Version 1.0.113 du 29/10/2026- New Features : A Research Turn Now Ends In A Dedicated Final Answer: After exploring, the assistant composes its reply in a separate step that has no tool access, so the answer you receive is a real reply instead of another round of tool calls.
- Improvements & Changes : A Turn Now Publishes Only What It Verified: Audit style specifics are withheld unless at least one file was actually read and the search actually scanned something, so an answer no longer presents plausible but unverified details as fact.
- Bug Fixes : A Codebase Audit No Longer Refuses On A Missing Server Entry: Asking an agent without an explicit codebase server entry to audit its codebase now works instead of being refused with a message pointing at a setting that does not exist.
- Bug Fixes : A Narrated Save With Leaked Tool JSON Is No Longer Shipped: An answer that narrated a file save while pasting raw tool call JSON, without ever reading or writing a file, is now rejected instead of being delivered.
- Bug Fixes : A Degraded Review No Longer Ships A Weak Answer: When the reviewer passes but the quality score is far below the shipped answer, the assistant now refuses instead of delivering a weak reply.
- Bug Fixes : A Correction After A False Save Promise Now Completes: When the assistant wrongly promised to save a memory it never stored, a follow up correction now runs to completion instead of failing review three times and falling back to a generic refusal.
- Bug Fixes : A Quality Verdict Returned As An Object Is Now Read Correctly: A verdict returned as an object is now coerced to its string before validation, so the quality check no longer drifts.
Version 1.0.112 du 28/10/2026- New Features : Autonomy Depth Presets For Each Agent: Each agent now has an Autonomy depth setting (Focused, Balanced, Deep) controlling how many tool calls, tool rounds, and how much time a turn may spend. Leave it blank to inherit the runtime default, or type your own values.
- New Features : Per Memory Kind Follow Switches For Workspace Agents: A workspace agent can now follow or ignore the global background, procedures, and preferences memories independently, so switching all three off fully isolates that agent from the shared store while its own memories still load.
- Improvements & Changes : Memory Preferences Now Have Their Own Card: The follow global preferences toggle moved out of Agent behavior into a dedicated Memory preferences card that is always shown, disabled on the system agent, so there is no longer a duplicate or hidden switch.
- Improvements & Changes : Stored Style Rules Survive Long Conversations: Style rules you saved (such as "no em-dashes") are now kept even when a long conversation trims older context, so the assistant keeps following them instead of forgetting.
- Improvements & Changes : Recovery No Longer Re-Asks What You Already Gave: A recovery answer no longer asks again for the role, cover letter, changelog, or agent details you already provided in your message.
- Bug Fixes : A Failed Turn Now Refuses Instead Of Shipping A Stub: When a turn fails its review several times in a row, the assistant now gives an explicit refusal instead of delivering a tiny leftover draft as though it were the answer.
- Bug Fixes : An Unreviewed Restored Draft Is No Longer Delivered: A draft brought back from a stash without passing review is no longer shipped as a degraded answer; the honest fallback is used instead.
- Bug Fixes : Copy No Longer Carries Citation Footers: Text copied from an answer no longer includes trailing citation footers.
(0 votes)
