Cross-deployment start() stamps the starter's spec version, not the executing deployment's

[pranaygp]

Problem

A run's specVersion is chosen in start() (packages/core/src/runtime/start.ts: const specVersion = opts.specVersion ?? world.specVersion) and written on run_created and the queue message. That is the spec version of the SDK calling start().

When start() is called without a deploymentId, the run executes on the same deployment, so the caller's version is the executor's version and everything is consistent.

When start({ deploymentId }) targets another deployment, the run is stamped with the caller's spec version even though a different deployment — possibly on an older or newer @workflow/core — will read and write its event log. Every consumer of the spec version then reasons about the wrong party:

• Reader gates that decide "does this run's runtime understand row X?" from run.specVersion (sealed-log noop, slot identity, and now the hook_disposed{forceClaimedBy} victim gate from feat(hooks): createHook({ experimental_force: true }) takes a held token over #4193 / vercel/workflow-server#980) can be wrong for cross-deployment runs: a new starter stamps ≥ N onto a run executed by an older deployment that does not read N.
• Writer behaviour keyed on the stamped version (compression ≥ 5, CBOR queue transport, attributes) is chosen by the starter but has to be honoured by the executor.

Deployments are pinned on Vercel, so there is no rolling-version window within one deployment; the discrepancy exists specifically between two different deployments.

Proposed fix

The executing deployment should determine the run's spec version. Options, roughly in order of preference:

1. start() already probes the target deployment via the health check for a cross-deployment start (healthCheck(world, { deploymentId, ... }) in start.ts; the response carries specVersion and workflowCoreVersion, parsed in runtime/helpers.ts). Use the probed specVersion (capped at what this SDK and the World can mint) instead of world.specVersion when deploymentId !== currentDeploymentId. Same-deployment starts are unchanged. If the probe fails or the target predates a versioned health response, fall back to the lowest version both sides are known to support rather than the caller's current version.
2. Alternatively let the target's run_started (written by the executor) settle/override the version, with Worlds treating run_created's version as provisional until then. Larger change; touches every World.

Either way, add mixed-version tests: new starter → older executor and older starter → newer executor, for the sealed-log and hook-force-claim reader gates.

Context

Surfaced while gating createHook({ experimental_force }) victims on specVersion >= 8 (#4193). An intermediate revision replaced that gate with a per-run executionContext.hookForceClaimReaderVersion attestation stamped from the probe (same pattern as hookResumeInputVersion); it was reverted to keep a single versioning mechanism and instead fix the stamping here for all gates at once. The reverted commit (0feb2fd on the hook-force-claim branch) shows how the probe value can be threaded into start().

🤖 Generated with Claude Code