Repository navigation
v0.38.0
A discovery run can now be stopped, and picked up where it stopped
Until this release a discovery run was all-or-nothing. If the process died — a proxy timeout, a pod eviction, an out-of-memory kill — everything it had paid for was gone, and the only way forward was to run it again from the first query. And the only way to stop a run early was Cancel, which reaped the agent and deleted the very state that would have made the run resumable.
v0.38.0 closes both. A run now checkpoints its work as it goes, and an interrupted run resumes instead of restarting.
What is durable now
| Phase | On resume |
|---|---|
| Exploration | restored whole — the agentic loop and its warehouse queries are not re-run |
| Analysis, per area | completed areas are restored; only the remaining ones are analysed |
| Analysis, per insight | an area that died mid-validation is finished, not re-analysed |
| Recommendations and their validation | generated once; only the unvalidated remainder is checked |
Exploration is where almost all of a run's cost sits — N agentic LLM calls plus N warehouse queries — and it was the phase with no durability at all. A run killed after validating four of six insights in an area now pays for two, and does not repeat the analysis call or the repair rounds behind it either.
Pause
Pause is a cooperative stop. The run finishes the step it is working on, keeps that step's checkpoint, and exits into a new status paused that Resume accepts. Use it to give the warehouse back, or to stop the clock, without throwing away what you have already paid for.
The stop travels through the run document the agent already reads before each phase to ask whether it may carry on — so it needs no new channel, and a run that is paused while its process is already gone is still paused.
Cancel is unchanged, and still means discard this.
Why it matters
A long discovery run against a large warehouse is a real bill. Making that work durable turns a crash from "pay again" into "carry on", and turns stopping a run from a forfeit into a decision you can reverse.
Discovery says fewer things that are not true
Twenty commits attacking false claims in the analysis agent at emission rather than at validation — validation can only reject a claim, and a rejected insight had nowhere to go.
The largest single defect: the query stored on a step was not the query that ran. When the self-healing fixer rewrote a rejected statement, the rows came from the rewrite while the recorded query kept the original proposal — 54 of 54 query steps in a real run — and four consumers read that field as evidence, including the verifier's evidence bundle and the live status feed.
Measured on TPC-H SF=1 against a DuckDB oracle built from the same dbgen bytes and checked against the published TPC-H Q1 answer file, so a disagreement is the agent's and not an engine's. The same language rule is carried into the seed packs, which are the copy that ships.
Also in this release
go-jose/v44.1.6 (GHSA-mgcc-cww5-fp2q).- The LLM provider guide is findable — it has a search title and description.
Images are cosign-signed; the tag is GPG-signed. See docs/verifying-artifacts.md.