feat(opencode): cache-friendly compaction via primary loop request path

Default compaction no longer uses a dedicated compaction agent. It inherits the session's active agent and delegates request assembly to the primary loop's shared envelope (same system prompt, instructions, skills, MCP context, and tool definitions), replaying the typed message head and appending the summary instruction as the final user message. This keeps the compaction request prefix-identical to the preceding turn so provider prompt/KV caches are reused instead of fully re-prefilled. An explicitly configured agent.compaction preserves the legacy behavior exactly: dedicated hidden agent, optional model override, and the serialized-transcript summary request.
docs(go): use responses API for Grok 4.5 (#42373 )
2026-08-14 07:19:49 -04:00 · 2026-08-13 23:10:07 -05:00 · 2026-08-14 01:28:55 +08:00 · 2026-08-13 13:27:41 -04:00 · 2026-08-13 16:51:43 +00:00 · 2026-08-13 16:30:12 +00:00
28 changed files with 332 additions and 126 deletions
@@ -40,7 +40,7 @@ already-loaded config until then.
 | Scope                         | Path                                                                                                                      |
 | ----------------------------- | ------------------------------------------------------------------------------------------------------------------------- |
 | Project config                | `./opencode.json`, `./opencode.jsonc`, or `.opencode/opencode.json` (opencode walks up from the cwd to the worktree root) |
-| Global config                 | `~/.config/opencode/opencode.json` (NOT `~/.opencode/`)                                                                   |
+| Global config                 | `~/.config/opencode/opencode.json` or `~/.config/opencode/opencode.jsonc` (NOT `~/.opencode/`)                            |
 | Project agents                | `.opencode/agent/<name>.md` or `.opencode/agents/<name>.md`                                                               |
 | Global agents                 | `~/.config/opencode/agent(s)/<name>.md`                                                                                   |
 | Project commands              | `.opencode/command/<name>.md` or `.opencode/commands/<name>.md`                                                           |
@@ -264,7 +264,10 @@ const layer = Layer.effect(
                }),
                line: match.line_number,
                offset: match.absolute_offset,
-                text: match.lines.text.length > 2_000 ? match.lines.text.slice(0, 2_000) + "..." : match.lines.text,
+                text:
+                  match.lines.text.length > 2_000
+                    ? match.lines.text.slice(0, 2_000).replace(/[\uD800-\uDBFF]$/, "") + "..."
+                    : match.lines.text,
                submatches: match.submatches.map((submatch) => ({
                  text: submatch.match.text,
                  start: submatch.start,
@@ -173,6 +173,18 @@ export const buildPrompt = (input: { readonly previousSummary?: string; readonly
  ].join("\n\n")
 }

+export const buildReplayPrompt = (previousSummary?: string) => {
+  const instruction =
+    "Create a new anchored summary from the conversation messages above so another coding agent can continue the work."
+  if (!previousSummary) return [instruction, SUMMARY_TEMPLATE].join("\n\n")
+  return [
+    instruction,
+    `Here is the summary of the conversation before the messages above:\n\n<prior-summary>\n${previousSummary}\n</prior-summary>`,
+    SUMMARY_UPDATE_INSTRUCTIONS,
+    SUMMARY_TEMPLATE,
+  ].join("\n\n")
+}
+
 export const make = (dependencies: Dependencies) => {
  const config = settings(dependencies.config)
  const compactAfterOverflow = Effect.fn("SessionCompaction.compactAfterOverflow")(function* (input: Input) {
@@ -62,4 +62,24 @@ describe("Ripgrep", () => {
      (tmp) => Effect.promise(() => tmp[Symbol.asyncDispose]()),
    ),
  )
+  it.live("does not split surrogate pairs in oversized line previews", () =>
+    Effect.acquireUseRelease(
+      Effect.promise(() => tmpdir()),
+      (tmp) =>
+        Effect.gen(function* () {
+          yield* Effect.promise(() =>
+            fs.writeFile(path.join(tmp.path, "unicode.txt"), `needle${"x".repeat(1_993)}😀\n`),
+          )
+
+          const matches = yield* (yield* Ripgrep.Service).grep({
+            cwd: tmp.path,
+            pattern: "needle",
+            limit: 10,
+          })
+
+          expect(matches[0]?.text).toBe(`needle${"x".repeat(1_993)}...`)
+        }),
+      (tmp) => Effect.promise(() => tmp[Symbol.asyncDispose]()),
+    ),
+  )
 })
@@ -216,21 +216,6 @@ const layer = Layer.effect(
            mode: "subagent",
            native: true,
          },
-          compaction: {
-            name: "compaction",
-            mode: "primary",
-            native: true,
-            hidden: true,
-            prompt: PROMPT_COMPACTION,
-            permission: Permission.merge(
-              defaults,
-              Permission.fromConfig({
-                "*": "deny",
-              }),
-              user,
-            ),
-            options: {},
-          },
          title: {
            name: "title",
            mode: "primary",
@@ -270,6 +255,26 @@ const layer = Layer.effect(
            continue
          }
          let item = agents[key]
+          // The compaction agent is not registered by default: compaction inherits
+          // the session's active agent so its request stays prefix-identical for
+          // prompt caching. An explicit `agent.compaction` config opts into the
+          // legacy dedicated agent, seeded here so overrides merge as before.
+          if (!item && key === "compaction")
+            item = agents[key] = {
+              name: "compaction",
+              mode: "primary",
+              native: true,
+              hidden: true,
+              prompt: PROMPT_COMPACTION,
+              permission: Permission.merge(
+                defaults,
+                Permission.fromConfig({
+                  "*": "deny",
+                }),
+                user,
+              ),
+              options: {},
+            }
          if (!item)
            item = agents[key] = {
              name: key,
@@ -20,7 +20,7 @@ import { RuntimeFlags } from "@/effect/runtime-flags"
 import { EventV2Bridge } from "@/event-v2-bridge"
 import { ProviderV2 } from "@opencode-ai/core/provider"
 import { ModelV2 } from "@opencode-ai/core/model"
-import { buildPrompt } from "@opencode-ai/core/session/compaction"
+import { buildPrompt, buildReplayPrompt } from "@opencode-ai/core/session/compaction"
 import { SessionCompactionEvent } from "@opencode-ai/schema/session-compaction-event"

 export const Event = SessionCompactionEvent
@@ -162,6 +162,15 @@ function splitTurn(input: {
  })
 }

+export type Run = (input: {
+  user: SessionV1.User
+  agent: Agent.Info
+  model: Provider.Model
+  processor: SessionProcessor.Handle
+  messages: SessionV1.WithParts[]
+  prompt: string
+}) => Effect.Effect<SessionProcessor.Result>
+
 export interface Interface {
  readonly isOverflow: (input: {
    tokens: SessionV1.Assistant["tokens"]
@@ -174,6 +183,7 @@ export interface Interface {
    sessionID: SessionID
    auto: boolean
    overflow?: boolean
+    run?: Run
  }) => Effect.Effect<"continue" | "stop">
  readonly create: (input: {
    sessionID: SessionID
@@ -322,6 +332,7 @@ const layer = Layer.effect(
      sessionID: SessionID
      auto: boolean
      overflow?: boolean
+      run?: Run
    }) {
      const parent = input.messages.findLast((m) => m.info.id === input.parentID)
      if (!parent || parent.info.role !== "user") {
@@ -355,11 +366,13 @@ const layer = Layer.effect(
        }
      }

-      const agent = yield* agents.get("compaction")
-      const model = agent.model
-        ? yield* provider.getModel(agent.model.providerID, agent.model.modelID).pipe(Effect.orDie)
-        : yield* provider.getModel(userMessage.model.providerID, userMessage.model.modelID).pipe(Effect.orDie)
      const cfg = yield* config.get()
+      const configured = cfg.agent?.compaction !== undefined
+      const agent = yield* agents.get(configured ? "compaction" : userMessage.agent)
+      const model =
+        configured && agent.model
+          ? yield* provider.getModel(agent.model.providerID, agent.model.modelID).pipe(Effect.orDie)
+          : yield* provider.getModel(userMessage.model.providerID, userMessage.model.modelID).pipe(Effect.orDie)
      const history = compactionPart && messages.at(-1)?.info.id === input.parentID ? messages.slice(0, -1) : messages
      const prior = completedCompactions(history)
      const hidden = new Set(prior.flatMap((item) => [item.userIndex, item.assistantIndex]))
@@ -377,18 +390,16 @@ const layer = Layer.effect(
      )
      const msgs = structuredClone(selected.head)
      yield* plugin.trigger("experimental.chat.messages.transform", {}, { messages: msgs })
-      const conversation = msgs.map(serialize).filter(Boolean).join("\n\n")
-      const nextPrompt =
+      // A configured compaction agent keeps the legacy serialized-transcript path.
+      const run = configured ? undefined : input.run
+      const conversation = run ? "" : msgs.map(serialize).filter(Boolean).join("\n\n")
+      const nextPrompt = [
        compacting.prompt ??
-        [
-          buildPrompt({
-            previousSummary,
-            context: [conversation],
-          }),
-          ...compacting.context,
-        ]
-          .filter(Boolean)
-          .join("\n\n")
+          (run ? buildReplayPrompt(previousSummary) : buildPrompt({ previousSummary, context: [conversation] })),
+        ...compacting.context,
+      ]
+        .filter(Boolean)
+        .join("\n\n")
      const ctx = yield* InstanceState.context
      const msg: SessionV1.Assistant = {
        id: MessageID.ascending(),
@@ -422,30 +433,43 @@ const layer = Layer.effect(
        sessionID: input.sessionID,
        model,
      })
-      const result = yield* processor.process({
-        user: userMessage,
-        agent,
-        sessionID: input.sessionID,
-        tools: {},
-        system: [],
-        messages: [
-          {
-            role: "user",
-            content: [
+      // The synthetic compaction marker carries no per-turn system/tools settings.
+      const active = msgs.findLast(
+        (message) => message.info.role === "user" && !message.parts.some((part) => part.type === "compaction"),
+      )
+      const result = run
+        ? yield* run({
+            user: active?.info.role === "user" ? active.info : userMessage,
+            agent,
+            model,
+            processor,
+            messages: msgs,
+            prompt: nextPrompt,
+          })
+        : yield* processor.process({
+            user: userMessage,
+            agent,
+            sessionID: input.sessionID,
+            tools: {},
+            system: [],
+            messages: [
              {
-                type: "text",
-                text: [
-                  nextPrompt,
-                  ...(compacting.prompt ? ["The following is the conversation history:", conversation] : []),
-                ]
-                  .filter(Boolean)
-                  .join("\n\n"),
+                role: "user",
+                content: [
+                  {
+                    type: "text",
+                    text: [
+                      nextPrompt,
+                      ...(compacting.prompt ? ["The following is the conversation history:", conversation] : []),
+                    ]
+                      .filter(Boolean)
+                      .join("\n\n"),
+                  },
+                ],
              },
            ],
-          },
-        ],
-        model,
-      })
+            model,
+          })

      if (result === "compact") {
        processor.message.error = new SessionV1.ContextOverflowError({
@@ -149,6 +149,45 @@ const layer = Layer.effect(
      } satisfies TaskPromptOps
    })

+    // One assembly path keeps compaction requests prefix-identical to normal turns for prompt caching.
+    const prepare = Effect.fn("SessionPrompt.prepare")(function* (input: {
+      session: Session.Info
+      agent: Agent.Info
+      model: Provider.Model
+      processor: SessionProcessor.Handle
+      messages: SessionV1.WithParts[]
+    }) {
+      const lastUserMsg = input.messages.findLast((message) => message.info.role === "user")
+      const tools = yield* SessionTools.resolve({
+        agent: input.agent,
+        session: input.session,
+        model: input.model,
+        processor: input.processor,
+        bypassAgentCheck: lastUserMsg?.parts.some((part) => part.type === "agent") ?? false,
+        messages: input.messages,
+        promptOps: yield* ops(),
+      }).pipe(
+        Effect.provideService(Plugin.Service, plugin),
+        Effect.provideService(Permission.Service, permission),
+        Effect.provideService(ToolRegistry.Service, registry),
+        Effect.provideService(MCP.Service, mcp),
+        Effect.provideService(Truncate.Service, truncate),
+        Effect.provideService(RuntimeFlags.Service, flags),
+      )
+      const [skills, env, instructions, mcpInstructions, modelMessages] = yield* Effect.all([
+        sys.skills(input.agent),
+        sys.environment(input.model),
+        instruction.system().pipe(Effect.orDie),
+        sys.mcp(input.agent, input.session.permission),
+        MessageV2.toModelMessagesEffect(input.messages, input.model),
+      ])
+      return {
+        tools,
+        modelMessages,
+        system: [...env, ...instructions, ...(mcpInstructions ? [mcpInstructions] : []), ...(skills ? [skills] : [])],
+      }
+    })
+
    const cancel = Effect.fn("SessionPrompt.cancel")(function* (sessionID: SessionID) {
      yield* Effect.logInfo("cancel", { "session.id": sessionID })
      yield* state.cancel(sessionID)
@@ -1153,6 +1192,21 @@ const layer = Layer.effect(
              sessionID,
              auto: task.auto,
              overflow: task.overflow,
+              run: ({ user, agent, model, processor, messages, prompt }) =>
+                Effect.gen(function* () {
+                  const prepared = yield* prepare({ session, agent, model, processor, messages })
+                  return yield* processor.process({
+                    user,
+                    agent,
+                    permission: session.permission,
+                    sessionID,
+                    parentSessionID: session.parentID,
+                    system: prepared.system,
+                    messages: [...prepared.modelMessages, { role: "user", content: prompt }],
+                    tools: prepared.tools,
+                    model,
+                  })
+                }),
            })
            if (result === "stop") break
            continue
@@ -1219,68 +1273,34 @@ const layer = Layer.effect(
            .pipe(Effect.onInterrupt(() => finalizeInterruptedAssistant))

          const outcome: "break" | "continue" = yield* Effect.gen(function* () {
-            const lastUserMsg = msgs.findLast((m) => m.info.role === "user")
-            const bypassAgentCheck = lastUserMsg?.parts.some((p) => p.type === "agent") ?? false
-            const promptOps = yield* ops()
-
-            const tools = yield* SessionTools.resolve({
-              agent,
-              session,
-              model,
-              processor: handle,
-              bypassAgentCheck,
-              messages: msgs,
-              promptOps,
-            }).pipe(
-              Effect.provideService(Plugin.Service, plugin),
-              Effect.provideService(Permission.Service, permission),
-              Effect.provideService(ToolRegistry.Service, registry),
-              Effect.provideService(MCP.Service, mcp),
-              Effect.provideService(Truncate.Service, truncate),
-              Effect.provideService(RuntimeFlags.Service, flags),
-            )
-
-            if (lastUser.format?.type === "json_schema") {
-              tools["StructuredOutput"] = createStructuredOutputTool({
-                schema: lastUser.format.schema,
-                onSuccess(output) {
-                  structured = output
-                },
-              })
-            }
-
            if (step === 1)
              yield* summary.summarize({ sessionID, messageID: lastUser.id }).pipe(Effect.ignore, Effect.forkIn(scope))

            yield* plugin.trigger("experimental.chat.messages.transform", {}, { messages: msgs })

-            const [skills, env, instructions, mcpInstructions, modelMsgs] = yield* Effect.all([
-              sys.skills(agent),
-              sys.environment(model),
-              instruction.system().pipe(Effect.orDie),
-              sys.mcp(agent, session.permission),
-              MessageV2.toModelMessagesEffect(msgs, model),
-            ])
-            const system = [
-              ...env,
-              ...instructions,
-              ...(mcpInstructions ? [mcpInstructions] : []),
-              ...(skills ? [skills] : []),
-            ]
+            const prepared = yield* prepare({ session, agent, model, processor: handle, messages: msgs })
            const format = lastUser.format ?? { type: "text" as const }
-            if (format.type === "json_schema") system.push(STRUCTURED_OUTPUT_SYSTEM_PROMPT)
+            if (format.type === "json_schema") {
+              prepared.system.push(STRUCTURED_OUTPUT_SYSTEM_PROMPT)
+              prepared.tools["StructuredOutput"] = createStructuredOutputTool({
+                schema: format.schema,
+                onSuccess(output) {
+                  structured = output
+                },
+              })
+            }
            const result = yield* handle.process({
              user: lastUser,
              agent,
              permission: session.permission,
              sessionID,
              parentSessionID: session.parentID,
-              system,
+              system: prepared.system,
              messages: [
-                ...modelMsgs,
+                ...prepared.modelMessages,
                ...(isLastStep ? [{ role: "assistant" as const, content: MAX_STEPS_PROMPT }] : []),
              ],
-              tools,
+              tools: prepared.tools,
              model,
              toolChoice: format.type === "json_schema" ? "required" : undefined,
            })
@@ -52,7 +52,7 @@ it.instance("returns default native agents when no config", () =>
    expect(names).toContain("plan")
    expect(names).toContain("general")
    expect(names).toContain("explore")
-    expect(names).toContain("compaction")
+    expect(names).not.toContain("compaction")
    expect(names).toContain("title")
    expect(names).toContain("summary")
  }),
@@ -170,17 +170,34 @@ it.instance("general agent denies todo tools", () =>
  }),
 )

-it.instance("compaction agent denies all permissions", () =>
+it.instance("compaction agent is absent without configuration", () =>
  Effect.gen(function* () {
    const compaction = yield* load((svc) => svc.get("compaction"))
-    expect(compaction).toBeDefined()
-    expect(compaction?.hidden).toBe(true)
-    expect(evalPerm(compaction, "bash")).toBe("deny")
-    expect(evalPerm(compaction, "edit")).toBe("deny")
-    expect(evalPerm(compaction, "read")).toBe("deny")
+    expect(compaction).toBeUndefined()
  }),
 )

+it.instance(
+  "configured compaction agent keeps native defaults",
+  () =>
+    Effect.gen(function* () {
+      const compaction = yield* load((svc) => svc.get("compaction"))
+      expect(compaction).toBeDefined()
+      expect(compaction?.native).toBe(true)
+      expect(compaction?.hidden).toBe(true)
+      expect(evalPerm(compaction, "bash")).toBe("deny")
+      expect(evalPerm(compaction, "edit")).toBe("deny")
+      expect(evalPerm(compaction, "read")).toBe("deny")
+    }),
+  {
+    config: {
+      agent: {
+        compaction: {},
+      },
+    },
+  },
+)
+
 it.instance(
  "custom agent from config creates new agent",
  () =>
@@ -710,6 +727,9 @@ it.instance(
  {
    config: {
      default_agent: "compaction",
+      agent: {
+        compaction: {},
+      },
    },
  },
 )
@@ -222,6 +222,11 @@ function cfg(compaction?: ConfigV1.Info["compaction"]) {
  return Layer.succeed(Config.Service, TestConfig.make({ get: () => Effect.succeed({ ...base, compaction }) }))
 }

+function cfgAgent(agent: NonNullable<ConfigV1.Info["agent"]>) {
+  const config = Schema.decodeUnknownSync(ConfigV1.Info)({ agent }) as ConfigV1.Info
+  return Layer.succeed(Config.Service, TestConfig.make({ get: () => Effect.succeed(config) }))
+}
+
 const defaultProvider = wide()
 const compactionTestNode = LayerNode.group([
  SessionCompaction.node,
@@ -812,6 +817,75 @@ describe("session.compaction.prune", () => {
 })

 describe("session.compaction.process", () => {
+  itCompaction.instance(
+    "inherits the last agent when compaction is not configured",
+    () => {
+      const stub = llm()
+      let captured: LLM.StreamInput | undefined
+      stub.push(reply("summary", (input) => (captured = input)))
+      return Effect.gen(function* () {
+        const ssn = yield* SessionNs.Service
+        const session = yield* ssn.create({})
+        const msg = yield* createUserMessage(session.id, "hello")
+        const msgs = yield* ssn.messages({ sessionID: session.id })
+
+        yield* SessionCompaction.use.process({
+          parentID: msg.id,
+          messages: msgs,
+          sessionID: session.id,
+          auto: false,
+        })
+
+        expect(captured?.agent.name).toBe("build")
+        expect(captured?.agent.prompt).toBe("custom build prompt")
+        expect(captured?.agent.temperature).toBe(0.37)
+        expect(String(captured?.model.providerID)).toBe("test")
+        expect(String(captured?.model.id)).toBe("test-model")
+      }).pipe(
+        withCompaction({
+          llm: stub.llmLayer,
+          config: cfgAgent({
+            build: { prompt: "custom build prompt", temperature: 0.37, model: "missing/missing" },
+          }),
+        }),
+      )
+    },
+    { git: true },
+  )
+
+  itCompaction.instance(
+    "preserves an explicitly configured compaction agent",
+    () => {
+      const stub = llm()
+      let captured: LLM.StreamInput | undefined
+      stub.push(reply("summary", (input) => (captured = input)))
+      return Effect.gen(function* () {
+        const ssn = yield* SessionNs.Service
+        const session = yield* ssn.create({})
+        const msg = yield* createUserMessage(session.id, "hello")
+        const msgs = yield* ssn.messages({ sessionID: session.id })
+
+        yield* SessionCompaction.use.process({
+          parentID: msg.id,
+          messages: msgs,
+          sessionID: session.id,
+          auto: false,
+        })
+
+        expect(captured?.agent.name).toBe("compaction")
+        expect(captured?.agent.prompt).toBe("custom compaction prompt")
+        expect(captured?.agent.temperature).toBe(0.81)
+        expect(captured?.agent.hidden).toBe(true)
+      }).pipe(
+        withCompaction({
+          llm: stub.llmLayer,
+          config: cfgAgent({ compaction: { prompt: "custom compaction prompt", temperature: 0.81 } }),
+        }),
+      )
+    },
+    { git: true },
+  )
+
  it.instance(
    "throws when parent is not a user message",
    Effect.gen(function* () {
@@ -444,6 +444,34 @@ const boot = Effect.fn("test.boot")(function* (input?: { title?: string }) {

 // Loop semantics

+it.instance("default compaction reuses the normal system and tools", () =>
+  Effect.gen(function* () {
+    const { llm } = yield* useServerConfig(providerCfg)
+    const { prompt, chat } = yield* boot()
+    const compact = yield* SessionCompaction.Service
+    yield* user(chat.id, "cache prefix message")
+    yield* llm.text("normal response")
+    yield* prompt.loop({ sessionID: chat.id })
+
+    yield* compact.create({ sessionID: chat.id, agent: "build", model: ref, auto: false })
+    yield* llm.text("compacted summary")
+    yield* prompt.loop({ sessionID: chat.id })
+
+    const inputs = yield* llm.inputs
+    expect(inputs).toHaveLength(2)
+    expect(inputs[1]?.tools).toEqual(inputs[0]?.tools)
+    const normalMessages = Array.isArray(inputs[0]?.messages) ? inputs[0].messages : []
+    const compactMessages = Array.isArray(inputs[1]?.messages) ? inputs[1].messages : []
+    const system = (message: unknown) =>
+      typeof message === "object" && message !== null && "role" in message && message.role === "system"
+    expect(compactMessages.find(system)).toEqual(normalMessages.find(system))
+    expect(JSON.stringify(compactMessages)).toContain("cache prefix message")
+    expect(JSON.stringify(compactMessages)).toContain("normal response")
+    expect(JSON.stringify(compactMessages)).toContain("Create a new anchored summary")
+    expect(JSON.stringify(compactMessages)).not.toContain("[User]: cache prefix message")
+  }),
+)
+
 noLLMServer.instance(
  "loop exits immediately when last assistant has stop finish",
  () =>
@@ -185,7 +185,7 @@ OpenCode Go هو اشتراك منخفض التكلفة — **$5 للشهر ال

 | Model             | Model ID          | Endpoint                                         | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ Također možete pristupiti Go modelima putem sljedećih API endpointa.

 | Model             | Model ID          | Endpoint                                         | AI SDK Paket                |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ Du kan også få adgang til Go-modeller gennem følgende API-endpoints.

 | Model             | Model ID          | Endpoint                                         | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -187,7 +187,7 @@ Du kannst auf die Go-Modelle auch über die folgenden API-Endpunkte zugreifen.

 | Modell            | Modell-ID         | Endpunkt                                         | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ También puedes acceder a los modelos de Go a través de los siguientes endpoint

 | Modelo            | ID del modelo     | Endpoint                                         | Paquete de AI SDK           |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ Vous pouvez également accéder aux modèles Go via les points de terminaison d'

 | Modèle            | ID de modèle      | Point de terminaison                             | Package AI SDK              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ You can also access Go models through the following API endpoints.

 | Model             | Model ID          | Endpoint                                         | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -195,7 +195,7 @@ Puoi anche accedere ai modelli Go tramite i seguenti endpoint API.

 | Modello           | ID Modello        | Endpoint                                         | Pacchetto AI SDK            |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ Goでは月額$10を支払い、その6倍の利用枠を提供することを

 | Model             | Model ID          | Endpoint                                         | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ Go에서는 월 $10를 지불하며, 저희는 그 6배의 사용량을 제공

 | 모델              | 모델 ID           | 엔드포인트                                       | AI SDK 패키지               |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ Du kan også få tilgang til Go-modeller gjennom følgende API-endepunkter.

 | Modell            | Modell-ID         | Endepunkt                                        | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -189,7 +189,7 @@ Możesz również uzyskać dostęp do modeli Go za pośrednictwem następującyc

 | Model             | ID modelu         | Punkt końcowy                                    | Pakiet AI SDK               |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ Você também pode acessar os modelos do Go através dos seguintes endpoints de

 | Modelo            | ID do Modelo      | Endpoint                                         | Pacote do AI SDK            |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -197,7 +197,7 @@ OpenCode Go включает следующие лимиты:

 | Модель            | ID модели         | Эндпоинт                                         | Пакет AI SDK                |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ OpenCode Go มีขีดจำกัดดังต่อไปนี้:

 | Model             | Model ID          | Endpoint                                         | AI SDK Package              |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ Go modellerine aşağıdaki API uç noktaları aracılığıyla da erişebilirsi

 | Model             | Model ID          | Uç Nokta                                         | AI SDK Paketi               |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ OpenCode Go 包含以下限制：

 | 模型              | 模型 ID           | 端点                                             | AI SDK 包                   |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
@@ -185,7 +185,7 @@ OpenCode Go 包含以下限制：

 | 模型              | 模型 ID           | 端點                                             | AI SDK 套件                 |
 | ----------------- | ----------------- | ------------------------------------------------ | --------------------------- |
-| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
+| Grok 4.5          | grok-4.5          | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GPT 5.6 Luna      | gpt-5.6-luna      | `https://opencode.ai/zen/go/v1/responses`        | `@ai-sdk/openai`            |
 | GLM-5.2           | glm-5.2           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
 | GLM-5.1           | glm-5.1           | `https://opencode.ai/zen/go/v1/chat/completions` | `@ai-sdk/openai-compatible` |
Author	SHA1	Message	Date
Aiden Cline	9e7b5f1701	feat(opencode): cache-friendly compaction via primary loop request path Default compaction no longer uses a dedicated compaction agent. It inherits the session's active agent and delegates request assembly to the primary loop's shared envelope (same system prompt, instructions, skills, MCP context, and tool definitions), replaying the typed message head and appending the summary instruction as the final user message. This keeps the compaction request prefix-identical to the preceding turn so provider prompt/KV caches are reused instead of fully re-prefilled. An explicitly configured agent.compaction preserves the legacy behavior exactly: dedicated hidden agent, optional model override, and the serialized-transcript summary request.	2026-08-13 23:10:07 -05:00
Jack	d0c2b41adf	docs(go): use responses API for Grok 4.5 (#42373 )	2026-08-14 01:28:55 +08:00
Frank	c7af47f9ed	update grok endpoint	2026-08-13 13:27:41 -04:00
Kit Langton	6c035e1fd7	fix(core): preserve unicode in grep previews (#42356 )	2026-08-13 16:51:43 +00:00
opencode-agent[bot]	ab7cbc808f	chore: generate	2026-08-13 16:30:12 +00:00
Aditya Sethi	62387f39d4	fix(skills): Update global config path in documentation (#42337 )	2026-08-13 18:27:29 +02:00