8 commits landed across Hermes in the last 3 hours, and the headline is a real UX upgrade for anyone living in the Kanban view.
What Changed
Done cards now show actionable results. feat(kanban): surface final_result for Done cards puts the run’s final result right on the card, while a follow-up fix(kanban): make Done-card results actionable and test(kanban): remove duplicate final-results footer (both from Teknium) clean up the rendering and kill a duplicated footer so the summary reads cleanly.
Approval verdicts get observable. fix(approval): emit observer hooks for smart verdicts (hellno) wires observer hooks into the approval path so “smart verdict” decisions can be watched and audited instead of happening in a black box.
Config values stop getting coerced. fix(config): preserve string-typed config values (liuhao1024) keeps string config entries as strings — no more silent type coercion corrupting settings.
Image-gen handles unsupported accounts. Teknium added fix(image-gen): classify unsupported Codex image accounts plus test(image-gen): cover Codex capability HTTP boundary, so accounts that can’t generate images are classified correctly and the HTTP capability boundary is tested.
Multi-model catalogs respect custom providers. fix(agent): honor custom-provider extra_body for multi-model catalogs makes sure a custom provider’s extra_body is passed through when you’re running a multi-model catalog — no more dropped parameters.
Why It Matters
The Kanban change is the one users will feel daily: finished runs now tell you what they produced without a click-through. Around it, the approval hooks, config preservation, and provider fixes harden the runtime so smart verdicts and custom catalogs behave predictably. For the gateway-side context work that pairs with these agent fixes, our Hermes gateway context-budget coverage tracks the broader runtime story.
Running multiple AI agents? Save on API costs and subscriptions at aiFiesta.