Six of NeuroLink's seven OpenAI-compatible providers read their model env var at registration time; Mistral's registration passes a bare literal instead, so MISTRAL_MODEL is silently ignored on the default `generate()` path.

A gzip decompression bomb in NeuroLink's ArchiveProcessor used to be caught only after the whole payload was inflated. Moving the bound into the decoder itself, via Node's `maxOutputLength`, cut peak RSS on a 400MB bomb from 411MB to 102MB.

client.callTool() in NeuroLink's realtime LiveKit agent used to run with no RequestOptions, so a stalled MCP server held a speech-to-speech turn open indefinitely. A 30-second toolTimeoutMs, sized against conversation latency rather than server latency, bounds the call and routes the failure through the handler's existing catch block instead of leaving the user in silence.

How NeuroLink's runIsolatedAgent() and registerAgentTool() turn the hand-copied "worker sub-agent" pattern from Curator and Yama into a real extension point, with delegation caps, depth limits, and leashed multi-leg runs.

validateAndResolveUrl resolved a hostname twice — once to validate, once to pin — leaving a TOCTOU window where DNS could answer differently between the two calls. The fix collapses both resolutions into one and pins the fetch to the entire validated address set instead of a single OS-picked IP.

How to write, store, and register a NeuroLink skill — JSON, frontmatter markdown, or a Claude-style SKILL.md directory — and how search_skills, list_skills, and the CLI's skills subcommand pick it up.

On Gemini-3, a pure tool-calling turn emits thoughtSignature and functionCall parts with no text at all, so NeuroLink's prose-preservation branch never fired and the loop ignored the caller's abortSignal for 18 minutes and 74 tool calls.

How NeuroLink's TaskManager survives process restarts: BullMQ vs Node timers for scheduling, Redis vs filesystem for persistence, and a nine-phase autoresearch loop.

How BaseServerAdapter, getFrameworkInstance(), and dynamic-import route groups let you mount NeuroLink MCP, proxy, and agent APIs onto a running Express or Fastify app without ceding server lifecycle control.

How NeuroLink models voice as TTS, STT, and full-duplex — ChunkedAudioStream backpressure, the realtime state machine, and OpenAI vs. Gemini Live divergence.