EPS
← All tasks·#2073Completed

workflow-fix: pin CLAUDE_CODE_AUTO_COMPACT_WINDOW in project settings + document the reduced-window thrash diagnostic

kind: infra#wf-fix#wf-fix-fp:77e9d6470b27
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-compaction fires on claude-fable-5 subagents at ~200-270k input+cache against a 1,000,000-token native window, three cycles in a row, after which the harness aborts the agent with "Autocompact is thrashing". Measured fleet-wide on 2026-08-04: 28 of 44 subagents across 12+ distinct sessions. Every thrash was claude-fable-5; zero on claude-opus-5 or claude-opus-4-7 (which reached 600,613 and 685,564 peaks the same day).

Root cause is upstream and known: a server-side GrowthBook experiment (tengu_amber_redwood) overrides the effective context window below native, and the autocompact threshold is derived from it. Resolution chain (from the issue): env CLAUDE_CODE_AUTO_COMPACT_WINDOW > user settings autoCompactWindow > GrowthBook > model native window. Enrollment is per-session and randomized, which is why some sessions were unaffected.

Goal

Make the override durable in the repo (not only in one operator's global settings), document the diagnostic, and correct the workflow guidance that attributes this failure to the wrong cause.

Workflow gap

  • Bug observed: Fable-5 subagents auto-compact at ~200-270k against a 1M native window and abort with the Autocompact-is-thrashing guard, fleet-wide (28 of 44 subagents on 2026-08-04 across 12+ sessions); the harness error text blames large reads but measured reads were <=12k tokens.
  • Why it is a workflow gap: the fix currently lives only in ~/.claude/settings.json (operator-global, unversioned, not in any clone), and .claude/rules/gotchas.md has no entry, so the next agent to hit this re-derives it from scratch. This session burned ~2h and falsified four wrong hypotheses before finding it.
  • Confidence (emitter): high — reproduced and verified end-to-end (below).
  • verified-at-filing: grep -rn --exclude-dir=worktrees "AUTO_COMPACT_WINDOW" .claude/rules .claude/skills .claude/agents .claude/settings.json CLAUDE.md scripts -> 0 hits; grep -rn --exclude-dir=worktrees -iE "tengu_amber|compact_boundary" <same paths> -> 0 hits; .claude/settings.json exists with no env block (2026-08-04 UTC). All three are ABSENCE-of-guard claims, so the 0-hit in-target result IS the evidence per .claude/rules/workflow-fix-on-bug.md clause (a). Per clause (a'), the absence claim was additionally probed semantically: the override is an exact env-var name with no shorter substring form, and the behavioural probe below confirms it was not otherwise in effect. Landed-fix history check: git log --oneline --since='7 days ago' -- .claude/settings.json .claude/rules/gotchas.md at compose time shows no autocompact-related commit. NOTE the --exclude-dir=worktrees scoping is load-bearing — a bare recursive grep of .claude/ traverses the /mnt/eps-data bind mount and times out (already documented in gotchas.md; it timed out once during this filing).

Evidence (measured this session, not recalled)

Baseline, the 7 wedged subagents of session f0789a8a — compact_boundary rows, trigger: auto:

preTokens = 202,059 / 206,725 / 206,835 / 206,857 / 208,690 / 211,323 /
            218,240 / 218,378 / 220,013 / 220,335 / 220,684 / 223,884 /
            225,161 / 235,834 / 248,696 / 249,587 / 251,958 / 252,062 /
            262,619 / 269,889

Controlled after-fix run, same model, override set:

beforeafter
auto-compact fired at202,059-269,889never
compactions3, then abort0
peak input+cache213,411 (killed)691,417
outcomeaborted after 4 tool calls7 full-file reads, exit 0, 13m34s, 80 requests

Live Models API: claude-fable-5 max_input_tokens=1,000,000. Delivery verified: a fresh claude -p with no CLI env var reports CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000, i.e. the settings.json env block propagates into the tool subprocess environment (it does NOT appear in /proc/<pid>/environ — that is a false-negative check, do not use it).

Proposed change

  1. .claude/settings.json — add the env block (file currently has none):

    "env": { "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000" }

    Verify against the model's real native window rather than hardcoding 1M blindly for a future smaller-window model; 1M is correct for the current fable-5/opus-5 fleet.

  2. .claude/rules/gotchas.md — new entry with: the symptom, the compact_boundary preTokens diagnostic (far below native = enrolled; near-native = normal), the override + its documented limitations (DISABLE_AUTO_COMPACT=1 does NOT affect it; CLAUDE_AUTOCOMPACT_PCT_OVERRIDE is capped by Math.min(); reported to fail on some arm64 native builds — we are x64), the /proc false-negative, and the fact that the harness error text is misleading ("a file being read or a tool output is likely too large" was false in every measured case here; largest read 12k tokens).

  3. Correct the wrong-cause guidance. CLAUDE.md's autocompact-thrash bullet and the -lean twin agent descriptions attribute this class to subagent fixed-overhead (spec + CLAUDE.md import tree + MCP schemas). That is NOT the cause here: the 977,125-peak fable-5 agent had a higher floor (233,432) than the victims, and #1336's own marker records a lean twin thrashing too. Add the reduced-window cause as the first thing to check, with the compact_boundary diagnostic to discriminate. Do not delete the fixed-overhead guidance — it remains a real (separate) class.

Scope / surfaces

  • Primary target: .claude/settings.json
  • Also: .claude/rules/gotchas.md, CLAUDE.md (autocompact-thrash bullet), .claude/rules/LESSONS.md if a new rule file is added instead of a gotchas entry, and the -lean agent descriptions under .claude/agents/.
  • OUT of scope: ~/.claude/settings.json (operator-global, already applied by hand this session; the repo copy is what this task makes durable).

Constraints / invariants

  • Workflow-surface only. No experiment code, configs/, or tasks/.
  • scripts/workflow_lint.py (no-flags) passes; ruff clean on touched files.
  • .claude/settings.json must stay valid JSON — validate with a parse after editing (a malformed global settings file breaks every future session).
  • Adding an env block must not disturb existing keys (permissions, hooks, enabledPlugins, model).
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 / a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard). Subagents still EMIT candidate blocks normally; the orchestrator parks them (.claude/rules/workflow-fix-on-bug.md § Recursion guard, and its brief-composer clause).

Upstream references (all read this session)

  • anthropics/claude-code#46331 — tengu_amber_redwood silently reduces the autocompact window; closed "not planned"; documents the resolution chain, the override, and its limitations.
  • anthropics/claude-code#43989 — v2.1.92 regression, threshold reduced to 400k on Opus 4.6 (1M context). OPEN.
  • anthropics/claude-code#60485 — thrashing on fresh sessions with ~41 KB of context; closed as duplicate. Confirms the error text misleads.
  • anthropics/claude-code#54056 — auto-compact at ~367K on Opus 4.7 [1m].
  • anthropics/claude-code#42394 — fires despite DISABLE_AUTO_COMPACT=1.
  • anthropics/claude-agent-sdk-python#958 — same thrash in the SDK.

Caveat to carry: the documented arm reduces to ~60% of native (1M -> 600k), but our observed threshold implies ~235-300k — a more aggressive arm than the issue describes. The override sits above the experiment in the chain either way, which is why the fix works regardless.

Provenance

  • workflow_fix_target: .claude/settings.json
  • fingerprint: 77e9d6470b27

Originating user instruction: "run the infra task now" (2026-08-04), following a live diagnosis + verified fix in this session.

Activity