mervexai 0.0.97

package/src/session/prompt.ts

@@ -0,0 +1,2151 @@
+import { LayerNode } from "@msrfteam/core/effect/layer-node"
+import { PermissionV1 } from "@msrfteam/core/v1/permission"
+import path from "path"
+import { SessionV1 } from "@msrfteam/core/v1/session"
+import os from "os"
+import { SessionID, MessageID, PartID } from "./schema"
+import { MessageV2 } from "./message-v2"
+import { SessionRetry } from "./retry"
+import { SessionRevert } from "./revert"
+import { Session } from "./session"
+import { Agent } from "../agent/agent"
+import { Provider } from "@/provider/provider"
+import { ProviderTransform } from "@/provider/transform"
+import { estimateInputTokens } from "./overflow"
+
+import { type Tool as AITool, tool, jsonSchema } from "ai"
+import type { JSONSchema7 } from "@ai-sdk/provider"
+import { SessionCompaction } from "./compaction"
+import { SystemPrompt } from "./system"
+import { repomapFromContext } from "../tool/repomap"
+import { getDiffContextImpl } from "../tool/diff-context"
+import { Instruction } from "./instruction"
+import { Plugin } from "../plugin"
+import { MAX_STEPS_PROMPT } from "@msrfteam/core/session/runner/max-steps"
+import { ToolRegistry } from "@/tool/registry"
+import { MCP } from "../mcp"
+import { LSP } from "@/lsp/lsp"
+import { ulid } from "ulid"
+import { ChildProcess, ChildProcessSpawner } from "effect/unstable/process"
+import { CrossSpawnSpawner } from "@msrfteam/core/cross-spawn-spawner"
+import * as Stream from "effect/Stream"
+import { Command } from "../command"
+import { pathToFileURL, fileURLToPath } from "url"
+import { Config } from "@/config/config"
+import { ConfigMarkdown } from "@/config/markdown"
+import { SessionSummary } from "./summary"
+import { NamedError } from "@msrfteam/core/util/error"
+import { SessionProcessor } from "./processor"
+import { Tool } from "@/tool/tool"
+import { Permission } from "@/permission"
+import { SessionStatus } from "./status"
+import { LLM } from "./llm"
+import { Shell } from "@msrfteam/core/shell"
+import { ShellID } from "@/tool/shell/id"
+import { FSUtil } from "@msrfteam/core/fs-util"
+import { Ripgrep } from "@msrfteam/core/ripgrep"
+import { Git } from "@msrfteam/core/git"
+import { AppProcess } from "@msrfteam/core/process"
+import { Truncate } from "@/tool/truncate"
+import { Image } from "@/image/image"
+import { decodeDataUrl } from "@/util/data-url"
+import { Process } from "@/util/process"
+import { Cause, Effect, Exit, Latch, Layer, Option, Scope, Context, Schema, Types } from "effect"
+import { detectSkills, dedupeSessionSkills } from "./auto-skill"
+import { Skill } from "../skill"
+import { InstanceState } from "@/effect/instance-state"
+import { TaskTool, type TaskPromptOps } from "@/tool/task"
+import { SessionRunState } from "./run-state"
+import { RuntimeFlags } from "@/effect/runtime-flags"
+import { GitAutoCommit } from "@/git/auto-commit"
+import { Hooks } from "@/hooks"
+import { SkillRestriction } from "@/skill/restriction"
+import { EventV2Bridge } from "@/event-v2-bridge"
+import { Database } from "@msrfteam/core/database/database"
+import { ModelV2 } from "@msrfteam/core/model"
+import { ProviderV2 } from "@msrfteam/core/provider"
+import { eq } from "drizzle-orm"
+import { SessionTable } from "@msrfteam/core/session/sql"
+import { SessionReminders } from "./reminders"
+import { SessionTools } from "./tools"
+import { LLMEvent } from "@msrfteam/llm"
+import { RateLimit } from "./rate-limit"
+import { Token } from "@/util/token"
+
+// @ts-ignore
+globalThis.AI_SDK_LOG_WARNINGS = false
+
+const decodeMessageInfo = Schema.decodeUnknownExit(SessionV1.Info)
+const decodeMessagePart = Schema.decodeUnknownExit(SessionV1.Part)
+const MAX_MCP_RESOURCE_BLOB_BYTES = 10 * 1024 * 1024
+const SUPPORTED_MCP_RESOURCE_ATTACHMENT_MIMES = new Set([
+  "application/pdf",
+  "image/gif",
+  "image/jpeg",
+  "image/png",
+  "image/webp",
+])
+
+const STRUCTURED_OUTPUT_DESCRIPTION = `Use this tool to return your final response in the requested structured format.
+
+IMPORTANT:
+- You MUST call this tool exactly once at the end of your response
+- The input must be valid JSON matching the required schema
+- Complete all necessary research and tool calls BEFORE calling this tool
+- This tool provides your final answer - no further actions are taken after calling it`
+
+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.`
+
+function mcpResourceBase64Size(value: string) {
+  const trimmed = value.replace(/\s/g, "")
+  const padding = trimmed.endsWith("==") ? 2 : trimmed.endsWith("=") ? 1 : 0
+  return Math.max(0, Math.floor((trimmed.length * 3) / 4) - padding)
+}
+
+function formatMcpResourceBytes(value: number) {
+  if (value < 1024) return `${value} B`
+  if (value < 1024 * 1024) return `${Math.ceil(value / 1024)} KB`
+  return `${Math.ceil(value / (1024 * 1024))} MB`
+}
+
+function isOrphanedInterruptedTool(part: SessionV1.ToolPart) {
+  // cleanup() marks abandoned tool_use blocks this way after retries/aborts.
+  // They are not pending work and must not trigger an assistant-prefill request.
+  return part.state.status === "error" && part.state.metadata?.interrupted === true
+}
+
+const AUTH_MESSAGE =
+  /401|403|unauthorized|unauthorised|forbidden|invalid[_-\s]?api[_-\s]?key|invalid[_-\s]?auth|authentication|access denied|account.*(suspend|deactiv|disabled)/i
+
+// Provider reliability: auth (401/403), rate-limit/overload, and context overflow errors are
+// eligible for automatic model fallback instead of killing the turn.
+function isFallbackEligible(error: SessionRetry.Err): boolean {
+  // Context overflow errors are eligible for fallback to a larger-context model
+  if (SessionV1.ContextOverflowError.isInstance(error)) return true
+  if (SessionV1.APIError.isInstance(error)) {
+    const status = error.data.statusCode
+    if (status === 401 || status === 403) return true
+    if (SessionRetry.retryable(error, "") !== undefined) return true
+    if (typeof error.data.message === "string" && AUTH_MESSAGE.test(error.data.message)) return true
+    if (typeof error.data.responseBody === "string" && AUTH_MESSAGE.test(error.data.responseBody)) return true
+    return false
+  }
+  try {
+    return AUTH_MESSAGE.test(JSON.stringify(error))
+  } catch {
+    return false
+  }
+}
+
+// Pinned files: absolute paths stored in session metadata and
+// injected as file parts into every new user message.
+const PINNED_KEY = "mervex.pinnedFiles"
+
+function readPinned(metadata: { readonly [key: string]: any } | undefined): string[] {
+  const value = metadata?.[PINNED_KEY]
+  if (!Array.isArray(value)) return []
+  return value.filter((item): item is string => typeof item === "string")
+}
+
+function pinnedFileParts(session: Session.Info, parts: PromptInput["parts"]) {
+  const existing = new Set<string>()
+  for (const part of parts) {
+    if (part.type === "file" && part.url.startsWith("file:")) {
+      existing.add(path.normalize(fileURLToPath(part.url)))
+    }
+  }
+  return readPinned(session.metadata)
+    .filter((file) => !existing.has(path.normalize(file)))
+    .map((file) => ({
+      type: "file" as const,
+      mime: "text/plain",
+      filename: path.basename(file),
+      url: pathToFileURL(file).href,
+    }))
+}
+
+export interface Interface {
+  readonly cancel: (sessionID: SessionID) => Effect.Effect<void>
+  readonly prompt: (input: PromptInput) => Effect.Effect<SessionV1.WithParts, Image.Error>
+  readonly loop: (input: LoopInput) => Effect.Effect<SessionV1.WithParts>
+  readonly shell: (input: ShellInput) => Effect.Effect<SessionV1.WithParts, Session.BusyError>
+  readonly command: (input: CommandInput) => Effect.Effect<SessionV1.WithParts, Image.Error>
+  readonly resolvePromptParts: (template: string) => Effect.Effect<PromptInput["parts"]>
+}
+
+export class Service extends Context.Service<Service, Interface>()("@mervex/SessionPrompt") {}
+
+const layer = Layer.effect(
+  Service,
+  Effect.gen(function* () {
+    const status = yield* SessionStatus.Service
+    const sessions = yield* Session.Service
+    const agents = yield* Agent.Service
+    const provider = yield* Provider.Service
+    const processor = yield* SessionProcessor.Service
+    const compaction = yield* SessionCompaction.Service
+    const plugin = yield* Plugin.Service
+    const commands = yield* Command.Service
+    const config = yield* Config.Service
+    const permission = yield* Permission.Service
+    const fsys = yield* FSUtil.Service
+    const ripgrep = yield* Ripgrep.Service
+    const git = yield* Git.Service
+    const mcp = yield* MCP.Service
+    const lsp = yield* LSP.Service
+    const registry = yield* ToolRegistry.Service
+    const truncate = yield* Truncate.Service
+    const image = yield* Image.Service
+    const spawner = yield* ChildProcessSpawner.ChildProcessSpawner
+    const scope = yield* Scope.Scope
+    const instruction = yield* Instruction.Service
+    const state = yield* SessionRunState.Service
+    const revert = yield* SessionRevert.Service
+    const summary = yield* SessionSummary.Service
+    const sys = yield* SystemPrompt.Service
+    const llm = yield* LLM.Service
+    const events = yield* EventV2Bridge.Service
+    const flags = yield* RuntimeFlags.Service
+    const gitAutoCommit = yield* GitAutoCommit.Service
+    const hooks = yield* Hooks.Service
+    const skillRestriction = yield* SkillRestriction.Service
+    const skill = yield* Skill.Service
+    const database = yield* Database.Service
+    const rateLimit = yield* RateLimit.Service
+    const { db } = database
+    const ops = Effect.fn("SessionPrompt.ops")(function* () {
+      return {
+        cancel: (sessionID: SessionID) => cancel(sessionID),
+        resolvePromptParts: (template: string) => resolvePromptParts(template),
+        prompt: (input: PromptInput) => prompt(input).pipe(Effect.catch(Effect.die)),
+      } satisfies TaskPromptOps
+    })
+
+    const cancel = Effect.fn("SessionPrompt.cancel")(function* (sessionID: SessionID) {
+      yield* Effect.logInfo("cancel", { "session.id": sessionID })
+      yield* state.cancel(sessionID)
+    })
+
+    const resolvePromptParts = Effect.fn("SessionPrompt.resolvePromptParts")(function* (template: string) {
+      const ctx = yield* InstanceState.context
+      const parts: Types.DeepMutable<PromptInput["parts"]> = [{ type: "text", text: template }]
+      const files = ConfigMarkdown.files(template)
+      const seen = new Set<string>()
+      yield* Effect.forEach(
+        files,
+        Effect.fnUntraced(function* (match) {
+          const name = match[1]
+          if (!name) return
+          if (seen.has(name)) return
+          seen.add(name)
+
+          const filepath = name.startsWith("~/")
+            ? path.join(os.homedir(), name.slice(2))
+            : path.resolve(ctx.worktree, name)
+
+          const info = yield* fsys.stat(filepath).pipe(Effect.option)
+          if (Option.isNone(info)) {
+            const found = yield* agents.get(name)
+            if (found) parts.push({ type: "agent", name: found.name })
+            return
+          }
+          const stat = info.value
+          parts.push({
+            type: "file",
+            url: pathToFileURL(filepath).href,
+            filename: name,
+            mime: stat.type === "Directory" ? "application/x-directory" : "text/plain",
+          })
+        }),
+        { concurrency: "unbounded", discard: true },
+      )
+      return parts
+    })
+
+    const title = Effect.fn("SessionPrompt.ensureTitle")(function* (input: {
+      session: Session.Info
+      history: SessionV1.WithParts[]
+      providerID: ProviderV2.ID
+      modelID: ModelV2.ID
+    }) {
+      yield* Effect.logDebug("ensureTitle called", { sessionID: input.session.id, title: input.session.title })
+      if (input.session.parentID) return
+      if (!Session.isDefaultTitle(input.session.title)) return
+
+      const real = (m: SessionV1.WithParts) =>
+        m.info.role === "user" && !m.parts.every((p) => "synthetic" in p && p.synthetic)
+      const idx = input.history.findIndex(real)
+      if (idx === -1) return
+      if (input.history.filter(real).length !== 1) return
+
+      const context = input.history.slice(0, idx + 1)
+      const firstUser = context[idx]
+      if (!firstUser || firstUser.info.role !== "user") return
+      const firstInfo = firstUser.info
+
+      const subtasks = firstUser.parts.filter((p): p is SessionV1.SubtaskPart => p.type === "subtask")
+      const onlySubtasks = subtasks.length > 0 && firstUser.parts.every((p) => p.type === "subtask")
+
+      const ag = yield* agents.get("title")
+      if (!ag) {
+        yield* Effect.logWarning("title agent not found, skipping title generation")
+        return
+      }
+      const mdl = ag.model
+        ? yield* provider.getModel(ag.model.providerID, ag.model.modelID)
+        : ((yield* provider.getSmallModel(input.providerID)) ??
+          (yield* provider.getModel(input.providerID, input.modelID)))
+      if (!mdl) {
+        yield* Effect.logWarning("no model available for title generation")
+        return
+      }
+      yield* Effect.logDebug("generating title", { provider: input.providerID, model: mdl.id })
+      const msgs = onlySubtasks
+        ? [{ role: "user" as const, content: subtasks.map((p) => p.prompt).join("\n") }]
+        : yield* MessageV2.toModelMessagesEffect(context, mdl)
+      const text = yield* llm
+        .stream({
+          agent: ag,
+          user: firstInfo,
+          system: [],
+          small: true,
+          tools: {},
+          model: mdl,
+          sessionID: input.session.id,
+          retries: 2,
+          messages: [{ role: "user", content: "Generate a title for this conversation:\n" }, ...msgs],
+        })
+        .pipe(
+          Stream.filter(LLMEvent.is.textDelta),
+          Stream.map((e) => e.text),
+          Stream.mkString,
+          Effect.catchCause((cause) =>
+            Effect.logError("title generation LLM failed", { error: Cause.squash(cause) }).pipe(
+              Effect.map(() => ""),
+            ),
+          ),
+        )
+      if (!text) return
+      const cleaned = text
+        .replace(/<think>[\s\S]*?<\/think>\s*/g, "")
+        .split("\n")
+        .map((line) => line.trim())
+        .find((line) => line.length > 0)
+      if (!cleaned) return
+      const t = cleaned.length > 100 ? cleaned.substring(0, 97) + "..." : cleaned
+      yield* Effect.logDebug("title generated", { title: t })
+      yield* sessions
+        .setTitle({ sessionID: input.session.id, title: t })
+        .pipe(Effect.catchCause((cause) => Effect.logError("failed to generate title", { error: Cause.squash(cause) })))
+    })
+
+    const handleSubtask = Effect.fn("SessionPrompt.handleSubtask")(function* (input: {
+      task: SessionV1.SubtaskPart
+      model: Provider.Model
+      lastUser: SessionV1.User
+      sessionID: SessionID
+      session: Session.Info
+      msgs: SessionV1.WithParts[]
+    }) {
+      const { task, model, lastUser, sessionID, session, msgs } = input
+      const ctx = yield* InstanceState.context
+      const promptOps = yield* ops()
+      const { task: taskTool } = yield* registry.named()
+      const taskModel = task.model ? yield* getModel(task.model.providerID, task.model.modelID, sessionID) : model
+      const assistantMessage: SessionV1.Assistant = yield* sessions.updateMessage({
+        id: MessageID.ascending(),
+        role: "assistant",
+        parentID: lastUser.id,
+        sessionID,
+        mode: task.agent,
+        agent: task.agent,
+        variant: lastUser.model.variant,
+        path: { cwd: ctx.directory, root: ctx.worktree },
+        cost: 0,
+        tokens: { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } },
+        modelID: taskModel.id,
+        providerID: taskModel.providerID,
+        time: { created: Date.now() },
+      })
+      let part: SessionV1.ToolPart = yield* sessions.updatePart({
+        id: PartID.ascending(),
+        messageID: assistantMessage.id,
+        sessionID: assistantMessage.sessionID,
+        type: "tool",
+        callID: ulid(),
+        tool: TaskTool.id,
+        state: {
+          status: "running",
+          input: {
+            prompt: task.prompt,
+            description: task.description,
+            subagent_type: task.agent,
+            command: task.command,
+          },
+          time: { start: Date.now() },
+        },
+      })
+      const taskArgs = {
+        prompt: task.prompt,
+        description: task.description,
+        subagent_type: task.agent,
+        command: task.command,
+      }
+      yield* plugin.trigger(
+        "tool.execute.before",
+        { tool: TaskTool.id, sessionID, callID: part.id },
+        { args: taskArgs },
+      )
+
+      const taskAgent = yield* agents.get(task.agent)
+      if (!taskAgent) {
+        const available = (yield* agents.list()).filter((a) => !a.hidden).map((a) => a.name)
+        const hint = available.length ? ` Available agents: ${available.join(", ")}` : ""
+        const error = new NamedError.Unknown({ message: `Agent not found: "${task.agent}".${hint}` })
+        yield* events.publish(Session.Event.Error, { sessionID, error: error.toObject() })
+        throw error
+      }
+
+      let error: Error | undefined
+      const taskAbort = new AbortController()
+      const result = yield* taskTool
+        .execute(taskArgs, {
+          agent: task.agent,
+          messageID: assistantMessage.id,
+          sessionID,
+          abort: taskAbort.signal,
+          callID: part.callID,
+          extra: { bypassAgentCheck: true, promptOps },
+          messages: msgs,
+          metadata: (val: { title?: string; metadata?: Record<string, any> }) =>
+            Effect.gen(function* () {
+              part = yield* sessions.updatePart({
+                ...part,
+                type: "tool",
+                state: { ...part.state, ...val },
+              } satisfies SessionV1.ToolPart)
+            }),
+          ask: (req: any) =>
+            permission
+              .ask({
+                ...req,
+                sessionID,
+                ruleset: Permission.merge(taskAgent.permission, session.permission ?? []),
+              })
+              .pipe(Effect.orDie),
+        })
+        .pipe(
+          Effect.catchCause((cause) => {
+            const defect = Cause.squash(cause)
+            error = defect instanceof Error ? defect : new Error(String(defect))
+            return Effect.logError("subtask execution failed", {
+              error,
+              agent: task.agent,
+              description: task.description,
+            })
+          }),
+          Effect.onInterrupt(() =>
+            Effect.gen(function* () {
+              taskAbort.abort()
+              assistantMessage.finish = "tool-calls"
+              assistantMessage.time.completed = Date.now()
+              yield* sessions.updateMessage(assistantMessage)
+              if (part.state.status === "running") {
+                yield* sessions.updatePart({
+                  ...part,
+                  state: {
+                    status: "error",
+                    error: "Cancelled",
+                    time: { start: part.state.time.start, end: Date.now() },
+                    metadata: part.state.metadata,
+                    input: part.state.input,
+                  },
+                } satisfies SessionV1.ToolPart)
+              }
+            }),
+          ),
+        )
+
+      const attachments = result?.attachments?.map((attachment) => ({
+        ...attachment,
+        id: PartID.ascending(),
+        sessionID,
+        messageID: assistantMessage.id,
+      }))
+
+      yield* plugin.trigger(
+        "tool.execute.after",
+        { tool: TaskTool.id, sessionID, callID: part.id, args: taskArgs },
+        result,
+      )
+
+      assistantMessage.finish = "tool-calls"
+      assistantMessage.time.completed = Date.now()
+      yield* sessions.updateMessage(assistantMessage)
+
+      if (result && part.state.status === "running") {
+        yield* sessions.updatePart({
+          ...part,
+          state: {
+            status: "completed",
+            input: part.state.input,
+            title: result.title,
+            metadata: result.metadata,
+            output: result.output,
+            attachments,
+            time: { ...part.state.time, end: Date.now() },
+          },
+        } satisfies SessionV1.ToolPart)
+      }
+
+      if (!result) {
+        yield* sessions.updatePart({
+          ...part,
+          state: {
+            status: "error",
+            error: error ? `Tool execution failed: ${error.message}` : "Tool execution failed",
+            time: {
+              start: part.state.status === "running" ? part.state.time.start : Date.now(),
+              end: Date.now(),
+            },
+            metadata: part.state.status === "pending" ? undefined : part.state.metadata,
+            input: part.state.input,
+          },
+        } satisfies SessionV1.ToolPart)
+      }
+
+      if (!task.command) return
+
+      const summaryUserMsg: SessionV1.User = {
+        id: MessageID.ascending(),
+        sessionID,
+        role: "user",
+        time: { created: Date.now() },
+        agent: lastUser.agent,
+        model: lastUser.model,
+      }
+      yield* sessions.updateMessage(summaryUserMsg)
+      yield* sessions.updatePart({
+        id: PartID.ascending(),
+        messageID: summaryUserMsg.id,
+        sessionID,
+        type: "text",
+        text: "Summarize the task tool output above and continue with your task.",
+        synthetic: true,
+      } satisfies SessionV1.TextPart)
+    })
+
+    const shellImpl = Effect.fn("SessionPrompt.shellImpl")(function* (input: ShellInput, ready?: Latch.Latch) {
+      return yield* Effect.uninterruptibleMask((restore) =>
+        Effect.gen(function* () {
+          const markReady = ready ? ready.open.pipe(Effect.asVoid) : Effect.void
+          const { msg, part, cwd } = yield* Effect.gen(function* () {
+            const ctx = yield* InstanceState.context
+            const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+            if (session.revert) {
+              yield* revert.cleanup(session)
+            }
+            const agent = yield* agents.get(input.agent)
+            if (!agent) {
+              const available = (yield* agents.list()).filter((a) => !a.hidden).map((a) => a.name)
+              const hint = available.length ? ` Available agents: ${available.join(", ")}` : ""
+              const error = new NamedError.Unknown({ message: `Agent not found: "${input.agent}".${hint}` })
+              yield* events.publish(Session.Event.Error, { sessionID: input.sessionID, error: error.toObject() })
+              throw error
+            }
+            const model = input.model ?? agent.model ?? (yield* currentModel(input.sessionID))
+            const userMsg: SessionV1.User = {
+              id: input.messageID ?? MessageID.ascending(),
+              sessionID: input.sessionID,
+              time: { created: Date.now() },
+              role: "user",
+              agent: input.agent,
+              model: { providerID: model.providerID, modelID: model.modelID },
+            }
+            yield* sessions.updateMessage(userMsg)
+            const userPart: SessionV1.Part = {
+              type: "text",
+              id: PartID.ascending(),
+              messageID: userMsg.id,
+              sessionID: input.sessionID,
+              text: "The following tool was executed by the user",
+              synthetic: true,
+            }
+            yield* sessions.updatePart(userPart)
+
+            const msg: SessionV1.Assistant = {
+              id: MessageID.ascending(),
+              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() },
+              role: "assistant",
+              tokens: { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } },
+              modelID: model.modelID,
+              providerID: model.providerID,
+            }
+            yield* sessions.updateMessage(msg)
+            const started = Date.now()
+            const part: SessionV1.ToolPart = {
+              type: "tool",
+              id: PartID.ascending(),
+              messageID: msg.id,
+              sessionID: input.sessionID,
+              tool: ShellID.ToolID,
+              callID: ulid(),
+              state: {
+                status: "running",
+                time: { start: started },
+                input: { command: input.command },
+              },
+            }
+            yield* sessions.updatePart(part)
+            return { msg, part, cwd: ctx.directory }
+          }).pipe(Effect.ensuring(markReady))
+
+          const cfg = yield* config.get()
+          const sh = Shell.preferred(cfg.shell)
+          const args = Shell.args(sh, input.command, cwd)
+          let output = ""
+          let aborted = false
+
+          const finish = Effect.uninterruptible(
+            Effect.gen(function* () {
+              if (aborted) {
+                output += "\n\n" + ["<metadata>", "User aborted the command", "</metadata>"].join("\n")
+              }
+              const completed = Date.now()
+              if (!msg.time.completed) {
+                msg.time.completed = completed
+                yield* sessions.updateMessage(msg)
+              }
+              if (part.state.status === "running") {
+                part.state = {
+                  status: "completed",
+                  time: { ...part.state.time, end: completed },
+                  input: part.state.input,
+                  title: "",
+                  metadata: { output },
+                  output,
+                }
+                yield* sessions.updatePart(part)
+              }
+            }),
+          )
+
+          const exit = yield* restore(
+            Effect.gen(function* () {
+              const shellEnv = yield* plugin.trigger(
+                "shell.env",
+                { cwd, sessionID: input.sessionID, callID: part.callID },
+                { env: {} },
+              )
+              const cmd = ChildProcess.make(sh, args, {
+                cwd,
+                extendEnv: true,
+                env: { ...shellEnv.env, TERM: "dumb" },
+                stdin: "ignore",
+                forceKillAfter: "3 seconds",
+              })
+              const handle = yield* spawner.spawn(cmd)
+              yield* Stream.runForEach(Stream.decodeText(handle.all), (chunk) =>
+                Effect.gen(function* () {
+                  output += chunk
+                  if (part.state.status === "running") {
+                    part.state.metadata = { output }
+                    yield* sessions.updatePart(part)
+                  }
+                }),
+              )
+              yield* handle.exitCode
+            }).pipe(Effect.scoped, Effect.orDie),
+          ).pipe(Effect.exit)
+
+          if (Exit.isFailure(exit) && Cause.hasInterrupts(exit.cause) && !Cause.hasDies(exit.cause)) {
+            aborted = true
+          }
+          yield* finish
+
+          if (Exit.isFailure(exit) && !aborted && !Cause.hasInterruptsOnly(exit.cause)) {
+            return yield* Effect.failCause(exit.cause)
+          }
+
+          return { info: msg, parts: [part] }
+        }),
+      )
+    })
+
+    const getModel = Effect.fn("SessionPrompt.getModel")(function* (
+      providerID: ProviderV2.ID,
+      modelID: ModelV2.ID,
+      sessionID: SessionID,
+    ) {
+      const exit = yield* provider.getModel(providerID, modelID).pipe(Effect.exit)
+      if (Exit.isSuccess(exit)) return exit.value
+      const err = Cause.squash(exit.cause)
+      if (Provider.ModelNotFoundError.isInstance(err)) {
+        const hint = err.suggestions?.length ? ` Did you mean: ${err.suggestions.join(", ")}?` : ""
+        yield* events.publish(Session.Event.Error, {
+          sessionID,
+          error: new NamedError.Unknown({
+            message: `Model not found: ${err.providerID}/${err.modelID}.${hint}`,
+          }).toObject(),
+        })
+      }
+      return yield* Effect.die(err)
+    })
+
+    const currentModel = Effect.fnUntraced(function* (sessionID: SessionID) {
+      const current = yield* db
+        .select({ model: SessionTable.model })
+        .from(SessionTable)
+        .where(eq(SessionTable.id, sessionID))
+        .get()
+        .pipe(Effect.orDie)
+      if (current?.model) {
+        return {
+          providerID: ProviderV2.ID.make(current.model.providerID),
+          modelID: ModelV2.ID.make(current.model.id),
+          ...(current.model.variant && current.model.variant !== "default" ? { variant: current.model.variant } : {}),
+        }
+      }
+      const match = yield* sessions
+        .findMessage(sessionID, (m) => m.info.role === "user" && !!m.info.model)
+        .pipe(Effect.orDie)
+      if (Option.isSome(match) && match.value.info.role === "user") return match.value.info.model
+      return yield* provider.defaultModel().pipe(Effect.orDie)
+    })
+
+    const createUserMessage = Effect.fn("SessionPrompt.createUserMessage")(function* (input: PromptInput) {
+      const agentName = input.agent
+      const ag = agentName ? yield* agents.get(agentName) : yield* agents.defaultInfo()
+      if (!ag) {
+        const available = (yield* agents.list()).filter((a) => !a.hidden).map((a) => a.name)
+        const hint = available.length ? ` Available agents: ${available.join(", ")}` : ""
+        const error = new NamedError.Unknown({ message: `Agent not found: "${agentName}".${hint}` })
+        yield* events.publish(Session.Event.Error, { sessionID: input.sessionID, error: error.toObject() })
+        throw error
+      }
+
+      const model = input.model ?? ag.model ?? (yield* currentModel(input.sessionID))
+      const same = ag.model && model.providerID === ag.model.providerID && model.modelID === ag.model.modelID
+      const full =
+        !input.variant && ag.variant && same
+          ? yield* provider
+              .getModel(model.providerID, model.modelID)
+              .pipe(Effect.catchIf(Provider.ModelNotFoundError.isInstance, () => Effect.succeed(undefined)))
+          : undefined
+      const variant = input.variant ?? (ag.variant && full?.variants?.[ag.variant] ? ag.variant : undefined)
+
+      const info: SessionV1.User = {
+        id: input.messageID ?? MessageID.ascending(),
+        role: "user",
+        sessionID: input.sessionID,
+        time: { created: Date.now() },
+        tools: input.tools,
+        agent: ag.name,
+        model: {
+          providerID: model.providerID,
+          modelID: model.modelID,
+          variant,
+        },
+        system: input.system,
+        format: input.format,
+      }
+
+      const current = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+      if (
+        current.agent !== info.agent ||
+        current.model?.providerID !== info.model.providerID ||
+        current.model?.id !== info.model.modelID ||
+        (current.model?.variant === "default" ? undefined : current.model?.variant) !== info.model.variant
+      ) {
+        yield* sessions.setAgentModel({
+          sessionID: input.sessionID,
+          agent: info.agent,
+          model: {
+            id: info.model.modelID,
+            providerID: info.model.providerID,
+            variant: info.model.variant ?? "default",
+          },
+          time: info.time.created,
+        })
+      }
+
+      yield* Effect.addFinalizer(() => instruction.clear(info.id))
+
+      type Draft<T> = T extends SessionV1.Part ? Omit<T, "id"> & { id?: string } : never
+      const assign = (part: Draft<SessionV1.Part>): SessionV1.Part => ({
+        ...part,
+        id: part.id ? PartID.make(part.id) : PartID.ascending(),
+      })
+
+      const resolvePart: (part: PromptInput["parts"][number]) => Effect.Effect<Draft<SessionV1.Part>[]> = Effect.fn(
+        "SessionPrompt.resolveUserPart",
+      )(function* (part) {
+        if (part.type === "file") {
+          if (part.source?.type === "resource") {
+            const { clientName, uri } = part.source
+            yield* Effect.logInfo("mcp resource", { clientName, uri, mime: part.mime })
+            const pieces: Draft<SessionV1.Part>[] = [
+              {
+                messageID: info.id,
+                sessionID: input.sessionID,
+                type: "text",
+                synthetic: true,
+                text: `Reading MCP resource: ${part.filename} (${uri})`,
+              },
+            ]
+            const exit = yield* mcp.readResource(clientName, uri).pipe(Effect.exit)
+            if (Exit.isSuccess(exit)) {
+              const content = exit.value
+              if (!content) throw new Error(`Resource not found: ${clientName}/${uri}`)
+              const items = Array.isArray(content.contents) ? content.contents : [content.contents]
+              for (const c of items) {
+                if (!c || typeof c !== "object") continue
+                if ("text" in c && typeof c.text === "string" && c.text) {
+                  pieces.push({
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: c.text,
+                  })
+                } else if ("blob" in c && typeof c.blob === "string" && c.blob) {
+                  const mime = "mimeType" in c && typeof c.mimeType === "string" ? c.mimeType : part.mime
+                  const filename = "uri" in c && typeof c.uri === "string" ? c.uri : part.filename
+                  const size = mcpResourceBase64Size(c.blob)
+                  if (!SUPPORTED_MCP_RESOURCE_ATTACHMENT_MIMES.has(mime)) {
+                    pieces.push({
+                      messageID: info.id,
+                      sessionID: input.sessionID,
+                      type: "text",
+                      synthetic: true,
+                      text: `[Binary MCP resource omitted: ${filename ?? uri} (${mime}, ${formatMcpResourceBytes(size)}) is not a supported attachment type]`,
+                    })
+                    continue
+                  }
+                  if (size > MAX_MCP_RESOURCE_BLOB_BYTES) {
+                    pieces.push({
+                      messageID: info.id,
+                      sessionID: input.sessionID,
+                      type: "text",
+                      synthetic: true,
+                      text: `[Binary MCP resource omitted: ${filename ?? uri} (${mime}, ${formatMcpResourceBytes(size)}) exceeds ${formatMcpResourceBytes(MAX_MCP_RESOURCE_BLOB_BYTES)}]`,
+                    })
+                    continue
+                  }
+                  pieces.push({
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: `[Binary MCP resource attached: ${filename ?? uri} (${mime})]`,
+                  })
+                  pieces.push({
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "file",
+                    mime,
+                    filename,
+                    url: `data:${mime};base64,${c.blob}`,
+                  })
+                }
+              }
+            } else {
+              const error = Cause.squash(exit.cause)
+              yield* Effect.logError("failed to read MCP resource", { error, clientName, uri })
+              const message = error instanceof Error ? error.message : String(error)
+              pieces.push({
+                messageID: info.id,
+                sessionID: input.sessionID,
+                type: "text",
+                synthetic: true,
+                text: `Failed to read MCP resource ${part.filename}: ${message}`,
+              })
+            }
+            return pieces
+          }
+          const url = new URL(part.url)
+          switch (url.protocol) {
+            case "data:":
+              if (part.mime === "text/plain") {
+                return [
+                  {
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: `Called the Read tool with the following input: ${JSON.stringify({ filePath: part.filename })}`,
+                  },
+                  {
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: decodeDataUrl(part.url),
+                  },
+                  { ...part, messageID: info.id, sessionID: input.sessionID },
+                ]
+              }
+              break
+            case "file:": {
+              yield* Effect.logInfo("file", { mime: part.mime })
+              const filepath = fileURLToPath(part.url)
+              const mime = (yield* fsys.isDir(filepath)) ? "application/x-directory" : part.mime
+
+              const { read } = yield* registry.named()
+              const execRead = (args: Parameters<typeof read.execute>[0], extra?: Tool.Context["extra"]) => {
+                const controller = new AbortController()
+                return read
+                  .execute(args, {
+                    sessionID: input.sessionID,
+                    abort: controller.signal,
+                    agent: input.agent!,
+                    messageID: info.id,
+                    extra: { bypassCwdCheck: true, ...extra },
+                    messages: [],
+                    metadata: () => Effect.void,
+                    ask: () => Effect.void,
+                  })
+                  .pipe(Effect.onInterrupt(() => Effect.sync(() => controller.abort())))
+              }
+
+              if (mime === "text/plain") {
+                let offset: number | undefined
+                let limit: number | undefined
+                const range = { start: url.searchParams.get("start"), end: url.searchParams.get("end") }
+                if (range.start != null) {
+                  const filePathURI = part.url.split("?")[0]
+                  let start = parseInt(range.start)
+                  let end = range.end ? parseInt(range.end) : undefined
+                  if (start === end) {
+                    const symbols = yield* lsp.documentSymbol(filePathURI).pipe(Effect.catch(() => Effect.succeed([])))
+                    for (const symbol of symbols) {
+                      let r: LSP.Range | undefined
+                      if ("range" in symbol) r = symbol.range
+                      else if ("location" in symbol) r = symbol.location.range
+                      if (r?.start?.line && r?.start?.line === start) {
+                        start = r.start.line
+                        end = r?.end?.line ?? start
+                        break
+                      }
+                    }
+                  }
+                  offset = Math.max(start, 1)
+                  if (end) limit = end - (offset - 1)
+                }
+                const args = { filePath: filepath, offset, limit }
+                const pieces: Draft<SessionV1.Part>[] = [
+                  {
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: `Called the Read tool with the following input: ${JSON.stringify(args)}`,
+                  },
+                ]
+                const exit = yield* provider.getModel(info.model.providerID, info.model.modelID).pipe(
+                  Effect.flatMap((mdl) => execRead(args, { model: mdl })),
+                  Effect.exit,
+                )
+                if (Exit.isSuccess(exit)) {
+                  const result = exit.value
+                  pieces.push({
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: result.output,
+                  })
+                  if (result.attachments?.length) {
+                    pieces.push(
+                      ...result.attachments.map((a) => ({
+                        ...a,
+                        synthetic: true,
+                        filename: a.filename ?? part.filename,
+                        messageID: info.id,
+                        sessionID: input.sessionID,
+                      })),
+                    )
+                  } else {
+                    pieces.push({ ...part, mime, messageID: info.id, sessionID: input.sessionID })
+                  }
+                } else {
+                  const error = Cause.squash(exit.cause)
+                  yield* Effect.logError("failed to read file", { error, filepath })
+                  const message = error instanceof Error ? error.message : String(error)
+                  yield* events.publish(Session.Event.Error, {
+                    sessionID: input.sessionID,
+                    error: new NamedError.Unknown({ message }).toObject(),
+                  })
+                  pieces.push({
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: `Read tool failed to read ${filepath} with the following error: ${message}`,
+                  })
+                }
+                return pieces
+              }
+
+              if (mime === "application/x-directory") {
+                const args = { filePath: filepath }
+                const exit = yield* execRead(args).pipe(Effect.exit)
+                if (Exit.isFailure(exit)) {
+                  const error = Cause.squash(exit.cause)
+                  yield* Effect.logError("failed to read directory", { error, filepath })
+                  const message = error instanceof Error ? error.message : String(error)
+                  yield* events.publish(Session.Event.Error, {
+                    sessionID: input.sessionID,
+                    error: new NamedError.Unknown({ message }).toObject(),
+                  })
+                  return [
+                    {
+                      messageID: info.id,
+                      sessionID: input.sessionID,
+                      type: "text",
+                      synthetic: true,
+                      text: `Read tool failed to read ${filepath} with the following error: ${message}`,
+                    },
+                  ]
+                }
+                return [
+                  {
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: `Called the Read tool with the following input: ${JSON.stringify(args)}`,
+                  },
+                  {
+                    messageID: info.id,
+                    sessionID: input.sessionID,
+                    type: "text",
+                    synthetic: true,
+                    text: exit.value.output,
+                  },
+                  { ...part, mime, messageID: info.id, sessionID: input.sessionID },
+                ]
+              }
+
+              return [
+                {
+                  messageID: info.id,
+                  sessionID: input.sessionID,
+                  type: "text",
+                  synthetic: true,
+                  text: `Called the Read tool with the following input: {"filePath":"${filepath}"}`,
+                },
+                {
+                  id: part.id,
+                  messageID: info.id,
+                  sessionID: input.sessionID,
+                  type: "file",
+                  url:
+                    `data:${mime};base64,` +
+                    Buffer.from(yield* fsys.readFile(filepath).pipe(Effect.catch(Effect.die))).toString("base64"),
+                  mime,
+                  filename: part.filename!,
+                  source: part.source,
+                },
+              ]
+            }
+          }
+        }
+
+        if (part.type === "agent") {
+          const perm = Permission.evaluate("task", part.name, ag.permission)
+          const hint = perm.action === "deny" ? " . Invoked by user; guaranteed to exist." : ""
+          return [
+            { ...part, messageID: info.id, sessionID: input.sessionID },
+            {
+              messageID: info.id,
+              sessionID: input.sessionID,
+              type: "text",
+              synthetic: true,
+              text:
+                " Use the above message and context to generate a prompt and call the task tool with subagent: " +
+                part.name +
+                hint,
+            },
+          ]
+        }
+
+        return [{ ...part, messageID: info.id, sessionID: input.sessionID }]
+      })
+
+      const resolvedParts = yield* Effect.forEach(
+        [...input.parts, ...pinnedFileParts(current, input.parts)],
+        resolvePart,
+        { concurrency: "unbounded" },
+      ).pipe(Effect.map((x) => x.flat().map(assign)))
+
+      yield* plugin.trigger(
+        "chat.message",
+        {
+          sessionID: input.sessionID,
+          agent: input.agent,
+          model: input.model,
+          messageID: input.messageID,
+          variant: input.variant,
+        },
+        { message: info, parts: resolvedParts },
+      )
+
+      const parts = yield* Effect.forEach(resolvedParts, (part) =>
+        part.type === "file" && part.mime.startsWith("image/")
+          ? image.normalize(part).pipe(
+              Effect.catchIf(
+                (error) => error instanceof Image.ResizerUnavailableError,
+                () => Effect.succeed(part),
+              ),
+            )
+          : Effect.succeed(part),
+      )
+
+      const parsed = decodeMessageInfo(info, { errors: "all", propertyOrder: "original" })
+      if (Exit.isFailure(parsed)) {
+        yield* Effect.logError("invalid user message before save", {
+          sessionID: input.sessionID,
+          messageID: info.id,
+          agent: info.agent,
+          model: info.model,
+          cause: Cause.pretty(parsed.cause),
+        })
+      }
+      for (const [index, part] of parts.entries()) {
+        const p = decodeMessagePart(part, { errors: "all", propertyOrder: "original" })
+        if (Exit.isSuccess(p)) continue
+        yield* Effect.logError("invalid user part before save", {
+          sessionID: input.sessionID,
+          messageID: info.id,
+          partID: part.id,
+          partType: part.type,
+          index,
+          cause: Cause.pretty(p.cause),
+          part,
+        })
+      }
+
+      yield* sessions.updateMessage(info)
+      for (const part of parts) yield* sessions.updatePart(part)
+
+      return { info, parts }
+    }, Effect.scoped)
+
+    const prompt: (input: PromptInput) => Effect.Effect<SessionV1.WithParts, Image.Error> = Effect.fn(
+      "SessionPrompt.prompt",
+    )(function* (input: PromptInput) {
+      const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+      yield* revert.cleanup(session)
+      const message = yield* createUserMessage(input)
+      yield* sessions.touch(input.sessionID)
+
+      const permissions: PermissionV1.Rule[] = []
+      for (const [t, enabled] of Object.entries(input.tools ?? {})) {
+        permissions.push({ permission: t, action: enabled ? "allow" : "deny", pattern: "*" })
+      }
+      if (permissions.length > 0) {
+        session.permission = permissions
+        yield* sessions.setPermission({ sessionID: session.id, permission: permissions })
+      }
+
+      if (input.noReply === true) return message
+      return yield* loop({ sessionID: input.sessionID })
+    })
+
+    const lastAssistant = Effect.fnUntraced(function* (sessionID: SessionID) {
+      const match = yield* sessions.findMessage(sessionID, (m) => m.info.role !== "user").pipe(Effect.orDie)
+      if (Option.isSome(match)) return match.value
+      const msgs = yield* sessions.messages({ sessionID, limit: 1 }).pipe(Effect.orDie)
+      if (msgs.length > 0) return msgs[0]
+      throw new Error("Impossible")
+    })
+
+    const runLoop = Effect.fn("SessionPrompt.run")(
+      function* (sessionID: SessionID) {
+        const ctx = yield* InstanceState.context
+        const cfg = yield* config.get()
+        let structured: unknown
+        let step = 0
+        let modelOverride: { providerID: ProviderV2.ID; modelID: ModelV2.ID } | undefined
+        const session = yield* sessions.get(sessionID).pipe(Effect.orDie)
+
+        while (true) {
+          yield* status.set(sessionID, { type: "busy" })
+          yield* Effect.logInfo("loop", { "session.id": sessionID, step })
+
+          let msgs = yield* MessageV2.filterCompactedEffect(sessionID).pipe(
+            Effect.provideService(Database.Service, database),
+          )
+
+          const { user: lastUser, assistant: lastAssistant, finished: lastFinished, tasks } = MessageV2.latest(msgs)
+
+          if (!lastUser) throw new Error("No user message found in stream. This should never happen.")
+
+          const lastAssistantMsg = msgs.findLast(
+            (msg) => msg.info.role === "assistant" && msg.info.id === lastAssistant?.id,
+          )
+          // Some providers return "stop" even when the assistant message contains
+          // tool calls. Keep the loop running so tool results can be sent back to
+          // the model, but ignore cleanup-marked interrupted orphans.
+          const hasToolCalls =
+            lastAssistantMsg?.parts.some(
+              (part) => part.type === "tool" && !part.metadata?.providerExecuted && !isOrphanedInterruptedTool(part),
+            ) ?? false
+
+          if (
+            lastAssistant?.finish &&
+            !["tool-calls", "unknown"].includes(lastAssistant.finish) &&
+            !hasToolCalls &&
+            lastAssistant.parentID === lastUser.id
+          ) {
+            const orphan = lastAssistantMsg?.parts.find(
+              (part): part is SessionV1.ToolPart => part.type === "tool" && isOrphanedInterruptedTool(part),
+            )
+            if (orphan) {
+              yield* Effect.logWarning("loop exit with orphaned interrupted tool", {
+                "session.id": sessionID,
+                messageID: lastAssistant.id,
+                tool: orphan.tool,
+                callID: orphan.callID,
+              })
+            }
+            yield* Effect.logInfo("exiting loop", { "session.id": sessionID })
+            break
+          }
+
+          step++
+          if (step === 1) {
+            yield* title({
+              session,
+              modelID: lastUser.model.modelID,
+              providerID: lastUser.model.providerID,
+              history: msgs,
+            }).pipe(Effect.ignore, Effect.forkIn(scope))
+
+            // Auto-load skills based on user context
+            const lastUserMsg = msgs.findLast((m) => m.info.role === "user")
+            if (lastUserMsg) {
+              const userText = lastUserMsg.parts
+                .filter((p) => p.type === "text")
+                .map((p) => p.text)
+                .join(" ")
+              if (userText) {
+                const allSkills = yield* skill.available()
+                const detected = detectSkills({
+                  userMessage: userText,
+                  availableSkills: allSkills.map((s) => ({ name: s.name, description: s.description })),
+                })
+                // Tejamkorlik: sessiyada allaqachon yuklangan skill'ni qayta yuklamaslik.
+                const fresh = dedupeSessionSkills(sessionID, detected)
+                for (const skillName of fresh) {
+                  yield* skill.require(skillName).pipe(Effect.catchTag("Skill.NotFoundError", () => Effect.void))
+                }
+              }
+            }
+          }
+
+          const ref = modelOverride ?? { providerID: lastUser.model.providerID, modelID: lastUser.model.modelID }
+          const model = yield* getModel(ref.providerID, ref.modelID, sessionID)
+          
+          // Check rate limits before making LLM call.
+          // Soft per-minute caps wait; hard daily caps and oversized inputs block.
+          const providerConfig = yield* provider.getProvider(model.providerID).pipe(Effect.catch(() => Effect.succeed(undefined)))
+          const rateLimitConfig = providerConfig?.rateLimit
+          if (rateLimitConfig) {
+            yield* rateLimit.waitIfExceeded(model.providerID, rateLimitConfig)
+            const estimatedInput = Token.estimate(JSON.stringify(msgs))
+            const blocked = yield* rateLimit.enforce(model.providerID, rateLimitConfig, estimatedInput)
+            if (blocked) {
+              const error = new NamedError.Unknown({ message: blocked })
+              yield* events.publish(Session.Event.Error, { sessionID, error: error.toObject() })
+              throw error
+            }
+          }
+          
+          const task = tasks.pop()
+
+          if (task?.type === "subtask") {
+            yield* handleSubtask({ task, model, lastUser, sessionID, session, msgs })
+            continue
+          }
+
+          if (task?.type === "compaction") {
+            const result = yield* compaction.process({
+              messages: msgs,
+              parentID: lastUser.id,
+              sessionID,
+              auto: task.auto,
+              overflow: task.overflow,
+            })
+            if (result === "stop") break
+            continue
+          }
+
+          if (
+            lastFinished &&
+            lastFinished.summary !== true &&
+            (yield* compaction.isOverflow({ tokens: lastFinished.tokens, model }))
+          ) {
+            yield* compaction.create({ sessionID, agent: lastUser.agent, model: lastUser.model, auto: true })
+            continue
+          }
+
+          const agent = yield* agents.get(lastUser.agent)
+          if (!agent) {
+            const available = (yield* agents.list()).filter((a) => !a.hidden).map((a) => a.name)
+            const hint = available.length ? ` Available agents: ${available.join(", ")}` : ""
+            const error = new NamedError.Unknown({ message: `Agent not found: "${lastUser.agent}".${hint}` })
+            yield* events.publish(Session.Event.Error, { sessionID, error: error.toObject() })
+            throw error
+          }
+          const maxSteps = agent.steps ?? Infinity
+          const isLastStep = step >= maxSteps
+          msgs = yield* SessionReminders.apply({ messages: msgs, agent, session }).pipe(
+            Effect.provideService(RuntimeFlags.Service, flags),
+            Effect.provideService(FSUtil.Service, fsys),
+            Effect.provideService(Session.Service, sessions),
+          )
+
+          const msg: SessionV1.Assistant = {
+            id: MessageID.ascending(),
+            parentID: lastUser.id,
+            role: "assistant",
+            mode: agent.name,
+            agent: agent.name,
+            variant: lastUser.model.variant,
+            path: { cwd: ctx.directory, root: ctx.worktree },
+            cost: 0,
+            tokens: { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } },
+            modelID: model.id,
+            providerID: model.providerID,
+            time: { created: Date.now() },
+            sessionID,
+          }
+          yield* sessions.updateMessage(msg)
+
+          const finalizeInterruptedAssistant = Effect.gen(function* () {
+            if (msg.time.completed) return
+            msg.error ??= MessageV2.fromError(new DOMException("Aborted", "AbortError"), {
+              providerID: msg.providerID,
+              aborted: true,
+            })
+            msg.time.completed = Date.now()
+            yield* sessions.updateMessage(msg)
+          })
+
+          const handle = yield* processor
+            .create({
+              assistantMessage: msg,
+              sessionID,
+              model,
+            })
+            .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),
+              Effect.provideService(Hooks.Service, hooks),
+              Effect.provideService(SkillRestriction.Service, skillRestriction),
+            )
+
+            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 repomap = yield* repomapFromContext(fsys, ripgrep, ctx.directory).pipe(
+              Effect.orDie,
+              Effect.catchCause(() => Effect.succeed(undefined)),
+            )
+            const diffCtx = yield* getDiffContextImpl(git)(ctx.directory).pipe(
+              Effect.orDie,
+              Effect.catchCause(() => Effect.succeed(undefined)),
+            )
+            const system = [
+              ...env,
+              ...(repomap ? [repomap] : []),
+              ...(diffCtx ? [diffCtx] : []),
+              ...instructions,
+              ...(mcpInstructions ? [mcpInstructions] : []),
+              ...(skills ? [skills] : []),
+            ]
+            const format = lastUser.format ?? { type: "text" as const }
+            if (format.type === "json_schema") system.push(STRUCTURED_OUTPUT_SYSTEM_PROMPT)
+            // Pre-flight guard: the overflow check at loop start relies on the
+            // last turn's reported usage, which can be missing or stale. Measure
+            // the actual outgoing request and compact before the provider can
+            // reject it for exceeding the model's combined token limit.
+            const preflightInput = estimateInputTokens({ system, messages: modelMsgs, tools })
+            const preflightMaxOut = ProviderTransform.maxOutputTokens(model, flags.outputTokenMax)
+            if (model.limit.context > 0 && preflightInput + preflightMaxOut > model.limit.context) {
+              yield* Effect.logWarning("preflight overflow, compacting before send", {
+                "session.id": sessionID,
+                estimated: preflightInput + preflightMaxOut,
+                limit: model.limit.context,
+              })
+              yield* compaction.create({
+                sessionID,
+                agent: lastUser.agent,
+                model: lastUser.model,
+                auto: true,
+              })
+              return "continue" as const
+            }
+            const result = yield* handle.process({
+              user: lastUser,
+              agent,
+              permission: session.permission,
+              sessionID,
+              parentSessionID: session.parentID,
+              system,
+              messages: [
+                ...modelMsgs,
+                ...(isLastStep ? [{ role: "assistant" as const, content: MAX_STEPS_PROMPT }] : []),
+              ],
+              tools,
+              model,
+              toolChoice: format.type === "json_schema" ? "required" : undefined,
+            })
+
+            if (structured !== undefined) {
+              handle.message.structured = structured
+              handle.message.finish = handle.message.finish ?? "stop"
+              yield* sessions.updateMessage(handle.message)
+              return "break" as const
+            }
+
+            const finished = handle.message.finish && !["tool-calls", "unknown"].includes(handle.message.finish)
+            if (finished && !handle.message.error) {
+              // Surface any content-filter finish (e.g. Anthropic stop_reason:
+              // refusal) as an error. These turns may have produced no visible
+              // output at all — previously the session went idle silently — or
+              // partial text that was cut off by the provider's filter.
+              if (handle.message.finish === "content-filter") {
+                handle.message.error = new SessionV1.ContentFilterError({
+                  message: "The response was blocked by the provider's content filter",
+                }).toObject()
+                yield* sessions.updateMessage(handle.message)
+                yield* events.publish(Session.Event.Error, { sessionID, error: handle.message.error })
+                return "break" as const
+              }
+              if (format.type === "json_schema") {
+                handle.message.error = new SessionV1.StructuredOutputError({
+                  message: "Model did not produce structured output",
+                  retries: 0,
+                }).toObject()
+                yield* sessions.updateMessage(handle.message)
+                return "break" as const
+              }
+            }
+
+            if (result === "stop") return "break" as const
+            if (result === "compact") {
+              yield* compaction.create({
+                sessionID,
+                agent: lastUser.agent,
+                model: lastUser.model,
+                auto: true,
+                overflow: !handle.message.finish,
+              })
+            }
+            return "continue" as const
+          }).pipe(
+            Effect.ensuring(instruction.clear(handle.message.id)),
+            Effect.onInterrupt(() => finalizeInterruptedAssistant),
+          )
+          // Track token spend for daily budgets (tpd). Uses real usage when
+          // reported, falls back to estimating the assistant text.
+          if (rateLimitConfig) {
+            const t = handle.message.tokens
+            const spent =
+              t.input + t.output + t.reasoning > 0
+                ? t.input + t.output + t.reasoning
+                : Token.estimate(JSON.stringify(handle.message))
+            yield* rateLimit.record(model.providerID, rateLimitConfig, spent).pipe(Effect.ignore)
+          }
+          if (outcome === "break") {
+            const failed = handle.message.error
+            if (failed && isFallbackEligible(failed)) {
+              const primary = { providerID: lastUser.model.providerID, modelID: lastUser.model.modelID }
+              const chain = [primary]
+              for (const entry of cfg.model_fallback ?? []) {
+                const parsed = Provider.parseModel(entry)
+                if (!parsed.providerID || !parsed.modelID) continue
+                if (chain.some((m) => m.providerID === parsed.providerID && m.modelID === parsed.modelID)) continue
+                chain.push(parsed)
+              }
+              const active = modelOverride ?? primary
+              const index = chain.findIndex(
+                (m) => m.providerID === active.providerID && m.modelID === active.modelID,
+              )
+              for (const candidate of index === -1 ? [] : chain.slice(index + 1)) {
+                const resolved = yield* provider
+                  .getModel(candidate.providerID, candidate.modelID)
+                  .pipe(Effect.catchIf(Provider.ModelNotFoundError.isInstance, () => Effect.succeed(undefined)))
+                if (!resolved) continue
+                modelOverride = { providerID: resolved.providerID, modelID: resolved.id }
+                yield* sessions
+                  .setAgentModel({
+                    sessionID,
+                    agent: lastUser.agent,
+                    model: {
+                      id: resolved.id,
+                      providerID: resolved.providerID,
+                      variant: lastUser.model.variant ?? "default",
+                    },
+                    time: Date.now(),
+                  })
+                  .pipe(Effect.orDie)
+                yield* Effect.logWarning("model fallback", {
+                  "session.id": sessionID,
+                  from: `${active.providerID}/${active.modelID}`,
+                  to: `${resolved.providerID}/${resolved.id}`,
+                })
+                break
+              }
+              if (modelOverride && (modelOverride.providerID !== active.providerID || modelOverride.modelID !== active.modelID)) {
+                continue
+              }
+            }
+            break
+          }
+
+          const gitCfg = cfg.git
+          if (gitCfg?.auto_commit) {
+            const lastUserMsg = msgs.findLast((m) => m.info.role === "user")
+            const summary = lastUserMsg
+              ? lastUserMsg.parts
+                  .filter((p): p is SessionV1.TextPart => p.type === "text")
+                  .map((p) => p.text)
+                  .join(" ")
+                  .slice(0, 72)
+              : ""
+            yield* gitAutoCommit
+              .commit({
+                worktree: ctx.worktree,
+                summary,
+              })
+              .pipe(Effect.ignore, Effect.forkIn(scope))
+          }
+          continue
+        }
+
+        yield* compaction.prune({ sessionID }).pipe(Effect.ignore, Effect.forkIn(scope))
+        return yield* lastAssistant(sessionID)
+      },
+    )
+
+    const loop: (input: LoopInput) => Effect.Effect<SessionV1.WithParts> = Effect.fn("SessionPrompt.loop")(function* (
+      input: LoopInput,
+    ) {
+      return yield* state.ensureRunning(input.sessionID, lastAssistant(input.sessionID), runLoop(input.sessionID))
+    })
+
+    const shell: (input: ShellInput) => Effect.Effect<SessionV1.WithParts, Session.BusyError> = Effect.fn(
+      "SessionPrompt.shell",
+    )(function* (input: ShellInput) {
+      const ready = yield* Latch.make()
+      return yield* state.startShell(input.sessionID, lastAssistant(input.sessionID), shellImpl(input, ready), ready)
+    })
+
+    const pinCommand = Effect.fn("SessionPrompt.pinCommand")(function* (input: CommandInput) {
+      const ctx = yield* InstanceState.context
+      const raw = input.arguments.match(argsRegex) ?? []
+      const args = raw.map((arg) => arg.replace(quoteTrimRegex, ""))
+      const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+      const current = readPinned(session.metadata)
+      const resolveArg = (arg: string) =>
+        path.isAbsolute(arg) ? path.normalize(arg) : path.resolve(ctx.worktree, arg)
+      const rel = (file: string) => path.relative(ctx.worktree, file) || file
+
+      const next = [...current]
+      const changed: string[] = []
+      const missing: string[] = []
+      if (input.command === "add") {
+        for (const arg of args) {
+          const file = resolveArg(arg)
+          const exists = yield* fsys.stat(file).pipe(Effect.option)
+          if (Option.isNone(exists)) {
+            missing.push(arg)
+            continue
+          }
+          if (next.includes(file)) continue
+          next.push(file)
+          changed.push(file)
+        }
+      } else {
+        for (const arg of args) {
+          const file = resolveArg(arg)
+          const index = next.findIndex((item) => item === file || item.endsWith(`/${arg}`))
+          if (index === -1) {
+            missing.push(arg)
+            continue
+          }
+          changed.push(next[index]!)
+          next.splice(index, 1)
+        }
+      }
+      yield* sessions
+        .setMetadata({ sessionID: input.sessionID, metadata: { ...(session.metadata ?? {}), [PINNED_KEY]: next } })
+        .pipe(Effect.orDie)
+
+      const lines: string[] = []
+      if (input.command === "add") {
+        for (const file of changed) lines.push(`Pinned ${rel(file)} — included in every reply context.`)
+        for (const arg of missing) lines.push(`Not found: ${arg}`)
+      } else {
+        for (const file of changed) lines.push(`Unpinned ${rel(file)}.`)
+        for (const arg of missing) lines.push(`Not pinned: ${arg}`)
+      }
+      if (next.length) {
+        lines.push("", "Pinned files:")
+        for (const file of next) lines.push(`- ${rel(file)}`)
+      } else {
+        lines.push("", "No pinned files. Use /add <file...> to pin files into context.")
+      }
+      const result = yield* prompt({
+        sessionID: input.sessionID,
+        agent: input.agent,
+        parts: [{ type: "text", text: lines.join("\n") }],
+        noReply: true,
+      })
+      yield* events.publish(Command.Event.Executed, {
+        name: input.command,
+        sessionID: input.sessionID,
+        arguments: input.arguments,
+        messageID: result.info.id,
+      })
+      return result
+    })
+
+    const commitCommand = Effect.fn("SessionPrompt.commitCommand")(function* (input: CommandInput) {
+      const ctx = yield* InstanceState.context
+      const worktree = ctx.worktree
+      const confirm = (text: string) =>
+        prompt({
+          sessionID: input.sessionID,
+          agent: input.agent,
+          parts: [{ type: "text", text }],
+          noReply: true,
+        }).pipe(
+          Effect.tap((result) =>
+            events.publish(Command.Event.Executed, {
+              name: input.command,
+              sessionID: input.sessionID,
+              arguments: input.arguments,
+              messageID: result.info.id,
+            }),
+          ),
+        )
+      const pending = yield* gitAutoCommit.pending(worktree)
+      if (pending.length === 0) {
+        const inside = yield* fsys.stat(path.join(worktree, ".git")).pipe(
+          Effect.map(() => true),
+          Effect.catch(() => Effect.succeed(false)),
+        )
+        return yield* confirm(
+          inside ? "Nothing to commit — working tree is clean." : "Cannot commit: not a git repository.",
+        )
+      }
+      const diff = yield* gitAutoCommit.patch(worktree)
+      const files = pending.map((s) => `${s.code} ${s.file}`).join("\n")
+      const current = yield* currentModel(input.sessionID)
+      const mdl =
+        (yield* provider.getSmallModel(current.providerID)) ??
+        (yield* provider
+          .getModel(current.providerID, current.modelID)
+          .pipe(Effect.catchIf(Provider.ModelNotFoundError.isInstance, () => Effect.succeed(undefined))))
+      let message = ""
+      const ag = yield* agents.defaultInfo()
+      if (mdl && ag) {
+        const user: SessionV1.User = {
+          id: MessageID.ascending(),
+          role: "user",
+          sessionID: input.sessionID,
+          time: { created: Date.now() },
+          agent: ag.name,
+          model: { providerID: mdl.providerID, modelID: mdl.id },
+        }
+        const raw = yield* llm
+          .stream({
+            agent: ag,
+            user,
+            system: [],
+            small: true,
+            tools: {},
+            model: mdl,
+            sessionID: input.sessionID,
+            retries: 2,
+            messages: [
+              {
+                role: "user",
+                content: `Write a single conventional commit message (one line like "feat(scope): summary", no body, no quotes, no markdown) for these changes:\n\n${files}\n\n${diff.slice(0, 12000)}`,
+              },
+            ],
+          })
+          .pipe(
+            Stream.filter(LLMEvent.is.textDelta),
+            Stream.map((e) => e.text),
+            Stream.mkString,
+            Effect.catchCause(() => Effect.succeed("")),
+          )
+        message =
+          raw
+            .split("\n")
+            .map((line) => line.trim().replace(/^["'`]|["'`]$/g, ""))
+            .find((line) => line.length > 0) ?? ""
+      }
+      if (!message) {
+        message = `feat: update ${pending.length} file(s): ${pending
+          .slice(0, 3)
+          .map((s) => s.file)
+          .join(", ")}`
+      }
+      const error = yield* gitAutoCommit.commitAll({ worktree, message })
+      if (error) {
+        return yield* confirm(`Commit failed: ${error}`)
+      }
+      return yield* confirm(`Committed ${pending.length} file(s):\n${message}`)
+    })
+
+    const doctorCommand = Effect.fn("SessionPrompt.doctorCommand")(function* (input: CommandInput) {
+      const ctx = yield* InstanceState.context
+      const cfg = yield* config.get()
+      const lines = ["Mervex doctor", ""]
+      const check = (label: string, ok: boolean, detail?: string) =>
+        lines.push(`${ok ? "ok" : "FAIL"} - ${label}${detail ? ` (${detail})` : ""}`)
+      const resolveModel = (ref: string) => {
+        const parsed = Provider.parseModel(ref)
+        return provider
+          .getModel(parsed.providerID, parsed.modelID)
+          .pipe(Effect.catchIf(Provider.ModelNotFoundError.isInstance, () => Effect.succeed(undefined)))
+      }
+
+      if (cfg.model) {
+        const mdl = yield* resolveModel(cfg.model)
+        check("model", !!mdl, cfg.model)
+      } else {
+        check("model", false, "not configured")
+      }
+      if (cfg.small_model) {
+        const mdl = yield* resolveModel(cfg.small_model)
+        check("small model", !!mdl, cfg.small_model)
+      }
+      for (const entry of cfg.model_fallback ?? []) {
+        const mdl = yield* resolveModel(entry)
+        check("fallback", !!mdl, entry)
+      }
+      const providers = yield* provider.list().pipe(Effect.catch(() => Effect.succeed({})))
+      const ids = Object.keys(providers)
+      check("providers", ids.length > 0, ids.length ? ids.join(", ") : "none configured")
+
+      const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+      check("pinned files", true, `${readPinned(session.metadata).length} pinned`)
+      check("snapshots", cfg.snapshot !== false)
+      check("directory", true, ctx.worktree)
+      const isRepo = yield* fsys
+        .stat(path.join(ctx.worktree, ".git"))
+        .pipe(
+          Effect.map(() => true),
+          Effect.catch(() => Effect.succeed(false)),
+        )
+      check("git repo", isRepo)
+
+      const result = yield* prompt({
+        sessionID: input.sessionID,
+        agent: input.agent,
+        parts: [{ type: "text", text: lines.join("\n") }],
+        noReply: true,
+      })
+      yield* events.publish(Command.Event.Executed, {
+        name: input.command,
+        sessionID: input.sessionID,
+        arguments: input.arguments,
+        messageID: result.info.id,
+      })
+      return result
+    })
+
+    const usageCommand = Effect.fn("SessionPrompt.usageCommand")(function* (input: CommandInput) {
+      const session = yield* sessions.get(input.sessionID).pipe(Effect.orDie)
+      const all = yield* sessions.messages({ sessionID: input.sessionID }).pipe(Effect.orDie)
+      const assistants = all.filter((m) => m.info.role === "assistant")
+      const users = all.filter((m) => m.info.role === "user")
+      const tokens = { input: 0, output: 0, reasoning: 0 }
+      for (const m of all) {
+        if (m.info.role !== "assistant") continue
+        const t = m.info.tokens
+        if (!t) continue
+        tokens.input += t.input ?? 0
+        tokens.output += t.output ?? 0
+        tokens.reasoning += t.reasoning ?? 0
+      }
+      const diffs = yield* sessions.diff(input.sessionID).pipe(Effect.catch(() => Effect.succeed([])))
+      const formatTokens = (n: number) => {
+        if (n >= 1000000) return `${(n / 1000000).toFixed(1)}M`
+        if (n >= 1000) return `${(n / 1000).toFixed(1)}K`
+        return `${n}`
+      }
+      const formatDuration = (ms: number) => {
+        if (ms < 60000) return `${Math.floor(ms / 1000)}s`
+        if (ms < 3600000) return `${Math.floor(ms / 60000)}m ${Math.floor((ms % 60000) / 1000)}s`
+        return `${Math.floor(ms / 3600000)}h ${Math.floor((ms % 3600000) / 60000)}m`
+      }
+      const lines = [
+        "Session usage",
+        "",
+        `Messages  ${users.length} user / ${assistants.length} assistant`,
+        `Tokens    ${formatTokens(tokens.input)} in / ${formatTokens(tokens.output)} out${tokens.reasoning ? ` / ${formatTokens(tokens.reasoning)} reasoning` : ""}`,
+      ]
+      if ((session.cost ?? 0) > 0) lines.push(`Cost      $${(session.cost ?? 0).toFixed(4)}`)
+      const created = session.time.created
+      const updated = session.time.updated
+      if (created && updated && updated > created) lines.push(`Duration  ${formatDuration(updated - created)}`)
+      lines.push(`Files     ${diffs.length} changed`)
+      const pinned = readPinned(session.metadata)
+      if (pinned.length) {
+        lines.push(`Pinned    ${pinned.length} file(s) in context`)
+      }
+      const result = yield* prompt({
+        sessionID: input.sessionID,
+        agent: input.agent,
+        parts: [{ type: "text", text: lines.join("\n") }],
+        noReply: true,
+      })
+      yield* events.publish(Command.Event.Executed, {
+        name: input.command,
+        sessionID: input.sessionID,
+        arguments: input.arguments,
+        messageID: result.info.id,
+      })
+      return result
+    })
+
+    const command = Effect.fn("SessionPrompt.command")(function* (input: CommandInput) {
+      if (input.command === "add" || input.command === "drop") return yield* pinCommand(input)
+      if (input.command === "commit") return yield* commitCommand(input)
+      if (input.command === "doctor") return yield* doctorCommand(input)
+      if (input.command === "usage") return yield* usageCommand(input)
+      yield* Effect.logInfo("command", {
+        "session.id": input.sessionID,
+        command: input.command,
+        agent: input.agent,
+      })
+      const cmd = yield* commands.get(input.command)
+      if (!cmd) {
+        const available = (yield* commands.list()).map((c) => c.name)
+        const hint = available.length ? ` Available commands: ${available.join(", ")}` : ""
+        const error = new NamedError.Unknown({ message: `Command not found: "${input.command}".${hint}` })
+        yield* events.publish(Session.Event.Error, { sessionID: input.sessionID, error: error.toObject() })
+        throw error
+      }
+      const agentName = cmd.agent ?? input.agent
+
+      const raw = input.arguments.match(argsRegex) ?? []
+      const args = raw.map((arg) => arg.replace(quoteTrimRegex, ""))
+      const templateCommand = yield* Effect.promise(async () => cmd.template)
+
+      const placeholders = templateCommand.match(placeholderRegex) ?? []
+      let last = 0
+      for (const item of placeholders) {
+        const value = Number(item.slice(1))
+        if (value > last) last = value
+      }
+
+      const withArgs = templateCommand.replaceAll(placeholderRegex, (_, index) => {
+        const position = Number(index)
+        const argIndex = position - 1
+        if (argIndex >= args.length) return ""
+        if (position === last) return args.slice(argIndex).join(" ")
+        return args[argIndex]
+      })
+      const usesArgumentsPlaceholder = templateCommand.includes("$ARGUMENTS")
+      let template = withArgs.replaceAll("$ARGUMENTS", input.arguments)
+
+      if (placeholders.length === 0 && !usesArgumentsPlaceholder && input.arguments.trim()) {
+        template = template + "\n\n" + input.arguments
+      }
+
+      const shellMatches = ConfigMarkdown.shell(template)
+      if (shellMatches.length > 0) {
+        const cfg = yield* config.get()
+        const sh = Shell.preferred(cfg.shell)
+        const results = yield* Effect.promise(() =>
+          Promise.all(
+            shellMatches.map(async ([, cmd]) => (await Process.text([cmd], { shell: sh, nothrow: true })).text),
+          ),
+        )
+        let index = 0
+        template = template.replace(bashRegex, () => results[index++])
+      }
+      template = template.trim()
+
+      const taskModel = yield* Effect.gen(function* () {
+        if (cmd.model) return Provider.parseModel(cmd.model)
+        if (cmd.agent) {
+          const cmdAgent = yield* agents.get(cmd.agent)
+          if (cmdAgent?.model) return cmdAgent.model
+        }
+        if (input.model) return Provider.parseModel(input.model)
+        return yield* currentModel(input.sessionID)
+      })
+
+      yield* getModel(taskModel.providerID, taskModel.modelID, input.sessionID)
+
+      const agent = agentName ? yield* agents.get(agentName) : yield* agents.defaultInfo()
+      if (!agent) {
+        const available = (yield* agents.list()).filter((a) => !a.hidden).map((a) => a.name)
+        const hint = available.length ? ` Available agents: ${available.join(", ")}` : ""
+        const error = new NamedError.Unknown({ message: `Agent not found: "${agentName}".${hint}` })
+        yield* events.publish(Session.Event.Error, { sessionID: input.sessionID, error: error.toObject() })
+        throw error
+      }
+
+      const templateParts = yield* resolvePromptParts(template)
+      const inputFiles = new Set(
+        input.parts?.filter((part) => new URL(part.url).protocol === "file:").map((part) => fileURLToPath(part.url)),
+      )
+      const uniqueTemplateParts = templateParts.filter(
+        (part) => part.type !== "file" || !inputFiles.has(fileURLToPath(part.url)),
+      )
+      const isSubtask = (agent.mode === "subagent" && cmd.subtask !== false) || cmd.subtask === true
+      const parts = isSubtask
+        ? [
+            {
+              type: "subtask" as const,
+              agent: agent.name,
+              description: cmd.description ?? "",
+              command: input.command,
+              model: { providerID: taskModel.providerID, modelID: taskModel.modelID },
+              prompt: templateParts.find((y) => y.type === "text")?.text ?? "",
+            },
+          ]
+        : [...uniqueTemplateParts, ...(input.parts ?? [])]
+
+      const userAgent = isSubtask ? (input.agent ?? (yield* agents.defaultInfo()).name) : agent.name
+      const userModel = isSubtask
+        ? input.model
+          ? Provider.parseModel(input.model)
+          : yield* currentModel(input.sessionID)
+        : taskModel
+
+      yield* plugin.trigger(
+        "command.execute.before",
+        { command: input.command, sessionID: input.sessionID, arguments: input.arguments },
+        { parts },
+      )
+
+      const result = yield* prompt({
+        sessionID: input.sessionID,
+        messageID: input.messageID,
+        model: userModel,
+        agent: userAgent,
+        parts,
+        variant: input.variant,
+      })
+      yield* events.publish(Command.Event.Executed, {
+        name: input.command,
+        sessionID: input.sessionID,
+        arguments: input.arguments,
+        messageID: result.info.id,
+      })
+      return result
+    })
+
+    return Service.of({
+      cancel,
+      prompt,
+      loop,
+      shell,
+      command,
+      resolvePromptParts,
+    })
+  }),
+)
+
+const ModelRef = Schema.Struct({
+  providerID: ProviderV2.ID,
+  modelID: ModelV2.ID,
+})
+
+export const PromptInput = Schema.Struct({
+  sessionID: SessionID,
+  messageID: Schema.optional(MessageID),
+  model: Schema.optional(ModelRef),
+  agent: Schema.optional(Schema.String),
+  noReply: Schema.optional(Schema.Boolean),
+  tools: Schema.optional(Schema.Record(Schema.String, Schema.Boolean)).annotate({
+    description:
+      "@deprecated tools and permissions have been merged, you can set permissions on the session itself now",
+  }),
+  format: Schema.optional(SessionV1.Format),
+  system: Schema.optional(Schema.String),
+  variant: Schema.optional(Schema.String),
+  parts: Schema.Array(
+    Schema.Union([
+      SessionV1.TextPartInput,
+      SessionV1.FilePartInput,
+      SessionV1.AgentPartInput,
+      SessionV1.SubtaskPartInput,
+    ]).annotate({ discriminator: "type" }),
+  ),
+})
+export type PromptInput = Schema.Schema.Type<typeof PromptInput>
+
+export class LoopInput extends Schema.Class<LoopInput>("SessionPrompt.LoopInput")({
+  sessionID: SessionID,
+}) {}
+
+export const ShellInput = Schema.Struct({
+  sessionID: SessionID,
+  messageID: Schema.optional(MessageID),
+  agent: Schema.String,
+  model: Schema.optional(ModelRef),
+  command: Schema.String,
+})
+export type ShellInput = Schema.Schema.Type<typeof ShellInput>
+
+export const CommandInput = Schema.Struct({
+  messageID: Schema.optional(MessageID),
+  sessionID: SessionID,
+  agent: Schema.optional(Schema.String),
+  model: Schema.optional(Schema.String),
+  arguments: Schema.String,
+  command: Schema.String,
+  variant: Schema.optional(Schema.String),
+  // Inlined (no identifier annotation) to keep the original SDK output — the
+  // PromptInput call site below references FilePartInput by ref via the
+  // Schema export in message-v2.ts.
+  parts: Schema.optional(
+    Schema.Array(
+      Schema.Union([
+        Schema.Struct({
+          id: Schema.optional(PartID),
+          type: Schema.Literal("file"),
+          mime: Schema.String,
+          filename: Schema.optional(Schema.String),
+          url: Schema.String,
+          source: Schema.optional(SessionV1.FilePartSource),
+        }),
+      ]).annotate({ discriminator: "type" }),
+    ),
+  ),
+})
+export type CommandInput = Schema.Schema.Type<typeof CommandInput>
+
+/** @internal Exported for testing */
+export function createStructuredOutputTool(input: {
+  schema: Record<string, any>
+  onSuccess: (output: unknown) => void
+}): AITool {
+  // Remove $schema property if present (not needed for tool input)
+  const { $schema: _, ...toolSchema } = input.schema
+
+  return tool({
+    description: STRUCTURED_OUTPUT_DESCRIPTION,
+    inputSchema: jsonSchema(toolSchema as JSONSchema7),
+    async execute(args) {
+      // AI SDK validates args against inputSchema before calling execute()
+      input.onSuccess(args)
+      return {
+        output: "Structured output captured successfully.",
+        title: "Structured Output",
+        metadata: { valid: true },
+      }
+    },
+    toModelOutput({ output }) {
+      return {
+        type: "text",
+        value: output.output,
+      }
+    },
+  })
+}
+const bashRegex = /!`([^`]+)`/g
+// Match [Image N] as single token, quoted strings, or non-space sequences
+const argsRegex = /(?:\[Image\s+\d+\]|"[^"]*"|'[^']*'|[^\s"']+)/gi
+const placeholderRegex = /\$(\d+)/g
+const quoteTrimRegex = /^["']|["']$/g
+
+export const node = LayerNode.make({
+  service: Service,
+  layer: layer,
+  deps: [
+    SessionStatus.node,
+    Session.node,
+    Agent.node,
+    Provider.node,
+    SessionProcessor.node,
+    SessionCompaction.node,
+    Plugin.node,
+    Command.node,
+    Config.node,
+    Permission.node,
+    FSUtil.node,
+    Ripgrep.node,
+    MCP.node,
+    LSP.node,
+    ToolRegistry.node,
+    Truncate.node,
+    Image.node,
+    CrossSpawnSpawner.node,
+    Instruction.node,
+    SessionRunState.node,
+    SessionRevert.node,
+    SessionSummary.node,
+    SystemPrompt.node,
+    LLM.node,
+    EventV2Bridge.node,
+    RuntimeFlags.node,
+    Database.node,
+    GitAutoCommit.node,
+    Hooks.node,
+    SkillRestriction.node,
+    Skill.node,
+    Git.node,
+    AppProcess.node,
+    RateLimit.node,
+  ],
+})
+
+export * as SessionPrompt from "./prompt"