backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
Never run Codex (or any coding agent) in full-access without sandbox + auto-review — GPT-5.6 has deleted $HOME by mistaking it for a temp dir — OpenAI confirmed the failure mode (full access, no sandbox/auto-review, $HOME env override then accidental delete); your CX peer seat must stay sandboxed with high-risk action gates, not “yolo full disk.” — x.com/reach_vb/status/2077631651203432508 (incident detail: x.com/thsottiaux/status/2077630111499882637)
Boris: highest-leverage eng work is encoding domain knowledge as agent infra (CLAUDE.md / REVIEW.md / skills / memories / lint+CI), not re-prompting one-offs; “loops” = automate whole classes of busywork forever — directly validates your hooks-over-prose + skills doctrine; add explicit
REVIEW.md(and treat missing auto-catchable patterns as automation failures, not human process debt). — x.com/bcherny/status/2077460395279692197EMERGING: “risk-based control gates” + open the CL early with design/prompt/plan before the agent implements — answer to exploding review load — Huntley’s reply to eng-leader review crisis: higher-power tooling, risk-tier gates, and design-first agent work instead of reviewing finished agent dumps. — x.com/GeoffreyHuntley/status/2077755631734599931 · parent: x.com/GergelyOrosz/status/2077694965300244793
EMERGING: “fresh-context merge judge” auto-merge path — AI LGTMs → second Claude with clean context on simplicity/blast-radius → auto-merge — Jarred’s near-term Bun endgame (human still merges today because CI/review rules aren’t good enough yet); maps cleanly onto your automated-gate + cross-provider tiers for low-blast work. — x.com/jarredsumner/status/2077504710739956019 · AI↔AI PR theater already live: x.com/GergelyOrosz/status/2077479764604883339
Claude Code 5h+weekly usage % reset; Codex usage also reset — but timers may not move with the % fill — temporary capacity on C1/C2/CX; don’t assume weekly clock restarted (Theo: % reset, existing weekly timer still firing). — x.com/theo/status/2077604800880087284 · x.com/theo/status/2077625416651800752 · x.com/theo/status/2077609569287823869
Code review as sole quality gate is collapsing under agent throughput; root-cause thinking is shifting to automation (tests/o11y) that should have gone red — eng leaders have no scaled human-review answer; strengthens keeping
no-mistakes/CI as floor and treating review as optional tier, not the control plane. — x.com/GergelyOrosz/status/2077694965300244793 · x.com/GergelyOrosz/status/2077696715893710915Agent-readable secrets should require Touch ID (1Password / macOS keyring) so privileged agent actions wait on a human biometric review — concrete pattern if any agent ever needs prod/infra credentials; pairs with the Codex full-access incident. — x.com/mitsuhiko/status/2077786098957230517 · x.com/mitsuhiko/status/2077797016298557739
Claude Code: prefix
!to run shell; stdout returns into the conversation for Claude to continue — small operator UX win for conductor sessions (force a command without burning a full agent tool-turn of intent). — x.com/delba_oliveira/status/2077786404956913756
Also noted (not ranked): OpenAI Dev Office Hours tomorrow 11:00 BST on GPT-5.6 / ChatGPT / Codex (https://x.com/reach_vb/status/2077796227651874830); Theo still HTML-pilled for plans/artifacts vs Codex inline HTML (https://x.com/theo/status/2077661371282714629); thread pressure toward universal AGENTS.md over tool-specific CLAUDE.md-only (replies under Boris).
Highest-leverage action today: audit every Codex/OpenClaw/full-access profile — sandbox on, auto-review on, no full-access; block $HOME/env override delete paths before any more CX runs.