[OwenKephart]

Summary

The api fleet's new parents could not use deployed eve@0.83.2 remote agents. This draft makes that pairing work end to end, and fixes a hang in concurrent ctx.agent sends. Three changes, each in its own commit:

1. Protocol negotiation. An eve@0.83.2 receiver rejects protocol 3 before admission and reports protocol 2. Negotiate once after that explicit refusal, with fresh credentials and the same callback and create-operation key. Never negotiate after acceptance, an ambiguous receipt, a rejection naming a session, or an unsupported version.
2. Protocol-2 human input (new parent → 0.83.2 child). The child's approvals, questions and sign-ins used to be refused with HTTP 400, failing its turn. They now reach the parent's person, and answers go back in the form the child accepts:
   • The two protocol-2 callback kinds are read as v27 relay facts, using the same v26 mappings the history reader applies (moved to protocol/v26-interactions.ts).
   • Answers go to the child as v26 inputResponses (requestId), keeping the forwarded responder. earlierProtocol now survives on the relay binding and in stored relay state, so replays keep the right form.
   • The child still settles its own approvals and sign-ins, and still decides who may answer. An answer it turns away leaves the approval open.
   • The child never reports how a question or budget prompt settles, so the parent settles those once the child accepts the answer. A Stop already settled by the parent is not settled twice.
   • The parent marks the callback URL it negotiates (eveRemoteProtocol=2) and reads protocol-2 input only on a marked URL. A parent run from before v27 would drop the translated relay. Its child's input therefore still fails visibly instead of waiting unseen.
   • Protocol 1 task callbacks are still refused.
3. Concurrent send ordering. Concurrent ctx.agent sends registered their reply waiters in send order, but deliveries could reach the child in another order. The newest call could then stay unsettled, leaving the serve loop waiting forever. Deliveries are now serialized per session handle. Turns are not, so steering still joins the running turn. Each send's auth and attribution are captured at send time.

This draft targets #4855's v27/lab-fixes branch. It does not add child-stream access for earlier peers, human input for a protocol-2 caller of a v27 child, a rollout, or a change to the api fleet's approved SDK pin. No issue was created.

Validation

• Real mixed-version loopback. A parent on this build and a child on the published eve@0.83.2 package, both as isolated CLI servers. Deterministic fixture models; no external tools, credentials or production state.
  • Approve: the parent mirrors the child's approval with its asked call (guard), and settles accepted. The child's tool runs as the forwarded responder (current/initiator: alice).
  • Deny: settles declined, and the child's tool reports "Tool execution was denied." without running.
  • Question: the child's ask_question receives {"answer":"iad1","status":"answered"}, and the parent's mirror settles accepted.
  • Cancel while the child waits: the mirror is withdrawn (asker-ended), the call is interrupted, and nothing is approved.
  • Before this change, the same approval failed TASK_FAILED with Remote input callback failed with HTTP 400.
• Regression check on the same build. Terminal results in both directions pass. Current-protocol scenarios pass: authorized child reads, Alice → Bob task reuse with Alice kept as initiator, a v27 approval relayed and answered through the parent, and turn-scoped cancellation.
• Unit tests.
  • Translation tests run protocol-2 bodies through the real relay/routeAnswer transitions.
  • Route tests cover marked, unmarked, malformed and protocol-1 bodies.
  • Also covered: the inputResponses answer wire, URL marking during negotiation and continuation, and stored-relay protocol persistence.
  • Ordering regressions reproduce the hang (admission seed → right → left) on the base.
• Results:
  • Full unit tier: 851 files, 8,103 passed, 1 skipped.
  • Integration: 2,099 passed. Its one failure is a model-auth test that only fails because the sandbox sets ANTHROPIC_BASE_URL; it passes 4/4 with that unset.
  • Scenario: 513 passed. Its 5 failures (eve init, Nitro pack, real-shell bash) fail identically on the untouched base 28de6fc8f.
  • Typecheck, scoped oxfmt/oxlint and docs:check (109 pages) are clean.
• This is not a deployed v/Slack journey.

Checklist

• 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]

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

Project	Deployment	Actions	Updated
eve-docs	Ready	Preview, v0	Oct 26, 2026 5:33pm UTC
eve-pkg	Ready	Preview, v0	Oct 26, 2026 5:33pm UTC

[github-actions]

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

Key takeaways

• No notable deltas vs v27/lab-fixes (28de6fc).

Delta vs v27/lab-fixes (28de6fc)

Area	Metric	Baseline	Current	Delta
Package	Packed tarball	7.82 MB	7.83 MB	+5.3 kB ⚠️
Package	Unpacked publish size	27.29 MB	27.31 MB	+19.5 kB ⚠️
Package	Installed footprint	75.61 MB	75.63 MB	+19.5 kB ⚠️
Package	Published files	4356	4360	+4
Package	Installed files	8349	8353	+4
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	17	17	0
Runtime	Unique function payloads	2	2	0
Runtime	Total function bytes	21.27 MB	21.31 MB	+40.4 kB ⚠️
Runtime	Public routes	18	18	0

Changed function payloads vs v27/lab-fixes (28de6fc) (2)

Function	Status	Baseline	Current	Delta	Route changes
functions/__server.func	changed	10.63 MB	10.66 MB	+20.2 kB ⚠️	none
functions/.well-known/workflow/v1/flow.func	changed	10.64 MB	10.66 MB	+20.2 kB ⚠️	none

eve init install

Metric	Baseline	Current	Delta
Installed footprint	113.99 MB	114.01 MB	+19.5 kB ⚠️
Installed packages	97	97	0
dependencies	4	4	0
devDependencies	2	2	0
Dependency package bytes	45.69 MB	45.71 MB	+19.5 kB ⚠️
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: 2026-10-26T17:33:20.554Z
• 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@0.83.2
• Package directory: packages/eve
• Tarball: 7.83 MB (eve-0.83.2.tgz)
• Unpacked payload: 27.31 MB across 4360 published files
• Installed footprint: 75.63 MB across 8353 installed files
• Installed root package: 27.31 MB
• Installed dependencies: 48.32 MB
• Installed package instances: 34
• Distinct installed package names: 32
• Installed dependency edges: 51
• Installed optional peer edges: 17
• Runtime dependencies: 2
• Peer dependencies: 15 (14 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: 27.31 MB (36.1%)
• @rolldown/binding-linux-x64-gnu: 19.89 MB (26.3%)
• ai: 8.37 MB (11.1%)
• zod: 6.15 MB (8.1%)
• undici: 3.58 MB (4.7%)

Publish payload breakdown

Published file size
🟠 dist/src/compiled/shadcn-registry/index.js       [#####...................] 3.85 MB 14.1%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [###.....................] 2.29 MB 8.4%
🟠 dist/src/compiled/run/index.js                   [#.......................] 1.00 MB 3.7%
🟡 dist/src/compiled/@vercel/blob/index.js          [#.......................] 604.9 kB 2.2%
🟡 dist/src/compiled/_chunks/workflow/undici-D-I... [#.......................] 521.9 kB 1.9%
🔴 Other published files                            [########################] 19.04 MB 69.7%

Installed footprint breakdown

Installed package size
🔴 eve                             [########################] 27.31 MB 36.1%
🔴 @rolldown/binding-linux-x64-gnu [#################.......] 19.89 MB 26.3%
🔴 ai                              [#######.................] 8.37 MB 11.1%
🔴 zod                             [#####...................] 6.15 MB 8.1%
🟠 undici                          [###.....................] 3.58 MB 4.7%
🟠 nitro                           [##......................] 1.89 MB 2.5%
🔴 Other installed packages        [#######.................] 8.44 MB 11.2%

Runtime dependencies (2)

Package	Range	Notes
nitro	3.0.260903-beta
undici	8.10.2

Peer dependencies (15)

Package	Range	Notes
@linqapp/sdk	^0.47.1	optional peer
@nuxt/kit	^4.0.0	optional peer
@nuxt/schema	^4.0.0	optional peer
@opentelemetry/api	^1.0.0	optional peer
@vercel/connect	^2.5.2	optional peer
ai	catalog:
braintrust	^3.0.0	optional peer
chat	^4.42.0	optional peer
dd-trace	^6.13.0	optional peer
just-bash	^3.6.0	optional peer
next	^16.0.0	optional peer
react	^19.0.0	optional peer
svelte	^5.0.0	optional peer
vite	^8.3.1	optional peer
vue	^3.5.0	optional peer

eve init install drill-down

eve init install details

• Command: eve init my-agent
• Package manager: npm
• Installed footprint: 114.01 MB across 10270 installed files
• Installed packages: 97 total (91 transitive-only)
• dependencies: 4 direct packages totaling 45.71 MB
• devDependencies: 2 direct packages totaling 5.11 MB
• Other transitive package files: 63.18 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 (24.5%)
• eve: 27.31 MB (24.0%)
• @rolldown/binding-linux-x64-gnu: 19.89 MB (17.4%)
• zod: 9.76 MB (8.6%)
• ai: 8.37 MB (7.3%)

Installed footprint breakdown

Installed package size
🔴 @typescript/typescript-linux-x64 [########################] 27.95 MB 24.5%
🔴 eve                              [#######################.] 27.31 MB 24.0%
🔴 @rolldown/binding-linux-x64-gnu  [#################.......] 19.89 MB 17.4%
🔴 zod                              [########................] 9.76 MB 8.6%
🔴 ai                               [#######.................] 8.37 MB 7.3%
🟠 undici                           [###.....................] 3.58 MB 3.1%
🔴 Other installed packages         [###############.........] 17.16 MB 15.1%

dependencies (4)

Package	Range	Installed size	Share
@vercel/connect	2.5.2	276.3 kB	0.2%
ai	^7.0.137	8.37 MB	7.3%
eve	file:eve-0.83.2.tgz	27.31 MB	24.0%
zod	4.5.4	9.76 MB	8.6%

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     [########################] 10.66 MB 50.0%
🔴 functions/__server.func                         [########################] 10.66 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 10.66 MB

Metric	Value
Public routes	/.well-known/workflow/v1/flow
Runtime	nodejs24.x
Handler	index.mjs
Payload	10.66 MB
Function files	10.66 MB across 124 files
Traced dependencies	0 B
Signal	🟠 Bundled file _chunks/vercel.web.mjs is 2.59 MB (24.4%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/vercel.web.mjs                         [#############...........] 2.59 MB 24.4%
🟡 _libs/undici.mjs                               [#####...................] 994.5 kB 9.3%
🟡 _chunks/sandbox2.mjs                           [####....................] 815.0 kB 7.6%
🟡 _chunks/compiled-artifacts-instrumentation.mjs [####....................] 730.7 kB 6.9%
🟡 _chunks/token-util-Bla5Z0G9.mjs                [####....................] 704.5 kB 6.6%
🔴 Other bundled files                            [########################] 4.82 MB 45.2%

🧾 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 • 10.66 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]/descendants/stream /eve/v1/session/[sessionId]/reset /eve/v1/session/[sessionId]/stream /eve/v1/session/[sessionId]/stubs
Internal aliases	/__server
Runtime	nodejs24.x
Handler	index.mjs
Payload	10.66 MB
Function files	10.66 MB across 124 files
Traced dependencies	0 B
Signal	🟠 Bundled file _chunks/vercel.web.mjs is 2.59 MB (24.4%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 _chunks/vercel.web.mjs                         [#############...........] 2.59 MB 24.4%
🟡 _libs/undici.mjs                               [#####...................] 994.5 kB 9.3%
🟡 _chunks/sandbox2.mjs                           [####....................] 815.0 kB 7.6%
🟡 _chunks/compiled-artifacts-instrumentation.mjs [####....................] 730.7 kB 6.9%
🟡 _chunks/token-util-Bla5Z0G9.mjs                [####....................] 704.5 kB 6.6%
🔴 Other bundled files                            [########################] 4.82 MB 45.2%

🧾 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: 3.96 s -> 3.86 s (-103.0 ms) vs v27/lab-fixes (28de6fc).
• Timing is informational: shared GitHub runners are too variable for a hard timing budget.

Detailed phase timings vs `v27/lab-fixes (28de6fc)`

Phase	Baseline	Current	Delta
extension.check	9.9 ms	20.1 ms	+10.2 ms
project.resolve	0.5 ms	9.6 ms	+9.1 ms
workspace.create	0.7 ms	1.1 ms	+0.4 ms
host.prepare	607.4 ms	544.8 ms	-62.6 ms
vercel.service-prefix.resolve	2.0 ms	2.1 ms	+0.1 ms
nitro.create	513.0 ms	452.5 ms	-60.5 ms
sandbox.prewarm	510.2 ms	570.4 ms	+60.2 ms
nitro.cache.prepare	0.3 ms	0.2 ms	-0.1 ms
nitro.prepare	0.8 ms	0.8 ms	0.0 ms
nitro.public-assets	0.8 ms	0.9 ms	+0.1 ms
nitro.prerender	0.5 ms	0.6 ms	+0.1 ms
nitro.bundle	2.16 s	2.11 s	-55.8 ms
nitro.cache.write	0.4 ms	0.4 ms	0.0 ms
vercel.workflow-function.materialize	56.0 ms	56.8 ms	+0.8 ms
agent-summary.emit	0.7 ms	0.7 ms	0.0 ms
connect-manifest.emit	0.4 ms	0.3 ms	-0.1 ms
nitro.close	0.1 ms	0.1 ms	0.0 ms
output.publish	3.6 ms	3.7 ms	+0.1 ms
workspace.remove	2.8 ms	3.0 ms	+0.2 ms

[roboeve]

Sending protocolVersion: 2 here tells the older child that this caller speaks protocol 2 in full. That includes the protocol-2 input relay, which this caller doesn't implement.

In the published eve@0.83.2 (dist/src/execution/forward-session-input.js, same as main), a child whose caller is on its own version posts every input.requested, approval.candidate, authorization.required, etc. to the callback URL as kind: "subagent-input-request" / "subagent-authorization-event". When the response isn't ok, it throws Remote input callback failed with HTTP ${status}. This branch's callback route only knows subagent-interaction / turn.completed / turn.failed, so it answers 400 Unsupported callback kind.

So after this change, any tool approval, ask, or sign-in in a 0.83.2 child fails that child's turn. As far as I can tell from main's finalization.ts, the step retries and then the parent's call fails with Remote input callback failed with HTTP 400. There's no protocol or upgrade guidance, unlike the reverse direction (formatEarlierRemoteCallerInputError). Before this PR, that only happened to children already in flight during an upgrade (docs L245-249). Now it's the normal path for every negotiated call. I traced this from the code and haven't run it end to end. The cheapest check is your loopback pair with a 0.83.2 child tool that needs approval.

Suggestion: have handleSessionCallbackRequest recognize those two protocol-2 kinds and settle the pending call with a failed subagent-result that carries an upgrade message, the way formatEarlierRemoteChildStreamError does for streams. Then add a wire-boundary test. At minimum, the new docs paragraph should say what the parent sees in this case.

[OwenKephart]

Agreed; result-only negotiation must not be presented as a working old-input relay. 1506fc6 now documents the HTTP-400 failure explicitly. I am checking the actual 0.83.2 approval path with the loopback pair. I am not synthesizing a child terminal outcome/usage from an input request: that would claim a lifecycle boundary the child did not produce. Leaving this thread open for the safe failure/upgrade design rather than claiming the docs alone solve it. The api fleet remains on the approved base artifact, not this candidate.

[OwenKephart]

Confirmed with the actual loopback pair and published eve@0.83.2 child: requesting approval produces no parent interaction; the child step retries three times, then the parent agent call fails with TASK_FAILED / Remote input callback failed with HTTP 400. The parent delivery itself completes, so checking only delivery success would miss it. No approval was answered and the fixture tool has no external effect. This remains explicitly unsupported and this thread stays open; terminal-result negotiation is not a complete old-HITL solution.

[OwenKephart]

960d290 implements the protocol-2 input relay, so negotiating protocol 2 now covers it.

• The callback route reads subagent-input-request and subagent-authorization-event as v27 relay facts, using the same v26 mappings as the history reader.
• Answers go back as v26 inputResponses with the forwarded responder.
• The child still settles its approvals and sign-ins, and an answer it turns away leaves the approval open. The parent settles questions and budget prompts once the child accepts the answer, because a protocol-2 child never reports how they settle.
• The negotiated callback URL is marked eveRemoteProtocol=2, and only marked URLs are translated. A pre-v27 parent run's child therefore still fails visibly instead of hanging on a relay that run can't read.

Same loopback pair with the published eve@0.83.2 child:

• Approve: settles accepted, and the tool runs as alice.
• Deny: settles declined, and the tool doesn't run.
• Question: the answer reaches the child's ask_question.
• Cancel: the mirror is withdrawn and nothing is approved.

The v27 relay, task reuse and cancellation scenarios still pass. Leaving this thread for you to resolve.

[vercel]

A protocol-2 budget prompt preserves the Stop option ID stop, but parent routing only recognizes deny, so choosing Stop leaves the parent turn running and records the prompt as accepted.

[OwenKephart]

Retested the published immutable build 960d2902a496401dd9953f72de1eb408081540bf against the original standalone concurrent-task reproducer. The requested ordering fix works in these checks:

• Unchanged original batch: 8/8 passed. Strengthened batch: 60/60 passed, all with the three deliveries sharing one child turn. Admission stayed seed → left → right; all three parent calls completed and every root became idle. The previous builds 619ea4ef and 28de6fc8 each had a hang in an eight-attempt batch. These use a deterministic model with real durable workflows, local World, HTTP client, and public streams—not a mocked admission function.
• Sequential sends still map to separate turns; overlapping sends share a turn; fan-out stays separate. agentCallActivity retains the expected causal links/shared-turn ownership. A deliberately lagging read prefix remains pending rather than looking settled.
• Separate loopback parent on this build → child on published npm eve@0.83.2: approval, denial, question/answer, and whole-parent-turn cancellation pass. The approved fixture returns its local-dev current/initiator; denial returns the tool-denied result. The question returns iad1/answered. Whole-turn cancellation settles the actual turn as cancelled and interrupts the mirrored input—distinct from merely accepting a cancel command.
• Following the old remote through SessionStore correctly reports child observation unavailable / unsupported while retaining the completed authoritative parent result.
• Real retained 0.80.1 recordings still read through the new server, preserving text/runtime metadata, cursor suffixes, and prompt ownerless-follow termination.

No SDK patches, private imports, resends, or application-manufactured settlement. This is bounded local evidence, not a claim about deployed grants, multi-user forwarding, real sign-in, or old unmarked callbacks surviving a rollout.