CLAUDE_AUTOCOMPACT_PCT_OVERRIDE silently stopped taking effect after 2026-07-14 (v2.1.220) #82761

Activity

1. bcherny

   added

   area:core

   on Aug 20, 2026
2. raddue commented on Sep 7, 2026

   

   raddue

   on Sep 7, 2026

   More actions

   Haven't tested CLAUDE_AUTOCOMPACT_PCT_OVERRIDE directly, so I can't confirm the exact regression window here — but I want to flag adjacent evidence for anyone reading this issue and considering "switch to the token-based autoCompactWindow knob instead" as a workaround: that one is also unreliable at fleet scale, independent of this specific env var.

   I parsed compactMetadata out of 92 trigger: "auto" compaction events across 6+ concurrent repos/sessions on my machine (2026-08-01 to 2026-09-07), all with a single autoCompactWindow value configured via user-level settings.json. Actual preTokens at auto-trigger ranged 370,799-828,136 (2.2x spread), with a dominant cluster around ~539-545k that never matches any value I configured (only ever 400000, more recently 300000). Full data and methodology on #75335 (comment: #75335 (comment)).

   Claude Code version 2.1.263 — well past the 2026-07-14/v2.1.220 break point reported here, so if the intent was "the token-window path still works even though the pct-override path broke," that doesn't match what I'm seeing on the window path either, on a materially later version. Worth someone checking whether these share a root cause (e.g. the same internal "effective window" resolution logic feeding both _ke/Gkn per the reverse-engineering in #70477) rather than treating them as two independent regressions.
3. Burlesque1 commented on Oct 19, 2026

   

   Burlesque1

   on Oct 19, 2026

   More actions

   Still reproducible on 2.1.299 (macOS, desktop app worker, but the mechanism below is version-level). Adding decompiled evidence that pinpoints where the override is dropped.

   The override is read and parsed correctly, and the threshold formula still implements it — but at least two call sites explicitly void it before use:

   // threshold computation still honors the override:
   function c5(effectiveWindow, opts) {
     const ceiling = effectiveWindow - 13000;              // summary buffer
     const pct = opts.testPctOverride;                     // from CLAUDE_AUTOCOMPACT_PCT_OVERRIDE
     if (pct !== undefined && !isNaN(pct) && pct > 0 && pct <= 100)
       return Math.min(Math.floor(effectiveWindow * (pct / 100)), ceiling);
     return ceiling;
   }
   
   // ...but the worker-state path passes testPctOverride: void 0 unconditionally:
   function resolveLevel(tokens, model, ctx, state) {
     if (state !== undefined)
       return gwe(tokens, state.effectiveWindow, {
         enabled: state.enabled,
         precomputeBufferFraction: XPE,
         testPctOverride: void 0,        // <-- override discarded
         testBlockingOverride: void 0
       }, ilr(model), state.threshold);
     ...
   }
   
   // ...and the fork path does the same on purpose (its own telemetry says so):
   function withoutOverride(model, ctx) {
     return c5(c9(model, ctx), { ...ope(model, ctx), testPctOverride: void 0 });
   }
   // telemetry message: "autocompact: not compacting the <agentType> fork:
   //  it ignores CLAUDE_AUTOCOMPACT_PCT_OVERRIDE, and ..."

   Observed behavior on 2.1.299 with CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80 and a 400000-token window: the session's autocompact_state frame reports threshold: 367000 — exactly the default window - min(maxOutputTokens, 20000) - 13000 formula with no override applied — while the same settings file and process env are confirmed correct (env inside the worker shows the variable; the threshold shown in the usage panel is 92%, not the configured 76%).

   So on current versions the variable is parsed but never reaches the compaction decision on the main worker path. The string table still documents it ("lowered further by CLAUDE_AUTOCOMPACT_PCT_OVERRIDE when set" in the autocompact_state schema description), which makes the silent ignore particularly misleading — the worker schema claims the override applies while the code path voids it.

   Happy to pull the exact call graph for other paths (remote/bridge, subagents) if useful — this matches #70477's observation that attached sessions ignore the override as well.