Skip to content

feat(eve): subagent sessions inherit the parent's remaining token quota#555

Merged
ruiconti merged 4 commits into
mainfrom
rui/subagent-quota-inheritance
Jul 7, 2026
Merged

feat(eve): subagent sessions inherit the parent's remaining token quota#555
ruiconti merged 4 commits into
mainfrom
rui/subagent-quota-inheritance

Conversation

@ruiconti

@ruiconti ruiconti commented Jul 6, 2026

Copy link
Copy Markdown
Contributor

Delegated subagent sessions got a fixed 5M input-token cap no matter what the delegating parent had left to spend, and a completed child's spend never counted against the parent's quota. Both halves were wrong in opposite directions: a parent near its limit could grant a child 5M it didn't have, and a parent nowhere near its limit could fan out children whose combined spend was unbounded by the root's configured budget. #488 surfaced child usage to callers; this PR closes the loop and makes the session quota coherent across a delegation tree.

What

  • Subagent dispatch computes the parent's remaining quota (configured limit minus accumulated session usage, per axis) and grants each child in the batch an even share of it: floor(remaining / delegationFanOut), where the fan-out counts the batch's local subagent calls. One batch is collectively bounded by the parent's remainder — N parallel children cannot each spend the full amount. The grant travels RunInput.subagentTokenBudget -> eve.subagentTokenBudget serialized context key -> child session creation, mirroring the existing subagentDepth plumbing. The fixed 5M default is gone.
  • A completed child's usage folds into the parent's session totals on resume (resolvePendingRuntimeActions), so the parent's limit check and the shares granted to later batches account for what the tree already spent. Turn totals stay untouched: child spend is attributed by the caller-side invoke_agent span from feat(eve): surface subagent usage to callers #488, so turn tags don't double-count.
  • maxInputTokensPerSession / maxOutputTokensPerSession accept false to uncap an axis explicitly. An uncapped parent delegates uncapped children regardless of fan-out; an authored child limit applies only when tighter than the parent's grant, and a child authoring false still gets the grant. A child can restrict itself, never exempt itself.
  • Durable sessions rehydrate persisted limits verbatim instead of re-resolving defaults, so an uncapped session stays uncapped across rehydration. Persisted limits are already resolved at creation; re-applying the root default on hydration would silently re-cap.

Why this shape

The alternative was keeping a smaller fixed subagent default. Any fixed number is wrong twice: too generous when the parent is nearly spent, too stingy for a deep task under a large budget. The parent's remainder split across the batch is the only grant with grounding at dispatch time, and #488 made it computable. This also moves toward #551 (uncapped sessions as the default) without touching the root 40M default — that removal stays with the research plan.

Remote-agent calls do not count toward the fan-out: they run on their own deployment under their own limits, so they never dilute the local shares.

Known slack

  • A child that underspends its share doesn't return the surplus to siblings in the same batch; the surplus becomes visible to the next batch after fold-back.
  • Failed child runs carry no usage (pre-existing in the result contract, see delegated-parent-notification.ts), so an errored child's spend doesn't draw down the parent.

Review notes

Rebased over #528 (maxSubagents) with one trivial conflict in the manifest limits schema. session.test.ts hydration expectations changed with the verbatim-rehydration semantics: no stored limits now means uncapped, not re-defaulted.

@vercel

vercel Bot commented Jul 6, 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 7, 2026 1:15am
eve-docs-1644 Ready Ready Preview, Comment, Open in v0 Jul 7, 2026 1:15am
eve-docs-4759 Ready Ready Preview, Comment, Open in v0 Jul 7, 2026 1:15am

@github-actions

github-actions Bot commented Jul 6, 2026

Copy link
Copy Markdown
Contributor

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

Key takeaways

  • No notable deltas vs main (5ef4ec6).

Delta vs main (5ef4ec6)

Area Metric Baseline Current Delta
Package Packed tarball 3.53 MB 3.53 MB +2.4 kB ⚠️
Package Unpacked publish size 12.81 MB 12.82 MB +8.4 kB ⚠️
Package Installed footprint 51.90 MB 51.91 MB +8.4 kB ⚠️
Package Published files 2437 2441 +4
Package Installed files 5664 5668 +4
Runtime Unique function payloads 2 2 0
Runtime Total function bytes 9.24 MB 9.25 MB +2.2 kB ⚠️
Runtime Public routes 9 9 0
Changed function payloads vs main (5ef4ec6) (2)
Function Status Baseline Current Delta Route changes
functions/.well-known/workflow/v1/flow.func changed 5.76 MB 5.77 MB +2.5 kB ⚠️ none
functions/__server.func changed 3.48 MB 3.48 MB -244 B ✅ none

eve init install

Metric Baseline Current Delta
Installed footprint 91.69 MB 91.69 MB +8.4 kB ⚠️
Installed packages 120 120 0
dependencies 4 4 0
devDependencies 2 2 0
Dependency package bytes 26.97 MB 26.98 MB +8.4 kB ⚠️
devDependency package bytes 4.81 MB 4.81 MB 0 B ➖
Build Metadata
  • Preset: vercel
  • Nitro: nitro@3.0.260610-beta
  • Output directory: apps/fixtures/weather-agent/.vercel/output
  • Build metadata timestamp: 2026-07-07T01:15:50.285Z
  • Route aliases: 9 public, 1 internal (10 total aliases)
  • Vercel routes in config: 10
  • Severity legend: 🔴 dominant/large, 🟠 notable, 🟡 watch, ⚪ small
Package Drill-Down

Package Details

  • Package: eve@0.20.0
  • Package directory: packages/eve
  • Tarball: 3.53 MB (eve-0.20.0.tgz)
  • Unpacked payload: 12.82 MB across 2441 published files
  • Installed footprint: 51.91 MB across 5668 installed files
  • Installed root package: 11.54 MB
  • Installed dependencies: 40.36 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

  • @rolldown/binding-linux-x64-gnu: 19.18 MB (36.9%)
  • eve: 11.54 MB (22.2%)
  • ai: 6.34 MB (12.2%)
  • zod: 5.04 MB (9.7%)
  • nitro: 2.41 MB (4.6%)
Publish payload breakdown
Published file size
🟠 dist/src/compiled/experimental-ai-sdk-code-mo... [####....................] 1.51 MB 11.7%
🟡 dist/src/compiled/@workflow/core/runtime.js      [##......................] 794.4 kB 6.2%
🟡 dist/src/compiled/@vercel/sandbox/index.js       [##......................] 632.2 kB 4.9%
🟡 dist/src/compiled/@chat-adapter/slack/index.js   [#.......................] 433.8 kB 3.4%
🟡 dist/src/compiled/_chunks/workflow/attribute-... [#.......................] 371.6 kB 2.9%
🔴 Other published files                            [########################] 9.08 MB 70.8%
Installed footprint breakdown
Installed package size
🔴 @rolldown/binding-linux-x64-gnu [########################] 19.18 MB 36.9%
🔴 eve                             [##############..........] 11.54 MB 22.2%
🔴 ai                              [########................] 6.34 MB 12.2%
🔴 zod                             [######..................] 5.04 MB 9.7%
🟠 nitro                           [###.....................] 2.41 MB 4.6%
🟡 rolldown                        [#.......................] 771.7 kB 1.5%
🔴 Other installed packages        [########................] 6.63 MB 12.8%
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: 91.69 MB across 7420 installed files
  • Installed packages: 120 total (114 transitive-only)
  • dependencies: 4 direct packages totaling 26.98 MB
  • devDependencies: 2 direct packages totaling 4.81 MB
  • Other transitive package files: 59.90 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: 29.69 MB (32.4%)
  • @rolldown/binding-linux-x64-gnu: 19.18 MB (20.9%)
  • eve: 11.54 MB (12.6%)
  • zod: 9.00 MB (9.8%)
  • ai: 6.34 MB (6.9%)
Installed footprint breakdown
Installed package size
🔴 @typescript/typescript-linux-x64 [########################] 29.69 MB 32.4%
🔴 @rolldown/binding-linux-x64-gnu  [###############.........] 19.18 MB 20.9%
🔴 eve                              [#########...............] 11.54 MB 12.6%
🔴 zod                              [#######.................] 9.00 MB 9.8%
🔴 ai                               [#####...................] 6.34 MB 6.9%
🟠 @types/node                      [##......................] 2.54 MB 2.8%
🔴 Other installed packages         [###########.............] 13.40 MB 14.6%
dependencies (4)
Package Range Installed size Share
@vercel/connect 0.2.2 98.6 kB 0.1%
ai ^7.0.0 6.34 MB 6.9%
eve file:eve-0.20.0.tgz 11.54 MB 12.6%
zod 4.4.3 9.00 MB 9.8%
devDependencies (2)
Package Range Installed size Share
@types/node 24.x 2.54 MB 2.8%
typescript 7.0.1-rc 2.27 MB 2.5%
Function Drill-Down

Payload Size Graph

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 5.77 MB 62.4%
🟠 functions/__server.func                         [##############..........] 3.48 MB 37.6%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 5.77 MB
Metric Value
Public routes /.well-known/workflow/v1/flow
Runtime nodejs24.x
Handler index.mjs
Payload 5.77 MB
Function files 5.77 MB across 26 files
Traced dependencies 0 B
Signal 🟠 Bundled file __eve_nitro_handler__.mjs is 1.99 MB (34.6%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 __eve_nitro_handler__.mjs              [########################] 1.99 MB 34.6%
🟠 _chunks/runtime.mjs                    [############............] 983.9 kB 17.1%
🟡 _chunks/sandbox.mjs                    [#########...............] 766.3 kB 13.3%
🟡 _chunks/attribute-changes-zAifvEhb.mjs [######..................] 473.4 kB 8.2%
🟡 _libs/@ai-sdk/gateway+[...].mjs        [#####...................] 413.5 kB 7.2%
🟠 Other bundled files                    [##############..........] 1.14 MB 19.7%

🧾 Vercel Config

{
  "handler": "index.mjs",
  "launcherType": "Nodejs",
  "shouldAddHelpers": false,
  "supportsResponseStreaming": true,
  "runtime": "nodejs24.x",
  "environment": {
    "NODE_OPTIONS": "--experimental-require-module"
  },
  "maxDuration": "max",
  "experimentalTriggers": [
    {
      "type": "queue/v2beta",
      "topic": "__eve776561746865722d6167656e74_wkf_workflow_*",
      "consumer": "default",
      "retryAfterSeconds": 5,
      "initialDelaySeconds": 0
    }
  ]
}

🟠 functions/__server.func • 8 public routes, 1 internal alias • 3.48 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]/stream
Internal aliases /__server
Runtime nodejs24.x
Handler index.mjs
Payload 3.48 MB
Function files 3.48 MB across 21 files
Traced dependencies 0 B
Signal 🟠 Bundled file _chunks/dev-runtime-artifacts.mjs is 914.5 kB (26.3%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/dev-runtime-artifacts.mjs      [########################] 914.5 kB 26.3%
🟠 _chunks/runtime.mjs                    [#######################.] 891.6 kB 25.6%
🟠 _chunks/sandbox.mjs                    [####################....] 766.3 kB 22.0%
🟡 _chunks/attribute-changes-zAifvEhb.mjs [############............] 449.0 kB 12.9%
⚪ _libs/zod.mjs                          [###.....................] 114.2 kB 3.3%
🟡 Other bundled files                    [#########...............] 343.4 kB 9.9%

🧾 Vercel Config

{
  "handler": "index.mjs",
  "launcherType": "Nodejs",
  "shouldAddHelpers": false,
  "supportsResponseStreaming": true,
  "runtime": "nodejs24.x"
}

Delegated subagent sessions previously received a fixed 5M input-token
cap regardless of what the delegating parent had left to spend, and a
child's spend never counted against the parent's quota — a parent could
fan out children whose combined spend was unbounded by the root's
configured limit.

Three changes make the session quota coherent across a delegation tree:

- Dispatch computes the parent's remaining quota (limit minus
  accumulated usage) and grants it to the child as its budget
  (RunInput.subagentTokenBudget -> eve.subagentTokenBudget -> child
  session creation). An authored child limit applies only when tighter;
  a child cannot self-exempt from the parent's grant. The fixed 5M
  default is removed: uncapped parents delegate uncapped children.
- Completed child usage (surfaced by #488) folds into the parent's
  session totals on resume, so limits and later grants account for what
  the tree already spent. Turn totals stay untouched — child spend is
  attributed by the caller-side invoke_agent span, not the turn tags.
- Session token limits accept false to uncap an axis explicitly, and
  durable sessions rehydrate persisted limits verbatim so an uncapped
  session stays uncapped instead of re-acquiring the root default.

Known slack: parallel children each receive the remaining quota at
their own dispatch (overshoot bounded by fan-out), and failed child
runs carry no usage (pre-existing in the result contract).

Signed-off-by: Rui Conti <ruiconti@gmail.com>
The parent's remaining quota previously traveled as a bespoke context
key (eve.subagentTokenBudget) serialized into the child's run context
and defensively re-parsed at session creation. Carry it instead as a
typed RunInput.limits field forwarded explicitly through the workflow
entry payload, reusing the limit field names authors already configure
(maxInputTokensPerSession / maxOutputTokensPerSession) with false
marking an uncapped axis. The serialization round-trip, its validation
helpers, and the SubagentTokenBudget type all go away.

Capping now happens in createSessionStep where the inherited limits
meet the resolved agent config; createSession no longer knows about
delegation. Subagent dispatch starts child runs via a new
startWorkflowRuntimeRun instead of constructing a throwaway Runtime
per action, since the run input (not runtime config) now carries the
per-child limits.

Behavior is unchanged: tighter of configured vs inherited wins, an
authored false cannot self-exempt from an inherited cap, and uncapped
parents still delegate uncapped children.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
RunSessionLimits grows maxSubagentDepth and maxSubagents, replacing the
standalone RunInput.subagentMaxDepth field so every inherited cap rides
the same typed limits channel as the token quota. A new
WorkflowMaxSubagentsKey carries the parent's effective Workflow
fan-out cap into delegated children: both dispatch paths and turnStep
resolve the child's configured limits.maxSubagents against the
inherited value, tighter wins, so a delegated child cannot run wider
Workflow fan-outs than the tree that spawned it allows.

The resolution helpers move to execution/run-session-limits.ts:
resolveInheritedTokenLimit (from create-session-step; false only
uncaps before inheritance, an inherited cap still binds) and the new
resolveInheritedCountLimit.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
Inherited limits now have one transport (the typed workflow-entry
payload) and one resolution point (createSessionStep), with effective
values materialized on the session:

- workflowMaxSubagents joins subagentMaxDepth as session state,
  resolved at creation and read from the session by turnStep and both
  dispatch steps. The WorkflowMaxSubagentsKey context channel and the
  per-step re-resolution against config go away, as does node-step's
  hidden config fallback.
- The SubagentMaxDepthKey context channel was a zombie after the
  payload migration: written into the child context and parsed on
  entry, but the parsed value was discarded. Key, write, and parse
  deleted; readSerializedSubagentSessionDepth narrows to
  readSerializedSubagentDepth.
- Behavior change: maxSubagentDepth now resolves min(configured,
  inherited) like every other axis, instead of inherited-wins, so a
  child agent can tighten its subtree's delegation depth. Each
  resolution site documents its axis semantics.
- workflow-runtime.ts returns to main's shape; the
  startWorkflowRuntimeRun extraction had one caller left after
  dispatch went back to createWorkflowRuntime. Its only remaining
  delta vs main is forwarding RunInput.limits into the entry payload.

Signed-off-by: Rui Conti <ruiconti@gmail.com>
@ruiconti ruiconti merged commit da2ec6c into main Jul 7, 2026
41 checks passed
@ruiconti ruiconti deleted the rui/subagent-quota-inheritance branch July 7, 2026 15:10
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