CODY BENCH V13 · 4 Fix Clusters · Anthropic Issue Intelligence

CODY BENCH V13 — Anthropic Issue Fix Clusters

The Cody Bench tracks Anthropic/claude-code issues where EOSE fleet = reference implementation or fix provider. 4 fix clusters identified from Day 96 Anthropic diamond pull. Each cluster = related issues sharing a single root cause. One architectural fix per cluster = all sibling issues resolve. Meek Flip Score = community traction × EOSE uniqueness (1–10).

TUI SCROLL / FLICKER

ASSESS Type B kinetic Effort M

#826 Console scrolling to top on every render 350💬 DUPE

#1913 Terminal Flickering during output 187💬

#57785 Duplicate lines appear on terminal resize 1💬

ROOT:Component re-renders trigger scroll reset. React state change → full re-render → scroll position lost. All three share this root.

FIX:Scroll-lock guard on render cycle. Preserve scroll position across state updates. One architectural guard = all three resolve.

EOSE:We hit this on msi01 WSL2. Fix = scroll-lock guard on render cycle. Terminal behavior confirmed reproducible in our environment.

Meek Flip Score /10 — fixable but not fleet-unique

MEMORY LEAKS

CLEARED (COI) Type G arch S+M

#8856 /tmp/claude-*-cwd never cleaned up on exit 106💬 HAS REPRO

#4953 Process grows to 120GB RAM then OOM killed 87💬

#11315 129GB RAM system freeze — machine unresponsive 53💬

#1421 JS Heap OOM while thinking (extended reasoning) 66💬

ROOT:No exit handler for tmpdir cleanup + no memory ceiling on thinking token buffer. Two-part root, same fix family.

FIX:process.on('exit') cleanup handler for /tmp/claude-* dirs + thinking token hard limit to prevent heap overflow.

EOSE:msi01 64GB WSL2 cap = we WILL hit #4953. #8856 has repro. MT-001 (Cluster Theorem): fixing #8856 alone without #4953 = zombie cluster. Fix both atomically.

Meek Flip Score /10 — has repro, EOSE hits it, fleet-critical

COMPACTION

CLEARED (COI) — PR: CompactionConfig Type G arch Effort L

#23 non-interactive /compact command — help wanted 16💬

#17428 Enhanced /compact with file-backed summaries 44💬

#20696 Compaction deadlocks / fails intermittently 59💬

ROOT:/compact is UI-only. No programmatic API surface for fleet-managed compaction. No CompactionConfig in SDK.

FIX:CompactionConfig in SDK — programmatic compaction trigger, threshold config, file-backed summary support. Fleet can call it without UI.

EOSE:OpenClaw fleet sessions need programmatic compaction. We have the design. PR = CompactionConfig in SDK. EOSE Labs Inc. is the entity contributor.

Meek Flip Score /10 — fleet-critical, we have the design, PR ready

AGENTS.MD / SKILLS DISCOVERY

CLEARED (COI) — HIGHEST PRIORITY

Type G arch Effort M EOSE PRIOR ART

#6235 AGENTS.md support for claude-code context loading 294💬 PRIOR ART

#9716 Skills not visible / discoverable in .claude/skills 69💬

ROOT:Claude Code has no standard way to load agent context files (AGENTS.md, SOUL.md, USER.md) or discover local skills (SKILL.md).

FIX:Structured context file loading spec — AGENTS.md as crew manifest, SKILL.md as skill descriptor, .claude/skills as discovery path.

EOSE:We WROTE AGENTS.md. OpenClaw loads SOUL.md + USER.md + MEMORY.md + SKILL.md. We ARE the reference implementation. #6235 has 294 comments = highest community demand. Our impl predates any spec. PRIOR ART CONFIRMED.

Meek Flip Score /10 — WE OWN THIS. Community wants it, we built it, prior art confirmed.

4

fix clusters

3

COI cleared

1

assess

11

total issues

All CLEARED clusters must be submitted under EOSE Labs Inc. entity. Never as individual (MT-003 COI Duality Theorem). Amani GC + Harvey CLO required for each PR.