits', presentation('credits', 'open-provider-settings')],
   ['usage_limit', presentation('usageLimit')],
@@ -117,6 +119,8 @@ export function chatErrorPresentation(input: ChatErrorPresentationInput): ChatEr
   }
   if (causes.includes('ensemble_multimodal_unsupported')) return presentation('ensembleImageUnsupported', 'choose-model')
   if (causes.includes('image_input_unsupported')) return presentation('imageUnsupported', 'choose-model')
+  if (causes.includes('provider_wire_payload_too_large')) return presentation('outboundRequestTooLarge')
+  if (causes.includes('provider_payload_too_large')) return presentation('requestTooLarge')
   if (causes.some(value => CONTEXT_LIMITS.has(value))) return presentation('contextLimit', 'choose-model')
   if (causes.some(value => ['attachment_capacity_unknown', 'attachment_capacity_unavailable'].includes(value))) {
     return presentation('modelUnavailable', 'open-model-settings')

opensquilla-webui/src/utils/chat/providerFailure.test.ts

@@ -14,6 +14,19 @@ describe('provider terminal metadata', () => {
     expect(normalizeTurnOutcome({ ...base, [key]: fields })).toEqual(normalizeTurnOutcome({ ...base, ...fields }))
   })
 
+  it('keeps remote and local request-size causes distinct in persisted outcomes', () => {
+    const remote = normalizeTurnOutcome({ turn_id: 'remote', failure_kind: 'payload_too_large' })
+    const local = normalizeTurnOutcome({ turn_id: 'local', failure_kind: 'wire_payload_too_large' })
+    expect(remote?.failureKind).toBe('payload_too_large')
+    expect(local?.failureKind).toBe('wire_payload_too_large')
+    expect(chatErrorPresentation({ failureKind: remote?.failureKind })).toEqual({
+      messageKey: 'chat.errorMessage.requestTooLarge',
+    })
+    expect(chatErrorPresentation({ failureKind: local?.failureKind })).toEqual({
+      messageKey: 'chat.errorMessage.outboundRequestTooLarge',
+    })
+  })
+
   it.each(['ABCDEF01', 'abcdef0', 'abcdef012', 'abcdef01 ', 123, null, ''])('rejects invalid id %s', errorId => {
     const outcome = normalizeTurnOutcome({ turn_id: 't', error_id: errorId })
     expect(outcome?.errorId).toBeNull()
@@ -38,7 +51,7 @@ describe('provider terminal metadata', () => {
   it.each(['en', 'zh-Hans', 'de', 'es', 'fr', 'ja'] as const)('localizes allowlisted causes in %s', async locale => {
     await loadLocaleMessages(locale)
     i18n.global.locale.value = locale
-    for (const kind of ['rate_limited', 'provider_overloaded', 'auth_invalid', 'context_overflow',
+    for (const kind of ['rate_limited', 'provider_overloaded', 'auth_invalid', 'context_overflow', 'payload_too_large', 'wire_payload_too_large',
       'unsupported_feature', 'insufficient_credits', 'model_not_found', 'transport_transient',
       'policy_refusal', 'empty_response', 'malformed_response', 'bad_request'] as const) {
       const message = localizedChatErrorMessage('429', 'safe fallback', false, kind)
@@ -97,6 +110,8 @@ describe('provider terminal metadata', () => {
     for (const [code, key] of [
       ['timeout', 'timeout'], ['llm_timeout', 'timeout'],
       ['provider_output_truncated', 'responseIncomplete'], ['provider_request_too_large', 'contextLimit'],
+      ['provider_payload_too_large', 'requestTooLarge'],
+      ['provider_wire_payload_too_large', 'outboundRequestTooLarge'],
     ]) {
       expect(localizedChatErrorMessage(code, 'specific guidance', false, 'rate_limited'))
         .toBe(i18n.global.t(`chat.errorMessage.${key}`))

opensquilla-webui/src/utils/chat/providerFailure.ts

@@ -1,6 +1,7 @@
 /** Stable public taxonomy only; never infer a provider cause from prose or HTTP codes. */
 const FAILURE_KINDS = new Set([
-  'rate_limited', 'provider_overloaded', 'auth_invalid', 'context_overflow',
+  'rate_limited', 'provider_overloaded', 'auth_invalid', 'context_overflow', 'payload_too_large',
+  'wire_payload_too_large',
   'unsupported_feature', 'insufficient_credits', 'usage_limit', 'usage_limit_reached', 'model_not_found', 'transport_transient',
   'policy_refusal', 'empty_response', 'malformed_response', 'bad_request',
 ])
@@ -17,6 +18,7 @@ export function diagnosticErrorId(value: unknown): string | undefined {
 const SPECIAL_TERMINALS = new Set([
   'timeout', 'llm_timeout', 'iteration_timeout', 'stream_idle_timeout', 'provider_output_truncated',
   'provider_request_too_large', 'provider_request_budget_exhausted', 'current_turn_context_exhausted',
+  'provider_payload_too_large', 'provider_wire_payload_too_large',
   'model_repetition_loop_detected', 'tool_run_budget_exhausted', 'llm_budget_exhausted',
   'turn_llm_call_budget_exceeded', 'turn_input_token_budget_exceeded',
   'turn_output_token_budget_exceeded', 'turn_billed_cost_budget_exceeded', 'max_iterations',