Skip to content

NextClaw v0.59.9 ​

This release continues the availability fixes for Issue #73. Version 0.59.8 removed the blanket tool restriction, but recent results could still be replaced prematurely with history previews, and older deferred flags could hide populated tool cards.

Read and use actual tool results ​

Recent tool results retain their content within the working-window and model-input budgets. Crossing a smaller display threshold no longer replaces those results with an archived placeholder. Agents can read feedback files, inspect command stdout, and use the actual SHA256 returned by a command.

Populated tool cards carrying older deferred flags can also be opened normally. Outputs exceeding the model budget still require bounded reads as indicated; history-list previews do not replace the current results needed by the model.

Continue after compaction and cropping ​

Fix empty leading round boundaries introduced by cropping, and stale round positions on active messages fully covered by compaction. These could produce Invalid assistant model round part offsets. Older saved empty leading rounds are supported, and older derived indexes are rebuilt automatically from original journals, without deleting conversations or manually editing records.

Original conversation content and actual write-failure protection remain intact. The first 0.59.1 journal failure trigger remains under investigation and the issue stays open. Content that was never saved cannot be recovered by this release.

Released under the MIT License.