dsh-fetch-timeouts

Raises Node's HTTP timeouts for the whole DeepSeek Harness process, so a slow local model is not cut off after 5 minutes of silence.

The problem it fixes

Node's built-in fetch (which dsh's model adapters use) gives up when a server sends no response headers for 300 seconds, or no body bytes for 300 seconds. dsh has no setting for those two timers: streamIdleTimeoutMs is dsh's own watchdog and timeoutMs is the SDK's request timer, so raising them changes the failure message from pi-ai stream idle timeout to Failure reason: terminated (UND_ERR_BODY_TIMEOUT / UND_ERR_HEADERS_TIMEOUT) at exactly 5:00.

Servers that stay silent that long include Ollama and LM Studio while a model thinks or generates a large tool call (for example the entire contents of a file for write), and any backend that does not send keepalive pings. llama.cpp's llama-server sends a ping every 30 seconds by default, so llama.cpp users usually do not need this plugin.

Install

dsh plugin --profile web add dsh-fetch-timeouts

That is enough: the defaults raise both timeouts to 30 minutes. To change them, add to ~/.dsh/profiles/web/cordis.patch.yml:

- id: fetch-timeouts
  config:
    headersTimeoutMs: 3600000   # time allowed before response headers arrive; 0 disables
    bodyTimeoutMs: 3600000      # time allowed between body chunks; 0 disables

Restart dsh web. One line confirms it at startup:

fetch-timeouts: headers 1800000 ms, body 1800000 ms (process-wide)

Also raise dsh's own watchdog on the provider route, or it will fire first:

llm-pi-ai:
  providers:
    ollama:
      streamIdleTimeoutMs: 1800000
      timeoutMs: 1800000

What you should know

• It is process-wide. Every fetch that uses Node's global dispatcher (model calls, web search, HTTP MCP servers, cloud providers) gets the same longer limits; web_fetch is not affected because it builds its own per-request agent. A genuinely dead connection therefore takes up to the configured time to be noticed, and once you have raised streamIdleTimeoutMs as well, dsh's idle watchdog is the only remaining backstop for a hung model server. Reasonable on a single-user machine; think twice on a shared host.
• It works by installing an undici Agent as Node's global fetch dispatcher. If NODE_USE_ENV_PROXY is set it installs undici's proxy-aware agent instead, so HTTP_PROXY, HTTPS_PROXY and NO_PROXY keep working. Tested on Node 22 with undici 8 (undici 8 requires Node 22.19 or newer). Confirmed by a user on Windows with Ollama on a 20 minute file write (discussion #4518).
• Loading the plugin's undici dependency already swaps Node's default dispatcher for undici's own (same 300 second defaults); the plugin then applies your timeouts. undici only installs its default when no global dispatcher exists yet, so another plugin loading undici later cannot replace the plugin's agent. Unloading the plugin returns to undici's default, not to Node's original object.
• It is a stopgap. When dsh exposes these timeouts itself (its pi-ai dependency already accepts a custom fetch), this plugin becomes unnecessary.

License

MIT