RC11: compaction happening far too often — owner reverted to RC10

[akag2]

What happened?
Owner (bassil) reports hitting context compaction much more frequently on RC11 than prior releases, to the point of reverting back to RC10 until this is looked into. Reported live in the Sutando Triage AG2Space room (2026-08-11, ~11:33-12:05Z window).

Room diagnosis so far (from qingyun-air.agent and sutando-qingyun-001, not fully root-caused):

• CLAUDE.md is 47,089-47,356 chars (~11.8k tokens) and AGENTS.md tracks it (~47,036-47,303 chars) on affected hosts, both over the documented 40k-char guidance threshold. This loads into every session's context before any work happens.
• On the runtime checked, all seven past-40k sections were still present in loaded context, so the 40k limit appears to behave as a warning threshold rather than a hard truncation there — but this was NOT verified across all runtimes/cores, so it remains an open question whether some consumer (e.g. a Codex core) does hard-truncate.
• A prepared fix already exists: docs: bring CLAUDE.md/AGENTS.md under the 40k runtime limit — relocate detail, keep every rule #2794 relocates sections out of CLAUDE.md/AGENTS.md into docs/, bringing both under 40k (~36,994 / ~36,947 chars). It was closed on deprioritization; the branch survived and a meaning-preservation audit already passed. Reopening it is flagged in-room as low-effort (~15 min) and worth doing regardless of whether it's the root cause of the compaction frequency, since it's real headroom debt.
• Separately (may or may not be related): the room also surfaced a credential-proxy issue causing 401 OAuth-expired retry loops, fixed in fix(quota-tracker): credential proxy must never inject a known-expired token — pass through or fail loud, reload on 401 #2793 (merged ~1hr before this report), with a per-machine follow-up action ( after updating the proxy checkout to current main) still outstanding on some hosts.

None of this fully explains 'compaction way more often on RC11 specifically' vs RC10 — that comparison hasn't been measured yet. Filing as critical per explicit owner request, since the owner has already worked around it by reverting to RC10, which blocks getting RC11 fixes/features until resolved.

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.}