fix(eve): preserve streamed tool inputs - #2395

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

Project	Deployment	Actions	Updated (UTC)
eve-docs	Ready	Preview, v0	Aug 27, 2026 1:47pm
eve-docs-4759	Ready	Preview, v0	Aug 27, 2026 1:47pm

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
eve-pkg	Skipped	v0	Aug 27, 2026 1:47pm

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

Key takeaways

• No notable deltas vs main (a77ae1b).

Delta vs main (a77ae1b)

eve init install

Published file size
🔴 dist/src/compiled/shadcn-registry/index.js       [##############..........] 9.76 MB 30.4%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [###.....................] 2.44 MB 7.6%
🟠 dist/src/compiled/experimental-ai-sdk-code-mo... [##......................] 1.51 MB 4.7%
🟡 dist/src/compiled/@vercel/blob/index.js          [#.......................] 901.8 kB 2.8%
🟡 dist/src/compiled/_chunks/workflow/undici-Dzn... [#.......................] 512.7 kB 1.6%
🔴 Other published files                            [########################] 17.01 MB 52.9%

Installed package size
🔴 eve                             [########################] 30.78 MB 40.6%
🔴 @rolldown/binding-linux-x64-gnu [###############.........] 19.33 MB 25.5%
🔴 ai                              [#####...................] 6.85 MB 9.0%
🔴 zod                             [####....................] 5.07 MB 6.7%
🟠 undici                          [###.....................] 3.50 MB 4.6%
🟠 nitro                           [##......................] 2.41 MB 3.2%
🔴 Other installed packages        [######..................] 7.93 MB 10.5%

Installed package size
🔴 eve                              [########################] 30.78 MB 26.9%
🔴 @typescript/typescript-linux-x64 [######################..] 27.95 MB 24.4%
🔴 @rolldown/binding-linux-x64-gnu  [###############.........] 19.33 MB 16.9%
🔴 zod                              [#######.................] 9.02 MB 7.9%
🔴 ai                               [#####...................] 6.85 MB 6.0%
🟠 undici                           [###.....................] 3.50 MB 3.1%
🔴 Other installed packages         [#############...........] 16.87 MB 14.8%

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 9.33 MB 50.0%
🔴 functions/__server.func                         [########################] 9.33 MB 50.0%

Bundled file size
🟠 index.mjs                   [###########.............] 1.93 MB 20.6%
🟡 _libs/undici.mjs            [#####...................] 980.5 kB 10.5%
🟡 _chunks/sandbox.mjs         [####....................] 781.9 kB 8.4%
🟡 _chunks/frames-Dl2Hcma4.mjs [####....................] 678.7 kB 7.3%
🟡 _chunks/chatgpt-model.mjs   [####....................] 666.1 kB 7.1%
🔴 Other bundled files         [########################] 4.30 MB 46.1%

{
  "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"
  }
}

Bundled file size
🟠 index.mjs                   [###########.............] 1.93 MB 20.6%
🟡 _libs/undici.mjs            [#####...................] 980.5 kB 10.5%
🟡 _chunks/sandbox.mjs         [####....................] 781.9 kB 8.4%
🟡 _chunks/frames-Dl2Hcma4.mjs [####....................] 678.7 kB 7.3%
🟡 _chunks/chatgpt-model.mjs   [####....................] 666.1 kB 7.1%
🔴 Other bundled files         [########################] 4.30 MB 46.1%

{
  "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: 3.33 s -> 2.48 s (-857.5 ms) vs main (a77ae1b).
• Timing is informational: shared GitHub runners are too variable for a hard timing budget.

We've built a downstream backport of this PR onto eve@0.27.2 (our pinned version). It is on a review branch, not yet merged or deployed, but we ran it against a live model to measure it — in case the numbers are useful input.

Retry/replay ordering matches what's documented here. A transient model retry re-runs the stream under a new callId, so a first attempt's input-streaming part can linger until the turn ends; our reducer (ported from this PR) is last-write-wins on inputTextSoFar, and our client always renders the last document-tool part in part order, which is the live one. No ordering surprises in live runs against a real model (Bedrock Claude).

Measured cost, real 15-page document turns, live model (not synthetic):

Because every event carries inputTextSoFar (the full accumulated input, not the delta), bytes grow with the square of document length — one 44k-character document cost 120 MB of durable log even with the interval on. The interval roughly halves it; it doesn't change the growth shape (that would need delta+offset encoding, a bigger change we haven't attempted).

We added a small, additive knob inside consumeStreamContent: a per-call minimum emit interval (env-configurable; we default to 100 ms, 0 reproduces this PR's behavior exactly) that time-gates tool-input-delta emission and flushes unconditionally on every other event, so ordering is unaffected — it uses the same barrier this PR already applies for providerActionBatch.flush(). It measurably halves durable write/storage cost with no reader-visible latency difference in our testing (first-visible-text and reload-catch-up times were within run-to-run noise between the two settings).

Two questions:

1. Any sense of merge timing? We're carrying this as a patch and would like to drop that maintenance cost.
2. Would a per-call emit-interval knob (upstream, defaulting to off/0 to preserve today's behavior) be something you'd take as a follow-up PR? Happy to open one.

Matching your event/field names verbatim made the backport a clean, isolated adapter on our side — thanks for the clear design.