Keep compaction boundaries aligned with raw user text

[cursor]

Problem

strip_compaction_preambles stored the compaction cut boundary so remap after API/TUI reload can find the first kept turn in SQLite. Two strips were too greedy:

1. rfind("\n\nQUERY: ") took the last orchestrator wrapper. A user turn that embeds that same delimiter (for example quoting docs) stored only the suffix. The DB row still has the full user text, so find_compaction_boundary misses and remap_compaction_onto_history fail-opens (covered_until=0: summary plus the entire hydrated history).
2. After that, any remaining leading [ + first blank line was treated as a pitfalls/intent envelope. User text that starts with a bracket block ([NOTE]\n\n…) or a [KNOWN PITFALLS] lesson body that contains \n\n was truncated the same way.

Transcript replay already documents the user-typed QUERY: case; compaction did not.

Fix

Strip only known orchestrator envelopes:

• [OPEN TODOS] … [END OPEN TODOS]
• [KNOWN PITFALLS … [END KNOWN PITFALLS]
• first \n\nQUERY: wrapper (index master)
• leading [INTENT] … through the following blank line (specialist ingress)

Do not rfind the QUERY mark and do not treat arbitrary [ + \n\n as a preamble.

Tests

unit_context_compaction — embedded QUERY delimiter, user text starting with [ plus an inner-blank pitfalls block, specialist INTENT-only strip, remap of the embedded-QUERY boundary. 30/30 passed locally.

Independently mergeable against main (0456350). Code hunks (context_compaction.cpp, test_context_compaction.cpp) do not overlap #321–#334. CHANGELOG [Unreleased] will need a rebase after whichever sibling merges first.

  

---

Note

Medium Risk
Changes how compaction boundaries are derived from stored user messages; incorrect stripping would still fail-open to full history, but wrong cuts could affect context windows after session reload.

Overview
Fixes compaction boundary stripping so stored cut points still match SQLite after reload.

strip_compaction_preambles no longer uses rfind on \n\nQUERY: (first wrapper only) and no longer treats any leading [ block before a blank line as orchestrator junk. It now removes only known envelopes: [OPEN TODOS], [KNOWN PITFALLS], the index \n\nQUERY: prefix, and a leading [INTENT] block. User turns that quote \n\nQUERY: or start with [NOTE]-style text keep full content, so boundary_content_matches / remap_compaction_onto_history do not fail-open to covered_until=0.

Unit tests cover embedded QUERY, bracket user text, INTENT-only strip, and remap with an embedded-QUERY boundary; CHANGELOG documents the fix.

^{Reviewed by Cursor Bugbot for commit b236841. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_71576940-fe63-41f0-9889-aaf1e1be5c92)