fix(core): don't resend a rejected conversation unchanged when compacting (#61015)

This commit is contained in:
Aiden Cline
2026-11-01 00:03:57 -05:00
committed by GitHub
Unverified
parent 01276f9061
commit 6597f6cd6c
2 changed files with 46 additions and 3 deletions
+14 -3
View File
@@ -518,7 +518,13 @@ export const layer = Layer.effect(
prepared,
budget - overhead,
send,
{ fit: fit - overhead, history: withoutTail(prepared.request, keepID), measured },
{
fit: fit - overhead,
history: withoutTail(prepared.request, keepID),
measured,
// The provider just rejected this conversation, so sending it unchanged again would be rejected too.
text: trigger.reason === "overflow",
},
background,
)
if (!keepID) return result
@@ -596,7 +602,12 @@ export const layer = Layer.effect(
prepared: Prepared,
target: number,
send: (request: LLMRequest, size?: Estimate) => Effect.Effect<Result, AIError | Failure>,
shrink: { readonly fit: number; readonly history: LLMRequest; readonly measured?: Estimate } = {
shrink: {
readonly fit: number
readonly history: LLMRequest
readonly measured?: Estimate
readonly text?: boolean
} = {
fit: target,
history: prepared.request,
},
@@ -608,7 +619,7 @@ export const layer = Layer.effect(
const policy = yield* SessionRunnerRetry.policy(context.session.id)
let rejections = 0
let rejected: number | undefined
let asText = false
let asText = shrink.text ?? false
let fit = shrink.fit
while (true) {
+32
View File
@@ -3191,6 +3191,38 @@ describe("SessionRunnerLLM", () => {
expect((yield* s.messages).find((message) => message.id === compaction.id)).toMatchObject({ status: "completed" })
})
scenario("does not resend a rejected conversation unchanged when compacting", function* (s) {
const filler = "context ".repeat(1_000)
// The provider counts far more than the characters suggest, so only its rejection shows the conversation is too long.
const answer = (index: number) =>
TestLLM.complete(
{
reason: { normalized: "stop" },
usage: { inputTokens: 60_000, nonCachedInputTokens: 60_000, outputTokens: 10 },
},
LLMEvent.textStart({ id: `answer-${index}` }),
LLMEvent.textDelta({ id: `answer-${index}`, text: `Answer ${index}` }),
LLMEvent.textEnd({ id: `answer-${index}` }),
)
yield* s.llm.push(...Array.from({ length: 4 }, (_, index) => answer(index)))
yield* Effect.forEach(
Array.from({ length: 4 }, (_, index) => index),
(index) => s.runPrompt(`Request ${index}: ${filler}`),
)
s.currentModel = recoveryModel
s.requests.length = 0
yield* s.llm.push(
TestLLM.failAfter(providerRejected({ message: "prompt too long", classification: "context-overflow" })),
TestLLM.text("## Objective\n- Recovered", "summary"),
TestLLM.text("Recovered", "recovered"),
)
yield* s.runPrompt("Continue")
expect(s.requests).toHaveLength(3)
expect(s.requests[1]?.tools).toEqual([])
expect(s.requests[1]?.messages.every((message) => message.role === "user")).toBeTrue()
})
scenario("aims the first overflow compaction below the rejected context", function* (s) {
const filler = "context ".repeat(1_000)
yield* s.llm.push(...Array.from({ length: 4 }, (_, index) => TestLLM.text(`Answer ${index}`, `answer-${index}`)))