fix: match openrouter model ids by parsed vendor, not raw prefix (#10497)

* fix: match openrouter model ids by parsed vendor, not raw prefix Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: anchor context-window matches so a version entry cannot swallow a longer version Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix: scope the context-window digit guard to version-suffixed entries Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs: state the thinking-suffix invariant without drafting history Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-10-05 08:02:28 +00:00 · 2026-08-04 10:35:56 +02:00
parent 99e7661a1b
commit 546b8d0769
4 changed files with 100 additions and 29 deletions
@@ -259,25 +259,31 @@ describe('OpenRouter prompt caching', () => {
 		{ role: 'tool', tool_call_id: 't1', content: 'tool output' }
 	]

-	it('breaks on the system prompt and the newest user turn for Anthropic models', async () => {
-		const { getCompletion } = await import('./lib')
-		h.currentModel = { provider: 'openrouter', model: 'anthropic/claude-sonnet-5' }
+	// The `~` form is OpenRouter's own floating alias for the same vendor, so it has
+	// to reach the same gate — a prefix match on the raw id silently misses it and
+	// puts the chat back on full price every turn.
+	it.each(['anthropic/claude-sonnet-5', '~anthropic/claude-sonnet-latest'])(
+		'breaks on the system prompt and the newest user turn for %s',
+		async (model) => {
+			const { getCompletion } = await import('./lib')
+			h.currentModel = { provider: 'openrouter', model }

-		await getCompletion([...conversation], new AbortController(), undefined, {
-			promptCaching: true
-		})
+			await getCompletion([...conversation], new AbortController(), undefined, {
+				promptCaching: true
+			})

-		const sent = openaiCreate.mock.calls[0][0].messages
-		const ephemeral = { type: 'ephemeral' }
-		expect(sent[0].content).toEqual([
-			{ type: 'text', text: 'system prompt', cache_control: ephemeral }
-		])
-		expect(sent[1].content).toEqual([{ type: 'text', text: 'first', cache_control: ephemeral }])
-		expect(sent[3].content).toBe('tool output')
-		// The chat replays this same history through the Anthropic path, which rejects
-		// unknown fields on its own blocks, so the originals must come back untouched.
-		expect(conversation[0].content).toBe('system prompt')
-	})
+			const sent = openaiCreate.mock.calls[0][0].messages
+			const ephemeral = { type: 'ephemeral' }
+			expect(sent[0].content).toEqual([
+				{ type: 'text', text: 'system prompt', cache_control: ephemeral }
+			])
+			expect(sent[1].content).toEqual([{ type: 'text', text: 'first', cache_control: ephemeral }])
+			expect(sent[3].content).toBe('tool output')
+			// The chat replays this same history through the Anthropic path, which rejects
+			// unknown fields on its own blocks, so the originals must come back untouched.
+			expect(conversation[0].content).toBe('system prompt')
+		}
+	)

 	it('leaves other OpenRouter upstreams alone', async () => {
 		const { getCompletion } = await import('./lib')
@@ -217,7 +217,12 @@ describe('model context windows', () => {
 		expect(getKnownModelContextWindow('claude-sonnet-4-6')).toBe(1000000)
 		expect(getKnownModelContextWindow('claude-opus-4-6')).toBe(1000000)
 		expect(getKnownModelContextWindow('claude-opus-4-8')).toBe(1000000)
+		expect(getKnownModelContextWindow('claude-opus-5')).toBe(1000000)
+		expect(getKnownModelContextWindow('claude-sonnet-5')).toBe(1000000)
 		expect(getKnownModelContextWindow('anthropic.claude-sonnet-4-6-v1:0')).toBe(1000000)
+		// OpenRouter dot-versions the same models; both spellings must resolve
+		expect(getKnownModelContextWindow('anthropic/claude-sonnet-4.6')).toBe(1000000)
+		expect(getKnownModelContextWindow('anthropic/claude-opus-4.8')).toBe(1000000)
 	})

 	it('keeps Haiku and older Claude models at 200K', () => {
@@ -242,6 +247,17 @@ describe('model context windows', () => {
 		expect(getKnownModelContextWindow('gpt-5.5')).toBe(1000000)
 	})

+	it('does not let a version entry claim a longer version number', () => {
+		// `gpt-4.1` and `gpt-4-1106-preview` collapse to the same prefix, but the
+		// latter is a 128K model and must stay unlisted
+		expect(getKnownModelContextWindow('gpt-4-1106-preview')).toBeUndefined()
+		expect(getKnownModelContextWindow('gpt-4.1-mini')).toBe(1000000)
+		// family fallbacks still catch a version welded onto the name (Ollama ids),
+		// which is what keeps compaction on for them
+		expect(getKnownModelContextWindow('llama3.1')).toBe(128000)
+		expect(getKnownModelContextWindow('llama3-70b-8192')).toBe(128000)
+	})
+
 	it('maps recent Gemini and DeepSeek models to the 1M window', () => {
 		expect(getKnownModelContextWindow('gemini-3.1-pro')).toBe(1000000)
 		expect(getKnownModelContextWindow('gemini-3-flash')).toBe(1000000)
@@ -1,5 +1,34 @@
 import type { AIProvider } from '$lib/gen'

+export type ParsedModelId = {
+	/** Vendor namespace when the id carries one (`anthropic/claude-sonnet-5`). */
+	vendor: string | undefined
+	/** Bare model id: no vendor prefix, no variant suffix. */
+	base: string
+}
+
+/**
+ * Split a model id into the parts the predicates below match on. Gateways decorate
+ * the vendor's id in ways a raw substring check misses: OpenRouter marks its own
+ * floating aliases with a `~` prefix (`~anthropic/claude-sonnet-latest`, distinct
+ * from the vendor-pinned `anthropic/claude-sonnet-5`) and appends `:variant`
+ * suffixes (`:free`, `:thinking`). Match on the parsed parts, never on the raw id.
+ *
+ * `base` is the last `/` segment: the deprecated `<model>/thinking` selection puts
+ * its marker where a gateway puts the model, so callers that must resolve one of
+ * those ids first run it through `stripLegacyThinkingSuffix`.
+ */
+export function parseModelId(model: string): ParsedModelId {
+	const normalized = model.toLowerCase().replace(/^~/, '')
+	const segments = normalized.split('/')
+	const last = segments[segments.length - 1]
+	const colon = last.indexOf(':')
+	return {
+		vendor: segments.length > 1 ? segments[0] : undefined,
+		base: colon > 0 ? last.slice(0, colon) : last
+	}
+}
+
 // Azure AI Foundry fronts multiple model families under one resource. Claude
 // deployments are served only through the Anthropic Messages API, so the chat must
 // route them like the native Anthropic provider (Anthropic SDK, message format)
@@ -19,23 +48,21 @@ export function usesAnthropicMessagesApi(provider: AIProvider, model: string): b
 // but only documents them for Anthropic-backed models, so the gate is on the routed
 // model rather than the provider alone.
 export function usesOpenRouterPromptCaching(provider: AIProvider, model: string): boolean {
-	return provider === 'openrouter' && model.toLowerCase().startsWith('anthropic/')
+	return provider === 'openrouter' && parseModelId(model).vendor === 'anthropic'
 }

 // gpt-5+ and o-series reasoning models reject the legacy `max_tokens` field on
 // the OpenAI/Azure Chat Completions API and require `max_completion_tokens`
-// instead. The check strips any provider prefix (e.g. OpenRouter's "openai/o3")
-// so it matches the bare model id, and the o-series match requires a digit after
-// the "o" (o1/o3/o4-mini) so it does not catch unrelated ids like Mistral's
-// "open-mistral-*" or "optimus-*".
+// instead. The check runs on the bare model id (so OpenRouter's "openai/o3"
+// matches), and the o-series match requires a digit after the "o" (o1/o3/o4-mini)
+// so it does not catch unrelated ids like Mistral's "open-mistral-*" or "optimus-*".
 export function requiresMaxCompletionTokens(model: string) {
-	const normalizedModel = model.toLowerCase()
-	const baseModel = normalizedModel.split('/').pop() ?? normalizedModel
+	const baseModel = parseModelId(model).base
 	return baseModel.startsWith('gpt-5') || /^o\d/.test(baseModel)
 }

 // Context windows of the models we know, most specific entry first — the first
-// name included in the model id wins, so provider-prefixed and date-suffixed
+// name found in the bare model id wins, so vendor-namespaced and date-suffixed
 // ids (anthropic.claude-sonnet-4-6-...-v1:0, gpt-5.2-2026-01-01) still resolve.
 // Conservative family fallbacks sit below the explicit entries; models not
 // listed at all resolve to undefined, which disables auto-trimming and the
@@ -45,6 +72,8 @@ const MODEL_CONTEXT_WINDOWS: [name: string, contextWindow: number][] = [
 	// Haiku, older Claude models (3.x, 4.0, 4.1, 4.5) and date-suffixed Claude 4
 	// base ids (claude-sonnet-4-20250514) fall through to 200K
 	['claude-fable-5', 1_000_000],
+	['claude-opus-5', 1_000_000],
+	['claude-sonnet-5', 1_000_000],
 	['claude-opus-4-8', 1_000_000],
 	['claude-opus-4-7', 1_000_000],
 	['claude-opus-4-6', 1_000_000],
@@ -74,8 +103,29 @@ const MODEL_CONTEXT_WINDOWS: [name: string, contextWindow: number][] = [
 	['codestral', 32_000]
 ]

+// Version separators differ by route to the same model: Anthropic writes
+// `claude-opus-4-8`, OpenRouter writes `anthropic/claude-opus-4.8`. Collapsing
+// dots to dashes on both sides keeps one table entry covering every route —
+// without it a dot-versioned id falls through to a coarser family entry.
+function normalizeVersionSeparators(model: string): string {
+	return model.replace(/\./g, '-')
+}
+
+// An entry that ends on a version digit must not run into a longer version:
+// `gpt-4.1` collapses to `gpt-4-1`, which would otherwise claim the 128K
+// `gpt-4-1106-preview` as a 1M model. Suffixes that continue with a separator
+// (`claude-opus-4-8` in `...-4-8-v1`, `gpt-5` in `gpt-5-mini`) still match.
+// Family fallbacks ending on a letter get no such guard — a version welded
+// straight onto the name (`llama3.1`) is exactly what they exist to catch.
+const MODEL_CONTEXT_WINDOW_MATCHERS: [matcher: RegExp, contextWindow: number][] =
+	MODEL_CONTEXT_WINDOWS.map(([name, contextWindow]) => {
+		const pattern = normalizeVersionSeparators(name).replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+		return [new RegExp(/\d$/.test(pattern) ? `${pattern}(?!\\d)` : pattern), contextWindow]
+	})
+
 export function getKnownModelContextWindow(model: string): number | undefined {
-	return MODEL_CONTEXT_WINDOWS.find(([name]) => model.includes(name))?.[1]
+	const id = normalizeVersionSeparators(parseModelId(model).base)
+	return MODEL_CONTEXT_WINDOW_MATCHERS.find(([matcher]) => matcher.test(id))?.[1]
 }

 export function getModelContextWindow(model: string) {
@@ -1,5 +1,5 @@
 import type { AIProvider, AIProviderModel } from '$lib/gen'
-import { usesAnthropicMessagesApi } from './modelConfig'
+import { parseModelId, usesAnthropicMessagesApi } from './modelConfig'

 /**
 * Reasoning effort is provider/model-specific. We never normalize a single
@@ -35,8 +35,7 @@ export function stripLegacyThinkingSuffix(model: string): string {

 /** Bare model id without any provider/gateway prefix (e.g. OpenRouter's `openai/o3`). */
 function baseModelId(model: string): string {
-	const normalized = model.toLowerCase()
-	return normalized.split('/').pop() ?? normalized
+	return parseModelId(model).base
 }

 /**