Summary

runPreflightCompactionIfNeeded decides whether to compact by taking the Math.max of three token estimates: the transcript-usage projection, the fresh persisted total, and a raw persisted total (stalePersistedPromptTokens). The first two respect the session freshness flag, but the third reads entry.totalTokens directly with no totalTokensFresh check:

const stalePersistedPromptTokens = hasPersistedTotalTokens
  ? Math.floor(persistedTotalTokens)
  : undefined;

When a compaction completes without a tokensAfter figure, incrementCompactionCount (src/auto-reply/reply/session-updates.ts:310) sets totalTokensFresh = false but deliberately keeps the old large totalTokens. resolveFreshSessionTotalTokens (src/config/sessions/types.ts:629) returns undefined in that state by design, so freshProjectedTokenCount correctly drops out, but the raw stalePersistedPromptTokens still feeds the Math.max gate. The stale large value then overrides the small fresh transcript estimate and force-fires another preflight compaction on a session that is actually small, producing an unnecessary (lossy, user-visible) summarization and a "compacting" stall on the next turn.

This is the persisted-total leg of the same symptom tracked in #81178. PR #81916 bounds the transcript-usage leg (usageProjectedTokenCount / snapshot tail bytes); it does not touch stalePersistedPromptTokens, so even with #81916 merged the stale persisted total still trips the gate. It is also the opposite direction of the rejected #82919, which wanted resolveFreshSessionTotalTokens to return stale values; this change instead keeps the existing freshness contract and extends it to the one consumer that bypassed it.

Fix

const stalePersistedPromptTokens =
  hasPersistedTotalTokens && entry.totalTokensFresh !== false
    ? Math.floor(persistedTotalTokens)
    : undefined;

The fresh-large path is unaffected: when totalTokensFresh is true or absent, the persisted total still counts, so a genuinely large session still compacts.

Real behavior proof

Behavior addressed: after a compaction that reports no tokensAfter (totalTokensFresh = false, large totalTokens retained), the next preflight check must not force another compaction when the real active transcript is small. A genuinely large fresh session must still compact.

Real environment tested: local OpenClaw source runtime at origin/main head 617c9d4b7f, Linux, Node 22. A standalone node --import tsx script drove the real runPreflightCompactionIfNeeded path against a real temporary JSONL transcript and session store; the compaction dependency was instrumented only to count whether the runtime attempted compaction.

Exact steps or command run after this patch: ran the proof script twice, once with the source fix reverted (git stash) and once with it applied, each invoking runPreflightCompactionIfNeeded with totalTokens=200000, a ~2KB transcript, and a 100000-token context window, for both totalTokensFresh: false (stale) and totalTokensFresh: true (fresh).

Evidence after fix (copied terminal output):

===== BEFORE FIX (origin/main behavior) =====
STALE (post-compaction, tokensAfter omitted): totalTokens=200000 totalTokensFresh=false small-transcript=~2KB -> compactAttempts=1
FRESH (genuinely large session)             : totalTokens=200000 totalTokensFresh=true small-transcript=~2KB -> compactAttempts=1
stale-case compactAttempts = 1 (want 0: small session must NOT re-compact)
RESULT: FAIL

===== AFTER FIX =====
STALE (post-compaction, tokensAfter omitted): totalTokens=200000 totalTokensFresh=false small-transcript=~2KB -> compactAttempts=0
FRESH (genuinely large session)             : totalTokens=200000 totalTokensFresh=true small-transcript=~2KB -> compactAttempts=1
stale-case compactAttempts = 0 (want 0)
fresh-case compactAttempts = 1 (want 1)
RESULT: PASS

Observed result after fix: the stale small session no longer compacts; the fresh large session still compacts.

Regression test: src/auto-reply/reply/agent-runner-memory.preflight-stale-tokens.test.ts. Case A (large but stale total, small transcript) fails before the fix and passes after; Case B (large fresh total) confirms the gate still fires. Run with node scripts/run-vitest.mjs.

What was not tested: no full pnpm check / build run (memory-constrained box). Scope is one expression in a single decision path; existing colocated suites were run as regression coverage.