fix(eve): keep client context out of session history

[vercel-gh-bot-4]

clientContext now reaches only the intended model call instead of replaying on every later turn. Fixes #2357.

Why

SendTurnOptions.clientContext promises request-scoped context that never enters durable history. The harness appended those messages to the same array later stored as session.history, so every turn retained stale context and increased future prompt size.

A focused reproduction sent three distinct context blocks across three turns. Before this change, each block remained in later prompts. The regression now asserts that each call sees only its own block and that history stores none of them.

Approach

Because the model prompt and durable history shared one array, building a temporary context-bearing projection should preserve the current call while keeping persistence context-free.

The harness now:

• inserts context into a temporary model view before the current delivery
• runs that view through the configured history projector
• sends only durable messages to compaction and session history
• includes context when deciding whether compaction is needed
• clears the compaction token baseline when provider usage includes ephemeral content that durable history cannot represent

Coverage includes repeated turns, custom projectors, tool continuations, automatic compaction, approval resume, and blank delivery messages.

Checks

• tool-loop.test.ts: 195 passed
• pnpm --filter eve run typecheck
• oxfmt --check and git diff --check
• correctness review approved

Two full parallel unit runs passed more than 6,930 tests but hit different 5-second timeouts in four unrelated import-heavy files. All four passed focused reruns.

Requested by: arielweinberger

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
eve-docs	Ready	Preview	Aug 21, 2026 12:39am
eve-docs-4759	Ready	Preview, v0	Aug 21, 2026 12:39am
eve-pkg	Ready	Preview	Aug 21, 2026 12:39am

[github-actions]

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

Key takeaways

• No notable deltas vs main (be9be27).

Delta vs main (be9be27)

Area	Metric	Baseline	Current	Delta
Package	Packed tarball	8.15 MB	8.15 MB	+440 B ⚠️
Package	Unpacked publish size	30.68 MB	30.68 MB	+360 B ⚠️
Package	Installed footprint	74.76 MB	74.76 MB	+360 B ⚠️
Package	Published files	3374	3374	0
Package	Installed files	7085	7085	0
Package	Installed package instances	35	35	0
Package	Distinct installed package names	34	34	0
Package	Installed dependency edges	53	53	0
Package	Installed optional peer edges	47	47	0
Runtime	Unique function payloads	2	2	0
Runtime	Total function bytes	18.10 MB	18.10 MB	+726 B ⚠️
Runtime	Public routes	16	16	0

Changed function payloads vs main (be9be27) (2)

Function	Status	Baseline	Current	Delta	Route changes
functions/__server.func	changed	9.05 MB	9.05 MB	+363 B ⚠️	none
functions/.well-known/workflow/v1/flow.func	changed	9.05 MB	9.05 MB	+363 B ⚠️	none

eve init install

Metric	Baseline	Current	Delta
Installed footprint	113.17 MB	113.17 MB	+360 B ⚠️
Installed packages	121	121	0
dependencies	4	4	0
devDependencies	2	2	0
Dependency package bytes	45.20 MB	45.20 MB	+360 B ⚠️
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: 2026-08-21T00:39:39.462Z
• Route aliases: 16 public, 1 internal (17 total aliases)
• Vercel routes in config: 19
• Severity legend: 🔴 dominant/large, 🟠 notable, 🟡 watch, ⚪ small

Package Drill-Down

Package Details

• Package: eve@0.42.0
• Package directory: packages/eve
• Tarball: 8.15 MB (eve-0.42.0.tgz)
• Unpacked payload: 30.68 MB across 3374 published files
• Installed footprint: 74.76 MB across 7085 installed files
• Installed root package: 29.32 MB
• Installed dependencies: 45.44 MB
• Installed package instances: 35
• Distinct installed package names: 34
• Installed dependency edges: 53
• Installed optional peer edges: 47
• Runtime dependencies: 2
• Peer dependencies: 5 (4 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: 29.32 MB (39.2%)
• @rolldown/binding-linux-x64-gnu: 19.91 MB (26.6%)
• ai: 6.71 MB (9.0%)
• zod: 5.07 MB (6.8%)
• undici: 3.50 MB (4.7%)

Publish payload breakdown

Published file size
🔴 dist/src/compiled/shadcn-registry/index.js       [###############.........] 9.76 MB 31.8%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [####....................] 2.44 MB 8.0%
🟠 dist/src/compiled/experimental-ai-sdk-code-mo... [##......................] 1.51 MB 4.9%
🟡 dist/src/compiled/_chunks/workflow/undici-Dzn... [#.......................] 512.7 kB 1.7%
🟡 dist/src/compiled/_chunks/node/undici-CUIybhQ... [#.......................] 508.8 kB 1.7%
🔴 Other published files                            [########################] 15.95 MB 52.0%

Installed footprint breakdown

Installed package size
🔴 eve                             [########################] 29.32 MB 39.2%
🔴 @rolldown/binding-linux-x64-gnu [################........] 19.91 MB 26.6%
🔴 ai                              [#####...................] 6.71 MB 9.0%
🔴 zod                             [####....................] 5.07 MB 6.8%
🟠 undici                          [###.....................] 3.50 MB 4.7%
🟠 nitro                           [##......................] 2.41 MB 3.2%
🔴 Other installed packages        [######..................] 7.84 MB 10.5%

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: 113.17 MB across 8961 installed files
• Installed packages: 121 total (115 transitive-only)
• dependencies: 4 direct packages totaling 45.20 MB
• devDependencies: 2 direct packages totaling 5.04 MB
• Other transitive package files: 62.93 MB

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

Heavy installed dependencies

• eve: 29.32 MB (25.9%)
• @typescript/typescript-linux-x64: 27.95 MB (24.7%)
• @rolldown/binding-linux-x64-gnu: 19.91 MB (17.6%)
• zod: 9.02 MB (8.0%)
• ai: 6.71 MB (5.9%)

Installed footprint breakdown

Installed package size
🔴 eve                              [########################] 29.32 MB 25.9%
🔴 @typescript/typescript-linux-x64 [#######################.] 27.95 MB 24.7%
🔴 @rolldown/binding-linux-x64-gnu  [################........] 19.91 MB 17.6%
🔴 zod                              [#######.................] 9.02 MB 8.0%
🔴 ai                               [#####...................] 6.71 MB 5.9%
🟠 undici                           [###.....................] 3.50 MB 3.1%
🔴 Other installed packages         [##############..........] 16.75 MB 14.8%

dependencies (4)

Package	Range	Installed size	Share
@vercel/connect	0.4.3	141.0 kB	0.1%
ai	^7.0.58	6.71 MB	5.9%
eve	file:eve-0.42.0.tgz	29.32 MB	25.9%
zod	4.4.3	9.02 MB	8.0%

devDependencies (2)

Package	Range	Installed size	Share
@types/node	24.x	2.54 MB	2.2%
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     [########################] 9.05 MB 50.0%
🔴 functions/__server.func                         [########################] 9.05 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 9.05 MB

Metric	Value
Public routes	/.well-known/workflow/v1/flow
Runtime	nodejs24.x
Handler	index.mjs
Payload	9.05 MB
Function files	9.05 MB across 50 files
Traced dependencies	0 B
Signal	🟠 Bundled file index.mjs is 2.01 MB (22.2%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                     [#################.......] 2.01 MB 22.2%
🟠 _chunks/runtime-artifacts.mjs [###############.........] 1.80 MB 19.9%
🟡 _libs/undici.mjs              [########................] 980.5 kB 10.8%
🟡 _chunks/sandbox.mjs           [#######.................] 781.9 kB 8.6%
🟡 _chunks/frames-Dl2Hcma4.mjs   [######..................] 678.7 kB 7.5%
🟠 Other bundled files           [########################] 2.80 MB 30.9%

🧾 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 • 15 public routes, 1 internal alias • 9.05 MB

Metric	Value
Public routes	/ /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	9.05 MB
Function files	9.05 MB across 50 files
Traced dependencies	0 B
Signal	🟠 Bundled file index.mjs is 2.01 MB (22.2%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                     [#################.......] 2.01 MB 22.2%
🟠 _chunks/runtime-artifacts.mjs [###############.........] 1.80 MB 19.9%
🟡 _libs/undici.mjs              [########................] 980.5 kB 10.8%
🟡 _chunks/sandbox.mjs           [#######.................] 781.9 kB 8.6%
🟡 _chunks/frames-Dl2Hcma4.mjs   [######..................] 678.7 kB 7.5%
🟠 Other bundled files           [########################] 2.80 MB 30.9%

🧾 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: 1.90 s -> 1.78 s (-114.4 ms) vs main (be9be27).
• Timing is informational: shared GitHub runners are too variable for a hard timing budget.

Detailed phase timings vs `main (be9be27)`

Phase	Baseline	Current	Delta
extension.check	9.7 ms	8.7 ms	-1.0 ms
project.resolve	0.3 ms	0.3 ms	0.0 ms
workspace.create	0.6 ms	0.6 ms	0.0 ms
host.prepare	332.9 ms	227.7 ms	-105.2 ms
vercel.service-prefix.resolve	1.5 ms	1.7 ms	+0.2 ms
nitro.create	179.8 ms	180.4 ms	+0.6 ms
sandbox.prewarm	214.9 ms	211.0 ms	-3.9 ms
nitro.cache.prepare	0.2 ms	0.2 ms	0.0 ms
nitro.prepare	0.5 ms	0.6 ms	+0.1 ms
nitro.public-assets	0.6 ms	0.6 ms	0.0 ms
nitro.prerender	0.4 ms	0.4 ms	0.0 ms
nitro.bundle	1.10 s	1.10 s	+1.0 ms
nitro.cache.write	0.2 ms	0.2 ms	0.0 ms
vercel.workflow-function.materialize	48.8 ms	42.7 ms	-6.1 ms
agent-summary.emit	0.4 ms	0.4 ms	0.0 ms
nitro.close	0.1 ms	0.1 ms	0.0 ms
output.publish	2.5 ms	2.4 ms	-0.1 ms
workspace.remove	1.5 ms	1.4 ms	-0.1 ms

[arielweinberger]

Reporter of #2357 here. Read through the diff, and the approach looks right to me: keeping the durable array context-free and building a separate model-facing projection is a cleaner fix than filtering at persist time, and the regression coverage lines up with what I reported (no replay across turns, context before the delivery message, nothing in session.history).

Two observations, neither a blocker.

1. The compaction token baseline is never established for clients that send context on every turn.

durableModelPromptMessageCount is undefined whenever contextMessages.length > 0, and createNextCompactionConfig only records usage when it is defined:

if (result.usage?.inputTokens !== undefined && durablePromptMessageCount !== undefined) {
  next.lastKnownInputTokens = result.usage.inputTokens;
  next.lastKnownPromptMessageCount = durablePromptMessageCount;
}

getInputTokenCount then takes the fallback branch for the life of the session:

return r === undefined || i === undefined || ...
  ? estimateTokens(e)
  : r + estimateTokens(e.slice(i));

For the use case that motivated the issue, an app attaching page or editor context to every message, clientContext is present on every turn, so the provider-reported input token count is never recorded once and compaction runs entirely on local estimation instead of an anchored count plus delta. Sessions that only sometimes attach context get a related but different shape: the anchor is retained from an earlier context-free turn and goes stale rather than being cleared.

I understand the reasoning, since the reported usage covers ephemeral content that durable history cannot reproduce, and skipping is clearly safer than recording a number that overstates durable size. It may still be worth deciding explicitly, because "always-on context" is probably the common case for this field rather than an edge case. One option that keeps an anchored count without overstating anything is to record the durable prefix length together with the usage attributable to it, if that is derivable, and otherwise this is worth a comment saying estimation is the intended path for these sessions.

2. Small unstated invariant in projectModelMessages.

const insertionIndex = deliveryMessage === undefined ? durableMessages.length : -1;

The -1 splice assumes the delivery message is the last entry of durableMessages. That holds at the first call site, and at the post-compaction call it depends on compaction preserving the tail. That is true today, so this is a note rather than a defect, but an assertion or a comment naming the invariant would make it harder to break later.

Thanks for turning this around so quickly.