Repository navigation
Run-detail and subagent polling fix
Fixes a regression introduced in v2026.10.9 and present in v2026.10.10: opening a run with a saved cost could return a server error, and MCP get_run could report “Could not reach the run service.” An orchestrator could poll successfully while a child ran, then fail its tool retry limit when the child completed and saved its cost. (#599, #600)
The API now converts the stored PostgreSQL numeric cost to the expected response type. Existing saved costs and outputs remain unchanged. Failed run reads also log the underlying database error with run and workspace IDs.
Database regression tests cover null, zero, and positive costs before and after completion. CI now runs the read and completion tests against PostgreSQL using the migration-defined cost type.
Upgrade notes
- No new database migrations relative to v2026.10.10.
- Pull both versioned images and update the API before the web service. Wait for API health before completing the upgrade.
- The tagged compose file and main default to
2026.10.11. Deployments with an explicitTAS_VERSIONoverride must set it to2026.10.11. - After upgrading, inspect saved subagent outputs before rerunning failed orchestrators. A child may have completed or performed external actions even though its parent could not read its result.
- This release fixes the deterministic read failure; it does not change model/tool retry policy.
Images
ghcr.io/tembo/tas-api:2026.10.11ghcr.io/tembo/tas-web:2026.10.11
Both images target Linux/amd64 and are signed with provenance and SBOM attestations.
Validation
The handler regression test reproduces HTTP 500 without the fix and passes with it. Local validation passed 91 regular Rust tests, both PostgreSQL integration tests, formatting, and Clippy (existing warnings only). Release checks also verify the production web build and repository CI.
Publication does not confirm that Carefeed's updater has run or that its instance has deployed the new version.
Full changelog: v2026.10.10...v2026.10.11