Long-Running Analyses Hold Up Better

This week was mostly about making Humm more dependable when threads get long, tool-heavy, or a little messy. We tightened the way Analyst handle reflection retries, large contexts, and OpenRouter reasoning replay so a recoverable mistake is less likely to surface as a broken answer.

That includes smarter retry handoffs for slash-command workflows, better handling when missing evidence actually requires another tool call, and safer behavior for large-context threads that used to skip repair entirely. We also removed request-time tool-history compaction that was distorting some long runs, which should make agent behavior more consistent across complex conversations.

Scheduled Reports Fail Less Abruptly

Scheduled reports are now sturdier when worker processes shut down mid-run or Python execution gets stuck during shutdown. Humm now caps report execution to the intended runtime budget, returns more cleanly from sandbox timeouts, and allows a bounded retry when the interruption is caused by worker shutdown rather than the report logic itself.

The practical change is simpler than the implementation: fewer report runs should sit around looking active for far too long, and more interrupted runs should recover automatically instead of quietly timing out.

Snowflake Queries Are More Reliable Again

Snowflake-backed SQL execution is back on firmer ground. We removed redundant query-path connection preflights, restored the outbound paths needed for OCSP validation, and added better timing metadata so slowdowns are easier to diagnose when they do happen.

For teams using Snowflake as part of analysis workflows, that should translate into fewer connection-related stalls before a query even starts, plus more predictable behavior when running SQL inside Humm.

Improvements

  • Analyst's prompt and runtime stack was overhauled for GPT-5.5, with tighter output contracts, clearer evidence rules, and more explicit stop conditions.
  • Optional null values on custom MCP tools are now treated as omitted when appropriate, which should reduce avoidable validation failures in tool execution.

Bug Fixes

  • Fixed Deep Agents OpenRouter reasoning replay so provider-specific reasoning details no longer break follow-up execution.
  • Fixed OpenRouter context handling so completed historical reasoning does not bloat future requests or trigger avoidable context failures.