hotsocket.fyi
This repository has no description
summarize via llm during compacts
author
nekomimi.pet
date
(Apr 18, 2026, 7:28 AM -0400)
commit
ddc2254c
ddc2254c5993d072075cf03a10326f68b0c0a03b
parent
8bc5aad9
8bc5aad95b22285c213c040aa5497da4d788fde6
+44
-1
+44
-1
src/runner/loop.ts
+44
-1
src/runner/loop.ts
···
15
15
import type { LoopHooks, LoopState } from "./types.js"
16
16
import {
17
17
API_BASE,
18
+
CONTEXT_COMPACT_TRIGGER_TOKENS,
18
19
FALLBACK_MODEL,
19
20
FALLBACK_TOOL_CHOICE,
20
21
FALLBACK_TOKEN_NUDGE_THRESHOLD,
···
36
37
shouldFallback,
37
38
summarizeConversationViaLLM,
38
39
} from "./util.js"
40
+
41
+
const LLM_POST_TURN_RECENT_MESSAGES = 40
39
42
40
43
type FunctionToolCall = OpenAI.Chat.ChatCompletionMessageToolCall & { type: "function" }
41
44
type ToolArgs = {
···
936
939
}
937
940
938
941
/**
942
+
* Post-turn compaction: when observed context crosses the compaction trigger,
943
+
* replace the middle of the conversation with a single LLM-generated summary so
944
+
* older history keeps its narrative/emotional texture instead of being reduced
945
+
* to bullet lines. Falls back to the heuristic compactor on failure.
946
+
*/
947
+
async function applyPostTurnCompaction(state: LoopState): Promise<boolean> {
948
+
if (state.contextSize >= CONTEXT_COMPACT_TRIGGER_TOKENS) {
949
+
const summaryClient = USE_FALLBACK ? fallbackClient : client
950
+
const summaryModel = USE_FALLBACK ? FALLBACK_MODEL : MODEL
951
+
if (summaryClient && summaryModel) {
952
+
const beforeCount = state.conversation.length
953
+
const beforeEstimate = estimatePromptTokens(state.conversation)
954
+
const summarized = await summarizeConversationViaLLM(
955
+
state.conversation,
956
+
summaryClient,
957
+
summaryModel,
958
+
{ recentKeep: LLM_POST_TURN_RECENT_MESSAGES },
959
+
)
960
+
if (summarized) {
961
+
const afterEstimate = estimatePromptTokens(summarized)
962
+
if (afterEstimate < beforeEstimate) {
963
+
state.conversation = summarized
964
+
state.contextSize = afterEstimate
965
+
console.log(
966
+
`[context] post-turn llm-summarized (${beforeCount} -> ${summarized.length} msgs, ${beforeEstimate} -> ${afterEstimate})`,
967
+
)
968
+
return true
969
+
}
970
+
console.warn(
971
+
`[context] post-turn llm summary not smaller (${beforeEstimate} -> ${afterEstimate}); falling back to heuristic`,
972
+
)
973
+
} else {
974
+
console.warn(`[context] post-turn llm summary unavailable; falling back to heuristic`)
975
+
}
976
+
}
977
+
}
978
+
return applyRollingCompaction(state, "post-turn")
979
+
}
980
+
981
+
/**
939
982
* Executes the main assistant/tool loop for an active conversation.
940
983
*
941
984
* The loop repeatedly:
···
957
1000
const outcome = await processAssistantTurn(convId, state, hooks)
958
1001
if (outcome === CycleOutcome.Rest) break
959
1002
960
-
applyRollingCompaction(state, "post-turn")
1003
+
await applyPostTurnCompaction(state)
961
1004
if (outcome !== CycleOutcome.NoTools) {
962
1005
applyContextNudge(state)
963
1006
}