Skip to content

feat(eve): run subagents as background tasks behind experimental.tasks - #1521

Open
ruiconti wants to merge 1 commit into
tasks/03-toolsfrom
tasks/04-delegated
Open

feat(eve): run subagents as background tasks behind experimental.tasks#1521
ruiconti wants to merge 1 commit into
tasks/03-toolsfrom
tasks/04-delegated

Conversation

@ruiconti

@ruiconti ruiconti commented Aug 1, 2026

Copy link
Copy Markdown
Contributor

Part 4/5 of the experimental.tasks stack, on top of #1519. The flag now means something: every subagent dispatch (fresh start, agentId continuation, local and remote) returns a task receipt and the parent turn keeps moving.

Data flows, before → after (flag on)

Flow Before After
Terminal child result child resumes the parent turn inbox token; the turn blocks in waitForRuntimeActionResults until every batch result arrives child's reply token is the task run's private hook; the originating call resolves immediately with { taskId, status: "working" }
Child HITL / authorization adapter forwards to the parent turn inbox same payloads land on the task run → input_required snapshot (authorization blocks, authorization.completed unblocks)
Ready notification n/a — parent was always waiting task run resumes the parent session hook with a deliver payload; a parked parent starts a turn, an active turn sees it at the next boundary. Terminal and input_required are the only wake triggers
Agent handle settled when the batch resolves with the child's outcome receipt carries a parked outcome, so the unchanged resolve path parks the handle at receipt; the address stays for follow-ups

The child wire itself is untouched — local notification, HITL forwarding, and the remote callback route all deliver to whatever token they were given. Delegation is a token-routing change, not a payload change.

Ordering

The task run is created before the child dispatch side effect, so a fast child always finds a live command hook. The command token is derived deterministically from (taskId, parent continuation token): a durable replay re-derives it and the duplicate run loses the hook claim.

Tool execution

  • task_peek/task_cancel run inside the dispatch step. Cancel commits cancelled on the run first, then propagates a cooperative abort through the handle address — a late child result can't revive the task.
  • task_await starts a small aggregation run that polls the selected snapshot streams and posts one tool-result to the turn's existing inbox wait. No new wait arm: the strict inbox filter only guards subagent-result kinds, so the result passes as-is. The run exits when the waiting turn's inbox is gone.
  • Parent finalization cancels live tasks before terminating children.

Scope

  • Flag off: dispatch behavior, wire shapes, and events unchanged — the agent-messaging scenario suite passes unmodified.
  • The flag implies conversation-mode children, so tasks and subagentPersistentSessions never produce a third mode.
  • Deferred: child usage accounting for delegated tasks (usageDelta reaches the task run but isn't folded into parent totals), progress/task_message.
  • subagent-start.ts is code motion out of the dispatch step (file-length cap).

Part of #1084.

@vercel

vercel Bot commented Aug 1, 2026

Copy link
Copy Markdown
Contributor

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
eve-docs Ready Ready Preview Aug 1, 2026 10:15pm
eve-docs-4759 Ready Ready Preview, v0 Aug 1, 2026 10:15pm

@github-actions

github-actions Bot commented Aug 1, 2026

Copy link
Copy Markdown
Contributor

Bundle + Package Summary: apps/fixtures/weather-agent

Key takeaways

  • No notable deltas vs tasks/03-tools (0906202).

Delta vs tasks/03-tools (0906202)

Area Metric Baseline Current Delta
Package Packed tarball 7.61 MB 7.62 MB +8.5 kB ⚠️
Package Unpacked publish size 28.52 MB 28.54 MB +26.3 kB ⚠️
Package Installed footprint 85.51 MB 85.53 MB +26.3 kB ⚠️
Package Published files 2990 3002 +12
Package Installed files 6785 6797 +12
Runtime Unique function payloads 2 2 0
Runtime Total function bytes 16.90 MB 16.96 MB +66.0 kB ⚠️
Runtime Public routes 11 11 0
Changed function payloads vs tasks/03-tools (0906202) (2)
Function Status Baseline Current Delta Route changes
functions/__server.func changed 8.45 MB 8.48 MB +33.0 kB ⚠️ none
functions/.well-known/workflow/v1/flow.func changed 8.45 MB 8.48 MB +33.0 kB ⚠️ none

eve init install

Metric Baseline Current Delta
Installed footprint 123.92 MB 123.94 MB +26.3 kB ⚠️
Installed packages 128 128 0
dependencies 4 4 0
devDependencies 2 2 0
Dependency package bytes 42.90 MB 42.93 MB +26.3 kB ⚠️
devDependency package bytes 5.04 MB 5.04 MB 0 B ➖
Build Metadata
  • Preset: vercel
  • Nitro: nitro@3.0.260610-beta
  • Output directory: apps/fixtures/weather-agent/.vercel/output
  • Build metadata timestamp: 2026-08-01T22:15:58.244Z
  • Route aliases: 11 public, 1 internal (12 total aliases)
  • Vercel routes in config: 14
  • Severity legend: 🔴 dominant/large, 🟠 notable, 🟡 watch, ⚪ small
Package Drill-Down

Package Details

  • Package: eve@0.29.4
  • Package directory: packages/eve
  • Tarball: 7.62 MB (eve-0.29.4.tgz)
  • Unpacked payload: 28.54 MB across 3002 published files
  • Installed footprint: 85.53 MB across 6797 installed files
  • Installed root package: 27.19 MB
  • Installed dependencies: 58.34 MB
  • Runtime dependencies: 2
  • Peer dependencies: 5 (4 optional)

Installed footprint is measured from an isolated temporary npm install of the packed tarball.

Heavy installed dependencies

  • eve: 27.19 MB (31.8%)
  • @rolldown/binding-linux-x64-gnu: 19.28 MB (22.5%)
  • @rolldown/binding-wasm32-wasi: 10.66 MB (12.5%)
  • ai: 6.57 MB (7.7%)
  • zod: 5.07 MB (5.9%)
Publish payload breakdown
Published file size
🔴 dist/src/compiled/shadcn-registry/index.js       [#################.......] 9.68 MB 33.9%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [####....................] 2.42 MB 8.5%
🟠 dist/src/compiled/experimental-ai-sdk-code-mo... [###.....................] 1.51 MB 5.3%
🟡 dist/src/compiled/_chunks/node/undici-DWL_MYm... [#.......................] 502.4 kB 1.8%
🟡 dist/src/compiled/_chunks/workflow/undici-DWL... [#.......................] 502.4 kB 1.8%
🔴 Other published files                            [########################] 13.93 MB 48.8%
Installed footprint breakdown
Installed package size
🔴 eve                             [########################] 27.19 MB 31.8%
🔴 @rolldown/binding-linux-x64-gnu [#################.......] 19.28 MB 22.5%
🔴 @rolldown/binding-wasm32-wasi   [#########...............] 10.66 MB 12.5%
🔴 ai                              [######..................] 6.57 MB 7.7%
🔴 zod                             [####....................] 5.07 MB 5.9%
🟠 undici                          [###.....................] 3.50 MB 4.1%
🔴 Other installed packages        [############............] 13.26 MB 15.5%
Runtime dependencies (2)
Package Range Notes
nitro 3.0.260610-beta
undici 8.9.0
Peer dependencies (5)
Package Range Notes
@opentelemetry/api ^1.0.0 optional peer
ai catalog:
braintrust ^3.0.0 optional peer
just-bash ^3.0.0 optional peer
microsandbox ^0.5.0 optional peer
eve init install drill-down

eve init install details

  • Command: eve init my-agent
  • Package manager: npm
  • Installed footprint: 123.94 MB across 8673 installed files
  • Installed packages: 128 total (122 transitive-only)
  • dependencies: 4 direct packages totaling 42.93 MB
  • devDependencies: 2 direct packages totaling 5.04 MB
  • Other transitive package files: 75.98 MB

Installed footprint is measured from an isolated temporary eve init my-agent using the current packed eve tarball.

Heavy installed dependencies

  • @typescript/typescript-linux-x64: 27.95 MB (22.5%)
  • eve: 27.19 MB (21.9%)
  • @rolldown/binding-linux-x64-gnu: 19.28 MB (15.6%)
  • @rolldown/binding-wasm32-wasi: 10.66 MB (8.6%)
  • zod: 9.02 MB (7.3%)
Installed footprint breakdown
Installed package size
🔴 @typescript/typescript-linux-x64 [########################] 27.95 MB 22.5%
🔴 eve                              [#######################.] 27.19 MB 21.9%
🔴 @rolldown/binding-linux-x64-gnu  [#################.......] 19.28 MB 15.6%
🔴 @rolldown/binding-wasm32-wasi    [#########...............] 10.66 MB 8.6%
🔴 zod                              [########................] 9.02 MB 7.3%
🔴 ai                               [######..................] 6.57 MB 5.3%
🔴 Other installed packages         [####################....] 23.27 MB 18.8%
dependencies (4)
Package Range Installed size Share
@vercel/connect 0.4.3 141.0 kB 0.1%
ai ^7.0.38 6.57 MB 5.3%
eve file:eve-0.29.4.tgz 27.19 MB 21.9%
zod 4.4.3 9.02 MB 7.3%
devDependencies (2)
Package Range Installed size Share
@types/node 24.x 2.54 MB 2.1%
typescript 7.0.2 2.50 MB 2.0%
Function Drill-Down

Payload Size Graph

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 8.48 MB 50.0%
🔴 functions/__server.func                         [########################] 8.48 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 8.48 MB
Metric Value
Public routes /.well-known/workflow/v1/flow
Runtime nodejs24.x
Handler index.mjs
Payload 8.48 MB
Function files 8.48 MB across 44 files
Traced dependencies 0 B
Signal 🟠 Bundled file index.mjs is 2.33 MB (27.4%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                       [########################] 2.33 MB 27.4%
🟠 _chunks/runtime-artifacts.mjs   [################........] 1.61 MB 18.9%
🟡 _libs/undici.mjs                [##########..............] 980.5 kB 11.6%
🟡 _chunks/sandbox.mjs             [########................] 769.0 kB 9.1%
🟡 _libs/@ai-sdk/gateway+[...].mjs [####....................] 432.8 kB 5.1%
🟠 Other bundled files             [########################] 2.37 MB 27.9%

🧾 Vercel Config

{
  "handler": "index.mjs",
  "launcherType": "Nodejs",
  "shouldAddHelpers": false,
  "supportsResponseStreaming": true,
  "runtime": "nodejs24.x",
  "maxDuration": "max",
  "experimentalTriggers": [
    {
      "type": "queue/v2beta",
      "topic": "__eve776561746865722d6167656e74_wkf_workflow_*",
      "consumer": "default",
      "retryAfterSeconds": 5,
      "initialDelaySeconds": 0
    }
  ],
  "environment": {
    "WORKFLOW_PRECONDITION_GUARD": "1"
  }
}

🟠 functions/__server.func • 10 public routes, 1 internal alias • 8.48 MB
Metric Value
Public routes /
/eve/v1/callback/[token]
/eve/v1/connections/[name]/callback/[token]
/eve/v1/health
/eve/v1/info
/eve/v1/session
/eve/v1/session/[sessionId]
/eve/v1/session/[sessionId]/cancel
/eve/v1/session/[sessionId]/stream
/eve/v1/session/reset
Internal aliases /__server
Runtime nodejs24.x
Handler index.mjs
Payload 8.48 MB
Function files 8.48 MB across 44 files
Traced dependencies 0 B
Signal 🟠 Bundled file index.mjs is 2.33 MB (27.4%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                       [########################] 2.33 MB 27.4%
🟠 _chunks/runtime-artifacts.mjs   [################........] 1.61 MB 18.9%
🟡 _libs/undici.mjs                [##########..............] 980.5 kB 11.6%
🟡 _chunks/sandbox.mjs             [########................] 769.0 kB 9.1%
🟡 _libs/@ai-sdk/gateway+[...].mjs [####....................] 432.8 kB 5.1%
🟠 Other bundled files             [########################] 2.37 MB 27.9%

🧾 Vercel Config

{
  "handler": "index.mjs",
  "launcherType": "Nodejs",
  "shouldAddHelpers": false,
  "supportsResponseStreaming": true,
  "runtime": "nodejs24.x"
}

Build Timing: e2e/fixtures/agent-tools-sandbox

This is an informational timing measurement inside eve build, from preflight through publication. Output-size measurement and profile writing are excluded.

Build mode: deployable Vercel build with sandbox template prewarm included.

  • Build pipeline: 2.03 s -> 1.94 s (-85.8 ms) vs tasks/03-tools (0906202).
  • Timing is informational: shared GitHub runners are too variable for a hard timing budget.
Detailed phase timings vs `tasks/03-tools (0906202)`
Phase Baseline Current Delta
extension.check 6.5 ms 6.0 ms -0.5 ms
project.resolve 6.5 ms 10.8 ms +4.3 ms
workspace.create 1.4 ms 1.4 ms 0.0 ms
host.prepare 292.0 ms 149.2 ms -142.8 ms
vercel.service-prefix.resolve 2.5 ms 2.4 ms -0.1 ms
nitro.create 203.3 ms 217.2 ms +13.9 ms
sandbox.prewarm 188.4 ms 180.3 ms -8.1 ms
nitro.cache.prepare 0.3 ms 0.3 ms 0.0 ms
nitro.prepare 1.0 ms 0.9 ms -0.1 ms
nitro.public-assets 0.7 ms 0.7 ms 0.0 ms
nitro.prerender 0.5 ms 0.5 ms 0.0 ms
nitro.bundle 1.29 s 1.34 s +47.8 ms
nitro.cache.write 0.4 ms 0.4 ms 0.0 ms
vercel.workflow-function.materialize 23.5 ms 24.0 ms +0.5 ms
agent-summary.emit 0.4 ms 0.5 ms +0.1 ms
nitro.close 0.1 ms 0.1 ms 0.0 ms
output.publish 3.6 ms 3.6 ms 0.0 ms
workspace.remove 2.8 ms 2.3 ms -0.5 ms

for (
let attempt = 0;
attempt < CANCEL_COMMIT_POLL_ATTEMPTS &&
!(view !== undefined && isReadyTaskStatus(view.status));

@vercel vercel Bot Aug 1, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

cancelOneTask's commit-poll loop stops on isReadyTaskStatus, which is true for input_required, so cancelling a HITL-blocked task exits before the cancelled state commits — skipping child-turn abort propagation and returning a stale view.

Fix on Vercel

With the flag on, every subagent dispatch (fresh start, agentId
continuation, local and remote) becomes delegated execution:

- the durable task run is created before the child dispatch side
  effect, so a fast child always finds a live command hook; a replay
  re-derives the same deterministic command token and the duplicate
  run loses the hook claim;
- the child's reply address is the task run's private hook instead of
  the parent turn's inbox, so the existing local notification, HITL
  forwarding, and remote callback wires deliver to the task unchanged;
- the originating call resolves immediately with a task receipt whose
  parked outcome settles the agent handle through the existing resolve
  path, keeping history provider-valid with exactly one result;
- a dispatch that never acknowledges a child fails its task out of
  band and returns the failure directly, so the model never sees a
  task id for work that never started.

The task run translates the child wire into lifecycle commands
(settled turn outcomes to complete/fail/cancel, HITL batches to
input_required, authorization events to blocked/unblocked) and wakes
the parent through the ordinary session delivery path when a task
becomes ready — a parked parent starts a turn, an active turn observes
the notification at its next safe boundary.

task_peek and task_cancel execute inside the dispatch step, which owns
the session task index and world access; task_cancel commits cancelled
before propagating a cooperative abort through the handle address.
task_await rides the turn's existing inbox wait: a small aggregation
run polls the selected snapshot streams and posts the one tool-result
the pending key expects, exiting quietly when the waiting turn is
gone. Parent finalization cancels live tasks before terminating
children.

The flag also implies conversation-mode children so experimental.tasks
and experimental.subagentPersistentSessions never produce a third
mode. Child usage accounting for delegated tasks is deferred; the
subagent-start machinery moved to its own module to keep the dispatch
step within the file-length cap. Without the flag, dispatch behavior,
wire shapes, and events are unchanged (agent-messaging scenario suite
passes unmodified).

Signed-off-by: Rui Conti <ruiconti@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant