fix(cli): resume subagents after session fork

This commit is contained in:
marius-kilocode
2026-08-06 11:16:17 +02:00
parent b5b022fda5
commit 4ea52f2dd1
6 changed files with 309 additions and 81 deletions
+5
View File
@@ -0,0 +1,5 @@
---
"@kilocode/cli": patch
---
Allow subagent tasks to be resumed after their parent session is forked.
+172 -49
View File
@@ -1,41 +1,180 @@
import { Effect } from "effect"
import type { Session } from "@/session/session"
import { MessageV2 } from "@/session/message-v2"
import { MessageID, PartID, SessionID } from "@/session/schema"
import { KiloPartLifecycle } from "./part-lifecycle"
const task = "task"
const stale = /^[ \t]*task_id:[^\r\n]*(?:(?:\r?\n){1,2}|$)/m
type Ops = Pick<Session.Interface, "get" | "messages" | "create" | "updateMessage" | "updatePart">
// Prepare a source part for a forked transcript copy: drop transient parts (returns undefined) and detach
// task calls into historical results. The caller assigns fresh ids and publishes via Session.updatePart.
// Keep terminal task references so the fork can give them private child sessions.
// In-flight jobs cannot be copied safely, so those remain historical errors.
export function prepareForkedPart(part: MessageV2.Part): MessageV2.Part | undefined {
if (KiloPartLifecycle.transient(part)) return undefined
return structuredClone(detachPart(part))
if (
part.type === "tool" &&
part.tool === task &&
(part.state.status === "pending" || part.state.status === "running")
) {
return structuredClone(detachPart(part))
}
return structuredClone(part)
}
function metadata(value: Record<string, unknown> | undefined) {
function childID(part: MessageV2.Part) {
if (part.type !== "tool" || part.tool !== task) return undefined
const state = part.state
const metadata = state.status === "pending" ? undefined : state.metadata
const values = [
metadata?.sessionId,
metadata?.sessionID,
part.metadata?.sessionId,
part.metadata?.sessionID,
state.input.task_id,
]
return values.find((value): value is string => typeof value === "string")
}
function mapRecord(value: Record<string, unknown> | undefined, map: Map<string, SessionID>, keys: string[]) {
if (!value) return value
const copy = { ...value }
delete copy.sessionId
delete copy.sessionID
for (const key of keys) {
const id = copy[key]
if (typeof id !== "string") continue
const replacement = map.get(id)
if (replacement) copy[key] = replacement
}
return copy
}
function input(value: Record<string, unknown>) {
const copy = { ...value }
delete copy.task_id
return copy
function output(value: string, map: Map<string, SessionID>) {
return [...map].reduce(
(text, [source, target]) =>
source === target
? text
: text
.replaceAll(source, target)
.replaceAll(`task id="${source}"`, `task id="${target}"`)
.replaceAll(`task_id="${source}"`, `task_id="${target}"`)
.replaceAll(`task_id: ${source}`, `task_id: ${target}`),
value,
)
}
function resumeHint(sessionID: string) {
return [
`This subagent session can be resumed: call the task tool again with task_id="${sessionID}"`,
"and a prompt describing how to continue or recover. Its prior context is preserved.",
].join(" ")
}
function remapPart(part: MessageV2.Part, map: Map<string, SessionID>) {
if (part.type === "text") {
const text = output(part.text, map)
return text === part.text ? part : { ...part, text }
}
if (part.type !== "tool" || part.tool !== task) return part
const next = structuredClone(part)
next.metadata = mapRecord(next.metadata, map, ["sessionId", "sessionID", "parentSessionId", "parentSessionID"])
const input = mapRecord(next.state.input, map, ["task_id"])
if (input) next.state.input = input
if (next.state.status !== "pending") {
const metadata = mapRecord(next.state.metadata, map, [
"sessionId",
"sessionID",
"parentSessionId",
"parentSessionID",
])
if (metadata) next.state.metadata = metadata
}
if (next.state.status === "completed") next.state.output = output(next.state.output, map)
if (next.state.status === "error") next.state.error = output(next.state.error, map)
return next
}
function copy(input: { source: Session.Info; parentID: SessionID; ops: Ops }) {
return Effect.gen(function* () {
const target = yield* input.ops.create({
parentID: input.parentID,
title: input.source.title,
agent: input.source.agent,
model: input.source.model,
metadata: structuredClone(input.source.metadata),
permission: input.source.permission ? [...input.source.permission] : undefined,
workspaceID: input.source.workspaceID,
})
const msgs = yield* input.ops.messages({ sessionID: input.source.id })
const ids = new Map<string, MessageID>()
for (const msg of msgs) {
const id = MessageID.ascending()
ids.set(msg.info.id, id)
const parentID = msg.info.role === "assistant" ? ids.get(msg.info.parentID) : undefined
const cloned = yield* input.ops.updateMessage({
...msg.info,
id,
sessionID: target.id,
...(msg.info.role === "assistant" && { cost: 0 }),
...(parentID && { parentID }),
})
for (const part of msg.parts) {
const prepared = prepareForkedPart(part)
if (!prepared) continue
const next: MessageV2.Part = {
...prepared,
id: PartID.ascending(),
messageID: cloned.id,
sessionID: target.id,
...(prepared.type === "step-finish" && { cost: 0 }),
}
if (next.type === "compaction" && next.tail_start_id) next.tail_start_id = ids.get(next.tail_start_id)
yield* input.ops.updatePart(next)
}
}
return target
})
}
export function remapChildren(input: {
sessionID: SessionID
ops: Ops
remapped?: Map<string, SessionID>
}): Effect.Effect<void, Session.NotFound> {
return Effect.gen(function* () {
const map = input.remapped ?? new Map<string, SessionID>()
const msgs = yield* input.ops.messages({ sessionID: input.sessionID })
const refs = msgs.flatMap((msg) =>
msg.parts.flatMap((part) => {
const child = childID(part)
return child ? [{ part, child }] : []
}),
)
for (const ref of refs) {
if (map.has(ref.child)) continue
const source = yield* input.ops.get(SessionID.make(ref.child)).pipe(Effect.orElseSucceed(() => undefined))
if (!source) continue
const target = yield* copy({ source, parentID: input.sessionID, ops: input.ops })
map.set(ref.child, target.id)
yield* remapChildren({ sessionID: target.id, ops: input.ops, remapped: map })
}
for (const msg of msgs) {
for (const part of msg.parts) {
const next = remapPart(part, map)
if (next !== part) yield* input.ops.updatePart(next)
}
}
})
}
/**
* Turns copied task calls into detached historical results.
*
* Child sessions are execution state, not conversation context. Their final
* result is already embedded in the parent task part, so a fork keeps that
* result while dropping references that could resume, stream, or route prompts
* to a child owned by the source session.
*/
function detachPart(part: MessageV2.Part): MessageV2.Part {
if (part.type !== "tool" || part.tool !== task) return part
const child = childID(part)
const hint = child ? `\n${resumeHint(child)}` : ""
const top = metadata(part.metadata)
const state = part.state
if (state.status === "pending") {
@@ -46,46 +185,30 @@ function detachPart(part: MessageV2.Part): MessageV2.Part {
state: {
status: "error",
input: input(state.input),
error: "Task was still pending when this session was forked.",
error: `Task was still pending when this session was forked.${hint}`,
time: { start: now, end: now },
},
}
}
if (state.status === "running") {
return {
...part,
metadata: top,
state: {
status: "error",
input: input(state.input),
error: "Task was still running when this session was forked.",
metadata: metadata(state.metadata),
time: { start: state.time.start, end: Date.now() },
},
}
}
if (state.status === "error") {
return {
...part,
metadata: top,
state: {
...state,
input: input(state.input),
metadata: metadata(state.metadata),
},
}
}
return {
...part,
metadata: top,
state: {
...state,
status: "error",
input: input(state.input),
output: state.output.replace(stale, ""),
metadata: metadata(state.metadata) ?? {},
error: `Task was still running when this session was forked.${hint}`,
metadata: metadata(state.metadata),
time: { start: state.time.start, end: Date.now() },
},
}
}
function metadata(value: Record<string, unknown> | undefined) {
if (!value) return value
return { ...value }
}
function input(value: Record<string, unknown>) {
return { ...value }
}
@@ -1,4 +1,4 @@
import { prepareForkedPart as _prepareForkedPart } from "./fork"
import { prepareForkedPart as _prepareForkedPart, remapChildren as _remapChildren } from "./fork"
import z from "zod"
import { Cause, Effect, Schema } from "effect"
import { Bus } from "@/bus"
@@ -437,6 +437,7 @@ export namespace KiloSession {
}
export const prepareForkedPart = _prepareForkedPart
export const remapChildren = _remapChildren
}
export { kiloSessionFork } from "./fork-command"
+7
View File
@@ -906,6 +906,13 @@ export const layer: Layer.Layer<
}
// kilocode_change - preserve imported/cumulative diffs when forking (self-contained Storage runtime keeps this shared file off the legacy Storage layer)
yield* carryForkDiff(input.sessionID, session.id)
// kilocode_change start - fork terminal task children under the new parent and remap their references
yield* KiloSession.remapChildren({
sessionID: session.id,
remapped: new Map([[input.sessionID, session.id]]),
ops: { get, messages, create, updateMessage, updatePart },
})
// kilocode_change end
return session
})
@@ -49,15 +49,12 @@ afterAll(async () => {
})
const sessions = {
create: (input?: Parameters<Session.Interface["create"]>[0]) =>
runtime.runPromise((svc) => svc.create(input)),
create: (input?: Parameters<Session.Interface["create"]>[0]) => runtime.runPromise((svc) => svc.create(input)),
get: (id: SessionID) => runtime.runPromise((svc) => svc.get(id)),
list: () => runtime.runPromise((svc) => svc.list()),
messages: (input: Parameters<Session.Interface["messages"]>[0]) =>
runtime.runPromise((svc) => svc.messages(input)),
updateMessage: <T extends MessageV2.Info>(msg: T) =>
runtime.runPromise((svc) => svc.updateMessage(msg)),
updatePart: <T extends MessageV2.Part>(part: T) =>
runtime.runPromise((svc) => svc.updatePart(part)),
messages: (input: Parameters<Session.Interface["messages"]>[0]) => runtime.runPromise((svc) => svc.messages(input)),
updateMessage: <T extends MessageV2.Info>(msg: T) => runtime.runPromise((svc) => svc.updateMessage(msg)),
updatePart: <T extends MessageV2.Part>(part: T) => runtime.runPromise((svc) => svc.updatePart(part)),
}
afterEach(async () => {
@@ -190,9 +187,9 @@ describe("Session.fork cost accounting", () => {
)
})
describe("Session.fork task detachment", () => {
describe("Session.fork task children", () => {
test(
"keeps completed task outcomes without cloning child sessions",
"clones completed task children under the forked parent",
async () => {
await using tmp = await tmpdir({ git: true })
await instance({
@@ -212,6 +209,13 @@ describe("Session.fork task detachment", () => {
const user = await userMsg(parent.id)
const assistant = await asstMsg(parent.id, user)
await sessions.updatePart(taskPart({ messageID: assistant, sessionID: parent.id, childSessionID: child.id }))
await sessions.updatePart({
id: PartID.ascending(),
messageID: assistant,
sessionID: parent.id,
type: "text",
text: `Subagent task ID: ${child.id}`,
} as MessageV2.TextPart)
const before = await sessions.list()
const server = HttpRouter.toWebHandler(HttpApiApp.routes, { disableLogger: true })
@@ -220,23 +224,36 @@ describe("Session.fork task detachment", () => {
directory: tmp.path,
fetch: ((request: Request) => server.handler(request, HttpApiApp.context)) as unknown as typeof fetch,
})
const { data: forked } = await client.session.fork(
{ sessionID: parent.id, directory: tmp.path },
{ throwOnError: true },
).finally(() => server.dispose())
const { data: forked } = await client.session
.fork({ sessionID: parent.id, directory: tmp.path }, { throwOnError: true })
.finally(() => server.dispose())
const after = await sessions.list()
expect(after).toHaveLength(before.length + 1)
expect(after).toHaveLength(before.length + 2)
const msgs = await sessions.messages({ sessionID: SessionID.make(forked.id) })
const tool = msgs.flatMap((msg) => msg.parts).find((part) => part.type === "tool") as MessageV2.ToolPart
expect(tool.state.status).toBe("completed")
if (tool.state.status !== "completed") throw new Error("expected completed task")
expect(tool.metadata).toEqual({ trace: "keep" })
expect(tool.state.metadata).toEqual({ model: { modelID: "test", providerID: "test" } })
expect(tool.state.input.task_id).toBeUndefined()
const clonedID = tool.state.input.task_id
expect(clonedID).not.toBe(child.id)
expect(tool.metadata).toEqual({ sessionId: clonedID, trace: "keep" })
expect(tool.state.metadata).toEqual({
sessionId: clonedID,
model: { modelID: "test", providerID: "test" },
})
if (typeof clonedID !== "string") throw new Error("expected a cloned task ID")
expect(tool.state.output).toBe(
"Background task completed: test task\r\n<task_result>\r\nchild outcome\r\n</task_result>",
`Background task completed: test task\r\n\ttask_id: ${clonedID} (for resuming to continue this task if needed)\r\n\r\n<task_result>\r\nchild outcome\r\n</task_result>`,
)
const clone = await sessions.get(SessionID.descending(clonedID))
expect(clone.parentID).toBe(SessionID.descending(forked.id))
expect((await sessions.messages({ sessionID: clone.id }))[0]?.parts).toContainEqual(
expect.objectContaining({ text: "child message content" }),
)
expect(msgs.flatMap((msg) => msg.parts)).toContainEqual(
expect.objectContaining({ text: `Subagent task ID: ${clonedID}` }),
)
const source = await sessions.messages({ sessionID: parent.id })
@@ -252,7 +269,7 @@ describe("Session.fork task detachment", () => {
)
test(
"turns copied running tasks into terminal historical errors",
"turns copied running tasks into resumable historical errors",
async () => {
await using tmp = await tmpdir({ git: true })
await instance({
@@ -278,15 +295,22 @@ describe("Session.fork task detachment", () => {
},
} as MessageV2.ToolPart)
const before = await sessions.list()
const forked = await Session.fork({ sessionID: parent.id })
const after = await sessions.list()
const msgs = await sessions.messages({ sessionID: forked.id })
const tool = msgs.flatMap((msg) => msg.parts).find((part) => part.type === "tool") as MessageV2.ToolPart
expect(tool.state.status).toBe("error")
if (tool.state.status !== "error") throw new Error("expected detached task error")
expect(tool.state.error).toContain("still running")
expect(tool.state.input.task_id).toBeUndefined()
expect(tool.state.metadata).toEqual({ variant: "high" })
expect(tool.metadata).toEqual({})
const clonedID = tool.state.input.task_id
if (typeof clonedID !== "string") throw new Error("expected a cloned task ID")
expect(clonedID).not.toBe(child.id)
expect(tool.state.error).toContain(`task_id="${clonedID}"`)
expect(tool.state.metadata).toEqual({ sessionId: clonedID, variant: "high" })
expect(tool.metadata).toEqual({ sessionId: clonedID })
expect(after).toHaveLength(before.length + 2)
expect((await sessions.get(SessionID.descending(clonedID))).parentID).toBe(forked.id)
},
})
},
@@ -294,7 +318,7 @@ describe("Session.fork task detachment", () => {
)
test(
"detaches pending and errored task references",
"detaches in-flight tasks and remaps errored task references",
async () => {
await using tmp = await tmpdir({ git: true })
await instance({
@@ -329,7 +353,7 @@ describe("Session.fork task detachment", () => {
state: {
status: "error",
input: { task_id: child.id },
error: "original error",
error: `original error; task_id="${child.id}"`,
metadata: { sessionID: child.id, detail: "keep" },
time: { start: Date.now(), end: Date.now() },
},
@@ -344,15 +368,20 @@ describe("Session.fork task detachment", () => {
expect(pending?.state.status).toBe("error")
if (!pending || pending.state.status !== "error") throw new Error("expected detached pending task")
expect(pending.state.error).toContain("still pending")
expect(pending.state.input.task_id).toBeUndefined()
expect(pending.metadata).toEqual({})
const pendingID = pending.state.input.task_id
if (typeof pendingID !== "string") throw new Error("expected a cloned pending task ID")
expect(pendingID).not.toBe(child.id)
expect(pending.state.error).toContain(`task_id="${pendingID}"`)
expect(pending.metadata).toEqual({ sessionID: pendingID })
expect(errored?.state.status).toBe("error")
if (!errored || errored.state.status !== "error") throw new Error("expected detached errored task")
expect(errored.state.error).toBe("original error")
expect(errored.state.input.task_id).toBeUndefined()
expect(errored.state.metadata).toEqual({ detail: "keep" })
expect(errored.metadata).toEqual({})
const clonedID = errored.state.input.task_id
if (typeof clonedID !== "string") throw new Error("expected a cloned task ID")
expect(clonedID).toBe(pendingID)
expect(errored.state.error).toBe(`original error; task_id="${clonedID}"`)
expect(errored.state.metadata).toEqual({ sessionID: clonedID, detail: "keep" })
expect(errored.metadata).toEqual({ sessionId: clonedID })
},
})
},
+63
View File
@@ -269,6 +269,69 @@ describe("tool.task", () => {
}),
)
// kilocode_change start - verify forked task children remain resumable
it.instance("execute resumes a cloned task session after the parent is forked", () =>
Effect.gen(function* () {
const sessions = yield* Session.Service
const { chat, assistant } = yield* seed()
const child = yield* sessions.create({ parentID: chat.id, title: "Existing child" })
yield* sessions.updatePart({
id: PartID.ascending(),
messageID: assistant.id,
sessionID: chat.id,
type: "tool",
callID: "call_1",
tool: "task",
metadata: { sessionId: child.id },
state: {
status: "completed",
input: { description: "inspect bug", prompt: "continue", task_id: child.id },
output: `<task id="${child.id}"><task_result>done</task_result></task>`,
title: "inspect bug",
metadata: { sessionId: child.id },
time: { start: Date.now(), end: Date.now() },
},
} as MessageV2.ToolPart)
const forked = yield* sessions.fork({ sessionID: chat.id })
const msgs = yield* sessions.messages({ sessionID: forked.id })
const part = msgs.flatMap((msg) => msg.parts).find((item) => item.type === "tool" && item.tool === "task") as
| MessageV2.ToolPart
| undefined
if (!part || part.state.status !== "completed") throw new Error("expected a completed task part")
const id = part.state.input.task_id
if (typeof id !== "string") throw new Error("expected a cloned task ID")
const parent = msgs.find((msg) => msg.info.role === "assistant")
if (!parent || parent.info.role !== "assistant") throw new Error("expected a forked assistant message")
const tool = yield* TaskTool
const def = yield* tool.init()
let seen: SessionPrompt.PromptInput | undefined
yield* def.execute(
{
description: "inspect bug",
prompt: "continue from the fork",
subagent_type: "general",
task_id: id,
},
{
sessionID: forked.id,
messageID: parent.info.id,
agent: "build",
abort: new AbortController().signal,
extra: { promptOps: stubOps({ onPrompt: (input) => (seen = input) }) },
messages: [],
metadata: () => Effect.void,
ask: () => Effect.void,
},
)
expect(seen?.sessionID).toBe(SessionID.descending(id))
expect((yield* sessions.get(SessionID.descending(id))).parentID).toBe(forked.id)
}),
)
// kilocode_change end
// kilocode_change start - resumed children rebuild parent platform attribution after restart
it.instance("execute preserves platform attribution when resuming a task", () =>
Effect.gen(function* () {