fix(ai): preserve compatible reasoning details (#37708)

This commit is contained in:
Aiden Cline
2026-07-19 22:30:55 -05:00
committed by GitHub
parent b9525b5878
commit 8b5655ed53
10 changed files with 858 additions and 53 deletions
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
+54 -1
View File
@@ -1,7 +1,7 @@
import { describe, expect } from "bun:test"
import { ConfigProvider, Effect, Schema } from "effect"
import { HttpClientRequest } from "effect/unstable/http"
import { LLM } from "../../src"
import { LLM, LLMEvent } from "../../src"
import { CloudflareAIGateway, CloudflareWorkersAI } from "../../src/providers/cloudflare"
import { LLMClient } from "../../src/route"
import { it } from "../lib/effect"
@@ -83,6 +83,59 @@ describe("Cloudflare", () => {
}),
)
it.effect("preserves reasoning details for AI Gateway continuation", () =>
Effect.gen(function* () {
const model = CloudflareAIGateway.configure({
accountId: "test-account",
gatewayId: "test-gateway",
apiKey: "test-token",
}).model("anthropic/claude-sonnet-4.6")
const details = [
{ type: "reasoning.text", text: "Think", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.text", text: "ing", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.text", signature: "signed", format: "anthropic-claude-v1", index: 0 },
]
const merged = [
{
type: "reasoning.text",
text: "Thinking",
signature: "signed",
format: "anthropic-claude-v1",
index: 0,
},
]
const response = yield* LLM.generate(LLM.request({ model, prompt: "Say hello." })).pipe(
Effect.provide(
dynamicResponse((input) =>
Effect.succeed(
input.respond(
sseEvents(
deltaChunk({ reasoning: "Think", reasoning_details: [details[0]] }),
deltaChunk({ reasoning: "ing", reasoning_details: [details[1]] }),
deltaChunk({ reasoning_details: [details[2]] }),
deltaChunk({ content: "Hello" }),
deltaChunk({}, "stop"),
),
{ headers: { "content-type": "text/event-stream" } },
),
),
),
),
)
expect(response.reasoning).toBe("Thinking")
expect(response.events.filter(LLMEvent.is.reasoningDelta)).toHaveLength(2)
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning", reasoningDetails: merged },
})
const replay = yield* LLMClient.prepare(LLM.request({ model, messages: [response.message] }))
expect(replay.body.messages).toEqual([
{ role: "assistant", content: "Hello", reasoning: "Thinking", reasoning_details: merged },
])
}),
)
it.effect("defaults AI Gateway id to default when omitted or blank", () =>
Effect.gen(function* () {
expect(
@@ -1,10 +1,12 @@
import { describe, expect } from "bun:test"
import { Effect } from "effect"
import { LLM, LLMEvent } from "../../src"
import { LLM, LLMEvent, LLMResponse } from "../../src"
import { OpenAIChat } from "../../src/protocols/openai-chat"
import * as OpenAICompatible from "../../src/providers/openai-compatible"
import * as OpenRouter from "../../src/providers/openrouter"
import { LLMClient } from "../../src/route"
import { recordedTests } from "../recorded-test"
import { expectWeatherToolLoop, goldenWeatherToolLoopRequest, runWeatherToolLoop } from "../recorded-scenarios"
const cases = [
{
@@ -15,6 +17,7 @@ const cases = [
}).model("anthropic/claude-sonnet-4.6"),
requires: ["OPENROUTER_API_KEY"],
cassette: "openrouter-reasoning",
structured: true,
},
{
name: "Vercel AI Gateway",
@@ -26,6 +29,7 @@ const cases = [
}).model("anthropic/claude-sonnet-4.6"),
requires: ["AI_GATEWAY_API_KEY"],
cassette: "vercel-ai-gateway-reasoning",
structured: true,
},
] as const
@@ -57,11 +61,82 @@ for (const item of cases) {
expect(response.text.replaceAll(",", "").trim()).toBe("37887")
expect(response.reasoning.length).toBeGreaterThan(0)
expect(response.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning" },
const metadata = response.message.content.find((part) => part.type === "reasoning")?.providerMetadata
expect(metadata?.openai?.reasoningField).toBe(item.structured ? "reasoning" : "reasoning_content")
expect(Array.isArray(metadata?.openai?.reasoningDetails)).toBe(item.structured)
if (!item.structured) return
const details = metadata?.openai?.reasoningDetails
if (!Array.isArray(details)) return
expect(
details.some(
(detail) =>
typeof detail === "object" &&
detail !== null &&
"signature" in detail &&
typeof detail.signature === "string" &&
detail.signature.length > 0,
),
).toBe(true)
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model: item.model, messages: [response.message] }),
)
expect(replay.body.messages).toMatchObject([
{ role: "assistant", content: response.text, reasoning: response.reasoning },
])
const replayDetails =
replay.body.messages[0]?.role === "assistant" ? replay.body.messages[0].reasoning_details : undefined
expect(Array.isArray(replayDetails)).toBe(true)
if (!Array.isArray(replayDetails)) return
expect(replayDetails).toEqual(details)
expect(replayDetails).toHaveLength(1)
expect(replayDetails[0]).toMatchObject({
type: "reasoning.text",
text: response.reasoning,
signature: expect.any(String),
})
}),
30_000,
)
recorded.effect.with(
"continues signed reasoning through a tool loop",
{ cassette: `${item.cassette}-tool-loop`, tags: ["continuation", "tool", "tool-loop"] },
() =>
Effect.gen(function* () {
const events = yield* runWeatherToolLoop(
goldenWeatherToolLoopRequest({
id: `${item.cassette}-tool-loop`,
model: item.model,
maxTokens: 1536,
temperature: false,
}),
)
expectWeatherToolLoop(events)
expect(
LLMResponse.text({
events: events.slice(events.findIndex(LLMEvent.is.stepFinish) + 1),
}).trim(),
).toMatch(/^Paris is sunny\.?$/)
const details = events
.filter(LLMEvent.is.reasoningEnd)
.map((event) => event.providerMetadata?.openai?.reasoningDetails)
.find(Array.isArray)
expect(Array.isArray(details)).toBe(item.structured)
if (!item.structured || !Array.isArray(details)) return
expect(
details.some(
(detail) =>
typeof detail === "object" &&
detail !== null &&
"signature" in detail &&
typeof detail.signature === "string" &&
detail.signature.length > 0,
),
).toBe(true)
}),
60_000,
)
})
}
@@ -570,6 +570,375 @@ describe("OpenAI Chat route", () => {
}),
)
it.effect("preserves and replays reasoning details alongside scalar reasoning", () =>
Effect.gen(function* () {
const details = [
{ type: "reasoning.text", text: "thinking", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.encrypted", data: "opaque", format: "anthropic-claude-v1", index: 1 },
]
const response = yield* LLMClient.generate(
LLM.updateRequest(request, {
tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
}),
).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning: "thinking", reasoning_details: [details[0]] } }] },
{ choices: [{ delta: { reasoning_details: [details[1]] } }] },
{
choices: [
{
delta: {
tool_calls: [
{ index: 0, id: "call_1", function: { name: "lookup", arguments: '{"query":"weather"}' } },
],
},
finish_reason: "tool_calls",
},
],
},
),
),
),
)
expect(response.reasoning).toBe("thinking")
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning", reasoningDetails: details },
})
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model, messages: [response.message] }),
)
expect(replay.body.messages).toEqual([
{
role: "assistant",
content: null,
reasoning: "thinking",
reasoning_details: details,
tool_calls: [
{
id: "call_1",
type: "function",
function: { name: "lookup", arguments: '{"query":"weather"}' },
},
],
},
])
}),
)
it.effect("uses reasoning details as display fallback without inventing a scalar replay field", () =>
Effect.gen(function* () {
const details = [
{ type: "reasoning.summary", summary: "thinking", format: "openai-responses-v1", index: 0 },
{ type: "reasoning.encrypted", data: "opaque", format: "openai-responses-v1", index: 1 },
]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning_details: [details[0]] } }] },
{ choices: [{ delta: { reasoning_details: [details[1]] } }] },
{ choices: [{ delta: { content: "Hello" } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("thinking")
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningDetails: details },
})
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model, messages: [response.message] }),
)
expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_details: details }])
}),
)
it.effect("preserves unknown reasoning details while using scalar display text", () =>
Effect.gen(function* () {
const details = [{ type: "reasoning.future", format: "provider-v2", state: { opaque: true } }]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning: "thinking", reasoning_details: details } }] },
{ choices: [{ delta: { content: "Hello" } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("thinking")
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning", reasoningDetails: details },
})
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model, messages: [response.message] }),
)
expect(replay.body.messages).toEqual([
{ role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: details },
])
}),
)
it.effect("uses scalar display text for signature-only reasoning details", () =>
Effect.gen(function* () {
const details = [{ type: "reasoning.text", signature: "signed", format: "provider-v2", index: 0 }]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning: "thinking", reasoning_details: details } }] },
{ choices: [{ delta: { content: "Hello" } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("thinking")
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning", reasoningDetails: details },
})
}),
)
it.effect("ignores scalar reasoning after content starts", () =>
Effect.gen(function* () {
const details = [{ type: "reasoning.text", text: "detail", format: "unknown", index: 0 }]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning_details: details } }] },
{ choices: [{ delta: { content: "Hello" } }] },
{ choices: [{ delta: { reasoning: "scalar" } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("detail")
expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningDetails: details },
})
}),
)
it.effect("preserves an explicitly empty reasoning details array", () =>
Effect.gen(function* () {
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning_details: [] } }] },
{ choices: [{ delta: { content: "Hello" } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("")
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningDetails: [] },
})
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model, messages: [response.message] }),
)
expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_details: [] }])
}),
)
it.effect("attaches signature-only details that arrive after content", () =>
Effect.gen(function* () {
const details = [
{ type: "reasoning.text", text: "thinking", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.text", signature: "signed", format: "anthropic-claude-v1", index: 0 },
]
const merged = [
{
type: "reasoning.text",
text: "thinking",
signature: "signed",
format: "anthropic-claude-v1",
index: 0,
},
]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning: "thinking", reasoning_details: [details[0]] } }] },
{ choices: [{ delta: { content: "Hello" } }] },
{ choices: [{ delta: { reasoning_details: [details[1]] } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("thinking")
expect(response.message.content.filter((part) => part.type === "reasoning")).toHaveLength(1)
expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning", reasoningDetails: merged },
})
expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
expect(response.events.filter(LLMEvent.is.reasoningDelta)).toHaveLength(1)
expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
expect(response.events.filter(LLMEvent.is.reasoningEnd).at(-1)?.providerMetadata).toEqual({
openai: { reasoningField: "reasoning", reasoningDetails: merged },
})
expect(response.events.findIndex(LLMEvent.is.reasoningEnd)).toBeLessThan(
response.events.findIndex(LLMEvent.is.textStart),
)
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model, messages: [response.message] }),
)
expect(replay.body.messages).toEqual([
{ role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: merged },
])
}),
)
it.effect("preserves metadata-only reasoning when the stream ends", () =>
Effect.gen(function* () {
const details = [{ type: "reasoning.encrypted", data: "opaque", format: "openai-responses-v1", index: 0 }]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning_details: details } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.message.content).toEqual([
{ type: "reasoning", text: "", providerMetadata: { openai: { reasoningDetails: details } } },
])
expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({ model, messages: [response.message] }),
)
expect(replay.body.messages).toEqual([{ role: "assistant", content: null, reasoning_details: details }])
}),
)
it.effect("flushes details-only display reasoning when the stream ends", () =>
Effect.gen(function* () {
const details = [{ type: "reasoning.summary", summary: "summary", format: "openai-responses-v1", index: 0 }]
const response = yield* LLMClient.generate(request).pipe(
Effect.provide(
fixedResponse(
sseEvents(
{ choices: [{ delta: { reasoning_details: details } }] },
{ choices: [{ delta: {}, finish_reason: "stop" }] },
),
),
),
)
expect(response.reasoning).toBe("summary")
expect(response.message.content).toEqual([
{ type: "reasoning", text: "summary", providerMetadata: { openai: { reasoningDetails: details } } },
])
}),
)
it.effect("replays details from multiple reasoning parts in order", () =>
Effect.gen(function* () {
const first = { type: "reasoning.text", text: "first", signature: "signed-0", index: 0 }
const second = { type: "reasoning.text", text: "second", signature: "signed-1", index: 1 }
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({
model,
messages: [
Message.assistant([
{
type: "reasoning",
text: "first",
providerMetadata: { openai: { reasoningDetails: [first] } },
},
{
type: "reasoning",
text: "second",
providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: [second] } },
},
]),
],
}),
)
expect(replay.body.messages).toEqual([
{ role: "assistant", content: null, reasoning: "firstsecond", reasoning_details: [first, second] },
])
}),
)
it.effect("retains scalar replay for mixed structured reasoning parts", () =>
Effect.gen(function* () {
const detail = { type: "reasoning.encrypted", data: "opaque", index: 0 }
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({
model,
messages: [
Message.assistant([
{
type: "reasoning",
text: "A",
providerMetadata: { openai: { reasoningDetails: [detail] } },
},
{ type: "reasoning", text: "B" },
]),
],
}),
)
expect(replay.body.messages).toEqual([
{ role: "assistant", content: null, reasoning_content: "AB", reasoning_details: [detail] },
])
}),
)
it.effect("replays native scalar reasoning alongside native details", () =>
Effect.gen(function* () {
const details = [{ type: "reasoning.encrypted", data: "opaque", index: 0 }]
const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
LLM.request({
model,
messages: [
Message.make({
role: "assistant",
content: [{ type: "reasoning", text: "thinking" }],
native: { openaiCompatible: { reasoning_content: "thinking", reasoning_details: details } },
}),
],
}),
)
expect(replay.body.messages).toEqual([
{ role: "assistant", content: null, reasoning_content: "thinking", reasoning_details: details },
])
}),
)
it.effect("assembles streamed tool call input", () =>
Effect.gen(function* () {
const body = sseEvents(
+99 -1
View File
@@ -1,6 +1,6 @@
import { describe, expect } from "bun:test"
import { Effect } from "effect"
import { LLM } from "../../src"
import { LLM, Message } from "../../src"
import { LLMClient } from "../../src/route"
import * as OpenRouter from "../../src/providers/openrouter"
import { it } from "../lib/effect"
@@ -53,4 +53,102 @@ describe("OpenRouter", () => {
})
}),
)
it.effect("preserves manually supplied reasoning details", () =>
Effect.gen(function* () {
const details = [
{ type: "reasoning.text", text: "Think", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.text", text: "ing", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.text", signature: "signed", format: "anthropic-claude-v1", index: 0 },
{ type: "reasoning.encrypted", data: "opaque", format: "openai-responses-v1", index: 1 },
]
const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
LLM.request({
model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
messages: [
Message.assistant([
{
type: "reasoning",
text: "Thinking",
providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
},
]),
],
}),
)
expect(prepared.body.messages).toEqual([
{
role: "assistant",
content: null,
reasoning: "Thinking",
reasoning_details: details,
},
])
}),
)
it.effect("preserves opaque and duplicate continuation details", () =>
Effect.gen(function* () {
const details = [
{ type: "reasoning.future", format: "provider-v2", state: { opaque: true } },
{ type: "reasoning.encrypted", id: "state", data: "opaque" },
{ type: "reasoning.encrypted", id: "state", data: "opaque" },
]
const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
LLM.request({
model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
messages: [
Message.assistant({
type: "reasoning",
text: "Thinking",
providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
}),
],
}),
)
expect(prepared.body.messages).toEqual([
{ role: "assistant", content: null, reasoning: "Thinking", reasoning_details: details },
])
}),
)
it.effect("does not merge distinct adjacent reasoning text blocks", () =>
Effect.gen(function* () {
const details = [
{ type: "reasoning.text", id: "first", index: 0, text: "A", opaque: "first" },
{ type: "reasoning.text", id: "second", index: 1, text: "B", opaque: "second" },
]
const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
LLM.request({
model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
messages: [
Message.assistant({
type: "reasoning",
text: "AB",
providerMetadata: { openai: { reasoningField: "reasoning", reasoningDetails: details } },
}),
],
}),
)
expect(prepared.body.messages).toEqual([
{ role: "assistant", content: null, reasoning: "AB", reasoning_details: details },
])
}),
)
it.effect("omits scalar reasoning without continuation details", () =>
Effect.gen(function* () {
const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
LLM.request({
model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
messages: [Message.assistant({ type: "reasoning", text: "Thinking" })],
}),
)
expect(prepared.body.messages).toEqual([{ role: "assistant", content: null }])
}),
)
})
+2 -23
View File
@@ -120,29 +120,8 @@ export const runWeatherToolLoop = (request: LLMRequest) =>
throw new Error("Weather tool loop exceeded 10 steps")
})
const assistantContent = (events: ReadonlyArray<LLMEvent>) => {
const content: ContentPart[] = []
for (const event of events) {
if (event.type === "text-delta" || event.type === "reasoning-delta") {
const type = event.type === "text-delta" ? "text" : "reasoning"
const last = content.at(-1)
if (last?.type === type) {
content[content.length - 1] = { ...last, text: `${last.text}${event.text}` }
} else {
content.push({ type, text: event.text })
}
continue
}
if (event.type === "text-end" || event.type === "reasoning-end") {
const type = event.type === "text-end" ? "text" : "reasoning"
const last = content.at(-1)
if (last?.type === type) content[content.length - 1] = { ...last, providerMetadata: event.providerMetadata }
continue
}
if (event.type === "tool-call") content.push(event)
}
return content
}
const assistantContent = (events: ReadonlyArray<LLMEvent>) =>
events.reduce(LLMResponse.reduce, LLMResponse.empty()).message.content
export const expectFinish = (
events: ReadonlyArray<LLMEvent>,