fix: wire onSpawn in runChildProcess call

[faizer1989]

Summary

• Passes onSpawn: ctx.onSpawn to the runChildProcess call in execute.ts so Paperclip can track the Hermes child PID.

Problem

runChildProcess was called without onSpawn, so the opts.onSpawn callback in @paperclipai/adapter-utils/server-utils.js was never invoked. As a result, Paperclip never called persistRunProcessMetadata, and heartbeat_runs.process_pid stayed NULL for every Hermes run.

This broke the orphan-process reaper's liveness check — without a PID, the reaper couldn't verify if the Hermes child was still alive. Combined with the missing hermes_local entry in SESSIONED_LOCAL_ADAPTERS (separate Paperclip PR), any Hermes run exceeding 5 minutes was falsely declared process_lost.

Verified

Live on ao-sh-derper: every historical Hermes run (including the successful 39-second AUT-19 run from 2026-04-23) showed process_pid = NULL.

Related

• Sister fix: add hermes_local to SESSIONED_LOCAL_ADAPTERS in paperclipai/paperclip (separate PR)
• Paperclip issue: AIW-185

[liransmartlighthouse]

Independent confirmation of this bug + fix from a separate deployment.

We hit this exact issue running hermes_local agents under a live Paperclip server (a multi-agent fleet). Same root cause described here: execute() never forwards ctx.onSpawn to runChildProcess, so heartbeat_runs.process_pid is never persisted (stays NULL). On any run slower than the reconcile/reaper window, Paperclip loses the in-memory handle, has no PID to verify liveness, marks the run failed/process_lost, and auto-retries it once — orphaning the still-live Hermes child (which keeps consuming the model and can collide on shared session/HOME state).

We applied the identical one-line change (onSpawn: ctx.onSpawn in the runChildProcess opts) as a local patch and verified end-to-end:

• run.processPid is now persisted,
• slow runs are no longer falsely marked process_lost,
• no more orphaned / auto-retried Hermes children.

Confirmed present in published 0.2.1, 0.3.0, and current main (onSpawn appears 0× in src/server/execute.ts). As you note, hermes_local being in SESSIONED_LOCAL_ADAPTERS is the sister Paperclip-side piece; with that in place, this PID-reporting gap is the only remaining cause. This bites real multi-agent deployments — a merge here would be appreciated. Thanks for the clean writeup; of the open PRs, #99 mirrors our analysis most closely.