keep recent turns during session compaction

Author: nexxeln

packages/opencode/src/agent/prompt/compaction.txt

@@ -1,6 +1,7 @@
 You are a helpful AI assistant tasked with summarizing conversations.
 
-When asked to summarize, provide a detailed but concise summary of the conversation.
+When asked to summarize, provide a detailed but concise summary of the older conversation history.
+The most recent turns may be preserved verbatim outside your summary, so focus on information that would still be needed to continue the work with that recent context available.
 Focus on information that would be helpful for continuing the conversation, including:
 - What was done
 - What is currently being worked on

packages/opencode/src/config/config.ts

@@ -1069,6 +1069,18 @@ export namespace Config {
         .object({
           auto: z.boolean().optional().describe("Enable automatic compaction when context is full (default: true)"),
           prune: z.boolean().optional().describe("Enable pruning of old tool outputs (default: true)"),
+          tail_turns: z
+            .number()
+            .int()
+            .min(0)
+            .optional()
+            .describe("Number of recent real user turns to keep verbatim during compaction (default: 2)"),
+          tail_tokens: z
+            .number()
+            .int()
+            .min(0)
+            .optional()
+            .describe("Token budget for retained recent turns during compaction"),
           reserved: z
             .number()
             .int()

packages/opencode/src/session/compaction.ts

@@ -4,6 +4,7 @@ import { Session } from "."
 import { SessionID, MessageID, PartID } from "./schema"
 import { Instance } from "../project/instance"
 import { Provider } from "../provider/provider"
+import { ProviderTransform } from "../provider/transform"
 import { MessageV2 } from "./message-v2"
 import z from "zod"
 import { Token } from "../util/token"
@@ -35,6 +36,24 @@ export namespace SessionCompaction {
   export const PRUNE_MINIMUM = 20_000
   export const PRUNE_PROTECT = 40_000
   const PRUNE_PROTECTED_TOOLS = ["skill"]
+  const DEFAULT_TAIL_TURNS = 2
+  const MIN_TAIL_TOKENS = 2_000
+  const MAX_TAIL_TOKENS = 8_000
+
+  function usable(input: { cfg: Config.Info; model: Provider.Model }) {
+    const reserved =
+      input.cfg.compaction?.reserved ?? Math.min(20_000, ProviderTransform.maxOutputTokens(input.model))
+    return input.model.limit.input
+      ? Math.max(0, input.model.limit.input - reserved)
+      : Math.max(0, input.model.limit.context - ProviderTransform.maxOutputTokens(input.model))
+  }
+
+  function tailBudget(input: { cfg: Config.Info; model: Provider.Model }) {
+    return (
+      input.cfg.compaction?.tail_tokens ??
+      Math.min(MAX_TAIL_TOKENS, Math.max(MIN_TAIL_TOKENS, Math.floor(usable(input) * 0.25)))
+    )
+  }
 
   export interface Interface {
     readonly isOverflow: (input: {
@@ -88,6 +107,55 @@ export namespace SessionCompaction {
         return overflow({ cfg: yield* config.get(), tokens: input.tokens, model: input.model })
       })
 
+      const estimate = Effect.fn("SessionCompaction.estimate")(function* (input: {
+        messages: MessageV2.WithParts[]
+        model: Provider.Model
+      }) {
+        const msgs = yield* MessageV2.toModelMessagesEffect(input.messages, input.model, { stripMedia: true })
+        return Token.estimate(JSON.stringify(msgs))
+      })
+
+      const select = Effect.fn("SessionCompaction.select")(function* (input: {
+        messages: MessageV2.WithParts[]
+        cfg: Config.Info
+        model: Provider.Model
+      }) {
+        const limit = input.cfg.compaction?.tail_turns ?? DEFAULT_TAIL_TURNS
+        if (limit <= 0) return { head: input.messages, tail_start_id: undefined }
+        const budget = tailBudget({ cfg: input.cfg, model: input.model })
+        const turns = input.messages.flatMap((msg, idx) =>
+          msg.info.role === "user" && !msg.parts.some((part) => part.type === "compaction") ? [idx] : [],
+        )
+        if (!turns.length) return { head: input.messages, tail_start_id: undefined }
+
+        let total = 0
+        let start = input.messages.length
+        let kept = 0
+
+        for (let i = turns.length - 1; i >= 0 && kept < limit; i--) {
+          const idx = turns[i]
+          const end = i + 1 < turns.length ? turns[i + 1] : input.messages.length
+          const size = yield* estimate({
+            messages: input.messages.slice(idx, end),
+            model: input.model,
+          })
+          if (kept === 0 && size > budget) {
+            log.info("tail fallback", { budget, size })
+            return { head: input.messages, tail_start_id: undefined }
+          }
+          if (total + size > budget) break
+          total += size
+          start = idx
+          kept++
+        }
+
+        if (kept === 0 || start === 0) return { head: input.messages, tail_start_id: undefined }
+        return {
+          head: input.messages.slice(0, start),
+          tail_start_id: input.messages[start]?.info.id,
+        }
+      })
+
       // goes backwards through parts until there are PRUNE_PROTECT tokens worth of tool
       // calls, then erases output of older tool calls to free context space
       const prune = Effect.fn("SessionCompaction.prune")(function* (input: { sessionID: SessionID }) {
@@ -150,6 +218,7 @@ export namespace SessionCompaction {
           throw new Error(`Compaction parent must be a user message: ${input.parentID}`)
         }
         const userMessage = parent.info
+        const compactionPart = parent.parts.find((part): part is MessageV2.CompactionPart => part.type === "compaction")
 
         let messages = input.messages
         let replay:
@@ -180,14 +249,22 @@ export namespace SessionCompaction {
         const model = agent.model
           ? yield* provider.getModel(agent.model.providerID, agent.model.modelID)
           : yield* provider.getModel(userMessage.model.providerID, userMessage.model.modelID)
+        const cfg = yield* config.get()
+        const history = compactionPart && messages.at(-1)?.info.id === input.parentID ? messages.slice(0, -1) : messages
+        const selected = yield* select({
+          messages: history,
+          cfg,
+          model,
+        })
         // Allow plugins to inject context or replace compaction prompt.
         const compacting = yield* plugin.trigger(
           "experimental.session.compacting",
           { sessionID: input.sessionID },
           { context: [], prompt: undefined },
         )
-        const defaultPrompt = `Provide a detailed prompt for continuing our conversation above.
-Focus on information that would be helpful for continuing the conversation, including what we did, what we're doing, which files we're working on, and what we're going to do next.
+        const defaultPrompt = `Summarize the older conversation history so another agent can continue the work with the retained recent turns.
+The most recent conversation turns will remain verbatim outside this summary, so focus on older context that is still needed to understand and continue the work.
+Include what we did, what we're doing, which files we're working on, and what we're going to do next.
 The summary that you construct will be used so that another agent can read it and continue the work.
 Do not call any tools. Respond only with the summary text.
 Respond in the same language as the user's messages in the conversation.
@@ -217,7 +294,7 @@ When constructing the summary, try to stick to this template:
 ---`
 
         const prompt = compacting.prompt ?? [defaultPrompt, ...compacting.context].join("\n\n")
-        const msgs = structuredClone(messages)
+        const msgs = structuredClone(selected.head)
         yield* plugin.trigger("experimental.chat.messages.transform", {}, { messages: msgs })
         const modelMessages = yield* MessageV2.toModelMessagesEffect(msgs, model, { stripMedia: true })
         const ctx = yield* InstanceState.context
@@ -280,6 +357,13 @@ When constructing the summary, try to stick to this template:
           return "stop"
         }
 
+        if (compactionPart && selected.tail_start_id && compactionPart.tail_start_id !== selected.tail_start_id) {
+          yield* session.updatePart({
+            ...compactionPart,
+            tail_start_id: selected.tail_start_id,
+          })
+        }
+
         if (result === "continue" && input.auto) {
           if (replay) {
             const original = replay.info

packages/opencode/src/session/message-v2.ts

@@ -208,6 +208,7 @@ export namespace MessageV2 {
     type: z.literal("compaction"),
     auto: z.boolean(),
     overflow: z.boolean().optional(),
+    tail_start_id: MessageID.zod.optional(),
   }).meta({
     ref: "CompactionPart",
   })
@@ -926,14 +927,21 @@ export namespace MessageV2 {
   export function filterCompacted(msgs: Iterable<MessageV2.WithParts>) {
     const result = [] as MessageV2.WithParts[]
     const completed = new Set<string>()
+    let retain: MessageID | undefined
     for (const msg of msgs) {
       result.push(msg)
-      if (
-        msg.info.role === "user" &&
-        completed.has(msg.info.id) &&
-        msg.parts.some((part) => part.type === "compaction")
-      )
-        break
+      if (retain) {
+        if (msg.info.id === retain) break
+        continue
+      }
+      if (msg.info.role === "user" && completed.has(msg.info.id)) {
+        const part = msg.parts.find((item): item is MessageV2.CompactionPart => item.type === "compaction")
+        if (!part) continue
+        if (!part.tail_start_id) break
+        retain = part.tail_start_id
+        if (msg.info.id === retain) break
+        continue
+      }
       if (msg.info.role === "assistant" && msg.info.summary && msg.info.finish && !msg.info.error)
         completed.add(msg.info.parentID)
     }