Skip to content

feat(eve): step state - persist delta returns for workflow replay - #3507

Open
AndrewBarba wants to merge 1 commit into
mainfrom
barba/step-state-deltas
Open

AndrewBarba wants to merge 1 commit into
mainfrom
barba/step-state-deltas

Conversation

@AndrewBarba

Copy link
Copy Markdown
Collaborator

Summary

Accumulating conversations repeat their full history and context in every turn-step input and output. Return state deltas and apply them in the owning workflow during normal replay, preserving cancellation and background-task checkpoints. Full inputs remain persisted; the 32-step serializer benchmark reduces combined state bytes by 47.8%. The state argument is separated from streams, signals, and delivery so adopting replay-derived inputs becomes turnStep.replayInputs = [0] after upgrading Workflow.

Deltas capture state before authored mutations and represent nested edits, appends, removals, and truncation. Aliased/cyclic subtrees travel together; rich values retain full replacement through Workflow's serializer. Initialization, handoffs, and other coordination steps keep their existing checkpoints. No new state store or runtime history retrieval is introduced. Total storage and diff work can still grow quadratically while full inputs remain.

Validation

  • pnpm build
  • pnpm test:unit; final session tests: pnpm --filter eve exec vitest run --config vitest.unit.config.ts src/execution/session (288 passed)
  • pnpm test:integration (885 passed); final serializer tests: pnpm --filter eve exec vitest run --config vitest.integration.config.ts src/execution/session/turn-step-delta.integration.test.ts (3 passed)
  • pnpm typecheck; final package check: pnpm --filter eve exec tsc -p tsconfig.json --noEmit
  • pnpm fmt, pnpm lint, pnpm guard:invariants, pnpm docs:check
  • Real session replay with Node's retained VM enabled/disabled, stored-output inspection, in-place authored state mutation, cancellation/background checkpoints, handoff/import coverage, interrupted-input retry simulation, and rich-value serialization. eve currently excludes the QuickJS runtime.
  • Expanded the existing reconnect e2e eval; e2e runs only in CI.

Fixed 2 KiB history additions, actual Workflow serialization with compression disabled (state payload bytes, excluding streams/events):

Steps Persisted inputs Full output baseline Delta outputs Combined reduction
8 61,891 78,497 20,377 41.4%
16 257,540 290,721 40,786 45.6%
32 1,050,116 1,116,449 81,618 47.8%

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)

Signed-off-by: Andrew Barba <barba@hey.com>
@vercel

vercel Bot commented Sep 18, 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 Sep 18, 2026 11:51am UTC
eve-pkg Ready Ready Preview, v0 Sep 18, 2026 11:51am UTC

@github-actions

Copy link
Copy Markdown
Contributor

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

Key takeaways

  • No notable deltas vs main (505601b).

Delta vs main (505601b)

Area Metric Baseline Current Delta
Package Packed tarball 8.77 MB 8.77 MB +4.0 kB ⚠️
Package Unpacked publish size 32.74 MB 32.75 MB +7.0 kB ⚠️
Package Installed footprint 78.82 MB 78.83 MB +7.0 kB ⚠️
Package Published files 4018 4022 +4
Package Installed files 7978 7982 +4
Package Installed package instances 33 33 0
Package Distinct installed package names 32 32 0
Package Installed dependency edges 51 51 0
Package Installed optional peer edges 8 8 0
Runtime Unique function payloads 2 2 0
Runtime Total function bytes 20.39 MB 20.44 MB +54.5 kB ⚠️
Runtime Public routes 18 18 0
Changed function payloads vs main (505601b) (2)
Function Status Baseline Current Delta Route changes
functions/__server.func changed 10.19 MB 10.22 MB +27.2 kB ⚠️ none
functions/.well-known/workflow/v1/flow.func changed 10.19 MB 10.22 MB +27.2 kB ⚠️ none

eve init install

Metric Baseline Current Delta
Installed footprint 117.05 MB 117.05 MB +7.0 kB ⚠️
Installed packages 113 113 0
dependencies 4 4 0
devDependencies 2 2 0
Dependency package bytes 49.41 MB 49.42 MB +7.0 kB ⚠️
devDependency package bytes 5.04 MB 5.04 MB 0 B ➖
Build Metadata
  • Preset: vercel
  • Nitro: nitro@3.0.260903-beta
  • Output directory: apps/fixtures/weather-agent/.vercel/output
  • Build metadata timestamp: 2026-09-18T11:51:35.626Z
  • Route aliases: 18 public, 1 internal (19 total aliases)
  • Vercel routes in config: 21
  • Severity legend: 🔴 dominant/large, 🟠 notable, 🟡 watch, ⚪ small
Package Drill-Down

Package Details

  • Package: eve@0.60.0
  • Package directory: packages/eve
  • Tarball: 8.77 MB (eve-0.60.0.tgz)
  • Unpacked payload: 32.75 MB across 4022 published files
  • Installed footprint: 78.83 MB across 7982 installed files
  • Installed root package: 31.28 MB
  • Installed dependencies: 47.55 MB
  • Installed package instances: 33
  • Distinct installed package names: 32
  • Installed dependency edges: 51
  • Installed optional peer edges: 8
  • Runtime dependencies: 2
  • Peer dependencies: 6 (5 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: 31.28 MB (39.7%)
  • @rolldown/binding-linux-x64-gnu: 19.10 MB (24.2%)
  • ai: 7.58 MB (9.6%)
  • zod: 6.76 MB (8.6%)
  • undici: 3.51 MB (4.5%)
Publish payload breakdown
Published file size
🔴 dist/src/compiled/shadcn-registry/index.js       [#############...........] 9.70 MB 29.6%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [###.....................] 2.29 MB 7.0%
🟠 dist/src/compiled/@ai-sdk/code-mode/index.js     [#.......................] 1.03 MB 3.1%
🟡 dist/src/compiled/@vercel/blob/index.js          [#.......................] 902.9 kB 2.8%
🟡 dist/src/compiled/_chunks/workflow/signal-exi... [#.......................] 514.5 kB 1.6%
🔴 Other published files                            [########################] 18.31 MB 55.9%
Installed footprint breakdown
Installed package size
🔴 eve                             [########################] 31.28 MB 39.7%
🔴 @rolldown/binding-linux-x64-gnu [###############.........] 19.10 MB 24.2%
🔴 ai                              [######..................] 7.58 MB 9.6%
🔴 zod                             [#####...................] 6.76 MB 8.6%
🟠 undici                          [###.....................] 3.51 MB 4.5%
🟠 nitro                           [#.......................] 1.89 MB 2.4%
🔴 Other installed packages        [#######.................] 8.71 MB 11.0%
Runtime dependencies (2)
Package Range Notes
nitro 3.0.260903-beta
undici 8.9.0
Peer dependencies (6)
Package Range Notes
@opentelemetry/api ^1.0.0 optional peer
ai catalog:
braintrust ^3.0.0 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: 117.05 MB across 9876 installed files
  • Installed packages: 113 total (107 transitive-only)
  • dependencies: 4 direct packages totaling 49.42 MB
  • devDependencies: 2 direct packages totaling 5.04 MB
  • Other transitive package files: 62.59 MB

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

Heavy installed dependencies

  • eve: 31.28 MB (26.7%)
  • @typescript/typescript-linux-x64: 27.95 MB (23.9%)
  • @rolldown/binding-linux-x64-gnu: 19.10 MB (16.3%)
  • zod: 10.37 MB (8.9%)
  • ai: 7.58 MB (6.5%)
Installed footprint breakdown
Installed package size
🔴 eve                              [########################] 31.28 MB 26.7%
🔴 @typescript/typescript-linux-x64 [#####################...] 27.95 MB 23.9%
🔴 @rolldown/binding-linux-x64-gnu  [###############.........] 19.10 MB 16.3%
🔴 zod                              [########................] 10.37 MB 8.9%
🔴 ai                               [######..................] 7.58 MB 6.5%
🟠 undici                           [###.....................] 3.51 MB 3.0%
🔴 Other installed packages         [#############...........] 17.26 MB 14.7%
dependencies (4)
Package Range Installed size Share
@vercel/connect 2.2.0 188.7 kB 0.2%
ai ^7.0.105 7.58 MB 6.5%
eve file:eve-0.60.0.tgz 31.28 MB 26.7%
zod 4.5.4 10.37 MB 8.9%
devDependencies (2)
Package Range Installed size Share
@types/node 24.x 2.54 MB 2.2%
typescript 7.0.2 2.50 MB 2.1%
Function Drill-Down

Payload Size Graph

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 10.22 MB 50.0%
🔴 functions/__server.func                         [########################] 10.22 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 10.22 MB
Metric Value
Public routes /.well-known/workflow/v1/flow
Runtime nodejs24.x
Handler index.mjs
Payload 10.22 MB
Function files 10.22 MB across 109 files
Traced dependencies 0 B
Signal 🟠 Bundled file index.mjs is 2.11 MB (20.7%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                        [##########..............] 2.11 MB 20.7%
🟡 _libs/undici.mjs                 [#####...................] 980.8 kB 9.6%
🟡 _chunks/sandbox.mjs              [####....................] 809.9 kB 7.9%
🟡 _chunks/chatgpt-model.mjs        [###.....................] 697.6 kB 6.8%
🟡 _chunks/signal-exit-B9U6kH7R.mjs [###.....................] 616.2 kB 6.0%
🔴 Other bundled files              [########################] 5.00 MB 48.9%

🧾 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 • 17 public routes, 1 internal alias • 10.22 MB
Metric Value
Public routes /
/.well-known/workflow/v1/webhook/[token]
/eve/v1/activity/[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 10.22 MB
Function files 10.22 MB across 109 files
Traced dependencies 0 B
Signal 🟠 Bundled file index.mjs is 2.11 MB (20.7%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                        [##########..............] 2.11 MB 20.7%
🟡 _libs/undici.mjs                 [#####...................] 980.8 kB 9.6%
🟡 _chunks/sandbox.mjs              [####....................] 809.9 kB 7.9%
🟡 _chunks/chatgpt-model.mjs        [###.....................] 697.6 kB 6.8%
🟡 _chunks/signal-exit-B9U6kH7R.mjs [###.....................] 616.2 kB 6.0%
🔴 Other bundled files              [########################] 5.00 MB 48.9%

🧾 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: 4.04 s -> 4.03 s (-15.2 ms) vs main (505601b).
  • Timing is informational: shared GitHub runners are too variable for a hard timing budget.
Detailed phase timings vs `main (505601b)`
Phase Baseline Current Delta
extension.check 0.3 ms 0.4 ms +0.1 ms
project.resolve 0.2 ms 0.2 ms 0.0 ms
workspace.create 0.4 ms 0.4 ms 0.0 ms
host.prepare 363.4 ms 406.4 ms +43.0 ms
vercel.service-prefix.resolve 1.2 ms 1.1 ms -0.1 ms
nitro.create 316.4 ms 320.9 ms +4.5 ms
sandbox.prewarm 337.0 ms 333.2 ms -3.8 ms
nitro.cache.prepare 0.2 ms 0.2 ms 0.0 ms
nitro.prepare 0.6 ms 0.5 ms -0.1 ms
nitro.public-assets 0.6 ms 0.5 ms -0.1 ms
nitro.prerender 0.3 ms 0.3 ms 0.0 ms
nitro.bundle 2.89 s 2.85 s -36.6 ms
nitro.cache.write 0.2 ms 0.3 ms +0.1 ms
vercel.workflow-function.materialize 65.2 ms 107.7 ms +42.5 ms
agent-summary.emit 0.6 ms 0.6 ms 0.0 ms
nitro.close 0.1 ms 0.1 ms 0.0 ms
output.publish 67.1 ms 2.9 ms -64.2 ms
workspace.remove 1.4 ms 1.2 ms -0.2 ms

}

/** Copy-on-write keeps the pre-step checkpoint available for cancellation and background work. */
export function applyValueDelta(before: unknown, delta: ValueDelta): unknown {

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.

What do you think of fuzzy testing this module APIs eg with fast-check? Ran this and found a bug in diff when reversing props

#3509

@ruiconti ruiconti left a comment

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.

Two performance observations from rerunning the serializer benchmark at this revision. These concern the scope of the storage savings and the evidence for net benefit; neither establishes another state-corruption defect.

(!isRecord(current) && !Array.isArray(current)) ||
Object.getOwnPropertySymbols(current).length > 0
) {
atomic.add("[]");

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.

One unchanged rich value disables delta savings for the entire projected state. atomic.add("[]") marks the root, so a Date anywhere in serialized context causes the context, session history, and metadata to be replaced together. This is broader than replacing only the rich subtree.

I reran the existing 32-step, 2 KiB append benchmark with compression disabled, adding only createdAt: new Date(0) to the initial serialized context and retaining it in subsequent results:

Case Input bytes Full-output bytes Delta-output bytes Combined reduction
Original plain state 1,050,116 1,116,449 81,618 47.76%
With one unchanged Date 1,051,907 1,118,209 1,119,905 -0.08%

All 32 deltas in the Date case were root replacements. The fallback preserves correctness, but this should be an explicit benchmark and documented optimization boundary. If narrowing the fallback, preserve aliases across the rich value and the rest of the graph.

globalThis,
false,
false,
false,

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.

The uncompressed result demonstrates reduced duplication, but it does not yet establish the production storage/computation tradeoff. I reran this same 32-step fixture with compression enabled for both step arguments and return values:

Compression Input bytes Full-output bytes Delta-output bytes Combined reduction
Disabled 1,050,116 1,116,449 81,618 47.76%
Enabled 14,127 13,893 9,620 15.25%

The fixture repeats "x" heavily, so neither percentage should be treated as an estimate for real conversations. Capture and diff also traverse accumulated state, adding computation that the byte benchmark does not measure.

Please add compressed measurements using representative conversation payloads, plus execution-time and peak-memory measurements for the full-result and delta paths. The decision criterion is storage saved versus added computation and maintenance cost, while preserving serialization behavior.

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