Version

v2026.7.1-beta.5 (commit b6387af), analysis of shipped dist code.

Environment

Linux gateway, OpenClaw running as a systemd user service; large-context primary with a 131k-context local terminal fallback.

Summary

Fallback candidate selection never considers the candidate model's context window. A session grown large on a big-context primary that fails over mid-turn to a smaller-window fallback immediately overflows, triggering compaction storms and risking unrecoverable compaction failure.

Code analysis

model-fallback-BV_Bq-Z8.js and fallbacks-*.js contain zero references to contextWindow/contextTokens. Selection is availability-ordered only; there is no check that the current prompt size fits the candidate's window, and no preemptive compaction before switching.

Observed impact

On our gateway: sessions grown on a 1M-context primary failed over mid-turn to a 131k-context local fallback, producing repeated overflow/compaction cycles with observed compaction give_up risk. 34 sessions on this gateway currently exceed 131k tokens — every one of them hits this on failover to the terminal fallback.

Expected behavior

Fallback selection should either (a) skip candidates whose contextWindow is smaller than the current prompt size, or (b) preemptively compact to fit before dispatching to that candidate — and log which path was taken.

Suggested fix

In the candidate loop, compare current prompt token estimate against each candidate's registered context window; skip too-small candidates (configurable), or invoke compaction targeting the candidate's window before the first request to it.

Related (not duplicates)

• Irreversible context compaction when swapping to a model with a smaller context window #98426 — irreversible compaction damage after swap to a smaller-window model (downstream effect; asks for context regrowth, not candidate skipping).
• Model switch can fail silently when carried-over session context is too large #58957 — manual model switch fails silently when carried-over context is too large.
• Feature: Trigger model fallback on context length exceeded #9986 — fallback triggered by context-length-exceeded (inverse direction).
• [Bug]: Self-hosted/custom OpenAI-compatible model discovery ignores top-level context_size/context_length field, under-reports context window (e.g. Cortecs GLM-5.2 shows 131k instead of 1M) #109367 / fix: self-hosted model discovery ignores top-level context_size/context_length field #109368 — self-hosted model discovery ignores context_size/context_length, so registered windows may be wrong; window-aware selection needs those fixed too.