v0.24.0

v0.24.0

Adds cross-project workspaces and searchable session history, fixes a Pi context-overflow, and removes several prompt-cache busts.

New: cross-project workspaces

Group related repositories (for example a set of microservices) into a workspace and share project memories across them. Member sessions see their own project's memories plus the memories other members choose to share — each one labelled by its source repo. Per-workspace checkboxes control which memory categories are shared (Constraints only, by default), so shared knowledge stays relevant and project-specific noise stays local.

Workspaces are managed from the dashboard (Workspaces panel): create a workspace, add member projects, pick the shared categories, and Save. A single project belongs to at most one workspace. Single-project users are completely unaffected — the rendered context is byte-identical to before.

New: searchable session history

ctx_search can now find older discussion by meaning, not just keywords. Magic Context embeds each compacted history segment so the agent can recall what was discussed weeks ago even when it doesn't remember the exact wording.

New history is embedded automatically as the session grows. To make an existing session's older history searchable, run /ctx-embed-history once — it backfills that session's compacted history in the background while you keep working. (You'll also see a reminder in the startup dialog.)

Search ranking was improved alongside this: verbatim matches now rank correctly, and common acronyms no longer flood results.

Fixes

• Pi: sessions could overflow the model context while still reporting moderate usage. Pi only refreshed its context measurement at the end of each turn, so a long tool-heavy turn could balloon the request past the model's window before the next measurement — the request was rejected even though the status line still read, say, 68%. Pi now floors its pressure with a live forward estimate (recomputed every step), so it sheds context before the request overflows. (OpenCode was never affected — it already measures per step.)

• GitHub Copilot: tool_use/tool_result adjacency error on fresh sessions (#135). Empty-content message markers that Anthropic silently drops were surviving onto the Copilot (openai-compatible) wire and breaking the required tool-call/result pairing. They're now stripped only for providers that drop them.

• Fewer prompt-cache busts. Three cases that were re-billing large prompt prefixes unnecessarily:

  • Editing a project doc (ARCHITECTURE.md / STRUCTURE.md) no longer rebuilds the cached history baseline — the new content folds in on the next natural cache turn instead.
  • A processed screenshot in an older message no longer loses its image data mid-conversation on a routine pass.
  • When the cached history baseline does have to rebuild (a memory change, a model switch, idle timeout), any pending tool-output cleanup is now folded into that same rebuild instead of causing a second bust a moment later.

Also in this release

• Setup wizard (npx @cortexkit/magic-context) now lists the models you actually have with type-ahead filtering, instead of a fixed recommendation list, and explains what the historian and dreamer do (neither needs a frontier model). Adds a --dry-run mode. (#144)
• Protected-tail snapshot is refreshed at history-summarization time, preventing a rare case where the summarizer could starve and stop reclaiming context.