Skip to content
MantleMantle docs

Release 0.118

v0.118.1 — 2026-07-06

Boot reconcile works on multi-admin brains again. Since the actor/anchor split (v0.111.0), a brain with more than one admin had several auth.users rows — and the boot reconcile’s “single owner” check read that as an unprovisioned install and silently skipped. Prompt, skill, and tool-group updates stopped reaching those brains on upgrade. Owner resolution now keys on the single anchor owner of the brain’s content (with the old single-user check as the fresh-install fallback), so upgrades propagate everywhere again.

v0.118.0 — 2026-07-06

Big page edits no longer die halfway. A large SOP restructure on a production brain exposed a chain of agent-editing failures, all fixed here:

  • Write batches are atomic. The tool-loop’s volume caps (40 calls/turn, 15/tool) used to trip mid-batch — a 10-delete batch got cut at 1-of-10 and left the draft half-edited. Caps now enforce at batch boundaries: a batch that starts under its caps always completes; when the budget ends the turn, the model is told explicitly so it reports what’s done vs what remains.
  • page_blocks_list no longer lies about drafts. It listed the published doc while the block-edit tools worked on the draft — so an agent looking at a broken draft saw a clean page and said so. The listing now reads the same editing baseline as the edit tools and flags has_draft / draft_updated_at; page_get flags the draft too.
  • Right tool for the job. The pages agent now picks its edit strategy by size: block tools for targeted fixes, one whole-body page_update_draft pass for big restructures (with the markdown table pitfalls documented — a # | … header row parses as a heading, not a table).
  • Per-agent tool budgets. memory_config.max_tool_calls / max_calls_per_tool override the flat caps; the pages agent ships with 100/40. Specialist memoryConfig now force-syncs on upgrade (like prompt/model/params), so existing brains get the new budgets.