Skip to content

feat(eve): send images to the model via toModelOutput content parts - #1250

Open
ruiconti wants to merge 7 commits into
rui/tool-model-output-content-partsfrom
rui/tool-model-output-content-parts-impl
Open

feat(eve): send images to the model via toModelOutput content parts#1250
ruiconti wants to merge 7 commits into
rui/tool-model-output-content-partsfrom
rui/tool-model-output-content-parts-impl

Conversation

@ruiconti

@ruiconti ruiconti commented Jul 27, 2026

Copy link
Copy Markdown
Contributor

Implements #456 per the plan in #1245 (this PR is stacked on it — base is rui/tool-model-output-content-parts, review only the top commit).

A tool that produces an image can now hand the pixels to a vision-capable model instead of describing them. toModelOutput accepts { type: "content", value } with text and file parts, and new toolOutput.text / toolOutput.json / toolOutput.content and toolOutputPart.text / toolOutputPart.file builders from eve/tools hide the SDK's tagged data: { type: "data", data } nesting.

How it works

  • ToolModelOutput widens to mirror the AI SDK's ToolResultOutput content parts, narrowed to the JSON-safe subset: file data is the tagged FileData union restricted to { type: "data" } with a base64 string. The AI SDK bakes the projection into the persisted tool-result part once, so a base64 payload survives durable replay verbatim.
  • normalizeToolModelOutput — the single funnel for authored projections — rejects raw bytes (Uint8Array/ArrayBuffer/Buffer) with a base64 pointer even though the SDK accepts them: JSON.stringify corrupts them silently at the durable boundary, the same failure class as File parts fail the model call with image.source.base64.data: Input should be a valid string under @workflow/world-postgres #497. The url/reference/text tags are rejected as not yet supported, keeping the tag space open.
  • Payloads above 3 MiB warn instead of failing. Unlike sandbox-ref attachments (ref-only in history, hydrated per call), content parts persist in history and are re-sent on every subsequent model call — the docs section says so.
  • execute stays JSON-only and action.result keeps the raw output; channels never see the projection.

Capability epochs

The widened union reaches both authoring surfaces, so tool bumps to epoch 3 and dynamicTool to epoch 4, both retained (additive change). The --update tooling deadlocks when two capabilities mismatch at once — each run refuses while the other is dirty — so the bump is applied by hand exactly as the script writes it: version table, retained-epoch fixtures, regenerated report JSONs. guard:invariants passes. That deadlock probably deserves its own issue.

Testing

Unit coverage in tools.test.ts (happy path, builders, byte/tag/empty/unknown-part rejections, error identity). The e2e eval adds a render-pixel fixture tool whose execute returns only a label + base64 string — the pixels reach the model exclusively through the content parts, so the model naming the color proves the file part arrived as vision input.

Out of scope, tracked in the plan: sandbox-ref file data, URL/provider-reference tags, video/audio policy (#543), and the #497 user-message bug itself.

@vercel

vercel Bot commented Jul 27, 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, Comment, Open in v0 Jul 27, 2026 9:56pm
eve-docs-1644 Ready Ready Preview, Comment, Open in v0 Jul 27, 2026 9:56pm
eve-docs-4759 Ready Ready Preview, Comment, Open in v0 Jul 27, 2026 9:56pm

@github-actions

github-actions Bot commented Jul 27, 2026

Copy link
Copy Markdown
Contributor

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

Key takeaways

  • No notable deltas vs rui/tool-model-output-content-parts (79f2889).

Delta vs rui/tool-model-output-content-parts (79f2889)

Area Metric Baseline Current Delta
Package Packed tarball 7.42 MB 7.42 MB +2.5 kB ⚠️
Package Unpacked publish size 27.93 MB 27.94 MB +8.6 kB ⚠️
Package Installed footprint 67.08 MB 67.09 MB +8.6 kB ⚠️
Package Published files 2786 2790 +4
Package Installed files 6026 6030 +4
Runtime Unique function payloads 2 2 0
Runtime Total function bytes 14.63 MB 14.64 MB +4.1 kB ⚠️
Runtime Public routes 11 11 0
Changed function payloads vs rui/tool-model-output-content-parts (79f2889) (2)
Function Status Baseline Current Delta Route changes
functions/__server.func changed 7.32 MB 7.32 MB +2.0 kB ⚠️ none
functions/.well-known/workflow/v1/flow.func changed 7.32 MB 7.32 MB +2.0 kB ⚠️ none

eve init install

Metric Baseline Current Delta
Installed footprint 105.49 MB 105.50 MB +8.6 kB ⚠️
Installed packages 122 122 0
dependencies 4 4 0
devDependencies 2 2 0
Dependency package bytes 42.21 MB 42.22 MB +8.6 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: 2026-07-27T21:56:56.983Z
  • 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@0.27.7
  • Package directory: packages/eve
  • Tarball: 7.42 MB (eve-0.27.7.tgz)
  • Unpacked payload: 27.94 MB across 2790 published files
  • Installed footprint: 67.09 MB across 6030 installed files
  • Installed root package: 26.65 MB
  • Installed dependencies: 40.44 MB
  • Runtime dependencies: 1
  • Peer dependencies: 5 (4 optional)

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

Heavy installed dependencies

  • eve: 26.65 MB (39.7%)
  • @rolldown/binding-linux-x64-gnu: 18.96 MB (28.3%)
  • ai: 6.43 MB (9.6%)
  • zod: 5.04 MB (7.5%)
  • nitro: 2.41 MB (3.6%)
Publish payload breakdown
Published file size
🔴 dist/src/compiled/shadcn-registry/index.js       [########################] 13.15 MB 47.1%
🟠 dist/src/compiled/experimental-ai-sdk-code-mo... [###.....................] 1.51 MB 5.4%
🟡 dist/src/compiled/@vercel/sandbox/index.js       [#.......................] 632.4 kB 2.3%
🟡 dist/src/compiled/_chunks/workflow/undici-C2Z... [#.......................] 502.4 kB 1.8%
🟡 dist/src/compiled/@chat-adapter/slack/index.js   [#.......................] 440.5 kB 1.6%
🔴 Other published files                            [#####################...] 11.70 MB 41.9%
Installed footprint breakdown
Installed package size
🔴 eve                             [########################] 26.65 MB 39.7%
🔴 @rolldown/binding-linux-x64-gnu [#################.......] 18.96 MB 28.3%
🔴 ai                              [######..................] 6.43 MB 9.6%
🔴 zod                             [#####...................] 5.04 MB 7.5%
🟠 nitro                           [##......................] 2.41 MB 3.6%
🟡 @ai-sdk/provider-utils          [#.......................] 852.0 kB 1.3%
🔴 Other installed packages        [######..................] 6.75 MB 10.1%
Runtime dependencies (1)
Package Range Notes
nitro 3.0.260610-beta
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: 105.50 MB across 7898 installed files
  • Installed packages: 122 total (116 transitive-only)
  • dependencies: 4 direct packages totaling 42.22 MB
  • devDependencies: 2 direct packages totaling 5.04 MB
  • Other transitive package files: 58.24 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 (26.5%)
  • eve: 26.65 MB (25.3%)
  • @rolldown/binding-linux-x64-gnu: 18.96 MB (18.0%)
  • zod: 9.00 MB (8.5%)
  • ai: 6.43 MB (6.1%)
Installed footprint breakdown
Installed package size
🔴 @typescript/typescript-linux-x64 [########################] 27.95 MB 26.5%
🔴 eve                              [#######################.] 26.65 MB 25.3%
🔴 @rolldown/binding-linux-x64-gnu  [################........] 18.96 MB 18.0%
🔴 zod                              [########................] 9.00 MB 8.5%
🔴 ai                               [######..................] 6.43 MB 6.1%
🟠 @types/node                      [##......................] 2.54 MB 2.4%
🔴 Other installed packages         [############............] 13.97 MB 13.2%
dependencies (4)
Package Range Installed size Share
@vercel/connect 0.4.2 135.8 kB 0.1%
ai ^7.0.34 6.43 MB 6.1%
eve file:eve-0.27.7.tgz 26.65 MB 25.3%
zod 4.4.3 9.00 MB 8.5%
devDependencies (2)
Package Range Installed size Share
@types/node 24.x 2.54 MB 2.4%
typescript 7.0.2 2.50 MB 2.4%
Function Drill-Down

Payload Size Graph

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 7.32 MB 50.0%
🔴 functions/__server.func                         [########################] 7.32 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 7.32 MB
Metric Value
Public routes /.well-known/workflow/v1/flow
Runtime nodejs24.x
Handler index.mjs
Payload 7.32 MB
Function files 7.32 MB across 43 files
Traced dependencies 0 B
Signal 🟠 Bundled file _chunks/runtime-artifacts.mjs is 1.57 MB (21.5%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/runtime-artifacts.mjs   [###############.........] 1.57 MB 21.5%
🟠 index.mjs                       [###########.............] 1.17 MB 16.0%
🟡 _chunks/world-vercel.mjs        [#########...............] 900.3 kB 12.3%
🟡 _chunks/sandbox.mjs             [#######.................] 767.6 kB 10.5%
🟡 _libs/@ai-sdk/gateway+[...].mjs [####....................] 432.5 kB 5.9%
🟠 Other bundled files             [########################] 2.47 MB 33.8%

🧾 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 • 7.32 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 7.32 MB
Function files 7.32 MB across 43 files
Traced dependencies 0 B
Signal 🟠 Bundled file _chunks/runtime-artifacts.mjs is 1.57 MB (21.5%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/runtime-artifacts.mjs   [###############.........] 1.57 MB 21.5%
🟠 index.mjs                       [###########.............] 1.17 MB 16.0%
🟡 _chunks/world-vercel.mjs        [#########...............] 900.3 kB 12.3%
🟡 _chunks/sandbox.mjs             [#######.................] 767.6 kB 10.5%
🟡 _libs/@ai-sdk/gateway+[...].mjs [####....................] 432.5 kB 5.9%
🟠 Other bundled files             [########################] 2.47 MB 33.8%

🧾 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: 1.94 s -> 1.85 s (-87.6 ms) vs rui/tool-model-output-content-parts (79f2889).
  • Timing is informational: shared GitHub runners are too variable for a hard timing budget.
Detailed phase timings vs `rui/tool-model-output-content-parts (79f2889)`
Phase Baseline Current Delta
extension.check 1.1 ms 1.2 ms +0.1 ms
project.resolve 0.6 ms 0.6 ms 0.0 ms
workspace.create 0.7 ms 0.9 ms +0.2 ms
host.prepare 386.4 ms 226.2 ms -160.2 ms
vercel.service-prefix.resolve 1.7 ms 1.8 ms +0.1 ms
nitro.create 212.4 ms 205.1 ms -7.3 ms
sandbox.prewarm 203.4 ms 216.1 ms +12.7 ms
nitro.cache.prepare 0.3 ms 0.2 ms -0.1 ms
nitro.prepare 0.8 ms 0.7 ms -0.1 ms
nitro.public-assets 0.7 ms 0.8 ms +0.1 ms
nitro.prerender 0.5 ms 0.5 ms 0.0 ms
nitro.bundle 1.12 s 1.19 s +67.0 ms
nitro.cache.write 0.4 ms 0.3 ms -0.1 ms
agent-summary.emit 0.5 ms 0.4 ms -0.1 ms
nitro.close 0.1 ms 0.1 ms 0.0 ms
output.publish 3.3 ms 3.3 ms 0.0 ms
workspace.remove 1.9 ms 2.0 ms +0.1 ms

ruiconti added 6 commits July 27, 2026 17:42
… parts

ToolModelOutput gains a content variant mirroring the AI SDK's
ToolResultOutput content parts narrowed to the JSON-safe subset: file
data is the tagged FileData union restricted to { type: "data" } with a
base64 string, so persisted tool results survive the durable JSON
boundary verbatim. normalizeToolModelOutput rejects raw bytes
(Uint8Array/ArrayBuffer/Buffer) with a base64 pointer — JSON.stringify
would corrupt them silently, the #497 failure class — and rejects the
url/reference/text FileData tags as not yet supported. Payloads above
the 3 MiB inline-image reference point warn: unlike sandbox-ref
attachments, content parts persist in history and are re-sent on every
subsequent model call.

New toolOutputPart.text/file builders ship from eve/tools so authors
skip the tagged nesting. execute stays JSON-only and action.result
keeps the raw output. The tool and dynamicTool capability epochs bump
(retained) since the widened union reaches both authoring surfaces.

Implements the plan in research/tool-model-output-content-parts.md
(#456).

Signed-off-by: Rui Conti <ruiconti@gmail.com>
Companion to toolOutputPart: toolOutput.text/json/content build the
model-facing ToolModelOutput union so authored projections compose as
toolOutput.content([toolOutputPart.text(...), toolOutputPart.file(...)]).
Pure sugar — hand-written literals remain valid. The module renames to
output-builders.ts since it now holds output builders, not just parts.

The tool capability epoch-3 report is regenerated in place rather than
bumped: epoch 3 is introduced by this branch and unreleased, so its
hash is not yet published contract history.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
… its own module

tools.ts is the tool-set assembly module; the content-parts change had
grown a second concern inside it — a ~165-line validation subsystem for
authored toModelOutput results. It moves to tool-model-output.ts next to
the serialization error machinery it throws.

Exporting the normalizer also deletes the test ceremony: the shape
matrix no longer builds a HarnessToolMap and fishes the toModelOutput
closure out of the SDK tool with casts per case — it calls the function.
tools.test.ts keeps one wiring test proving buildToolSet applies the
normalizer.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
The hand-rolled probe-cast guards in the normalizer are replaced by a
zod discriminated union (vendored #compiled/zod + formatValidationError,
the established untrusted-boundary pattern). Structural rejections now
carry path-labeled messages like:

  Invalid discriminator value. Expected 'text' | 'file' at "value[0].type"

Two author mistakes stay hand-written pre-checks because the schema
cannot explain them: raw bytes (zod would report a missing data.type
key on a Uint8Array; the real fix is base64-encoding — and silent
JSON.stringify corruption is the #497 failure class) and the not-yet-
supported url/reference/text FileData tags. As a side effect the byte
rejection is now checked once for both payload positions.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
The hand-written 1x1 base64 was corrupt; OpenAI and Anthropic both
rejected it at the model call ('image data does not represent a valid
image'). Replaced with a programmatically generated 64x64 solid-red PNG
(132 bytes) verified to decode — 64x64 also clears any provider
minimum-size rules. The unit-test constant is swapped to the same valid
payload so the corrupt string cannot be copy-pasted from the repo.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
The discover and build-extension tests hardcoded tool contract v2/v3
literals, so the epoch bump in this branch aged them out (and any
future bump would again). The unsupported-version fixture now requires
current + 1 — unsupported by construction — and the stamped-manifest
assertion tracks EXTENSION_CAPABILITY_VERSIONS; the real assertions
(rejection diagnostics, the set of stamped capabilities) are unchanged.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
assertSupportedFilePartData says what holds after the call; the old
rejectTeachableFilePartMistakes described the error strategy instead.

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.

2 participants