Skip to content

Fix stale model budgets and add staged lossless remote compaction #256

Description

@andrei-hasna

Problem

Recent sanitized Codewith logs show gpt-5.6-sol remote compaction using model_context_window_tokens=258400 while failing single-pass compact requests carry roughly 1.75-2.78 MB of model-visible history. PR #245 preserves history when overflow occurs, but it intentionally does not implement staged recovery.

Expected

Use authoritative model context metadata and compact oversized histories in bounded, provenance-preserving stages.

Acceptance

  • Correct known current OpenAI model context-window metadata without changing unrelated provider caps.
  • Budget the serialized request including instructions, tools, and history.
  • Summarize bounded chunks before final compaction.
  • Preserve unsummarized history and checkpoint state on every failure.
  • Do not loop after deterministic overflow.
  • Return a typed, actionable recovery result.
  • Add large-history, resume, model-switch, cached-input, and tool-output regressions.

Related PR: #245
Related Todos: e6164e35, 5f509275, b1ce3ef0, aaa695d2

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions