@epoch-ai/cli 2.2.5 → 2.2.7

package/test/session/llm.test.ts

@@ -0,0 +1,1308 @@
+import { afterAll, beforeAll, beforeEach, describe, expect, test } from "bun:test"
+import path from "path"
+import { tool, type ModelMessage } from "ai"
+import { Cause, Exit, Stream } from "effect"
+import z from "zod"
+import { makeRuntime } from "../../src/effect/run-service"
+import { LLM } from "../../src/session/llm"
+import { Instance } from "../../src/project/instance"
+import { Provider } from "../../src/provider/provider"
+import { ProviderTransform } from "../../src/provider/transform"
+import { ModelsDev } from "../../src/provider/models"
+import { ProviderID, ModelID } from "../../src/provider/schema"
+import { Filesystem } from "../../src/util/filesystem"
+import { tmpdir } from "../fixture/fixture"
+import type { Agent } from "../../src/agent/agent"
+import type { MessageV2 } from "../../src/session/message-v2"
+import { SessionID, MessageID } from "../../src/session/schema"
+
+describe("session.llm.hasToolCalls", () => {
+  test("returns false for empty messages array", () => {
+    expect(LLM.hasToolCalls([])).toBe(false)
+  })
+
+  test("returns false for messages with only text content", () => {
+    const messages: ModelMessage[] = [
+      {
+        role: "user",
+        content: [{ type: "text", text: "Hello" }],
+      },
+      {
+        role: "assistant",
+        content: [{ type: "text", text: "Hi there" }],
+      },
+    ]
+    expect(LLM.hasToolCalls(messages)).toBe(false)
+  })
+
+  test("returns true when messages contain tool-call", () => {
+    const messages = [
+      {
+        role: "user",
+        content: [{ type: "text", text: "Run a command" }],
+      },
+      {
+        role: "assistant",
+        content: [
+          {
+            type: "tool-call",
+            toolCallId: "call-123",
+            toolName: "bash",
+          },
+        ],
+      },
+    ] as ModelMessage[]
+    expect(LLM.hasToolCalls(messages)).toBe(true)
+  })
+
+  test("returns true when messages contain tool-result", () => {
+    const messages = [
+      {
+        role: "tool",
+        content: [
+          {
+            type: "tool-result",
+            toolCallId: "call-123",
+            toolName: "bash",
+          },
+        ],
+      },
+    ] as ModelMessage[]
+    expect(LLM.hasToolCalls(messages)).toBe(true)
+  })
+
+  test("returns false for messages with string content", () => {
+    const messages: ModelMessage[] = [
+      {
+        role: "user",
+        content: "Hello world",
+      },
+      {
+        role: "assistant",
+        content: "Hi there",
+      },
+    ]
+    expect(LLM.hasToolCalls(messages)).toBe(false)
+  })
+
+  test("returns true when tool-call is mixed with text content", () => {
+    const messages = [
+      {
+        role: "assistant",
+        content: [
+          { type: "text", text: "Let me run that command" },
+          {
+            type: "tool-call",
+            toolCallId: "call-456",
+            toolName: "read",
+          },
+        ],
+      },
+    ] as ModelMessage[]
+    expect(LLM.hasToolCalls(messages)).toBe(true)
+  })
+})
+
+describe("session.llm.parseGroundTruthRules", () => {
+  test("extracts and formats operational facts correctly", () => {
+    const raw = `
+ZONE 1 & 3: OPERATIONAL FACTS (Anchors for high-attention regions)
+These are immutable truths extracted by the MCP scan or defined by the engine.
+operational_facts[22]{fact_id, zone, circumstance, content}:
+fact_01, "zone_1", "Always", "The environment context limit is strictly 32K tokens."
+fact_02, "zone_1", "Always", "Always use parallel tools."
+ZONE 2: BEHAVIORAL RULE PACKS
+`
+    const result = LLM.parseGroundTruthRules(raw, ["new_feature_pack"])
+    expect(result.operationalFacts).toContain("ZONE 1 & 3: OPERATIONAL FACTS")
+    expect(result.operationalFacts).toContain("- The environment context limit is strictly 32K tokens.")
+    expect(result.operationalFacts).toContain("- Always use parallel tools.")
+    expect(result.operationalFacts).not.toContain("fact_01")
+    expect(result.operationalFacts).not.toContain("zone_1")
+  })
+
+  test("replaces [CONTEXT_LIMIT] placeholder in operational facts", () => {
+    const raw = `
+ZONE 1 & 3: OPERATIONAL FACTS
+fact_01, "zone_1", "Always", "The environment context limit is strictly [CONTEXT_LIMIT]."
+`
+    const result = LLM.parseGroundTruthRules(raw, [], 128000)
+    expect(result.operationalFacts).toContain("- The environment context limit is strictly 128K tokens.")
+  })
+
+  test("filters behavioral rules based on active agent pack", () => {
+    const raw = `
+ZONE 1 & 3: OPERATIONAL FACTS
+fact_01, "zone_1", "Always", "Fact 1"
+ZONE 2: BEHAVIORAL RULE PACKS
+rule_packs:
+new_feature_pack: [domains.epistemic.epi_01, domains.reasoning.reas_01]
+context_mgmt_pack: [domains.memory.mem_01]
+
+ZONE 2: RULE LIBRARY
+domains:
+epistemic:
+epi_01, "Trigger 1", "Behaviour 1", "Example 1"
+epi_02, "Trigger 2", "Behaviour 2", "Example 2"
+reasoning:
+reas_01, "Trigger 3", "Behaviour 3", "Example 3"
+memory:
+mem_01, "Trigger 4", "Behaviour 4", "Example 4"
+ZONE 3: PROJECT-SPECIFIC RULES
+Some specific rules
+`
+
+    // Test with "build" agent (maps to new_feature_pack)
+    const buildResult = LLM.parseGroundTruthRules(raw, ["new_feature_pack"])
+    expect(buildResult.behavioralRules).toContain("Trigger 1")
+    expect(buildResult.behavioralRules).toContain("Behaviour 1")
+    expect(buildResult.behavioralRules).toContain("Trigger 3")
+    expect(buildResult.behavioralRules).not.toContain("Trigger 2")
+    expect(buildResult.behavioralRules).not.toContain("Trigger 4")
+
+    // Test with "explore" agent (maps to context_mgmt_pack)
+    const exploreResult = LLM.parseGroundTruthRules(raw, ["context_mgmt_pack"])
+    expect(exploreResult.behavioralRules).toContain("Trigger 4")
+    expect(exploreResult.behavioralRules).not.toContain("Trigger 1")
+  })
+
+  test("combines multiple rule packs correctly", () => {
+    const raw = `
+ZONE 2: BEHAVIORAL RULE PACKS
+rule_packs:
+debugging_pack: [domains.epi_01]
+refactoring_pack: [domains.reas_01]
+
+ZONE 2: RULE LIBRARY
+domains:
+epi_01, "Trigger 1", "Behaviour 1", "Example 1"
+reas_01, "Trigger 2", "Behaviour 2", "Example 2"
+`
+    const result = LLM.parseGroundTruthRules(raw, ["debugging_pack", "refactoring_pack"])
+    expect(result.behavioralRules).toContain("Trigger 1")
+    expect(result.behavioralRules).toContain("Trigger 2")
+  })
+
+  test("falls back to raw text when markers are missing", () => {
+    const raw = `Just some plain text without any zone markers.`
+    const result = LLM.parseGroundTruthRules(raw, ["new_feature_pack"])
+
+    expect(result.operationalFacts).toBe("")
+    expect(result.projectSpecific).toBe("")
+    expect(result.behavioralRules).toBe(raw)
+  })
+
+  test("extracts correctly if project-specific rules are missing", () => {
+    const raw = `
+ZONE 1 & 3: OPERATIONAL FACTS
+fact_01, "zone_1", "Always", "Fact 1"
+ZONE 2: BEHAVIORAL RULE PACKS
+Rule 1`
+
+    const result = LLM.parseGroundTruthRules(raw, ["new_feature_pack"])
+
+    expect(result.operationalFacts).toContain("Fact 1")
+    expect(result.behavioralRules).toContain("Rule 1") // Falls back to raw zone 2 string if pack matches fail
+    expect(result.projectSpecific).toBe("")
+  })
+})
+
+type Capture = {
+  url: URL
+  headers: Headers
+  body: Record<string, unknown>
+}
+
+const state = {
+  server: null as ReturnType<typeof Bun.serve> | null,
+  queue: [] as Array<{
+    path: string
+    response: Response | ((req: Request, capture: Capture) => Response)
+    resolve: (value: Capture) => void
+  }>,
+}
+
+function deferred<T>() {
+  const result = {} as { promise: Promise<T>; resolve: (value: T) => void }
+  result.promise = new Promise((resolve) => {
+    result.resolve = resolve
+  })
+  return result
+}
+
+function waitRequest(pathname: string, response: Response) {
+  const pending = deferred<Capture>()
+  state.queue.push({ path: pathname, response, resolve: pending.resolve })
+  return pending.promise
+}
+
+function timeout(ms: number) {
+  return new Promise<never>((_, reject) => {
+    setTimeout(() => reject(new Error(`timed out after ${ms}ms`)), ms)
+  })
+}
+
+function waitStreamingRequest(pathname: string) {
+  const request = deferred<Capture>()
+  const requestAborted = deferred<void>()
+  const responseCanceled = deferred<void>()
+  const encoder = new TextEncoder()
+
+  state.queue.push({
+    path: pathname,
+    resolve: request.resolve,
+    response(req: Request) {
+      req.signal.addEventListener("abort", () => requestAborted.resolve(), { once: true })
+
+      return new Response(
+        new ReadableStream<Uint8Array>({
+          start(controller) {
+            controller.enqueue(
+              encoder.encode(
+                [
+                  `data: ${JSON.stringify({
+                    id: "chatcmpl-abort",
+                    object: "chat.completion.chunk",
+                    choices: [{ delta: { role: "assistant" } }],
+                  })}`,
+                ].join("\n\n") + "\n\n",
+              ),
+            )
+          },
+          cancel() {
+            responseCanceled.resolve()
+          },
+        }),
+        {
+          status: 200,
+          headers: { "Content-Type": "text/event-stream" },
+        },
+      )
+    },
+  })
+
+  return {
+    request: request.promise,
+    requestAborted: requestAborted.promise,
+    responseCanceled: responseCanceled.promise,
+  }
+}
+
+beforeAll(() => {
+  state.server = Bun.serve({
+    port: 0,
+    async fetch(req) {
+      const next = state.queue.shift()
+      if (!next) {
+        console.error("UNEXPECTED REQUEST:", req.method, req.url)
+        return new Response("unexpected request", { status: 500 })
+      }
+
+      const url = new URL(req.url)
+      const body = (await req.json()) as Record<string, unknown>
+      console.log("MOCK SERVER RECEIVED:", req.method, url.pathname, "messages:", (body.messages as any[])?.length)
+      next.resolve({ url, headers: req.headers, body })
+
+      if (!url.pathname.endsWith(next.path)) {
+        return new Response("not found", { status: 404 })
+      }
+
+      return typeof next.response === "function"
+        ? next.response(req, { url, headers: req.headers, body })
+        : next.response
+    },
+  })
+})
+
+beforeEach(() => {
+  state.queue.length = 0
+})
+
+afterAll(() => {
+  state.server?.stop()
+})
+
+function createChatStream(text: string) {
+  const payload =
+    [
+      `data: ${JSON.stringify({
+        id: "chatcmpl-1",
+        object: "chat.completion.chunk",
+        choices: [{ delta: { role: "assistant" } }],
+      })}`,
+      `data: ${JSON.stringify({
+        id: "chatcmpl-1",
+        object: "chat.completion.chunk",
+        choices: [{ delta: { content: text } }],
+      })}`,
+      `data: ${JSON.stringify({
+        id: "chatcmpl-1",
+        object: "chat.completion.chunk",
+        choices: [{ delta: {}, finish_reason: "stop" }],
+      })}`,
+      "data: [DONE]",
+    ].join("\n\n") + "\n\n"
+
+  const encoder = new TextEncoder()
+  return new ReadableStream<Uint8Array>({
+    start(controller) {
+      controller.enqueue(encoder.encode(payload))
+      controller.close()
+    },
+  })
+}
+
+async function loadFixture(providerID: string, modelID: string) {
+  const fixturePath = path.join(import.meta.dir, "../tool/fixtures/models-api.json")
+  const data = await Filesystem.readJson<Record<string, ModelsDev.Provider>>(fixturePath)
+  const provider = data[providerID]
+  if (!provider) {
+    throw new Error(`Missing provider in fixture: ${providerID}`)
+  }
+  const model = provider.models[modelID]
+  if (!model) {
+    throw new Error(`Missing model in fixture: ${modelID}`)
+  }
+  return { provider, model }
+}
+
+function createEventStream(chunks: unknown[], includeDone = false) {
+  const lines = chunks.map((chunk) => `data: ${typeof chunk === "string" ? chunk : JSON.stringify(chunk)}`)
+  if (includeDone) {
+    lines.push("data: [DONE]")
+  }
+  const payload = lines.join("\n\n") + "\n\n"
+  const encoder = new TextEncoder()
+  return new ReadableStream<Uint8Array>({
+    start(controller) {
+      controller.enqueue(encoder.encode(payload))
+      controller.close()
+    },
+  })
+}
+
+function createEventResponse(chunks: unknown[], includeDone = false) {
+  return new Response(createEventStream(chunks, includeDone), {
+    status: 200,
+    headers: { "Content-Type": "text/event-stream" },
+  })
+}
+
+describe("session.llm.stream", () => {
+  test("sends temperature, tokens, and reasoning options for openai-compatible models", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const providerID = "vivgrid"
+    const modelID = "gemini-3.1-pro-preview"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+
+    const request = waitRequest(
+      "/chat/completions",
+      new Response(createChatStream("Hello"), {
+        status: 200,
+        headers: { "Content-Type": "text/event-stream" },
+      }),
+    )
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-1")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+          temperature: 0.4,
+          topP: 0.8,
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-1"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make(providerID), modelID: resolved.id, variant: "high" },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {},
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        const body = capture.body
+        const headers = capture.headers
+        const url = capture.url
+
+        expect(url.pathname.startsWith("/v1/")).toBe(true)
+        expect(url.pathname.endsWith("/chat/completions")).toBe(true)
+        expect(headers.get("Authorization")).toBe("Bearer test-key")
+
+        expect(body.model).toBe(resolved.api.id)
+        expect(body.temperature).toBe(0.4)
+        expect(body.top_p).toBe(0.8)
+        expect(body.stream).toBe(true)
+
+        const maxTokens = (body.max_tokens as number | undefined) ?? (body.max_output_tokens as number | undefined)
+        const expectedMaxTokens = ProviderTransform.maxOutputTokens(resolved)
+        expect(maxTokens).toBe(expectedMaxTokens)
+
+        const reasoning = (body.reasoningEffort as string | undefined) ?? (body.reasoning_effort as string | undefined)
+        expect(reasoning).toBe("high")
+      },
+    })
+  })
+
+  test("raw stream abort signal cancels provider response body promptly", async () => {
+    const server = state.server
+    if (!server) throw new Error("Server not initialized")
+
+    const providerID = "alibaba"
+    const modelID = "qwen-plus"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+    const pending = waitStreamingRequest("/chat/completions")
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-raw-abort")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+        const user = {
+          id: MessageID.make("user-raw-abort"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make(providerID), modelID: resolved.id },
+        } satisfies MessageV2.User
+
+        const ctrl = new AbortController()
+        const result = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: ctrl.signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {},
+        })
+
+        const iter = result.fullStream[Symbol.asyncIterator]()
+        await pending.request
+        await iter.next()
+        ctrl.abort()
+
+        await Promise.race([pending.responseCanceled, timeout(500)])
+        await Promise.race([pending.requestAborted, timeout(500)]).catch(() => undefined)
+        await iter.return?.()
+      },
+    })
+  })
+
+  test("service stream cancellation cancels provider response body promptly", async () => {
+    const server = state.server
+    if (!server) throw new Error("Server not initialized")
+
+    const providerID = "alibaba"
+    const modelID = "qwen-plus"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+    const pending = waitStreamingRequest("/chat/completions")
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-service-abort")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+        const user = {
+          id: MessageID.make("user-service-abort"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make(providerID), modelID: resolved.id },
+        } satisfies MessageV2.User
+
+        const ctrl = new AbortController()
+        const { runPromiseExit } = makeRuntime(LLM.Service, LLM.defaultLayer)
+        const run = runPromiseExit(
+          (svc) =>
+            svc
+              .stream({
+                user,
+                sessionID,
+                model: resolved,
+                agent,
+                system: { zone1: [], zone2: ["You are a helpful assistant."] },
+                messages: [{ role: "user", content: "Hello" }],
+                tools: {},
+              })
+              .pipe(Stream.runDrain),
+          { signal: ctrl.signal },
+        )
+
+        await pending.request
+        ctrl.abort()
+
+        await Promise.race([pending.responseCanceled, timeout(500)])
+        const exit = await run
+        expect(Exit.isFailure(exit)).toBe(true)
+        if (Exit.isFailure(exit)) {
+          expect(Cause.hasInterrupts(exit.cause)).toBe(true)
+        }
+        await Promise.race([pending.requestAborted, timeout(500)]).catch(() => undefined)
+      },
+    })
+  })
+
+  test("keeps tools enabled by prompt permissions", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const providerID = "alibaba"
+    const modelID = "qwen-plus"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+
+    const request = waitRequest(
+      "/chat/completions",
+      new Response(createChatStream("Hello"), {
+        status: 200,
+        headers: { "Content-Type": "text/event-stream" },
+      }),
+    )
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-tools")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "question", pattern: "*", action: "deny" }],
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-tools"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make(providerID), modelID: resolved.id },
+          tools: { question: true },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          permission: [{ permission: "question", pattern: "*", action: "allow" }],
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {
+            question: tool({
+              description: "Ask a question",
+              inputSchema: z.object({}),
+              execute: async () => ({ output: "" }),
+            }),
+          },
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        const tools = capture.body.tools as Array<{ function?: { name?: string } }> | undefined
+        expect(tools?.some((item) => item.function?.name === "question")).toBe(true)
+      },
+    })
+  })
+
+  test("sends responses API payload for OpenAI models", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const source = await loadFixture("openai", "gpt-5.2")
+    const model = source.model
+
+    const responseChunks = [
+      {
+        type: "response.created",
+        response: {
+          id: "resp-1",
+          created_at: Math.floor(Date.now() / 1000),
+          model: model.id,
+          service_tier: null,
+        },
+      },
+      {
+        type: "response.output_text.delta",
+        item_id: "item-1",
+        delta: "Hello",
+        logprobs: null,
+      },
+      {
+        type: "response.completed",
+        response: {
+          incomplete_details: null,
+          usage: {
+            input_tokens: 1,
+            input_tokens_details: null,
+            output_tokens: 1,
+            output_tokens_details: null,
+          },
+          service_tier: null,
+        },
+      },
+    ]
+    const request = waitRequest("/responses", createEventResponse(responseChunks, true))
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: ["openai"],
+            provider: {
+              openai: {
+                name: "OpenAI",
+                env: ["OPENAI_API_KEY"],
+                npm: "@ai-sdk/openai",
+                api: "https://api.openai.com/v1",
+                models: {
+                  [model.id]: model,
+                },
+                options: {
+                  apiKey: "test-openai-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.openai, ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-2")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+          temperature: 0.2,
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-2"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make("openai"), modelID: resolved.id, variant: "high" },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {},
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        const body = capture.body
+
+        expect(capture.url.pathname.endsWith("/responses")).toBe(true)
+        expect(body.model).toBe(resolved.api.id)
+        expect(body.stream).toBe(true)
+        expect((body.reasoning as { effort?: string } | undefined)?.effort).toBe("high")
+
+        const maxTokens = body.max_output_tokens as number | undefined
+        expect(maxTokens).toBe(undefined) // match codex cli behavior
+      },
+    })
+  })
+
+  test("accepts user image attachments as data URLs for OpenAI models", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const source = await loadFixture("openai", "gpt-5.2")
+    const model = source.model
+    const chunks = [
+      {
+        type: "response.created",
+        response: {
+          id: "resp-data-url",
+          created_at: Math.floor(Date.now() / 1000),
+          model: model.id,
+          service_tier: null,
+        },
+      },
+      {
+        type: "response.output_text.delta",
+        item_id: "item-data-url",
+        delta: "Looks good",
+        logprobs: null,
+      },
+      {
+        type: "response.completed",
+        response: {
+          incomplete_details: null,
+          usage: {
+            input_tokens: 1,
+            input_tokens_details: null,
+            output_tokens: 1,
+            output_tokens_details: null,
+          },
+          service_tier: null,
+        },
+      },
+    ]
+    const request = waitRequest("/responses", createEventResponse(chunks, true))
+    const image = `data:image/png;base64,${Buffer.from(
+      await Bun.file(path.join(import.meta.dir, "../tool/fixtures/large-image.png")).arrayBuffer(),
+    ).toString("base64")}`
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: ["openai"],
+            provider: {
+              openai: {
+                name: "OpenAI",
+                env: ["OPENAI_API_KEY"],
+                npm: "@ai-sdk/openai",
+                api: "https://api.openai.com/v1",
+                models: {
+                  [model.id]: model,
+                },
+                options: {
+                  apiKey: "test-openai-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.openai, ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-data-url")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-data-url"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make("openai"), modelID: resolved.id },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: new AbortController().signal,
+          messages: [
+            {
+              role: "user",
+              content: [
+                { type: "text", text: "Describe this image" },
+                {
+                  type: "file",
+                  mediaType: "image/png",
+                  filename: "large-image.png",
+                  data: image,
+                },
+              ],
+            },
+          ] as ModelMessage[],
+          tools: {},
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        expect(capture.url.pathname.endsWith("/responses")).toBe(true)
+      },
+    })
+  })
+
+  test("sends messages API payload for Anthropic Compatible models", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const providerID = "minimax"
+    const modelID = "MiniMax-M2.5"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+
+    const chunks = [
+      {
+        type: "message_start",
+        message: {
+          id: "msg-1",
+          model: model.id,
+          usage: {
+            input_tokens: 3,
+            cache_creation_input_tokens: null,
+            cache_read_input_tokens: null,
+          },
+        },
+      },
+      {
+        type: "content_block_start",
+        index: 0,
+        content_block: { type: "text", text: "" },
+      },
+      {
+        type: "content_block_delta",
+        index: 0,
+        delta: { type: "text_delta", text: "Hello" },
+      },
+      { type: "content_block_stop", index: 0 },
+      {
+        type: "message_delta",
+        delta: { stop_reason: "end_turn", stop_sequence: null, container: null },
+        usage: {
+          input_tokens: 3,
+          output_tokens: 2,
+          cache_creation_input_tokens: null,
+          cache_read_input_tokens: null,
+        },
+      },
+      { type: "message_stop" },
+    ]
+    const request = waitRequest("/messages", createEventResponse(chunks))
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-anthropic-key",
+                  baseURL: `${server.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-3")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+          temperature: 0.4,
+          topP: 0.9,
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-3"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make("minimax"), modelID: ModelID.make("MiniMax-M2.5") },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {},
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        const body = capture.body
+
+        expect(capture.url.pathname.endsWith("/messages")).toBe(true)
+        expect(body.model).toBe(resolved.api.id)
+        expect(body.max_tokens).toBe(ProviderTransform.maxOutputTokens(resolved))
+        expect(body.temperature).toBe(0.4)
+        expect(body.top_p).toBe(0.9)
+      },
+    })
+  })
+
+  test("sends Google API payload for Gemini models", async () => {
+    const server = state.server
+    if (!server) {
+      throw new Error("Server not initialized")
+    }
+
+    const providerID = "google"
+    const modelID = "gemini-2.5-flash"
+    const fixture = await loadFixture(providerID, modelID)
+    const provider = fixture.provider
+    const model = fixture.model
+    const pathSuffix = `/v1beta/models/${model.id}:streamGenerateContent`
+
+    const chunks = [
+      {
+        candidates: [
+          {
+            content: {
+              parts: [{ text: "Hello" }],
+            },
+            finishReason: "STOP",
+          },
+        ],
+        usageMetadata: {
+          promptTokenCount: 1,
+          candidatesTokenCount: 1,
+          totalTokenCount: 2,
+        },
+      },
+    ]
+    const request = waitRequest(pathSuffix, createEventResponse(chunks))
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            $schema: "https://epochcli.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-google-key",
+                  baseURL: `${server.url.origin}/v1beta`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-test-4")
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+          temperature: 0.3,
+          topP: 0.8,
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-4"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make(providerID), modelID: resolved.id },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: ["You are a helpful assistant."] },
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {},
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        const body = capture.body
+        const config = body.generationConfig as
+          | { temperature?: number; topP?: number; maxOutputTokens?: number }
+          | undefined
+
+        expect(capture.url.pathname).toBe(pathSuffix)
+        expect(config?.temperature).toBe(0.3)
+        expect(config?.topP).toBe(0.8)
+        expect(config?.maxOutputTokens).toBe(ProviderTransform.maxOutputTokens(resolved))
+      },
+    })
+  })
+
+  test("injects Internal State Check block with 'model' role and dynamic agent name", async () => {
+    const providerID = "vivgrid"
+    const modelID = "gemini-3.1-pro-preview"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+
+    const request = new Promise<Capture>((resolve) =>
+      state.queue.push({
+        path: "/chat/completions",
+        response: new Response(createChatStream("Hello"), {
+          status: 200,
+          headers: { "Content-Type": "text/event-stream" },
+        }),
+        resolve,
+      }),
+    )
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "epochcli.json"),
+          JSON.stringify({
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-key",
+                  baseURL: `${state.server!.url.origin}/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(ProviderID.make(providerID), ModelID.make(model.id))
+        const sessionID = SessionID.make("session-state-check")
+        const agent = {
+          name: "build",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+        } satisfies Agent.Info
+
+        const user = {
+          id: MessageID.make("user-state-check"),
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID: ProviderID.make(providerID), modelID: resolved.id },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: { zone1: [], zone2: [] },
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Hello" }],
+          tools: {},
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        const messages = capture.body.messages as any[]
+
+        // Find the system messages
+        const systemMessages = messages.filter((m) => m.role === "system")
+        expect(systemMessages.length).toBeGreaterThanOrEqual(1)
+
+        // Find the user message containing state check
+        const userMsg = messages.find((m) => m.role === "user" && m.content.includes("[INTERNAL STATE CHECK]"))
+
+        expect(userMsg).toBeDefined()
+        expect(userMsg.content).toContain("Hello")
+        expect(userMsg.content).toContain("Role: Assigned to [BUILD].")
+        expect(userMsg.content).toContain("<|channel>thought")
+      },
+    })
+  })
+})
+
+describe("session.llm.toolExecutionWrapper", () => {
+  test("Auto-Fallback catches prerequisite rejection and executes sc_guidance", async () => {
+    // This is a placeholder test for the Auto-Fallback logic.
+    // Setting up the full LLM context for originalExecute wrapper is complex,
+    // so we assert that the structural logic is present in the codebase.
+    const fs = require("fs")
+    const llmCode = fs.readFileSync("src/session/llm.ts", "utf-8")
+    expect(llmCode).toContain("Auto-fallback triggered for missing prerequisite sc_guidance")
+    expect(llmCode).toContain("System overriding sc_approve. Prerequisite missing. Auto-executing sc_guidance.")
+  })
+
+  test("Clerk Interceptor catches generic prerequisite errors and generates directive", async () => {
+    // This is a placeholder test for the Clerk Interceptor logic.
+    const fs = require("fs")
+    const llmCode = fs.readFileSync("src/session/llm.ts", "utf-8")
+    expect(llmCode).toContain("Triggering Clerk Interceptor for prerequisite error")
+    expect(llmCode).toContain("CRITICAL SYSTEM DIRECTIVE:")
+  })
+})