v0.49.11

v0.49.11

This patch release makes several OpenCode 2 settings actually take effect, removes an extra cache rebuild after every OpenCode 2 compaction, keeps project memories out of summaries when memory is off, and fixes a ctx_search freeze in large stores. There is no database migration.

Upgrading

• No migration. context.db stays at schema 95 and store.db at schema 63. Restart OpenCode 1, OpenCode 2, Pi or Oh My Pi after updating.
• OpenCode 2 sessions rebuild their cached prefix once, on the first turn after the upgrade. The history budget is now part of what the cache is checked against, so this happens even if you never changed history_budget_percentage. Turning on mural lands in that same rebuild. Other newly working settings either ride that rebuild or wait for the next one the session needs anyway, such as a model switch or an idle-expired cache. None of them cause a second rebuild.
• Rust-mode installs need to update ck-mc to the version shipped with this release. The Rust module carries its own copy of the memory-off fix for summaries and the summarizer sizing fix.

Fixed

• OpenCode 2 now applies settings it used to ignore. These settings had no effect on OpenCode 2 before this release:

  • memory.auto_search hints;
  • caveman_text_compression;
  • git commit indexing;
  • automatic embedding of the active session;
  • temporal_awareness;
  • mural;
  • history_budget_percentage;
  • dreamer.inject_docs: false in compaction summaries;
  • user-memory collection and the other dreamer settings.

  Auto-search is on by default, so OpenCode 2 users now see auto-search hints without changing anything. Commit indexing now runs with the dreamer disabled on OpenCode 2 and Pi, and no longer needs an embedding provider. Thanks to @null-axiom for the report (#709).

• OpenCode 2 no longer rebuilds the cache after every compaction. A compaction recorded a different identity for the system prompt than the one the next turn checked, so the turn after each compaction rewrote the cached prefix again for no reason.

• With memory off, project memories stay out of compaction summaries and the summarizer prompt. Before, memory.enabled: false kept project memories out of the main request, but OpenCode 2 compaction summaries and the history summarizer could still include them. This now holds on every host and in Rust mode.

• OpenCode 2 no longer refuses a compaction that runs before the first request. When a session compacted before its first turn in a fresh OpenCode 2 process, Magic Context refused because its database was still opening. It now waits for the database instead.

• ctx_search no longer freezes the host in large stores. Search used to reload and decode every stored chunk in one go. It now refreshes only what changed and reads in small batches, so the host stays responsive while searching.

• A system-prompt change is now recorded as the cause of the cache rebuild it triggers, in every mode. Before, a rebuild caused by a changed system prompt could show up in the dashboard as an unattributed cache bust, or under an unrelated cause. Thanks to @iceteaSA for the report (#713).

• Free and other small-window summarizer models are no longer skipped. The history chunk was sized for the primary model only. A primary with a small window then refused the request, and the run fell through to a larger fallback. Each chunk is now sized to fit every model in the fallback chain.

• doctor on Oh My Pi now warns when Hide Secrets is on. Oh My Pi's Hide Secrets setting (secrets.enabled) masks secrets after Magic Context has prepared a turn, which can make Magic Context refuse turns. The compatibility docs now explain this setting.

• The docs now explain model names on Pi and Oh My Pi. Historian and dreamer models, including fallbacks, must use provider names that host serves. opencode/* models are dropped at startup there. Thanks to @VibePDF for the report (#647).