Skip to content

feat: track model benchmark results across models and eve versions - #4470

Draft
prha wants to merge 1 commit into
mainfrom
prha/eval-bench
Draft

prha wants to merge 1 commit into
mainfrom
prha/eval-bench

Conversation

@prha

@prha prha commented Oct 7, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Currently, we cannot see how eval correctness, latency, and cost change across models and eve versions. CI keeps eval artifacts only on failure, for 7 days. We only run our evals on a couple models, to find regressions on PRs.

This PR adds an eval-benchmark workflow. It runs fixture evals on registry models and on mock-model worlds, nightly and on manual dispatch, on main only. It derives a versioned metric set from the captured event streams and publishes immutable records to private Blob for an internal dashboard.

The only change to published eve is that eval artifacts now record each eval's start and completion time.

Validation

  • pnpm test:eval-benchmark (33 tests) and node --test .github/scripts/*.test.mjs pass.
  • pnpm --filter eve exec vitest run --config vitest.unit.config.ts src/evals/runner/artifacts.test.ts passes.
  • pnpm fmt, pnpm lint, pnpm guard:invariants, pnpm docs:check, pnpm check:deps, and pnpm --filter eve typecheck are clean.
  • Local mock-model eve eval runs of agent-tools and agent-subagents produced full metric sets. Trimmed copies of those captures are the golden test fixtures.
  • plan.mjs resolved every registry model against the live AI Gateway catalog.
  • A temporary, since-removed pull request trigger ran the workflow end to end and published to Blob: one model on agent-tools (runs 37681411702 and 37682053298), then all 10 registry models × 3 fixtures × 1 attempt (run 37705854499, 21.5 minutes, $8.49, 0 gaps).

Checklist

  • This change was requested or approved by a maintainer
  • I ran the relevant checks from CONTRIBUTING.md
  • I added tests and documentation where relevant
  • I added a changeset if this touches the published eve package
  • DCO sign-off passes for every commit (git commit --signoff)

@vercel

vercel Bot commented Oct 7, 2026 •

Copy link
Copy Markdown
Contributor

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

Project Deployment Actions Updated
eve-docs Ready Ready Preview, v0 Oct 8, 2026 6:31am UTC
eve-pkg Ready Ready Preview, v0 Oct 8, 2026 6:31am UTC

@prha prha added the run-eval-benchmark Run the model benchmark on this PR (temporary, eval-benchmark.yml) label Oct 7, 2026
@prha
prha force-pushed the prha/eval-bench branch from 62c62e5 to a2c2af2 Compare October 7, 2026 20:21
@prha prha added run-eval-benchmark Run the model benchmark on this PR (temporary, eval-benchmark.yml) and removed run-eval-benchmark Run the model benchmark on this PR (temporary, eval-benchmark.yml) labels Oct 7, 2026
@socket-security

socket-security Bot commented Oct 7, 2026 •

Copy link
Copy Markdown

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff Package Supply Chain
Security
Vulnerability Quality Maintenance License
Addedgithub/​actions/​download-artifact@​3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c10010010010080

View full report

Comment thread scripts/eval-benchmark/records.mjs Outdated
);
const missingReason = dir === undefined ? "job-missing" : "eval-missing";
const expectedIds = expected ?? fixture.eval_ids;
for (const evalId of new Set([...expectedIds, ...results.keys()]))

@vercel vercel Bot Oct 7, 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.

buildRecords falls back to the full fixture.eval_ids (which includes real-model-tagged evals) for mock legs that lack an expected.json, fabricating spurious gap records for evals the mock track never runs.

Fix on Vercel

@github-actions

github-actions Bot commented Oct 7, 2026 •

Copy link
Copy Markdown
Contributor

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

Key takeaways

  • No notable deltas vs main (89704cb).

Delta vs main (89704cb)

Area Metric Baseline Current Delta
Package Packed tarball 7.49 MB 7.49 MB +291 B ⚠️
Package Unpacked publish size 26.39 MB 26.39 MB +172 B ⚠️
Package Installed footprint 74.06 MB 74.06 MB +172 B ⚠️
Package Published files 4159 4159 0
Package Installed files 8160 8160 0
Package Installed package instances 34 34 0
Package Distinct installed package names 32 32 0
Package Installed dependency edges 51 51 0
Package Installed optional peer edges 9 9 0
Runtime Unique function payloads 2 2 0
Runtime Total function bytes 19.92 MB 19.92 MB -120 B ✅
Runtime Public routes 17 17 0
Changed function payloads vs main (89704cb) (2)
Function Status Baseline Current Delta Route changes
functions/__server.func changed 9.96 MB 9.96 MB -60 B ✅ none
functions/.well-known/workflow/v1/flow.func changed 9.96 MB 9.96 MB -60 B ✅ none

eve init install

Metric Baseline Current Delta
Installed footprint 112.35 MB 112.35 MB +172 B ⚠️
Installed packages 103 103 0
dependencies 4 4 0
devDependencies 2 2 0
Dependency package bytes 44.04 MB 44.04 MB +172 B ⚠️
devDependency package bytes 5.11 MB 5.11 MB 0 B ➖
Build Metadata
  • Preset: vercel
  • Nitro: nitro@3.0.260903-beta
  • Output directory: apps/fixtures/weather-agent/.vercel/output
  • Build metadata timestamp: 2026-10-08T06:32:00.620Z
  • Route aliases: 17 public, 1 internal (18 total aliases)
  • Vercel routes in config: 20
  • Severity legend: 🔴 dominant/large, 🟠 notable, 🟡 watch, ⚪ small
Package Drill-Down

Package Details

  • Package: eve@0.74.0
  • Package directory: packages/eve
  • Tarball: 7.49 MB (eve-0.74.0.tgz)
  • Unpacked payload: 26.39 MB across 4159 published files
  • Installed footprint: 74.06 MB across 8160 installed files
  • Installed root package: 26.09 MB
  • Installed dependencies: 47.97 MB
  • Installed package instances: 34
  • Distinct installed package names: 32
  • Installed dependency edges: 51
  • Installed optional peer edges: 9
  • Runtime dependencies: 2
  • Peer dependencies: 7 (6 optional)

Installed footprint is measured from an isolated temporary npm install of the packed tarball.
Graph metrics read only package.json files in package directories directly beneath a node_modules boundary, including nested boundaries. Each directory is one package instance; distinct names come from those manifests. Dependency edges count each unique name in dependencies or optionalDependencies per instance; optional peer edges count peerDependencies marked optional.

Heavy installed dependencies

  • eve: 26.09 MB (35.2%)
  • @rolldown/binding-linux-x64-gnu: 19.62 MB (26.5%)
  • ai: 8.01 MB (10.8%)
  • zod: 6.14 MB (8.3%)
  • undici: 3.56 MB (4.8%)
Publish payload breakdown
Published file size
🟠 dist/src/compiled/shadcn-registry/index.js       [#####...................] 3.85 MB 14.6%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [###.....................] 2.29 MB 8.7%
🟠 dist/src/compiled/@ai-sdk/code-mode/index.js     [#.......................] 1.03 MB 3.9%
🟡 dist/src/compiled/@vercel/blob/index.js          [#.......................] 604.9 kB 2.3%
🟡 dist/src/compiled/_chunks/workflow/signal-exi... [#.......................] 521.9 kB 2.0%
🔴 Other published files                            [########################] 18.10 MB 68.6%
Installed footprint breakdown
Installed package size
🔴 eve                             [########################] 26.09 MB 35.2%
🔴 @rolldown/binding-linux-x64-gnu [##################......] 19.62 MB 26.5%
🔴 ai                              [#######.................] 8.01 MB 10.8%
🔴 zod                             [######..................] 6.14 MB 8.3%
🟠 undici                          [###.....................] 3.56 MB 4.8%
🟠 nitro                           [##......................] 1.89 MB 2.6%
🔴 Other installed packages        [########................] 8.75 MB 11.8%
Runtime dependencies (2)
Package Range Notes
nitro 3.0.260903-beta
undici 8.10.2
Peer dependencies (7)
Package Range Notes
@opentelemetry/api ^1.0.0 optional peer
ai catalog:
braintrust ^3.0.0 optional peer
chat ^4.41.0 optional peer
dd-trace ^6.13.0 optional peer
just-bash ^3.1.0 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: 112.35 MB across 10055 installed files
  • Installed packages: 103 total (97 transitive-only)
  • dependencies: 4 direct packages totaling 44.04 MB
  • devDependencies: 2 direct packages totaling 5.11 MB
  • Other transitive package files: 63.20 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 (24.9%)
  • eve: 26.09 MB (23.2%)
  • @rolldown/binding-linux-x64-gnu: 19.62 MB (17.5%)
  • zod: 9.76 MB (8.7%)
  • ai: 8.01 MB (7.1%)
Installed footprint breakdown
Installed package size
🔴 @typescript/typescript-linux-x64 [########################] 27.95 MB 24.9%
🔴 eve                              [######################..] 26.09 MB 23.2%
🔴 @rolldown/binding-linux-x64-gnu  [#################.......] 19.62 MB 17.5%
🔴 zod                              [########................] 9.76 MB 8.7%
🔴 ai                               [#######.................] 8.01 MB 7.1%
🟠 undici                           [###.....................] 3.56 MB 3.2%
🔴 Other installed packages         [###############.........] 17.37 MB 15.5%
dependencies (4)
Package Range Installed size Share
@vercel/connect 2.2.0 188.7 kB 0.2%
ai ^7.0.128 8.01 MB 7.1%
eve file:eve-0.74.0.tgz 26.09 MB 23.2%
zod 4.5.4 9.76 MB 8.7%
devDependencies (2)
Package Range Installed size Share
@types/node 24.x 2.61 MB 2.3%
typescript 7.0.2 2.50 MB 2.2%
Function Drill-Down

Payload Size Graph

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 9.96 MB 50.0%
🔴 functions/__server.func                         [########################] 9.96 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 9.96 MB
Metric Value
Public routes /.well-known/workflow/v1/flow
Runtime nodejs24.x
Handler index.mjs
Payload 9.96 MB
Function files 9.96 MB across 127 files
Traced dependencies 0 B
Signal 🟠 Bundled file _chunks/vercel.web.mjs is 2.03 MB (20.4%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/vercel.web.mjs                         [###########.............] 2.03 MB 20.4%
🟠 _libs/undici.mjs                               [#####...................] 1.01 MB 10.1%
🟡 _chunks/sandbox3.mjs                           [####....................] 815.4 kB 8.2%
🟡 _chunks/compiled-artifacts-instrumentation.mjs [####....................] 762.7 kB 7.7%
🟡 _chunks/token-util-CddTZ_ef.mjs                [####....................] 704.7 kB 7.1%
🔴 Other bundled files                            [########################] 4.63 MB 46.5%

🧾 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 • 16 public routes, 1 internal alias • 9.96 MB
Metric Value
Public routes /
/.well-known/workflow/v1/webhook/[token]
/eve/v1/callback/[token]
/eve/v1/connections/[name]/callback/[attemptId]/[token]
/eve/v1/connections/[name]/callback/[token]
/eve/v1/health
/eve/v1/info
/eve/v1/session
/eve/v1/session/[parentSessionId]/subagents/[callId]/[childSessionId]/stream
/eve/v1/session/[sessionId]
/eve/v1/session/[sessionId]/cancel
/eve/v1/session/[sessionId]/clear
/eve/v1/session/[sessionId]/compact
/eve/v1/session/[sessionId]/reset
/eve/v1/session/[sessionId]/stream
/eve/v1/task-input/[token]
Internal aliases /__server
Runtime nodejs24.x
Handler index.mjs
Payload 9.96 MB
Function files 9.96 MB across 127 files
Traced dependencies 0 B
Signal 🟠 Bundled file _chunks/vercel.web.mjs is 2.03 MB (20.4%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/vercel.web.mjs                         [###########.............] 2.03 MB 20.4%
🟠 _libs/undici.mjs                               [#####...................] 1.01 MB 10.1%
🟡 _chunks/sandbox3.mjs                           [####....................] 815.4 kB 8.2%
🟡 _chunks/compiled-artifacts-instrumentation.mjs [####....................] 762.7 kB 7.7%
🟡 _chunks/token-util-CddTZ_ef.mjs                [####....................] 704.7 kB 7.1%
🔴 Other bundled files                            [########################] 4.63 MB 46.5%

🧾 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: 7.20 s -> 3.47 s (-3.74 s) vs main (89704cb).
  • Timing is informational: shared GitHub runners are too variable for a hard timing budget.
Detailed phase timings vs `main (89704cb)`
Phase Baseline Current Delta
extension.check 0.5 ms 0.5 ms 0.0 ms
project.resolve 0.3 ms 0.3 ms 0.0 ms
workspace.create 0.5 ms 0.5 ms 0.0 ms
host.prepare 4.52 s 820.1 ms -3.70 s
vercel.service-prefix.resolve 1.4 ms 1.5 ms +0.1 ms
nitro.create 562.3 ms 562.7 ms +0.4 ms
sandbox.prewarm 305.6 ms 360.2 ms +54.6 ms
nitro.cache.prepare 0.3 ms 0.3 ms 0.0 ms
nitro.prepare 0.7 ms 0.7 ms 0.0 ms
nitro.public-assets 0.8 ms 0.8 ms 0.0 ms
nitro.prerender 0.5 ms 0.5 ms 0.0 ms
nitro.bundle 1.66 s 1.57 s -91.5 ms
nitro.cache.write 0.2 ms 0.3 ms +0.1 ms
vercel.workflow-function.materialize 70.6 ms 70.7 ms +0.1 ms
agent-summary.emit 0.7 ms 0.7 ms 0.0 ms
connect-manifest.emit 0.3 ms 0.3 ms 0.0 ms
nitro.close 0.1 ms 0.1 ms 0.0 ms
output.publish 3.8 ms 3.7 ms -0.1 ms
workspace.remove 2.6 ms 2.5 ms -0.1 ms

Adds the eval-benchmark workflow: event-derived measurement primitives,
eval artifact timing, the standard per-attempt metric set and gap
classification, the schema_version 1 record contract, registry planning
against the AI Gateway catalog, nightly and manual runs on main, and
immutable publication to private Blob.

Signed-off-by: prha <prha@vercel.com>

This branch was successfully deployed

2 active deployments
Preview – eve-pkg — 33c14917 Deployed Oct 8, 2026 by vercel[bot]
Preview – eve-docs — 33c14917 Deployed Oct 8, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

run-eval-benchmark Run the model benchmark on this PR (temporary, eval-benchmark.yml)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant