Design OpenClaw so a parent process that spawns it inside a sandbox can observe startup deterministically from process launch through gateway bind, sidecar startup, channel/account startup, and final operational readiness.
This guide is intentionally high-level. It describes the product and architecture we should build, not a file-by-file implementation checklist.
Documentation is a pull system. You go find what you need. A plugin is a push system. Knowledge finds you at the moment you need it. The difference is the same gap that separates polling an API every five seconds and subscribing to a webhook.
Every system for getting information to a consumer falls into one of two categories: pull or push. In a pull system, the consumer initiates the request. They know what they need, they know where it lives, and they go get it. In a push system, the producer initiates the delivery. The consumer declares what they care about, and relevant information arrives when it becomes available.
The Vercel Plugin for Claude Code now ships a hybrid prompt matching engine that combines exact phrase matching with full-text lexical search, synonym expansion, and suffix stemming. The result: skills are injected when the developer means something, not just when they say the exact magic words.
Three variations, each covering a different angle. All draw from the same product (Workflow DevKit) and the same proof points, but target different reader motivations.
Format: Deep-dive (1200-1500 words) Primary keyword: durable AI streaming
Key differences between Claude Code and Cursor hook formats:
| Concept | Claude Code | Cursor |
|---|---|---|
| Event name | hook_event_name |
hook_event_name (same) |
| Session ID | session_id |
conversation_id |
| Working dir | cwd + CLAUDE_PROJECT_ROOT env |
workspace_roots[] + CURSOR_PROJECT_DIR env |
| Tool name | tool_name |
tool_name (same for preToolUse/postToolUse) |
The Vercel Sandbox API rejects resources: { vcpus: 1 } with a 400 error, but multiple documentation pages show or imply 1 vCPU as a valid configuration. The actual minimum is 2 vCPUs (even numbers only).
This matters for I/O-bound workloads. We measured CPU utilization inside OpenClaw sandboxes over a 2-hour live test: <2% CPU even at 162 msgs/hr sustained. A 1 vCPU / 2 GB tier would halve memory costs, which are 97-99% of total sandbox cost for this workload.
OpenClaw sandboxes on Vercel needed a manually-configured AI_GATEWAY_API_KEY environment variable to authenticate with Vercel's AI Gateway for LLM inference. This was a deployment friction point — every new deployment or team member needed the key configured.
Replaced the static API key with Vercel's OIDC (OpenID Connect) token federation. The dashboard's Vercel Functions automatically receive an OIDC JWT that proves they're running on Vercel infrastructure. AI Gateway accepts this JWT as a bearer token — no static key needed.