backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
Claude Code cut its system prompt ~80%; smarter models need less direction and fewer examples — trim your global skills/hooks/CLAUDE.md bloat on model upgrades so Fable/executor seats get room to run free-form instead of over-constrained mimicry of old examples — x.com/trq212/status/2078901672441790818
Hard-gate swarm harness features in user-level
AGENTS.mdbefore Fable can auto-spawn — Kun Chen burned full quota when Fable spun 52 Fable subagents on simple PR scans; his shared rule now: before “dynamic workflows” / “ultra code” / any large subagent swarm, explain tradeoffs and require explicit approval (dotfiles: github.com/kunchenguid/dotfiles/) — x.com/kunchenguid/status/2078584596371997106Do not route security/vuln analysis to Fable; Vercel stealth cyber evals: Fable 100% refuse, Sol best thoroughness (~7× cost), Kimi K3 best price/recall, use
deepsecharness — re-pin solidity-auditor / CSO / defensive review to CX Sol (baseline) or Kimi continuous, never conductor-Fable — x.com/rauchg/status/2078647648307880209Fable-on-subscription is capacity-restored permanent intent, not another soft extension — trq212 (in-window): goal remains Fable as a regular Max subscription part once capacity holds; keep Fable conductor doctrine past the Jul 19 soft cliff, still respect capacity/50%-of-limits constraints from the official rollout thread he amplified — x.com/trq212/status/2078595829150224663
Your installed
claudebinary already ships unreleased Bun-in-Rust (reports as Bun 1.4.0 / canary; public Bun still ~1.3.x) — treat Claude Code runtime upgrades as opaque binary embeds (startup/tooling quirks,BUN_OPTIONS=--preloadworks;BUN_BE_BUNdisabled on claude per mitsuhiko) — x.com/simonw/status/2078692298301587758 · simonwillison.net/2026/Jul/19/claude-code-in-bun-in-rust/EMERGING: “ChatGPT Work” ≈ OpenAI’s Claw-class agent product (desktop skins Codex; mobile ≠ Chat) — if you peer CX with OpenAI consumer surfaces, map Work to agent runtime not chat mode; expect more “Claw” framing as category name — x.com/simonw/status/2078650408747331709
Unexpected prompt-cache misses on the Codex backend (mitsuhiko; may be client-side) — watch CX token burn on multi-turn peer review; don’t assume cache hits until you measure — x.com/mitsuhiko/status/2078754048887451763
EMERGING: “loop people” as a social class whose opinions track their agents; vocabulary joke-shift loops → graphs → florps — treat loop-engineering dogma as agent-amplified fashion unless gated by your verifier/gate scripts; don’t rename harnesses for hype — x.com/mitsuhiko/status/2078768091996082294 · x.com/steipete/status/2078586663270592776
Highest-leverage action today: add Kun-style explicit-approval language for swarm/subagent fan-out to global agent policy, then prune one overloaded skill/CLAUDE.md section per trq212’s 80% trim lesson.