Preserve image inputs across configured model routes

.github/scripts/windows_test_assignments.json

@@ -81,6 +81,7 @@
       "tests/test_engine/test_agent_provider_identity.py",
       "tests/test_engine/test_agent_transactional_tool_publication.py",
       "tests/test_engine/test_attachment_aware_routing.py",
+      "tests/test_engine/test_attachment_replay_ownership.py",
       "tests/test_engine/test_auxiliary_usage_accounting.py",
       "tests/test_engine/test_cancelled_turn_segments.py",
       "tests/test_engine/test_coding_mode.py",
@@ -105,6 +106,7 @@
       "tests/test_engine/test_resolve_turn_thinking.py",
       "tests/test_engine/test_route_plan.py",
       "tests/test_engine/test_router_calibration.py",
+      "tests/test_engine/test_router_configured_image_policy.py",
       "tests/test_engine/test_router_decision_event.py",
       "tests/test_engine/test_routing_policy_parity.py",
       "tests/test_engine/test_runtime_agent_iteration_timeout.py",
@@ -158,6 +160,7 @@
       "tests/test_provider/test_context_profile_parity.py",
       "tests/test_provider/test_error_secret_boundary.py",
       "tests/test_provider/test_failure_injection.py",
+      "tests/test_provider/test_image_projection.py",
       "tests/test_provider/test_list_models_provider_identity.py",
       "tests/test_provider/test_live_catalog.py",
       "tests/test_provider/test_preset_registry.py",
@@ -857,6 +860,7 @@
       "tests/test_session/test_agent_task_storage.py",
       "tests/test_session/test_artifact_session_lifecycle.py",
       "tests/test_session/test_assistant_message_identity.py",
+      "tests/test_session/test_attachment_manifest.py",
       "tests/test_session/test_branch_media_copy.py",
       "tests/test_session/test_cascade_pagination.py",
       "tests/test_session/test_compaction.py",

.github/scripts/windows_test_durations.json

@@ -335,6 +335,7 @@
     "tests/test_engine/test_artifact_routing_policy.py": 0.01,
     "tests/test_engine/test_attachment_aware_routing.py": 9.013,
     "tests/test_engine/test_attachment_messages.py": 1.073,
+    "tests/test_engine/test_attachment_replay_ownership.py": 0.01,
     "tests/test_engine/test_auxiliary_usage_accounting.py": 0.028,
     "tests/test_engine/test_bootstrap_snapshot_invalidation.py": 0.478,
     "tests/test_engine/test_cache_break_monitor.py": 0.106,
@@ -400,6 +401,7 @@
     "tests/test_engine/test_router_budget_gate.py": 0.043,
     "tests/test_engine/test_router_calibration.py": 0.144,
     "tests/test_engine/test_router_calibration_service.py": 3.359,
+    "tests/test_engine/test_router_configured_image_policy.py": 0.01,
     "tests/test_engine/test_router_control.py": 0.062,
     "tests/test_engine/test_router_decision_event.py": 0.07,
     "tests/test_engine/test_router_decision_record.py": 0.786,
@@ -955,6 +957,7 @@
     "tests/test_provider/test_error_secret_boundary.py": 0.194,
     "tests/test_provider/test_failure_classification_parity.py": 0.353,
     "tests/test_provider/test_failure_injection.py": 0.066,
+    "tests/test_provider/test_image_projection.py": 0.01,
     "tests/test_provider/test_live_catalog.py": 0.089,
     "tests/test_provider/test_list_models_provider_identity.py": 0.01,
     "tests/test_provider/test_models_dev_snapshot_costs.py": 0.01,
@@ -1205,6 +1208,7 @@
     "tests/test_session/test_agent_task_storage.py": 4.576,
     "tests/test_session/test_artifact_session_lifecycle.py": 0.01,
     "tests/test_session/test_assistant_message_identity.py": 0.062,
+    "tests/test_session/test_attachment_manifest.py": 0.01,
     "tests/test_session/test_branch_media_copy.py": 0.88,
     "tests/test_session/test_cascade_pagination.py": 0.078,
     "tests/test_session/test_compaction.py": 0.282,

desktop/electron/scripts/test-onboarding-flow.mjs

@@ -1135,20 +1135,21 @@ try {
   assert.equal(credential.routerTiers.c1.model, 'deepseek-v4-pro-0813')
   assert.equal(credential.routerTiers.c2.model, 'kimi-k2.7-code')
   assert.equal(credential.routerTiers.c3.model, 'glm-5.2')
-  assert.equal(credential.routerTiers.c0.supportsImage, false)
-  assert.equal(credential.routerTiers.c1.supportsImage, false)
-  assert.equal(credential.routerTiers.c2.supportsImage, false)
-  assert.equal(credential.routerTiers.c3.supportsImage, false)
+  assert.equal(Object.hasOwn(credential.routerTiers.c0, 'supportsImage'), false)
+  assert.equal(Object.hasOwn(credential.routerTiers.c1, 'supportsImage'), false)
+  assert.equal(Object.hasOwn(credential.routerTiers.c2, 'supportsImage'), false)
+  assert.equal(Object.hasOwn(credential.routerTiers.c3, 'supportsImage'), false)
   assert.equal(credential.routerTiers.c3.ensembleEnabled, true)
   assert.equal(credential.routerTiers.image_model.model, 'kimi-k2.6')
-  assert.equal(credential.routerTiers.image_model.supportsImage, true)
+  assert.equal(Object.hasOwn(credential.routerTiers.image_model, 'supportsImage'), false)
   assert.match(config, /\[squilla_router\]\nenabled = true/)
   assert.match(config, /\[llm\][\s\S]*?model = "deepseek-v4-pro-0813"/)
   assert.match(config, /\[squilla_router\.tiers\.c0\]\nprovider = "tokenrhythm"\nmodel = "deepseek-v4-flash-0731"/)
   assert.match(config, /\[squilla_router\.tiers\.c1\]\nprovider = "tokenrhythm"\nmodel = "deepseek-v4-pro-0813"/)
   assert.match(config, /\[squilla_router\.tiers\.c2\]\nprovider = "tokenrhythm"\nmodel = "kimi-k2.7-code"/)
   assert.match(config, /\[squilla_router\.tiers\.c3\][\s\S]*?model = "glm-5.2"[\s\S]*?ensemble_enabled = true/)
   assert.doesNotMatch(config, /thinking_level\s*=/)
+  assert.doesNotMatch(config, /supports_image\s*=/)
   assert.match(config, /\[llm_ensemble\]\nenabled = false/)
   assert.equal(successfulProbeServer.requests.length, 1)
   assert.equal(successfulProbeServer.requests[0].url, '/v1/chat/completions')

desktop/electron/scripts/test-router-tier-normalization.mjs

@@ -3,10 +3,10 @@ import { strict as assert } from 'node:assert'
 import { normalizeRouterTiers } from '../dist/router-tier-normalization.js'
 
 const currentFallback = {
-  c0: { provider: 'tokenrhythm', model: 'deepseek-v4-flash-0731' },
+  c0: { provider: 'tokenrhythm', model: 'deepseek-v4-flash-0731', supportsImage: false },
   c1: { provider: 'tokenrhythm', model: 'deepseek-v4-pro-0813' },
   c2: { provider: 'tokenrhythm', model: 'kimi-k2.7-code' },
-  c3: { provider: 'tokenrhythm', model: 'glm-5.2', ensembleEnabled: true },
+  c3: { provider: 'tokenrhythm', model: 'glm-5.2', supportsImage: true, ensembleEnabled: true },
 }
 
 const legacyCredentialTiers = {
@@ -22,15 +22,32 @@ assert.deepEqual(
   Object.fromEntries(Object.entries(legacyCredentialTiers).map(([name, tier]) => [name, tier.model])),
 )
 assert.equal(Object.hasOwn(loaded.c3, 'ensembleEnabled'), false)
+assert.equal(Object.hasOwn(loaded.c0, 'supportsImage'), false)
+assert.equal(Object.hasOwn(loaded.c3, 'supportsImage'), false)
 
 // saveDesktopCredential normalizes the already-loaded same-provider ladder a
 // second time. The missing legacy opt-in must remain missing on that pass.
 const resaved = normalizeRouterTiers(loaded, currentFallback)
 assert.deepEqual(resaved, loaded)
 assert.equal(Object.hasOwn(resaved.c3, 'ensembleEnabled'), false)
+assert.equal(Object.hasOwn(resaved.c0, 'supportsImage'), false)
+
+for (const key of ['supports_image', 'supportsImage']) {
+  for (const value of [false, true]) {
+    const legacyCapability = normalizeRouterTiers(
+      { ...legacyCredentialTiers, c0: { ...legacyCredentialTiers.c0, [key]: value } },
+      currentFallback,
+    )
+    assert.equal(Object.hasOwn(legacyCapability.c0, 'supportsImage'), false)
+    assert.equal(Object.hasOwn(legacyCapability.c0, 'supports_image'), false)
+    assert.equal(legacyCapability.c0.model, legacyCredentialTiers.c0.model)
+  }
+}
 
 const fresh = normalizeRouterTiers(undefined, currentFallback)
 assert.equal(fresh.c3.ensembleEnabled, true)
+assert.equal(Object.hasOwn(fresh.c0, 'supportsImage'), false)
+assert.equal(Object.hasOwn(fresh.c3, 'supportsImage'), false)
 
 const explicitSnakeCase = normalizeRouterTiers(
   { ...legacyCredentialTiers, c3: { ...legacyCredentialTiers.c3, ensemble_enabled: true } },

desktop/electron/src/main.ts

@@ -1914,18 +1914,18 @@ function minimaxRouterProfile(provider: string): Record<string, RouterTier> {
 
 const ROUTER_PROFILES: Record<string, Record<string, RouterTier>> = {
   tokenrhythm: {
-    c0: { provider: 'tokenrhythm', model: 'deepseek-v4-flash-0731', description: 'Fast DeepSeek V4 Flash 0731 route for simple work', supportsImage: false },
-    c1: { provider: 'tokenrhythm', model: 'deepseek-v4-pro-0813', description: 'Default DeepSeek V4 Pro 0813 route for normal agent work', supportsImage: false },
-    c2: { provider: 'tokenrhythm', model: 'kimi-k2.7-code', description: 'Strong Kimi 2.7 Code route for harder coding and analysis', supportsImage: false },
-    c3: { provider: 'tokenrhythm', model: 'glm-5.2', description: 'Highest tier: shared B5 fusion; GLM 5.2 is retained for single-model C3 mode', supportsImage: false, ensembleEnabled: true },
-    image_model: { provider: 'tokenrhythm', model: 'kimi-k2.6', description: 'Vision route for image attachments', supportsImage: true, imageOnly: true },
+    c0: { provider: 'tokenrhythm', model: 'deepseek-v4-flash-0731', description: 'Fast DeepSeek V4 Flash 0731 route for simple work' },
+    c1: { provider: 'tokenrhythm', model: 'deepseek-v4-pro-0813', description: 'Default DeepSeek V4 Pro 0813 route for normal agent work' },
+    c2: { provider: 'tokenrhythm', model: 'kimi-k2.7-code', description: 'Strong Kimi 2.7 Code route for harder coding and analysis' },
+    c3: { provider: 'tokenrhythm', model: 'glm-5.2', description: 'Highest tier: shared B5 fusion; GLM 5.2 is retained for single-model C3 mode', ensembleEnabled: true },
+    image_model: { provider: 'tokenrhythm', model: 'kimi-k2.6', description: 'Vision route for image attachments', imageOnly: true },
   },
   openrouter: {
     c0: { provider: 'openrouter', model: 'deepseek/deepseek-v4-flash', description: 'Fast everyday work', thinkingLevel: 'high' },
     c1: { provider: 'openrouter', model: 'deepseek/deepseek-v4-pro', description: 'Balanced agent work', thinkingLevel: 'high' },
     c2: { provider: 'openrouter', model: 'z-ai/glm-5.2', description: 'Complex reasoning', thinkingLevel: 'high' },
     c3: { provider: 'openrouter', model: 'anthropic/claude-opus-4.8', description: 'Highest quality review and planning', thinkingLevel: 'high' },
-    image_model: { provider: 'openrouter', model: 'moonshotai/kimi-k2.6', description: 'Vision route for image attachments', supportsImage: true, imageOnly: true, thinkingLevel: 'medium' },
+    image_model: { provider: 'openrouter', model: 'moonshotai/kimi-k2.6', description: 'Vision route for image attachments', imageOnly: true, thinkingLevel: 'medium' },
   },
   openai: {
     c0: { provider: 'openai', model: 'gpt-5.4-nano', description: 'Fast simple work', thinkingLevel: 'none' },
@@ -1952,10 +1952,10 @@ const ROUTER_PROFILES: Record<string, Record<string, RouterTier>> = {
     c3: { provider: 'gemini', model: 'gemini-3.1-pro-preview', description: 'Deep reasoning', thinkingLevel: 'high' },
   },
   moonshot: {
-    c0: { provider: 'moonshot', model: 'kimi-k2.6', description: 'Fast multimodal work', supportsImage: true, thinkingLevel: 'low' },
-    c1: { provider: 'moonshot', model: 'kimi-k2.6', description: 'Balanced multimodal work', supportsImage: true, thinkingLevel: 'medium' },
-    c2: { provider: 'moonshot', model: 'kimi-k2.6', description: 'Complex text and image work', supportsImage: true, thinkingLevel: 'medium' },
-    c3: { provider: 'moonshot', model: 'kimi-k2.7-code', description: 'Code-heavy deep reasoning', supportsImage: true, thinkingLevel: 'high' },
+    c0: { provider: 'moonshot', model: 'kimi-k2.6', description: 'Fast multimodal work', thinkingLevel: 'low' },
+    c1: { provider: 'moonshot', model: 'kimi-k2.6', description: 'Balanced multimodal work', thinkingLevel: 'medium' },
+    c2: { provider: 'moonshot', model: 'kimi-k2.6', description: 'Complex text and image work', thinkingLevel: 'medium' },
+    c3: { provider: 'moonshot', model: 'kimi-k2.7-code', description: 'Code-heavy deep reasoning', thinkingLevel: 'high' },
   },
   kimi_coding_openai: textRouterProfile(
     'kimi_coding_openai',
@@ -2139,7 +2139,6 @@ function routerTierTomlLines(name: string, tier: RouterTier): string[] {
     `model = ${tomlString(tier.model)}`,
   ]
   if (tier.description) lines.push(`description = ${tomlString(tier.description)}`)
-  if (tier.supportsImage !== undefined) lines.push(`supports_image = ${tier.supportsImage ? 'true' : 'false'}`)
   if (tier.imageOnly !== undefined) lines.push(`image_only = ${tier.imageOnly ? 'true' : 'false'}`)
   if (tier.thinkingLevel) lines.push(`thinking_level = ${tomlString(tier.thinkingLevel)}`)
   if (tier.ensembleEnabled !== undefined) lines.push(`ensemble_enabled = ${tier.ensembleEnabled ? 'true' : 'false'}`)

desktop/electron/src/router-tier-normalization.ts

@@ -2,6 +2,7 @@ export interface RouterTier {
   provider: string
   model: string
   description?: string
+  /** Read compatibility only; the gateway resolves model capability. */
   supportsImage?: boolean
   imageOnly?: boolean
   thinkingLevel?: string
@@ -20,7 +21,11 @@ function canonicalTierKey(name: string): string {
 }
 
 function cloneRouterTiers(tiers: Record<string, RouterTier>): Record<string, RouterTier> {
-  return Object.fromEntries(Object.entries(tiers).map(([name, tier]) => [name, { ...tier }]))
+  return Object.fromEntries(Object.entries(tiers).map(([name, tier]) => {
+    const copy = { ...tier }
+    delete copy.supportsImage
+    return [name, copy]
+  }))
 }
 
 function normalizeBooleanSetting(raw: unknown, fallback: boolean): boolean {
@@ -58,18 +63,18 @@ export function normalizeRouterTiers(
     const hasEnsembleEnabled = Object.prototype.hasOwnProperty.call(tier, 'ensembleEnabled')
       || Object.prototype.hasOwnProperty.call(tier, 'ensemble_enabled')
     const ensembleEnabled = tier.ensembleEnabled ?? tier.ensemble_enabled
-    out[name] = {
+    const normalizedTier: RouterTier = {
       ...out[name],
       provider,
       model,
       description: String(tier.description || out[name]?.description || ''),
-      supportsImage: Boolean(tier.supportsImage ?? tier.supports_image ?? out[name]?.supportsImage),
       imageOnly: Boolean(tier.imageOnly ?? tier.image_only ?? out[name]?.imageOnly),
       thinkingLevel: String(tier.thinkingLevel ?? tier.thinking_level ?? out[name]?.thinkingLevel ?? ''),
       ...(hasEnsembleEnabled
         ? { ensembleEnabled: normalizeBooleanSetting(ensembleEnabled, false) }
         : {}),
     }
+    out[name] = normalizedTier
   }
   return out
 }

docs/features/squilla-router.md

@@ -78,10 +78,22 @@ Packaged static B5 lineups use a 120-second total budget per proposer and a
 180-second aggregator idle budget. Operator-authored `custom_b5` lineups use
 300 and 480 seconds respectively unless explicitly configured otherwise.
 
-C3 fusion itself is excluded from image routing, but the dedicated
-`image_model` tier remains eligible and is preferred for image requests. If it
-is unavailable, another non-C3 tier with `supports_image = true` may handle the
-request.
+Image routing considers only the configured C0–C3 single-model deployments;
+C3 fusion is excluded. Image input capability is resolved automatically from
+the shared provider model catalog, including API-provided deployment metadata
+and offline catalog fallback. The web and desktop clients do not expose a
+manual image-capability switch. Legacy tier `supports_image` values remain
+readable but do not override the deployment's capability, and `image_model`
+is retained for compatibility without creating an executable fifth route.
+
+When capability is unknown, the configured model may receive a native image
+request. A recognized image-input rejection can advance through the remaining
+configured tiers before falling back to explicit not-analyzed/analysis-failed
+text markers. Direct retries only its configured model with markers; Ensemble
+uses text markers for all members. These retries stop after visible output or
+tool effects. Projection never deletes the canonical attachment, so switching
+back to a vision-capable model can recover earlier images after history
+compression.
 
 Disable routing and use the configured provider/model directly:
 

opensquilla-webui/src/components/setup/SetupModelStrategyPanel.test.ts

@@ -317,6 +317,9 @@ describe('SetupModelStrategyPanel', () => {
     expect(el.querySelector('[aria-label="c0 request entry"]')).toBeTruthy()
     expect(el.querySelector('.setup-tier-table__row.is-head')?.textContent)
       .toContain('Request entry')
+    expect(el.querySelector('[aria-label$="supports image"]')).toBeNull()
+    expect(el.querySelector('[data-testid="router-image-capability-hint"]')?.textContent)
+      .toContain('detected automatically from provider metadata and request results')
 
     app.unmount()
   })

opensquilla-webui/src/components/setup/SetupModelStrategyPanel.vue

@@ -107,7 +107,7 @@ const emit = defineEmits<{
   updateFixedModel: [value: string]
   updateRouterDefaultTier: [value: string]
   updateRouterVisualMode: [value: string]
-  updateTierField: [name: string, key: 'provider' | 'model' | 'thinkingLevel' | 'supportsImage' | 'ensembleEnabled' | 'ensembleSelectionMode', value: string | boolean]
+  updateTierField: [name: string, key: 'provider' | 'model' | 'thinkingLevel' | 'ensembleEnabled' | 'ensembleSelectionMode', value: string | boolean]
   updateEnsembleScheme: [value: 'preset' | 'custom']
   addEnsembleCandidate: [provider: string, model: string, role: EnsembleCandidateRole]
   removeEnsembleCandidate: [candidate: EnsembleCandidateView]
@@ -825,6 +825,9 @@ function credentialLabel(candidate: EnsembleCandidateView): string {
           @update-tier-field="(name, key, value) => emit('updateTierField', name, key, value)"
           @migrate-legacy-ensemble="emit('migrateEnsembleLegacy')"
         />
+        <p class="control-section__desc" data-testid="router-image-capability-hint">
+          {{ t('setup.router.imageCapabilityAutomatic') }}
+        </p>
 
         <details
           class="setup-model-strategy__runtime setup-model-strategy__advanced"