backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
EMERGING: RPI (Research → Plan → Implement) with throwaway tactical docs is winning over durable “spec-driven development.” — Durable specs become a second source of truth that drifts; for your loops, plans should be execution artifacts you discard after ship and re-research next time (tokens cheap, your time isn’t). — x.com/GergelyOrosz/status/2080917955404062887
Opus 5 is the daily driver but needs a tight leash: it concludes fast/confidently wrong unless you pre-load evidence; Fable becomes the occasional “Winston Wolf.” — Matches your conductor routing; raise the bar on context-before-commit and independent verification for Opus-led runs. — x.com/kunchenguid/status/2080823426261115075
EMERGING: multi-round “autoreview” as a skill/loop (66 rounds on a hard refactor), not a one-shot PR glance. — Strong pattern for gnarly refactors in your fleet: budget multi-pass review skills instead of single-pass “looks good.” — x.com/steipete/status/2080899298838098034
Default Codex to auto-review sandboxing, not full access. — OpenAI staff guidance for weekend/full-access thrash; lower blast radius on the CX peer seat unless the workdir is already isolated. — x.com/reach_vb/status/2081044499879649681
Agent escape risk is real (HF eval sandbox); defend with sandboxed agent runtimes + defensive scanners (deepsec framing). — Your multi-agent host fleet is the surface; don’t run high-permission agents on the main machine without isolation. — x.com/rauchg/status/2081047912008872293
Daily scheduled Claude loop that finds and deletes dead code (cross-platform dead-everywhere constraint). — Cheap unattended GC pattern for large multi-target codebases; fits your standing-goal / loop engine. — x.com/jarredsumner/status/2080912636225695835
Claude Code now has different system prompts per model; file-harness tools are expected to fade as models shell more. — Stop assuming one CLAUDE.md/skill stack behaves identically across Opus/Fable seats; don’t overbuild on File-tool abstractions. — x.com/trq212/status/2081043974450450775 · x.com/trq212/status/2080786587009307009
Codex is a fully open-source harness (auditable prompts + open-weights capable) — and had a same-day elevated-error incident that was mitigated. — Peer-seat transparency win; keep CX as failover-aware (status.openai.com) not “always up.” — x.com/reach_vb/status/2081058669144510787 · x.com/reach_vb/status/2080958657487917263
Highest-leverage action for today: Default the Codex peer seat to auto-review (not full access) and treat implementation plans as disposable RPI artifacts—delete after ship, re-research next time.