feat(core): compact v2 session context

packages/core/src/config/compaction.ts

@@ -4,7 +4,6 @@ import { Schema } from "effect"
 import { NonNegativeInt } from "../schema"
 
 export class Keep extends Schema.Class<Keep>("ConfigV2.Compaction.Keep")({
-  turns: NonNegativeInt.pipe(Schema.optional),
   tokens: NonNegativeInt.pipe(Schema.optional),
 }) {}
 

packages/core/src/session/compaction.ts

@@ -0,0 +1,224 @@
+export * as SessionCompaction from "./compaction"
+
+import { LLM, LLMError, LLMEvent, Message, type LLMRequest, type Model } from "@opencode-ai/llm"
+import { DateTime, Effect, Stream } from "effect"
+import type { Config } from "../config"
+import type { EventV2 } from "../event"
+import { SessionEvent } from "./event"
+import { SessionMessage } from "./message"
+import { SessionSchema } from "./schema"
+import { Token } from "../util/token"
+
+const DEFAULT_BUFFER = 20_000
+const DEFAULT_KEEP_TOKENS = 8_000
+const TOOL_OUTPUT_MAX_CHARS = 2_000
+const SUMMARY_OUTPUT_TOKENS = 4_096
+const SUMMARY_TEMPLATE = `Output exactly the Markdown structure shown inside <template> and keep the section order unchanged. Do not include the <template> tags in your response.
+<template>
+## Goal
+- [single-sentence task summary]
+
+## Constraints & Preferences
+- [user constraints, preferences, specs, or "(none)"]
+
+## Progress
+### Done
+- [completed work or "(none)"]
+
+### In Progress
+- [current work or "(none)"]
+
+### Blocked
+- [blockers or "(none)"]
+
+## Key Decisions
+- [decision and why, or "(none)"]
+
+## Next Steps
+- [ordered next actions or "(none)"]
+
+## Critical Context
+- [important technical facts, errors, open questions, or "(none)"]
+
+## Relevant Files
+- [file or directory path: why it matters, or "(none)"]
+</template>
+
+Rules:
+- Keep every section, even when empty.
+- Use terse bullets, not prose paragraphs.
+- Preserve exact file paths, commands, error strings, and identifiers when known.
+- Do not mention the summary process or that context was compacted.`
+
+type Entry = {
+  readonly seq: number
+  readonly message: SessionMessage.Message
+}
+
+type Settings = {
+  readonly auto: boolean
+  readonly buffer: number
+  readonly tokens: number
+}
+
+type Dependencies = {
+  readonly events: EventV2.Interface
+  readonly llm: {
+    readonly stream: (request: LLMRequest) => Stream.Stream<LLMEvent, LLMError>
+  }
+  readonly config: readonly Config.Entry[]
+}
+
+const estimate = (value: unknown) => Token.estimate(JSON.stringify(value))
+
+const truncate = (value: string) =>
+  value.length <= TOOL_OUTPUT_MAX_CHARS ? value : `${value.slice(0, TOOL_OUTPUT_MAX_CHARS)}\n[truncated]`
+
+const serialize = (message: SessionMessage.Message) => {
+  if (message.type === "user") {
+    const files = message.files?.map((file) => `[Attached ${file.mime}: ${file.name ?? file.uri}]`) ?? []
+    return [`[User]: ${message.text}`, ...files].join("\n")
+  }
+  if (message.type === "assistant") {
+    return message.content
+      .flatMap((part) => {
+        if (part.type === "text") return [`[Assistant]: ${part.text}`]
+        if (part.type === "reasoning") return part.text ? [`[Assistant reasoning]: ${part.text}`] : []
+        const input = typeof part.state.input === "string" ? part.state.input : JSON.stringify(part.state.input)
+        if (part.state.status === "completed")
+          return [
+            `[Assistant tool call]: ${part.name}(${input})`,
+            `[Tool result]: ${truncate(JSON.stringify(part.state.content))}`,
+          ]
+        if (part.state.status === "error")
+          return [`[Assistant tool call]: ${part.name}(${input})`, `[Tool error]: ${part.state.error.message}`]
+        return [`[Assistant tool call]: ${part.name}(${input})`]
+      })
+      .join("\n")
+  }
+  if (message.type === "system") return `[System update]: ${message.text}`
+  if (message.type === "synthetic") return `[Synthetic context]: ${message.text}`
+  if (message.type === "shell") return `[Shell]: ${message.command}\n${truncate(message.output)}`
+  return ""
+}
+
+const settings = (documents: readonly Config.Entry[]) => {
+  const configured = documents
+    .filter((entry): entry is Config.Document => entry.type === "document")
+    .flatMap((entry) => (entry.info.compaction ? [entry.info.compaction] : []))
+  return configured.reduce<Settings>(
+    (result, current) => ({
+      auto: current.auto ?? result.auto,
+      buffer: current.buffer ?? result.buffer,
+      tokens: current.keep?.tokens ?? result.tokens,
+    }),
+    { auto: true, buffer: DEFAULT_BUFFER, tokens: DEFAULT_KEEP_TOKENS },
+  )
+}
+
+const select = (
+  entries: readonly Entry[],
+  tokens: number,
+): { readonly head: string; readonly recent: string } | undefined => {
+  const conversation = entries
+    .filter((entry) => entry.message.type !== "compaction")
+    .map((entry) => serialize(entry.message))
+    .filter(Boolean)
+  if (conversation.length === 0) return
+  let total = 0
+  let split = conversation.length
+  let splitPrefix = ""
+  let splitSuffix = ""
+  for (let index = conversation.length - 1; index >= 0; index--) {
+    const next = total + Token.estimate(conversation[index])
+    if (next > tokens) {
+      const remaining = Math.max(0, tokens - total) * 4
+      if (remaining > 0) {
+        splitPrefix = conversation[index].slice(0, -remaining)
+        splitSuffix = conversation[index].slice(-remaining)
+        split = index + 1
+      }
+      break
+    }
+    total = next
+    split = index
+  }
+  return {
+    head: [...conversation.slice(0, split), splitPrefix].filter(Boolean).join("\n\n"),
+    recent: [splitSuffix, ...conversation.slice(split)].filter(Boolean).join("\n\n"),
+  }
+}
+
+export const buildPrompt = (input: { readonly previousSummary?: string; readonly context: readonly string[] }) =>
+  [
+    input.previousSummary
+      ? `Update the anchored summary below using the conversation history above.\nPreserve still-true details, remove stale details, and merge in the new facts.\n<previous-summary>\n${input.previousSummary}\n</previous-summary>`
+      : "Create a new anchored summary from the conversation history.",
+    SUMMARY_TEMPLATE,
+    ...input.context,
+  ].join("\n\n")
+
+export const make = (dependencies: Dependencies) => {
+  const config = settings(dependencies.config)
+  return Effect.fn("SessionCompaction.compactIfNeeded")(function* (input: {
+    readonly sessionID: SessionSchema.ID
+    readonly entries: readonly Entry[]
+    readonly model: Model
+    readonly request: LLMRequest
+  }) {
+    const context = input.model.route.defaults.limits?.context
+    if (!config.auto || context === undefined || context <= 0) return false
+    const output = input.request.generation?.maxTokens ?? input.model.route.defaults.limits?.output ?? 0
+    if (
+      estimate({ system: input.request.system, messages: input.request.messages, tools: input.request.tools }) <=
+      context - Math.max(output, config.buffer)
+    )
+      return false
+
+    const selected = select(input.entries, config.tokens)
+    const previousSummary = input.entries.find((entry) => entry.message.type === "compaction")?.message
+    if (!selected || (selected.head.length === 0 && previousSummary?.type !== "compaction")) return false
+    const summaryPrompt = buildPrompt({
+      previousSummary: previousSummary?.type === "compaction" ? previousSummary.summary : undefined,
+      context: [previousSummary?.type === "compaction" ? previousSummary.recent : "", selected.head].filter(Boolean),
+    })
+    const summaryOutput = Math.min(output || SUMMARY_OUTPUT_TOKENS, SUMMARY_OUTPUT_TOKENS)
+    if (Token.estimate(summaryPrompt) > context - summaryOutput) return false
+    const messageID = SessionMessage.ID.create()
+    yield* dependencies.events.publish(SessionEvent.Compaction.Started, {
+      sessionID: input.sessionID,
+      messageID,
+      timestamp: yield* DateTime.now,
+      reason: "auto",
+    })
+
+    const chunks: string[] = []
+    yield* dependencies.llm
+      .stream(
+        LLM.request({
+          model: input.model,
+          messages: [Message.user(summaryPrompt)],
+          tools: [],
+          generation: { maxTokens: summaryOutput },
+        }),
+      )
+      .pipe(
+        Stream.runForEach((event) => {
+          if (!LLMEvent.is.textDelta(event)) return Effect.void
+          chunks.push(event.text)
+          return Effect.void
+        }),
+      )
+    const summary = chunks.join("")
+    if (!summary.trim()) return yield* Effect.die("Compaction returned an empty summary")
+    yield* dependencies.events.publish(SessionEvent.Compaction.Ended, {
+      sessionID: input.sessionID,
+      messageID,
+      timestamp: yield* DateTime.now,
+      reason: "auto",
+      text: summary,
+      recent: selected.recent,
+    })
+    return true
+  })
+}

packages/core/src/session/event.ts

@@ -435,15 +435,16 @@ export namespace Compaction {
 
   export const Delta = EventV2.define({
     type: "session.next.compaction.delta",
-    ...options,
     schema: {
       ...Base,
+      messageID: SessionMessageID.ID,
       text: Schema.String,
     },
   })
   export type Delta = typeof Delta.Type
 
-  export const Ended = EventV2.define({
+  // Retain the unpublished v1 decoder so stored beta events remain replayable.
+  export const EndedV1 = EventV2.define({
     type: "session.next.compaction.ended",
     ...options,
     schema: {
@@ -452,6 +453,18 @@ export namespace Compaction {
       include: Schema.String.pipe(Schema.optional),
     },
   })
+
+  export const Ended = EventV2.define({
+    type: "session.next.compaction.ended",
+    sync: { aggregate: "sessionID", version: 2 },
+    schema: {
+      ...Base,
+      messageID: SessionMessageID.ID,
+      reason: Started.data.fields.reason,
+      text: Schema.String,
+      recent: Schema.String,
+    },
+  })
   export type Ended = typeof Ended.Type
 }
 
@@ -482,10 +495,9 @@ const DurableDefinitions = [
   Reasoning.Ended,
   Retried,
   Compaction.Started,
-  Compaction.Delta,
   Compaction.Ended,
 ] as const
-const EphemeralDefinitions = [Text.Delta, Tool.Input.Delta, Reasoning.Delta] as const
+const EphemeralDefinitions = [Text.Delta, Tool.Input.Delta, Reasoning.Delta, Compaction.Delta] as const
 
 export const Durable = Schema.Union(DurableDefinitions, { mode: "oneOf" }).pipe(Schema.toTaggedUnion("type"))
 export type DurableEvent = typeof Durable.Type

packages/core/src/session/history.ts

@@ -12,7 +12,7 @@ const decode = Schema.decodeUnknownEffect(SessionMessage.Message)
 
 const latestCompaction = Effect.fnUntraced(function* (db: DatabaseService, sessionID: SessionSchema.ID) {
   return yield* db
-    .select({ seq: SessionMessageTable.seq })
+    .select()
     .from(SessionMessageTable)
     .where(and(eq(SessionMessageTable.session_id, sessionID), eq(SessionMessageTable.type, "compaction")))
     .orderBy(desc(SessionMessageTable.seq))
@@ -27,7 +27,7 @@ const messageRows = Effect.fnUntraced(function* (
   compaction: { readonly seq: number } | undefined,
   baselineSeq?: number,
 ) {
-  return yield* db
+  const rows = yield* db
     .select()
     .from(SessionMessageTable)
     .where(
@@ -49,6 +49,7 @@ const messageRows = Effect.fnUntraced(function* (
     .orderBy(asc(SessionMessageTable.seq))
     .all()
     .pipe(Effect.orDie)
+  return rows
 })
 
 const decodeMessageRow = (row: typeof SessionMessageTable.$inferSelect) =>
@@ -83,9 +84,17 @@ export const loadForRunner = Effect.fn("SessionHistory.loadForRunner")(function*
   sessionID: SessionSchema.ID,
   baselineSeq: number,
 ) {
-  return yield* Effect.forEach(
-    yield* messageRows(db, sessionID, yield* latestCompaction(db, sessionID), baselineSeq),
-    decodeMessageRow,
+  return (yield* entriesForRunner(db, sessionID, baselineSeq)).map((entry) => entry.message)
+})
+
+export const entriesForRunner = Effect.fn("SessionHistory.entriesForRunner")(function* (
+  db: DatabaseService,
+  sessionID: SessionSchema.ID,
+  baselineSeq: number,
+) {
+  const rows = yield* messageRows(db, sessionID, yield* latestCompaction(db, sessionID), baselineSeq)
+  return yield* Effect.forEach(rows, (row) =>
+    decodeMessageRow(row).pipe(Effect.map((message) => ({ seq: row.seq, message }))),
   )
 })