v5.10.6

• dsh's web_search no longer dies with tool call timed out after 60000ms when the first engine is slow (#35). --timeout limits each engine, and every fallback engine used to start with a fresh budget, so a slow first engine could use up the host's whole 60 seconds before the fallback answered. The new --deadline <ms> limits the whole run: each engine gets the smaller of --timeout and the time left, and an engine whose turn comes after the deadline is skipped and listed in attempts. The dsh plugin now passes 40 seconds per engine and a 55 second deadline for web_search (configurable as providerTimeoutMs and providerDeadlineMs), and 120 and 180 seconds for x_search and read_page.
• Engine timeouts now cover the whole request. Tavily, Exa, and Firecrawl stopped timing once the response headers arrived, so a slowly streamed body could run past the timeout. Firecrawl's DNS check before a page fetch, and the local fetcher's DNS lookup, now give up when the timeout fires.
• An engine that ignores its own timeout is now stopped as well. A command-line engine (agy, Grok Build) is killed 5 seconds past its timeout under a deadline, and 30 seconds past it without one, as before. An in-process engine is abandoned at the same point.
• The CLI now exits once its result is written, after making sure any engine that ignored SIGTERM has been sent SIGKILL. Before, a run could leave such an engine running, or wait on leftover work after printing its result.