diff --git a/packages/opencode/src/command/index.ts b/packages/opencode/src/command/index.ts
index 057754cd9ef8..42ed0b945003 100644
--- a/packages/opencode/src/command/index.ts
+++ b/packages/opencode/src/command/index.ts
@@ -8,6 +8,7 @@ import { Config } from "@/config/config"
 import { MCP } from "../mcp"
 import { Skill } from "../skill"
 import PROMPT_INITIALIZE from "./template/initialize.txt"
+import PROMPT_LOOP from "./template/loop.txt"
 import PROMPT_REVIEW from "./template/review.txt"
 import { LegacyEvent } from "@opencode-ai/schema/legacy-event"
 
@@ -45,6 +46,8 @@ export function hints(template: string) {
 
 export const Default = {
   INIT: "init",
+  LOOP: "loop",
+  PROACTIVE: "proactive",
   REVIEW: "review",
 } as const
 
@@ -86,6 +89,24 @@ const layer = Layer.effect(
         subtask: true,
         hints: hints(PROMPT_REVIEW),
       }
+      commands[Default.LOOP] = {
+        name: Default.LOOP,
+        description: "run a prompt repeatedly until it finishes or you stop it",
+        source: "command",
+        get template() {
+          return PROMPT_LOOP
+        },
+        hints: ["$ARGUMENTS"],
+      }
+      commands[Default.PROACTIVE] = {
+        name: Default.PROACTIVE,
+        description: "alias for /loop",
+        source: "command",
+        get template() {
+          return PROMPT_LOOP
+        },
+        hints: ["$ARGUMENTS"],
+      }
 
       for (const [name, command] of Object.entries(cfg.command ?? {})) {
         commands[name] = {
diff --git a/packages/opencode/src/command/template/loop.txt b/packages/opencode/src/command/template/loop.txt
new file mode 100644
index 000000000000..69d8ea053b6a
--- /dev/null
+++ b/packages/opencode/src/command/template/loop.txt
@@ -0,0 +1,8 @@
+Run a prompt repeatedly while the session stays open.
+
+Usage:
+- `/loop <prompt>` starts a self-paced loop.
+- `/loop 5m <prompt>` starts a scheduled loop.
+- `/loop stop` stops the active loop for the session.
+
+If no prompt is provided, use `.claude/loop.md` from the project when present.
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
index eb116f6b960f..0a4ac6d551bb 100644
--- a/packages/opencode/src/session/prompt.ts
+++ b/packages/opencode/src/session/prompt.ts
@@ -81,6 +81,22 @@ IMPORTANT:
 
 const STRUCTURED_OUTPUT_SYSTEM_PROMPT = `IMPORTANT: The user has requested structured output. You MUST use the StructuredOutput tool to provide your final response. Do NOT respond with plain text - you MUST call the StructuredOutput tool with your answer formatted according to the schema.`
 
+const LOOP_CONTINUE_MARKER = "<loop:continue>"
+const LOOP_STOP_MARKER = "<loop:stop>"
+const LOOP_SELF_PACED_DELAY_MS = 1000
+
+type LoopCommandParsed =
+  | { action: "stop" }
+  | {
+      action: "start"
+      intervalMs?: number
+      prompt: string
+    }
+
+type LoopState = {
+  stop: () => void
+}
+
 function mcpResourceBase64Size(value: string) {
   const trimmed = value.replace(/\s/g, "")
   const padding = trimmed.endsWith("==") ? 2 : trimmed.endsWith("=") ? 1 : 0
@@ -151,9 +167,180 @@ const layer = Layer.effect(
 
     const cancel = Effect.fn("SessionPrompt.cancel")(function* (sessionID: SessionID) {
       yield* Effect.logInfo("cancel", { "session.id": sessionID })
+      stopLoop(sessionID)
       yield* state.cancel(sessionID)
     })
 
+    const activeLoops = new Map<SessionID, LoopState>()
+
+    yield* Effect.addFinalizer(() =>
+      Effect.sync(() => {
+        for (const loop of activeLoops.values()) loop.stop()
+        activeLoops.clear()
+      }),
+    )
+
+    const createSyntheticReply = Effect.fn("SessionPrompt.createSyntheticReply")(function* (input: {
+      sessionID: SessionID
+      userText: string
+      assistantText: string
+      agent: string
+      model: { providerID: ProviderV2.ID; modelID: ModelV2.ID }
+      variant?: string
+    }) {
+      const ctx = yield* InstanceState.context
+      const userMsg: SessionV1.User = {
+        id: MessageID.ascending(),
+        role: "user",
+        sessionID: input.sessionID,
+        time: { created: Date.now() },
+        agent: input.agent,
+        model: {
+          providerID: input.model.providerID,
+          modelID: input.model.modelID,
+          ...(input.variant ? { variant: input.variant } : {}),
+        },
+      }
+      yield* sessions.updateMessage(userMsg)
+      yield* sessions.updatePart({
+        id: PartID.ascending(),
+        type: "text",
+        messageID: userMsg.id,
+        sessionID: input.sessionID,
+        text: input.userText,
+      })
+
+      const assistant: SessionV1.Assistant = {
+        id: MessageID.ascending(),
+        role: "assistant",
+        sessionID: input.sessionID,
+        parentID: userMsg.id,
+        mode: input.agent,
+        agent: input.agent,
+        cost: 0,
+        path: { cwd: ctx.directory, root: ctx.worktree },
+        time: { created: Date.now(), completed: Date.now() },
+        tokens: { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } },
+        modelID: input.model.modelID,
+        providerID: input.model.providerID,
+        ...(input.variant ? { variant: input.variant } : {}),
+        finish: "stop",
+      }
+      yield* sessions.updateMessage(assistant)
+      const part: SessionV1.TextPart = {
+        id: PartID.ascending(),
+        type: "text",
+        messageID: assistant.id,
+        sessionID: input.sessionID,
+        text: input.assistantText,
+      }
+      yield* sessions.updatePart(part)
+      return { info: assistant, parts: [part] }
+    })
+
+    const stopLoop = (sessionID: SessionID) => {
+      const loop = activeLoops.get(sessionID)
+      if (!loop) return false
+      loop.stop()
+      activeLoops.delete(sessionID)
+      return true
+    }
+
+    const readLoopPrompt = Effect.fn("SessionPrompt.readLoopPrompt")(function* () {
+      const ctx = yield* InstanceState.context
+      const filepath = path.join(ctx.worktree, ".claude", "loop.md")
+      const exists = yield* fsys.exists(filepath).pipe(Effect.orDie)
+      if (!exists) return undefined
+      const content = (yield* fsys.readFileString(filepath).pipe(Effect.orDie)).trim()
+      return content || undefined
+    })
+
+    const normalizeLoopResult = Effect.fn("SessionPrompt.normalizeLoopResult")(function* (result: SessionV1.WithParts) {
+      const textPart = result.parts.findLast((part): part is SessionV1.TextPart => part.type === "text")
+      if (!textPart) return { continueLoop: false, result }
+      const parsed = extractLoopControl(textPart.text)
+      if (parsed.text === textPart.text) return { continueLoop: false, result }
+      textPart.text = parsed.text
+      yield* sessions.updatePart(textPart)
+      return { continueLoop: parsed.action === "continue", result }
+    })
+
+    const loopStep = Effect.fn("SessionPrompt.loopStep")(function* (input: {
+      sessionID: SessionID
+      agent: string
+      model: { providerID: ProviderV2.ID; modelID: ModelV2.ID }
+      variant?: string
+      promptText: string
+      parts?: CommandInput["parts"]
+    }) {
+      const result = yield* prompt({
+        sessionID: input.sessionID,
+        messageID: MessageID.ascending(),
+        agent: input.agent,
+        model: input.model,
+        variant: input.variant,
+        parts: [{ type: "text", text: input.promptText }, ...(input.parts ?? [])],
+      })
+      return yield* normalizeLoopResult(result)
+    })
+
+    const startLoop = Effect.fn("SessionPrompt.startLoop")(function* (input: {
+      sessionID: SessionID
+      agent: string
+      model: { providerID: ProviderV2.ID; modelID: ModelV2.ID }
+      variant?: string
+      promptText: string
+      intervalMs?: number
+    }) {
+      stopLoop(input.sessionID)
+      let stopped = false
+
+      const delay = input.intervalMs ?? LOOP_SELF_PACED_DELAY_MS
+
+      const fiber = yield* Effect.gen(function* () {
+        while (!stopped) {
+          const exit = yield* loopStep({
+            sessionID: input.sessionID,
+            agent: input.agent,
+            model: input.model,
+            variant: input.variant,
+            promptText: input.promptText,
+          }).pipe(Effect.exit)
+
+          if (stopped) return
+
+          if (Exit.isSuccess(exit)) {
+            if (!exit.value.continueLoop) {
+              stopLoop(input.sessionID)
+              return
+            }
+          } else {
+            const error = Cause.squash(exit.cause)
+            if (error instanceof Session.BusyError) {
+              // continue waiting
+            } else {
+              stopLoop(input.sessionID)
+              yield* events.publish(Session.Event.Error, {
+                sessionID: input.sessionID,
+                error: new NamedError.Unknown({
+                  message: error instanceof Error ? error.message : String(error),
+                }).toObject(),
+              })
+              return
+            }
+          }
+
+          yield* Effect.sleep(delay)
+        }
+      }).pipe(Effect.forkIn(scope))
+
+      activeLoops.set(input.sessionID, {
+        stop: () => {
+          stopped = true
+        },
+      })
+    })
+
     const resolvePromptParts = Effect.fn("SessionPrompt.resolvePromptParts")(function* (template: string) {
       const ctx = yield* InstanceState.context
       const parts: Types.DeepMutable<PromptInput["parts"]> = [{ type: "text", text: template }]
@@ -1367,7 +1554,60 @@ const layer = Layer.effect(
         yield* events.publish(Session.Event.Error, { sessionID: input.sessionID, error: error.toObject() })
         throw error
       }
-      const agentName = cmd.agent ?? input.agent
+      const agentName = cmd.agent ?? input.agent ?? (yield* agents.defaultInfo()).name
+
+      if (input.command === Command.Default.LOOP || input.command === Command.Default.PROACTIVE) {
+        const taskModel = input.model ? Provider.parseModel(input.model) : yield* currentModel(input.sessionID)
+        yield* getModel(taskModel.providerID, taskModel.modelID, input.sessionID)
+        const parsed = parseLoopCommandArguments(input.arguments)
+        if (parsed.action === "stop") {
+          const stopped = stopLoop(input.sessionID)
+          const result = yield* createSyntheticReply({
+            sessionID: input.sessionID,
+            userText: `/${input.command} stop`,
+            assistantText: stopped ? "Stopped the active loop." : "No active loop is running for this session.",
+            agent: agentName,
+            model: taskModel,
+            variant: input.variant,
+          }).pipe(Effect.orDie)
+          yield* events.publish(Command.Event.Executed, {
+            name: input.command,
+            sessionID: input.sessionID,
+            arguments: input.arguments,
+            messageID: result.info.id,
+          })
+          return result
+        }
+
+        const loopPrompt = parsed.prompt || (yield* readLoopPrompt().pipe(Effect.orDie)) || DEFAULT_LOOP_PROMPT
+        const first = yield* loopStep({
+          sessionID: input.sessionID,
+          agent: agentName,
+          model: taskModel,
+          variant: input.variant,
+          promptText: loopPrompt,
+          parts: input.parts,
+        }).pipe(Effect.orDie)
+        if (first.continueLoop) {
+          yield* startLoop({
+            sessionID: input.sessionID,
+            agent: agentName,
+            model: taskModel,
+            variant: input.variant,
+            promptText: loopPrompt,
+            intervalMs: parsed.intervalMs,
+          }).pipe(Effect.orDie)
+        } else {
+          stopLoop(input.sessionID)
+        }
+        yield* events.publish(Command.Event.Executed, {
+          name: input.command,
+          sessionID: input.sessionID,
+          arguments: input.arguments,
+          messageID: first.result.info.id,
+        })
+        return first.result
+      }
 
       const raw = input.arguments.match(argsRegex) ?? []
       const args = raw.map((arg) => arg.replace(quoteTrimRegex, ""))
@@ -1561,6 +1801,70 @@ export const CommandInput = Schema.Struct({
 })
 export type CommandInput = Schema.Schema.Type<typeof CommandInput>
 
+const durationRegex = /^([1-9]\d*)(s|m|h|d)$/i
+export const DEFAULT_LOOP_PROMPT = [
+  "You are running in loop mode.",
+  "",
+  "Perform proactive maintenance for the current session and project.",
+  "If there is nothing useful to do right now, say what you checked and stop.",
+  "",
+  `End your final text response with exactly one control marker on its own line: ${LOOP_CONTINUE_MARKER} if another loop iteration should run, or ${LOOP_STOP_MARKER} if the work is complete or blocked until the user or an external system changes state.`,
+].join("\n")
+
+export function parseLoopCommandArguments(input: string): LoopCommandParsed {
+  const trimmed = input.trim()
+  if (!trimmed) return { action: "start", prompt: "" }
+  if (/^(stop|off)$/i.test(trimmed)) return { action: "stop" }
+  const [first, ...rest] = trimmed.split(/\s+/)
+  const match = first.match(durationRegex)
+  if (!match) return { action: "start", prompt: buildLoopPrompt(trimmed) }
+  const value = Number(match[1])
+  const unit = match[2].toLowerCase()
+  const intervalMs =
+    unit === "s"
+      ? value * 1000
+      : unit === "m"
+        ? value * 60 * 1000
+        : unit === "h"
+          ? value * 60 * 60 * 1000
+          : value * 24 * 60 * 60 * 1000
+  return {
+    action: "start",
+    intervalMs,
+    prompt: rest.length > 0 ? buildLoopPrompt(rest.join(" ")) : "",
+  }
+}
+
+export function buildLoopPrompt(goal: string) {
+  const trimmed = goal.trim()
+  return [
+    "You are running in loop mode.",
+    "",
+    trimmed ? `Goal:\n${trimmed}` : "Perform proactive maintenance for the current session and project.",
+    "",
+    "Work autonomously until the goal is complete or you are blocked by the user, permissions, or an external dependency.",
+    "Re-check relevant state before stopping so you do not miss new changes.",
+    `End your final text response with exactly one control marker on its own line: ${LOOP_CONTINUE_MARKER} if another loop iteration should run, or ${LOOP_STOP_MARKER} if the work is complete or blocked until the user or an external system changes state.`,
+  ].join("\n")
+}
+
+export function extractLoopControl(text: string): { action: "continue" | "stop"; text: string } {
+  const trimmedEnd = text.replace(/\s+$/, "")
+  if (trimmedEnd.endsWith(LOOP_CONTINUE_MARKER)) {
+    return {
+      action: "continue",
+      text: trimmedEnd.slice(0, -LOOP_CONTINUE_MARKER.length).replace(/\s+$/, ""),
+    }
+  }
+  if (trimmedEnd.endsWith(LOOP_STOP_MARKER)) {
+    return {
+      action: "stop",
+      text: trimmedEnd.slice(0, -LOOP_STOP_MARKER.length).replace(/\s+$/, ""),
+    }
+  }
+  return { action: "stop", text }
+}
+
 /** @internal Exported for testing */
 export function createStructuredOutputTool(input: {
   schema: Record<string, any>
diff --git a/packages/opencode/test/session/prompt.test.ts b/packages/opencode/test/session/prompt.test.ts
index 491ad06aaf47..a12aede8a15c 100644
--- a/packages/opencode/test/session/prompt.test.ts
+++ b/packages/opencode/test/session/prompt.test.ts
@@ -5,7 +5,7 @@ import { LayerNode } from "@opencode-ai/core/effect/layer-node"
 import { SessionProjector } from "@opencode-ai/core/session/projector"
 import { eq } from "drizzle-orm"
 import { EventV2Bridge } from "@/event-v2-bridge"
-import { expect } from "bun:test"
+import { expect, test } from "bun:test"
 import { Cause, Deferred, Duration, Effect, Exit, Fiber, Layer } from "effect"
 import path from "path"
 import { fileURLToPath } from "url"
@@ -877,6 +877,74 @@ it.instance("loop continues when finish is stop but assistant has tool parts", (
   }),
 )
 
+it.instance("loop command self-paces until assistant signals stop", () =>
+  Effect.gen(function* () {
+    const { llm } = yield* useServerConfig(providerCfg)
+    const prompt = yield* SessionPrompt.Service
+    const sessions = yield* Session.Service
+    const session = yield* sessions.create({
+      title: "Pinned",
+      permission: [{ permission: "*", pattern: "*", action: "allow" }],
+    })
+
+    yield* llm.text("first iteration\n<loop:continue>")
+    yield* llm.text("second iteration\n<loop:stop>")
+
+    const result = yield* prompt.command({
+      sessionID: session.id,
+      command: "loop",
+      arguments: "check whether the deploy finished",
+    })
+
+    expect(result.parts.some((part) => part.type === "text" && part.text === "first iteration")).toBe(true)
+    let texts: string[] = []
+    for (let i = 0; i < 20; i++) {
+      yield* Effect.sleep(75)
+      const messages = yield* sessions.messages({ sessionID: session.id })
+      texts = messages.flatMap((message) =>
+        message.info.role === "assistant"
+          ? message.parts.flatMap((part) => (part.type === "text" ? [part.text] : []))
+          : [],
+      )
+      if (texts.includes("second iteration")) break
+    }
+
+    expect(texts).toContain("first iteration")
+    expect(texts).toContain("second iteration")
+    expect(texts.join("\n")).not.toContain("<loop:")
+    expect(yield* llm.calls).toBe(2)
+  }),
+)
+
+it.instance("loop command stop cancels the pending next iteration", () =>
+  Effect.gen(function* () {
+    const { llm } = yield* useServerConfig(providerCfg)
+    const prompt = yield* SessionPrompt.Service
+    const sessions = yield* Session.Service
+    const session = yield* sessions.create({
+      title: "Pinned",
+      permission: [{ permission: "*", pattern: "*", action: "allow" }],
+    })
+
+    yield* llm.text("first iteration\n<loop:continue>")
+
+    yield* prompt.command({
+      sessionID: session.id,
+      command: "loop",
+      arguments: "watch the deploy",
+    })
+    const stop = yield* prompt.command({
+      sessionID: session.id,
+      command: "loop",
+      arguments: "stop",
+    })
+
+    expect(stop.parts.some((part) => part.type === "text" && part.text === "Stopped the active loop.")).toBe(true)
+    yield* Effect.sleep(1200)
+    expect(yield* llm.calls).toBe(1)
+  }),
+)
+
 it.instance("failed subtask preserves metadata on error tool state", () =>
   Effect.gen(function* () {
     const { llm } = yield* useServerConfig((url) => ({
@@ -2400,3 +2468,31 @@ noLLMServer.instance(
     }),
   30_000,
 )
+
+test("parseLoopCommandArguments handles stop, interval, and prompt text", () => {
+  expect(SessionPrompt.parseLoopCommandArguments("stop")).toEqual({ action: "stop" })
+  expect(SessionPrompt.parseLoopCommandArguments("5m check deploy")).toEqual({
+    action: "start",
+    intervalMs: 5 * 60 * 1000,
+    prompt: SessionPrompt.buildLoopPrompt("check deploy"),
+  })
+  expect(SessionPrompt.parseLoopCommandArguments("check deploy")).toEqual({
+    action: "start",
+    prompt: SessionPrompt.buildLoopPrompt("check deploy"),
+  })
+})
+
+test("extractLoopControl strips loop markers from assistant text", () => {
+  expect(SessionPrompt.extractLoopControl("still working\n<loop:continue>")).toEqual({
+    action: "continue",
+    text: "still working",
+  })
+  expect(SessionPrompt.extractLoopControl("all done\n<loop:stop>")).toEqual({
+    action: "stop",
+    text: "all done",
+  })
+  expect(SessionPrompt.extractLoopControl("no marker")).toEqual({
+    action: "stop",
+    text: "no marker",
+  })
+})