fix(core): ensure compaction is more reliable, add reserve token buffer to ensure that input window has enough room to compact (#12924)

Co-authored-by: James Lal <james@littlebearlabs.io>

packages/opencode/src/session/compaction.ts

@@ -6,7 +6,6 @@ import { Instance } from "../project/instance"
 import { Provider } from "../provider/provider"
 import { MessageV2 } from "./message-v2"
 import z from "zod"
-import { SessionPrompt } from "./prompt"
 import { Token } from "../util/token"
 import { Log } from "../util/log"
 import { SessionProcessor } from "./processor"
@@ -14,6 +13,7 @@ import { fn } from "@/util/fn"
 import { Agent } from "@/agent/agent"
 import { Plugin } from "@/plugin"
 import { Config } from "@/config/config"
+import { ProviderTransform } from "@/provider/transform"
 
 export namespace SessionCompaction {
   const log = Log.create({ service: "session.compaction" })
@@ -27,15 +27,22 @@ export namespace SessionCompaction {
     ),
   }
 
+  const COMPACTION_BUFFER = 20_000
+
   export async function isOverflow(input: { tokens: MessageV2.Assistant["tokens"]; model: Provider.Model }) {
     const config = await Config.get()
     if (config.compaction?.auto === false) return false
     const context = input.model.limit.context
     if (context === 0) return false
-    const count = input.tokens.input + input.tokens.cache.read + input.tokens.output
-    const output = Math.min(input.model.limit.output, SessionPrompt.OUTPUT_TOKEN_MAX) || SessionPrompt.OUTPUT_TOKEN_MAX
-    const usable = input.model.limit.input || context - output
-    return count > usable
+
+    const count =
+      input.tokens.total ||
+      input.tokens.input + input.tokens.output + input.tokens.cache.read + input.tokens.cache.write
+
+    const reserved =
+      config.compaction?.reserved ?? Math.min(COMPACTION_BUFFER, ProviderTransform.maxOutputTokens(input.model))
+    const usable = input.model.limit.input ? input.model.limit.input - reserved : context - reserved
+    return count >= usable
   }
 
   export const PRUNE_MINIMUM = 20_000
@@ -139,8 +146,34 @@ export namespace SessionCompaction {
       { 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 considering new session will not have access to our conversation."
+    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.
+The summary that you construct will be used so that another agent can read it and continue the work.
+
+When constructing the summary, try to stick to this template:
+---
+## Goal
+
+[What goal(s) is the user trying to accomplish?]
+
+## Instructions
+
+- [What important instructions did the user give you that are relevant]
+- [If there is a plan or spec, include information about it so next agent can continue using it]
+
+## Discoveries
+
+[What notable things were learned during this conversation that would be useful for the next agent to know when continuing the work]
+
+## Accomplished
+
+[What work has been completed, what work is still in progress, and what work is left?]
+
+## Relevant files / directories
+
+[Construct a structured list of relevant files that have been read, edited, or created that pertain to the task at hand. If all the files in a directory are relevant, include the path to the directory.]
+---`
+
     const promptText = compacting.prompt ?? [defaultPrompt, ...compacting.context].join("\n\n")
     const result = await processor.process({
       user: userMessage,
@@ -181,7 +214,7 @@ export namespace SessionCompaction {
         sessionID: input.sessionID,
         type: "text",
         synthetic: true,
-        text: "Continue if you have next steps",
+        text: "Continue if you have next steps, or stop and ask for clarification if you are unsure how to proceed.",
         time: {
           start: Date.now(),
           end: Date.now(),

packages/opencode/src/session/index.ts

@@ -4,7 +4,7 @@ import { BusEvent } from "@/bus/bus-event"
 import { Bus } from "@/bus"
 import { Decimal } from "decimal.js"
 import z from "zod"
-import { type LanguageModelUsage, type ProviderMetadata } from "ai"
+import { type ProviderMetadata } from "ai"
 import { Config } from "../config/config"
 import { Flag } from "../flag/flag"
 import { Identifier } from "../id/id"
@@ -22,6 +22,8 @@ import { Snapshot } from "@/snapshot"
 import type { Provider } from "@/provider/provider"
 import { PermissionNext } from "@/permission/next"
 import { Global } from "@/global"
+import type { LanguageModelV2Usage } from "@ai-sdk/provider"
+import { iife } from "@/util/iife"
 
 export namespace Session {
   const log = Log.create({ service: "session" })
@@ -439,34 +441,58 @@ export namespace Session {
   export const getUsage = fn(
     z.object({
       model: z.custom<Provider.Model>(),
-      usage: z.custom<LanguageModelUsage>(),
+      usage: z.custom<LanguageModelV2Usage>(),
       metadata: z.custom<ProviderMetadata>().optional(),
     }),
     (input) => {
-      const cacheReadInputTokens = input.usage.cachedInputTokens ?? 0
-      const cacheWriteInputTokens = (input.metadata?.["anthropic"]?.["cacheCreationInputTokens"] ??
-        // @ts-expect-error
-        input.metadata?.["bedrock"]?.["usage"]?.["cacheWriteInputTokens"] ??
-        // @ts-expect-error
-        input.metadata?.["venice"]?.["usage"]?.["cacheCreationInputTokens"] ??
-        0) as number
-
-      const excludesCachedTokens = !!(input.metadata?.["anthropic"] || input.metadata?.["bedrock"])
-      const adjustedInputTokens = excludesCachedTokens
-        ? (input.usage.inputTokens ?? 0)
-        : (input.usage.inputTokens ?? 0) - cacheReadInputTokens - cacheWriteInputTokens
       const safe = (value: number) => {
         if (!Number.isFinite(value)) return 0
         return value
       }
+      const inputTokens = safe(input.usage.inputTokens ?? 0)
+      const outputTokens = safe(input.usage.outputTokens ?? 0)
+      const reasoningTokens = safe(input.usage.reasoningTokens ?? 0)
+
+      const cacheReadInputTokens = safe(input.usage.cachedInputTokens ?? 0)
+      const cacheWriteInputTokens = safe(
+        (input.metadata?.["anthropic"]?.["cacheCreationInputTokens"] ??
+          // @ts-expect-error
+          input.metadata?.["bedrock"]?.["usage"]?.["cacheWriteInputTokens"] ??
+          // @ts-expect-error
+          input.metadata?.["venice"]?.["usage"]?.["cacheCreationInputTokens"] ??
+          0) as number,
+      )
+
+      // OpenRouter provides inputTokens as the total count of input tokens (including cached).
+      // AFAIK other providers (OpenRouter/OpenAI/Gemini etc.) do it the same way e.g. vercel/ai#8794 (comment)
+      // Anthropic does it differently though - inputTokens doesn't include cached tokens.
+      // It looks like OpenCode's cost calculation assumes all providers return inputTokens the same way Anthropic does (I'm guessing getUsage logic was originally implemented with anthropic), so it's causing incorrect cost calculation for OpenRouter and others.
+      const excludesCachedTokens = !!(input.metadata?.["anthropic"] || input.metadata?.["bedrock"])
+      const adjustedInputTokens = safe(
+        excludesCachedTokens ? inputTokens : inputTokens - cacheReadInputTokens - cacheWriteInputTokens,
+      )
+
+      const total = iife(() => {
+        // Anthropic doesn't provide total_tokens, also ai sdk will vastly undercount if we
+        // don't compute from components
+        if (
+          input.model.api.npm === "@ai-sdk/anthropic" ||
+          input.model.api.npm === "@ai-sdk/amazon-bedrock" ||
+          input.model.api.npm === "@ai-sdk/google-vertex/anthropic"
+        ) {
+          return adjustedInputTokens + outputTokens + cacheReadInputTokens + cacheWriteInputTokens
+        }
+        return input.usage.totalTokens
+      })
 
       const tokens = {
-        input: safe(adjustedInputTokens),
-        output: safe(input.usage.outputTokens ?? 0),
-        reasoning: safe(input.usage?.reasoningTokens ?? 0),
+        total,
+        input: adjustedInputTokens,
+        output: outputTokens,
+        reasoning: reasoningTokens,
         cache: {
-          write: safe(cacheWriteInputTokens),
-          read: safe(cacheReadInputTokens),
+          write: cacheWriteInputTokens,
+          read: cacheReadInputTokens,
         },
       }
 

packages/opencode/src/session/llm.ts

@@ -25,8 +25,7 @@ import { Auth } from "@/auth"
 
 export namespace LLM {
   const log = Log.create({ service: "llm" })
-
-  export const OUTPUT_TOKEN_MAX = Flag.OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX || 32_000
+  export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
 
   export type StreamInput = {
     user: MessageV2.User
@@ -149,14 +148,7 @@ export namespace LLM {
     )
 
     const maxOutputTokens =
-      isCodex || provider.id.includes("github-copilot")
-        ? undefined
-        : ProviderTransform.maxOutputTokens(
-            input.model.api.npm,
-            params.options,
-            input.model.limit.output,
-            OUTPUT_TOKEN_MAX,
-          )
+      isCodex || provider.id.includes("github-copilot") ? undefined : ProviderTransform.maxOutputTokens(input.model)
 
     const tools = await resolveTools(input)
 

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

@@ -210,6 +210,7 @@ export namespace MessageV2 {
     snapshot: z.string().optional(),
     cost: z.number(),
     tokens: z.object({
+      total: z.number().optional(),
       input: z.number(),
       output: z.number(),
       reasoning: z.number(),
@@ -383,6 +384,7 @@ export namespace MessageV2 {
     summary: z.boolean().optional(),
     cost: z.number(),
     tokens: z.object({
+      total: z.number().optional(),
       input: z.number(),
       output: z.number(),
       reasoning: z.number(),

packages/opencode/src/session/processor.ts

@@ -342,6 +342,9 @@ export namespace SessionProcessor {
               stack: JSON.stringify(e.stack),
             })
             const error = MessageV2.fromError(e, { providerID: input.model.providerID })
+            if (MessageV2.ContextOverflowError.isInstance(error)) {
+              // TODO: Handle context overflow error
+            }
             const retry = SessionRetry.retryable(error)
             if (retry !== undefined) {
               attempt++

packages/opencode/src/session/prompt.ts

@@ -52,7 +52,6 @@ globalThis.AI_SDK_LOG_WARNINGS = false
 
 export namespace SessionPrompt {
   const log = Log.create({ service: "session.prompt" })
-  export const OUTPUT_TOKEN_MAX = Flag.OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX || 32_000
 
   const state = Instance.state(
     () => {

packages/opencode/src/session/retry.ts

@@ -59,9 +59,8 @@ export namespace SessionRetry {
   }
 
   export function retryable(error: ReturnType<NamedError["toObject"]>) {
-    // DO NOT retry context overflow errors
+    // context overflow errors should not be retried
     if (MessageV2.ContextOverflowError.isInstance(error)) return undefined
-
     if (MessageV2.APIError.isInstance(error)) {
       if (!error.data.isRetryable) return undefined
       if (error.data.responseBody?.includes("FreeUsageLimitError"))