Expose context composition and compaction controls #50

Open
opened 2026-10-24 23:26:38 +00:00 by TheComputerGuy · 0 comments

Problem

Token totals alone cannot explain model behavior. Users need to know which instructions, memories, skills, messages, and tool results entered a request, what was omitted, and what compaction changed.

Proposed outcome

Add a context inspection view and configurable compaction policy. Record provenance and token estimates for components, plus a digest and summary of every compaction. Permit bounded pinning of critical instructions without overriding organisation policy.

Acceptance criteria

  • Effective context can be inspected for each provider request without exposing secrets.
  • Compaction records what was retained, summarized, or dropped.
  • Budget and context estimates explain why compaction occurred.
  • Pinned content has bounded size and explicit precedence.
  • Replay reconstructs the same context provenance.
### Problem Token totals alone cannot explain model behavior. Users need to know which instructions, memories, skills, messages, and tool results entered a request, what was omitted, and what compaction changed. ### Proposed outcome Add a context inspection view and configurable compaction policy. Record provenance and token estimates for components, plus a digest and summary of every compaction. Permit bounded pinning of critical instructions without overriding organisation policy. ### Acceptance criteria - Effective context can be inspected for each provider request without exposing secrets. - Compaction records what was retained, summarized, or dropped. - Budget and context estimates explain why compaction occurred. - Pinned content has bounded size and explicit precedence. - Replay reconstructs the same context provenance.
Sign in to join this conversation.
No description provided.