fix(core): cap session output tokens

packages/core/src/session/runner/llm.ts

@@ -40,6 +40,8 @@ import { Snapshot } from "../../snapshot"
 import { makeLocationNode } from "../../effect/app-node"
 import { llmClient } from "../../effect/app-node-platform"
 
+const MAX_OUTPUT_TOKENS = 32_000
+
 /**
  * Runs one durable coding-agent Session until it settles.
  *
@@ -210,6 +212,9 @@ const layer = Layer.effect(
           .map(SystemPart.make),
         messages: [...toLLMMessages(context, model), ...(isLastStep ? [Message.assistant(MAX_STEPS_PROMPT)] : [])],
         tools: toolMaterialization?.definitions ?? [],
+        generation: {
+          maxTokens: Math.min(model.route.defaults.limits?.output ?? 0, MAX_OUTPUT_TOKENS) || MAX_OUTPUT_TOKENS,
+        },
         toolChoice: isLastStep ? "none" : undefined,
       })
       if (yield* compaction.compactIfNeeded({ sessionID: session.id, entries, model, request }))

packages/core/test/session-runner.test.ts

@@ -108,6 +108,11 @@ const recoveryModel = Model.make({
   provider: "fake",
   route: OpenAIChat.route.with({ limits: { context: 20_000, output: 1_000 } }),
 })
+const fullContextOutputModel = Model.make({
+  id: "full-context-output",
+  provider: "fake",
+  route: OpenAIChat.route.with({ limits: { context: 500_000, output: 500_000 } }),
+})
 const authorizations: Tool.Context[] = []
 const executions: string[] = []
 const permission = Layer.succeed(
@@ -655,6 +660,22 @@ describe("SessionRunnerLLM", () => {
     }),
   )
 
+  it.effect("caps output when the catalog output limit consumes the full context window", () =>
+    Effect.gen(function* () {
+      yield* setup
+      const session = yield* SessionV2.Service
+      currentModel = fullContextOutputModel
+      yield* session.prompt({ sessionID, prompt: Prompt.make({ text: "hi" }), resume: false })
+      requests.length = 0
+      response = []
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(1)
+      expect(requests[0]?.generation).toEqual({ maxTokens: 32_000 })
+    }),
+  )
+
   it.effect("retries the first provider turn after system context becomes available", () =>
     Effect.gen(function* () {
       yield* setup