feat(eve): replace Workflow with durable code mode

[ruiconti]

Summary

Tool orchestration needs to await subagents and resume completed work after a crash. Previously, the experimental Workflow wrapper only orchestrated agents; this replaces it with a root-only code_mode workflow tool that discovers the full advertised catalog and invokes eligible tools and subagents. Enable it with experimental.codeMode: { mode: "eager" | "lazy", maxSubagents?: number }. Related to #2508.

Programs pin their catalog and subagent budget when dispatched. Ordinary calls execute in separate durable steps; subagent calls use the owning session's agent-invoke channel and await the child result. Calls issued together settle concurrently. Authorization can park and resume a call, and individual failures remain catchable by the program.

• Discovery includes every advertised tool in both modes. Approval-gated tools, provider tools, framework controls, ordinary background tools, and authored workflow tools remain direct and are marked requiresDirectCall.
• Connection discovery updates the next model step's catalog. Dynamic tools retain the captured values from the dispatching step.
• Each program permits 100 subagent invocations by default. Retries and continuations count; excess calls reject before dispatch. maxSubagents changes the limit.
• Breaking migration: remove the file exporting experimental_workflow() from eve/tools/workflow, then enable code mode in agent.ts. The old helper and harness-specific Workflow continuation machinery are removed. Authored "use workflow" tools remain supported.
• Workflow compilation preserves availability metadata, so child sessions keep their ordinary tools and never receive the root-only code_mode wrapper.

Programs previously ran guest catch/finally handlers while parking a tool call, and deterministic program failures consumed workflow retries. The pinned Run patch now treats suspension as host control flow and marks guest-source errors separately from infrastructure errors. eve returns known program failures from the step as data, then reports the tool error to the model so it can revise the source. The one-program orchestration sentence is restored; eager still exposes eligible direct tools. The vendoring cache now fingerprints the installed worker so a same-version patch rebuilds the shipped runtime. Readable patch and reproduction.

Scripted Code Mode E2Es cover execution contracts. The real-model eager/lazy evals check complete pagination, overlapping independent reads, handling an unavailable source, and correct saved reports. They log program grouping and repeated page reads for diagnosis; they do not require a single program or benchmark savings against direct tools.

Optional tool arguments now retain their omission semantics at the provider boundary. eve explicitly disables implicit strict-mode normalization and still validates inputs before execution.

Validation

• Rebased onto 4b5fad41b at ed2894a46: all 47 workspace build/typecheck tasks, 141 focused unit tests, 13 compiled-SDK integration tests, and the installed-extension scenario pass. Frozen installation, formatting/lint, docs and invariant checks pass. Published compatibility reports are preserved: new tool epoch 31 and dynamic-tool epoch 30 replace the removed Workflow contracts; subagent epoch 10 retains main's epoch 9. The new steering fixture now imports the cancellation tool from its own fixture after the sleeper fixture moved. GitHub reports no merge conflicts; required CI checks are pending.

• Suspension/recovery repair at 0c77a2645: 63 focused eve unit tests, 13 compiled-SDK integration tests, and the packaged Code Mode worker-assets scenario pass. The underlying Run repair passes all 336 tests and its typechecks. eve runtime build, package and scripted-fixture typechecks, changed-file formatting/lint, docs checks and invariant checks pass. Independent source review found no blocking lifecycle/classification defect. The same compiled-SDK catch probe that timed out at 30s on the baseline now completes in 16.7ms; catch/finally and the exact 120,418-byte payload with eight distinct leaf calls pass. This is an in-process interruption/resolution probe, not a full workflow-crash or model-performance benchmark. Two scripted E2Es were added; CI is running on the rebased head.

• Optional-argument fix: 41 tool unit tests and three input-validation integration tests pass; the new provider-boundary regression failed before the fix. Package and fixture typechecks, lint, docs, invariant checks, and the runtime build pass. The four-call Terra probe produced 0/2 valid inputs with strict omitted and 2/2 with strict: false. A real-model fixture eval was added; no eve test workflow was recorded for the benchmark baseline 143b0e5a7.

• pnpm --filter eve exec vitest run --config vitest.unit.config.ts --maxWorkers=4: 7,985 passed, one skipped. The final child-visibility test adjustment also passed its focused harness run.

• Targeted integration runs: 40 passed across code-mode sandbox/connection execution, turn and session-limit cancellation, cancelled-turn settlement, and authored workflow tools.

• Targeted packaged scenarios: seven passed, covering local/remote child continuation, cancellation, code-mode compilation and sandbox assets, and installed extension compatibility.

• pnpm --filter eve build:js, package tsc --noEmit, changed-file formatting/lint, pnpm guard:invariants, and pnpm docs:check passed.

• pnpm --filter agent-code-mode --filter agent-cancellation --filter agent-subagents typecheck passed, including fixture builds. E2E evals were added/migrated and are left to CI.

• New planning fixtures: both builds and typechecks passed with EVE_E2E_MODEL=mock; five assertion unit tests and two audit scenarios passed. Matrix discovery includes both fixtures on both registered models. Live-model eval results are pending CI; no E2E suite was run locally.

• After rebasing onto eb5949a62: frozen installation, all 46 workspace build/typecheck tasks, invariant checks, changed-file formatting/lint, and docs checks passed. Focused coverage passed: 68 compiler/Code Mode unit tests, 16 Code Mode integration tests, the installed-extension scenario, and seven planning-fixture tests. Published extension history is preserved; the tool and dynamic-tool contracts now use epoch 28.

Benchmark diagnosis: GPT-5.6 Terra, 2026-09-06

At baseline 143b0e5a7, optional eager never selected Code Mode in 20 controlled attempts. Restoring the one-program instruction reduced lazy root-turn time by 33%, with 8/8 correct answers in both conditions. Forcing Code Mode was not uniformly beneficial.

The tables below are measurements of that baseline and its explicitly described experimental variants. The repair at 0c77a2645 has deterministic regression validation above; it has not been rerun through these model benchmarks.

The old Fable pilot and the newer Terra run are not a controlled comparison. The task files are unchanged (594 manifest files verified), but the model, eager tool availability, orchestration wording and runtime artifact changed. The pilot used an uncommitted experimental package with a local suspension fix; the benchmark baseline lacked that repair. This does not identify an upstream commit that introduced the bug.

Controlled selection experiment: 80 attempts. Same Terra model with high reasoning, runtime engine, tools and ten frozen cases; two repetitions across five task shapes and two seeds. Each column has 20 attempts. These are attempt success counts, not best-of-two scores. Every attempt and recovery stays in the timing totals.

Direct disables Code Mode. Optional eager is the implementation at baseline 143b0e5a7. Forced eager hides eligible direct tools. Guidance retains both tool surfaces and repeats the current selection/orchestration text in the main instructions; that intervention changes repetition and placement together. The latter two are experimental variants, not changes included in this PR.

Measurement	Direct	Optional eager	Forced eager	Guidance
Agent lifecycle, mean seconds	20.72	22.97	19.86	18.83
Root turn, mean seconds	12.80	14.96	11.91	10.75
Input tokens, total	529,059	816,375	404,617	533,075
Cached input, included above	405,467	665,433	341,995	436,711
Output tokens, total	23,583	26,456	18,646	20,397
Model steps	78	76	61	52
Code Mode programs	0	0	43	12
Failed dispatched actions	0	0	2	0
Root-model cost, USD	0.6730	0.8278	0.4486	0.5729
Attempts using Code Mode	0/20	0/20	20/20	12/20
Original passes	19/20	20/20	19/20	19/20
Passes after separate grading audit	20/20	20/20	19/20	19/20

Root-turn duration by shape; four attempts per condition in each row:

Shape	Direct	Optional eager	Forced eager	Guidance
Single account	3.44s	4.20s	4.12s	3.71s
One-project summary	6.46s	7.15s	7.04s	7.28s
Eight-project summary	8.05s	10.92s	8.54s	7.42s
32-project summary	16.88s	19.71s	13.94s	12.42s
32-project inventory	29.15s	32.84s	25.90s	22.92s

Optional eager used Code Mode zero times and added 3,673 input tokens to the first request in every first-repetition pair. The tool construction exposes both direct schemas and the Code Mode description. Forced Code Mode won three of four inventory comparisons; the fourth recovered from two syntax errors and took 50.63s versus direct’s 38.93s.

The direct large-task traces emit 32 domain calls within one model step, as separate streamed requests. This already avoids 32 separate model round trips; it does not prove simultaneous network execution. Code Mode can cross dependency boundaries and reduce intermediate data. These local synthetic fixtures do not establish production API latency or a universal fanout threshold.

Orchestration experiment: 24 attempts. All conditions use lazy mode, the same engine and discovery tools, and one/32-project summaries with two seeds and two repetitions. One intervention restores only the first orchestration sentence; the other restores the complete old three-sentence block from 287c074.

Measurement	Current wording	First sentence restored	Full old block restored
Agent lifecycle, mean seconds	20.46	16.13	19.74
Root turn, mean seconds	12.83	8.62	12.23
Input tokens	113,697	71,838	99,470
Cached input, included above	92,015	49,582	71,204
Output tokens	4,265	4,322	5,621
Model steps	41	26	33
Discovery-only programs	10	8	10
Work programs	23	10	15
Failed dispatched actions	0	0	1
Root-model cost, USD	0.1237	0.1174	0.1523
Original passes	8/8	8/8	8/8

Restoring only the first sentence reduced work programs from 23 to 10 and model steps from 41 to 26, with nearly unchanged fixture work (148 versus 150 calls). Root-turn duration fell from 12.83s to 8.62s; seven of eight matched pairs were faster and all answers passed. The full old block averaged 12.23s and recovered from one invalid program after four workflow attempts (3.177s of failed-action time, retained). This supports a small orchestration change; it does not solve eager selection. The first-sentence intervention says to complete the task in one execution program and retains the current reuse instruction. Restoring the full block also adds discovery/retry wording and removes that reuse sentence; its individual clauses are not isolated. Two repetitions on fixed cases establish observed behavior, not a universal policy or statistical significance.

Confirmed failures and remaining defects. One direct answer was correct but the grader extracted an array instead of an object; its correction is reported separately. Forced and guidance each failed one summary by aggregating only the first 100 of 725 global domains. Native replay of those programs reproduces the wrong result, so these are agent pagination mistakes.

The two invalid programs in the recovered inventory trial each received four workflow attempts. Their failed-action intervals totaled 6.482s, including the first attempts and retry delays. The repair at 0c77a2645 returns known deterministic program failures to the model without repeating unchanged source. The old experimental error classifier did not cover the RUN_ERROR returned by these syntax errors, so restoring that classifier alone is insufficient.

A separate suspension defect reproduced in the baseline packaged SDK: certain try/catch or finally programs time out before a leaf tool executes. The catch reproducer timed out at 30s in the baseline package and completes in 17ms with the pilot’s repaired package; ordinary execution controls pass. None of the 55 Code Mode programs in the 80-trial selection run uses that pattern, and no timeout occurs there. The repair above addresses this defect; it was not the cause of those observed timings.

Measurement and provenance. Agent lifecycle includes startup and cleanup; root turn excludes both. Cleanup averages roughly 5.4s. Environment setup, building and grading are excluded from both durations. Input totals include cached context; output totals include reasoning. Provider caching stays enabled. Trials run sequentially with shuffled case order and rotated condition order; there are no full-trial retries. Internal workflow retries remain measured.

Recorded model + grader spend for these 104 attempts: $3.4635 ($3.0398 selection; $0.4237 orchestration). Combined recorded-spend stop $5; per-session stop $0.25. Stops are checked after charges, not hard billing caps.

Runtime commit: 143b0e5a7dc8a7335063b6d9dc2c3b54845cb1f9. Base archive SHA-256: 93a537f044ea2a45808b5e539e812c376b21bebf2db44a72a11045711f327bc8. Dataset SHA-256: 107296bae88bfad6b6e26b4aa8f4047f479702e2fe88ce8b54d711d93eafecd5. Harbor: 2fe1615503fed39ad82b7ce09b22497996b30f1f. Saved runtime archives, dependency locks, instructions, agent definitions and adapters match each experiment’s manifest. Raw grades remain unchanged.

Earlier full 18-case Terra snapshot at the same commit

This snapshot has one attempt per case/mode; its aggregate covers a different case mix from the controlled ten-case diagnostic above.

Measurement	Direct	Eager	Lazy
Agent lifecycle, mean seconds	21.0	21.9	24.2
Root turn, mean seconds	13.4	14.3	16.5
Input tokens	377,864	762,582	269,641
Output tokens	18,812	19,405	17,990
Code Mode programs	0	1	71
Root-model cost, USD	0.5154	0.7033	0.4365
Original passes	18/18	18/18	17/18
Passes after separate audit	18/18	18/18	18/18

The lazy false negative was an extracted status label. Total recorded model + grader spend was $1.9670, separate from the 104-attempt diagnostic. The earlier Mini run had an invalid direct control and cannot establish a Code Mode speedup. The optional-argument schema fix at this commit was separately reproduced and validated; no real-model CI E2E result is claimed here.

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)

Diff size

Docs — 11 files · +179 / -25

Documents Code Mode configuration, migration, selection guidance, and optional tool arguments.

Implementation — 87 files · +2033 / -1299

Replaces the Workflow wrapper with durable Code Mode execution and its tool/catalog integration; preserves optional arguments at the provider boundary. Includes the generated Run worker patch and its readable upstream source/test patch.

Tests — 124 files · +3987 / -1358

Covers runtime contracts, cancellation, tool visibility, real-model planning fixtures, optional-argument validation, suspension and program-error recovery. Counts fixture files with tests; the upstream regression tests embedded in the readable dependency patch are counted with that patch.

[vercel]

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

Project	Deployment	Actions	Updated
eve-docs	Ready	Preview, v0	Sep 7, 2026 3:53am UTC
eve-docs-4759	Ready	Preview, v0	Sep 7, 2026 3:53am UTC
eve-pkg	Ready	Preview, v0	Sep 7, 2026 3:53am UTC

[github-actions]

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

Key takeaways

• Runtime delta: function payloads 20.72 MB -> 21.10 MB (+380.3 kB ⚠️).

Delta vs main (4b5fad4)

Area	Metric	Baseline	Current	Delta
Package	Packed tarball	8.65 MB	8.65 MB	+2.6 kB ⚠️
Package	Unpacked publish size	32.59 MB	32.60 MB	+8.5 kB ⚠️
Package	Installed footprint	77.67 MB	77.68 MB	+8.5 kB ⚠️
Package	Published files	3847	3845	-2
Package	Installed files	7745	7743	-2
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	7	7	0
Runtime	Unique function payloads	2	2	0
Runtime	Total function bytes	20.72 MB	21.10 MB	+380.3 kB ⚠️
Runtime	Public routes	18	18	0

Changed function payloads vs main (4b5fad4) (2)

Function	Status	Baseline	Current	Delta	Route changes
functions/__server.func	changed	10.36 MB	10.55 MB	+190.1 kB ⚠️	none
functions/.well-known/workflow/v1/flow.func	changed	10.36 MB	10.55 MB	+190.1 kB ⚠️	none

eve init install

Metric	Baseline	Current	Delta
Installed footprint	116.11 MB	116.12 MB	+8.5 kB ⚠️
Installed packages	113	113	0
dependencies	4	4	0
devDependencies	2	2	0
Dependency package bytes	48.72 MB	48.73 MB	+8.5 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: 2026-09-07T03:54:09.218Z
• 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@0.52.2
• Package directory: packages/eve
• Tarball: 8.65 MB (eve-0.52.2.tgz)
• Unpacked payload: 32.60 MB across 3845 published files
• Installed footprint: 77.68 MB across 7743 installed files
• Installed root package: 31.17 MB
• Installed dependencies: 46.51 MB
• Installed package instances: 33
• Distinct installed package names: 32
• Installed dependency edges: 51
• Installed optional peer edges: 7
• 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: 31.17 MB (40.1%)
• @rolldown/binding-linux-x64-gnu: 19.31 MB (24.9%)
• ai: 7.01 MB (9.0%)
• zod: 6.41 MB (8.3%)
• undici: 3.51 MB (4.5%)

Publish payload breakdown

Published file size
🔴 dist/src/compiled/shadcn-registry/index.js       [#############...........] 9.76 MB 29.9%
🟠 dist/src/compiled/@photon-ai/chat-adapter-ime... [###.....................] 2.27 MB 7.0%
🟠 dist/src/compiled/@ai-sdk/code-mode/index.js     [#.......................] 1.03 MB 3.1%
🟡 dist/src/compiled/@vercel/blob/index.js          [#.......................] 901.4 kB 2.8%
🟡 dist/src/compiled/_chunks/workflow/signal-exi... [#.......................] 514.6 kB 1.6%
🔴 Other published files                            [########################] 18.14 MB 55.6%

Installed footprint breakdown

Installed package size
🔴 eve                             [########################] 31.17 MB 40.1%
🔴 @rolldown/binding-linux-x64-gnu [###############.........] 19.31 MB 24.9%
🔴 ai                              [#####...................] 7.01 MB 9.0%
🔴 zod                             [#####...................] 6.41 MB 8.3%
🟠 undici                          [###.....................] 3.51 MB 4.5%
🟠 nitro                           [#.......................] 1.89 MB 2.4%
🔴 Other installed packages        [######..................] 8.36 MB 10.8%

Runtime dependencies (2)

Package	Range	Notes
nitro	3.0.260903-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	^3.1.0	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: 116.12 MB across 9639 installed files
• Installed packages: 113 total (107 transitive-only)
• dependencies: 4 direct packages totaling 48.73 MB
• devDependencies: 2 direct packages totaling 5.04 MB
• Other transitive package files: 62.35 MB

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

Heavy installed dependencies

• eve: 31.17 MB (26.8%)
• @typescript/typescript-linux-x64: 27.95 MB (24.1%)
• @rolldown/binding-linux-x64-gnu: 19.31 MB (16.6%)
• zod: 10.37 MB (8.9%)
• ai: 7.01 MB (6.0%)

Installed footprint breakdown

Installed package size
🔴 eve                              [########################] 31.17 MB 26.8%
🔴 @typescript/typescript-linux-x64 [######################..] 27.95 MB 24.1%
🔴 @rolldown/binding-linux-x64-gnu  [###############.........] 19.31 MB 16.6%
🔴 zod                              [########................] 10.37 MB 8.9%
🔴 ai                               [#####...................] 7.01 MB 6.0%
🟠 undici                           [###.....................] 3.51 MB 3.0%
🔴 Other installed packages         [#############...........] 16.79 MB 14.5%

dependencies (4)

Package	Range	Installed size	Share
@vercel/connect	1.0.0	167.9 kB	0.1%
ai	^7.0.82	7.01 MB	6.0%
eve	file:eve-0.52.2.tgz	31.17 MB	26.8%
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	2.2%

Function Drill-Down

Payload Size Graph

Unique function payload size and share of total
🔴 functions/.well-known/workflow/v1/flow.func     [########################] 10.55 MB 50.0%
🔴 functions/__server.func                         [########################] 10.55 MB 50.0%

Top Function Payloads

🟠 functions/.well-known/workflow/v1/flow.func • 1 public route • 10.55 MB

Metric	Value
Public routes	/.well-known/workflow/v1/flow
Runtime	nodejs24.x
Handler	index.mjs
Payload	10.55 MB
Function files	10.55 MB across 116 files
Traced dependencies	0 B
Signal	🟠 Bundled file index.mjs is 2.02 MB (19.2%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                        [#########...............] 2.02 MB 19.2%
🟡 _libs/undici.mjs                 [####....................] 980.8 kB 9.3%
🟡 _chunks/esm-Fqlolk7e.mjs         [###.....................] 723.3 kB 6.9%
🟡 _chunks/chatgpt-model.mjs        [###.....................] 693.8 kB 6.6%
🟡 _chunks/signal-exit-Dsy-TT0V.mjs [###.....................] 616.2 kB 5.8%
🔴 Other bundled files              [########################] 5.51 MB 52.3%

🧾 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 • 10.55 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	10.55 MB
Function files	10.55 MB across 116 files
Traced dependencies	0 B
Signal	🟠 Bundled file index.mjs is 2.02 MB (19.2%)

🟠 🔎 Dependency Analysis

📦 Bundled files:

Bundled file size
🟠 index.mjs                        [#########...............] 2.02 MB 19.2%
🟡 _libs/undici.mjs                 [####....................] 980.8 kB 9.3%
🟡 _chunks/esm-Fqlolk7e.mjs         [###.....................] 723.3 kB 6.9%
🟡 _chunks/chatgpt-model.mjs        [###.....................] 693.8 kB 6.6%
🟡 _chunks/signal-exit-Dsy-TT0V.mjs [###.....................] 616.2 kB 5.8%
🔴 Other bundled files              [########################] 5.51 MB 52.2%

🧾 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: 6.19 s -> 6.13 s (-59.0 ms) vs main (4b5fad4).
• Timing is informational: shared GitHub runners are too variable for a hard timing budget.

Detailed phase timings vs `main (4b5fad4)`

Phase	Baseline	Current	Delta
extension.check	12.4 ms	0.9 ms	-11.5 ms
project.resolve	0.4 ms	0.4 ms	0.0 ms
workspace.create	0.6 ms	0.6 ms	0.0 ms
host.prepare	855.0 ms	671.9 ms	-183.1 ms
vercel.service-prefix.resolve	2.6 ms	2.7 ms	+0.1 ms
nitro.create	572.7 ms	618.8 ms	+46.1 ms
sandbox.prewarm	250.6 ms	293.7 ms	+43.1 ms
nitro.cache.prepare	0.3 ms	0.3 ms	0.0 ms
nitro.prepare	0.9 ms	0.8 ms	-0.1 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	4.44 s	4.48 s	+41.5 ms
nitro.cache.write	0.3 ms	0.4 ms	+0.1 ms
vercel.workflow-function.materialize	47.3 ms	51.7 ms	+4.4 ms
agent-summary.emit	0.7 ms	0.7 ms	0.0 ms
nitro.close	0.2 ms	0.1 ms	-0.1 ms
output.publish	3.4 ms	3.7 ms	+0.3 ms
workspace.remove	2.0 ms	2.3 ms	+0.3 ms