1.0.0-beta.13

published 16 September 2026

• Removed the run governor that ended a turn for repeating a tool call with nothing written in between, and the repetition breaker that ended a turn on repeated identical results: a session is never stopped for repeating itself, and an empty provider reply is announced and the request sent again once
• Fixed the live token counter reading near zero while a tool call's input streams: it counts the streamed input, and the wide layout's count keeps pace with the stream
• Fixed an edit refused for lines the model had not read costing a read and a retry: the refusal carries those lines, numbered with their anchor, and a read past the window cap answers its first window with the offset to continue
• Added a prune of superseded file results when the context reaches forty percent of the model's window, a setting, so a long session no longer carries dead results until it overflows
• Fixed the lease verbs disagreeing on their argument's name: claim and release both take paths, read the older word as the same list, and a claim with neither is refused naming both
• Fixed an empty shell command accepted as a background task: it is refused before it spawns
• Fixed a background agent's background command recorded from defaults, missing from the tasks board and never reported back to the agent: the record carries its launch facts, the board lists it, and the agent is told when it finishes
• Fixed a lookup's empty answer from grep, pgrep, cmp, diff or test reported as a failure: it is a result naming the exit code, and a real failure keeps its error
• Fixed the JavaScript kernel's environment going unstated: the Eval tool says it is an ES module without require, where Node's crypto lives and how long an idle kernel is kept; a failed cell names the fact and the import; a declaration whose initializer threw is no longer reported as surviving, and a declaration after a block's closing brace persists
• Fixed a script run from inside a Mercury session refusing the session's own environment stamp as foreign
• Fixed a failed Workshop cell reporting the runtime's plumbing instead of its error: a syntax error carries the parser's diagnosis with line and column, a failed bridge call names the call and the cell line, and the transcript's error wrapper no longer rides inside the cell's message
• Updated the eleven bundled skills, rewritten from current documentation at about half their length
• Fixed the DeepSeek route sending the reasoning effort inside the thinking object where the API takes it at the top level
• Fixed the summary line under a turn counting repeated reads of one file as files read: it counts reads
• Fixed the Boot face's MCPs & Skills page calling the session empty while the bundled skills and the built-in server load: its empty states name what they list
• Fixed --model and --effort on the command line losing to a resumed session's saved values, and /model forgetting its choice at the next boot: an explicit flag wins, the resume card names which won, and /model saves the choice as the default
• Fixed a rejected credential retried ten times with growing waits: it is refreshed once, then the request settles as a sign-in blocker naming the account and the remedy
• Fixed a run recording completed with deliverables still open: no run reports complete while requested deliverables stay open, and an older record that does reopens on resume