[sentry-junior]

Stacked on #2602 (base compaction-summary-main-model). Retarget to main once #2602 merges. Depends on the event schemas from #2600, which are already released.

Reads compaction_prepared events and swaps a ready summary into history without a model call. A separate PR adds the job that writes these events. Until it merges, nothing in prod writes them, so this PR changes no behavior there.

When a swap runs

• Turn start (the agent's compaction step before its first model request): swaps whenever a ready summary exists.
• Between model calls: swaps only above the 90% threshold, and only in place of the blocking summary call.
• A summary is used only if all of these hold:
  • it is the newest background compaction event in the current History version;
  • its outcome is ready;
  • its historyVersion and modelProfile match the current ones;
  • the replacement fits under the compaction target.
• Otherwise the Turn carries on as today, including the blocking fallback.

Replacement history

• In order: the latest session context and AGENTS.md from the summarized part, retained user messages (up to ~20k), the summary, every message after markSeq (with sourceEventSeq), then pending messages.
• Written with replaceHistory(..., { expectedHistoryVersion }). On a conflict, the summary is discarded and nothing is written.
• feat(conversations): Read background compaction events and check history version on replace #2600's schemas are used exactly as merged.

Tool-result trimming (tool-result-trimming.ts)

• Applies to carried-over results before the last user message.
• A result over 2k chars becomes a header, its first and last 500 chars, and readConversationEvent({ seq }).
• These stay whole: the latest Turn's results, error results up to 8k chars, results with no source event, and short results.
• Images become [image cleared: <tool>, seq N].
• Tool call ids don't change, so every call keeps its result.
• The blocking path's replacement carries no tool results today, so trimming isn't hooked in there.

Telemetry: a swap gets its own agent.context_compaction span. Its attributes are:

• app.compaction.outcome: applied / not_ready / discarded;
• app.compaction.discard_reason;
• the prepared summary's age;
• the count and estimate of carried messages;
• the trimmed result count and chars removed.

The mode (blocking / active) gives the swap point. The details are in TELEMETRY.md.

Decisions for review

• split_tool_call is an extra discard reason. It applies when the first message after the mark is a tool result, because the mark falls between a tool call and its result.
• Turn start now runs two small indexed queries per Turn to look for a background event.
• A stale prepared event (its historyVersion doesn't match, but it was appended in the current version) is reported as discarded on every Turn until the next replacement.
• The read-back eval in continuity.eval.ts can't pass yet. Its compaction is the blocking path, which keeps no tool results. It needs the background job to write compaction_prepared first.

via @thomasjiangcy.

--

View Session [Sentry]

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
junior-docs	Ready	Preview	Nov 3, 2026 2:01pm UTC

[cursor]

Stale Bugbot comment from a previous run.

[github-actions]

1 screenshot change — 1 changed · 0 added · 0 removed

Review screenshots in Frameshift

Settings Mcp · Desktop Changed

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e1775f3. Configure here.}