RC11: compaction firing far too often (mid-task, well under context limit)

[akag2]

What happened?
Multiple independent reports converging on the same session, Sutando Triage room, 2026-08-11: RC11 (desktop/AG2Space-embedded runtime) is compacting far more often than expected — one measured instance compacted at 19m43s / 47.3k tokens / only 29% into a single task.

What's been verified so far (peer agents qingyun-001 and air, cross-checked against each other and against this session):

• CLAUDE.md is 47,089 chars (~11,800 tokens); AGENTS.md is another 47,036 chars if both load — a fixed per-session tax before any work happens, and over the documented 40k-char warning threshold.
• The earlier theory that RC11 silently TRUNCATES content past the 40k mark (a correctness bug, not just a cost bug) was raised, then measured and retracted by its own author (qingyun-001): on both qingyun-001's and air's runtimes, all past-40k sections (incl. the tail Skills/manifest section) are present and loaded, not cut. So on the runtimes actually measured, the 40k limit behaves as a warning + token tax, not truncation.
• Open / unmeasured: whether RC11's specific embedded runtime (the desktop app session) behaves differently from the two runtimes above — nobody has confirmed load-vs-truncate specifically there yet.
• A prepared fix already exists: docs: bring CLAUDE.md/AGENTS.md under the 40k runtime limit — relocate detail, keep every rule #2794 relocates sections of CLAUDE.md into docs/, cutting it to ~36,994 chars (~2,500 tokens back per session). It was dropped tonight to focus elsewhere; the branch is preserved (owned by qingyun-001) and can be reopened cheaply.
• One unconfirmed hypothesis floated in the room: that this may be a Claude Code-side compaction-trigger issue rather than a Sutando-side context-size issue — not verified, flagging as an open question rather than a conclusion.

Reported per explicit owner request (Slack + AG2Space, both 2026-08-11) to file as a critical bug via this skill.

Steps to reproduce

1. (from reporter — to confirm)

Expected behavior
(to confirm)

Severity (reporter): critical

---

_{🤖 Filed automatically from in-app user feedback — bug / severity critical. Edit freely; this is a starting point, not a fixed record.}