superscalarlisted

Use BEFORE dispatching multiple sub-agents in parallel (Agent tool fan-outs, Workflow.parallel/pipeline, multi-lane Edit operations). Consult the issue_width formula to bound concurrency, apply the cost-benefit gate to decide spawn-vs-inline, honor the irreversibility barrier (write/deploy/send retire-gated), enforce in-order retire, and respect the v0.4 nested-repo worktree limitation. **v0.4.1 §3.1 Hyperbrief interlock** — for write/deploy/send lanes that pass the cost-benefit gate, also invoke `hyperbrief-trigger-check` (4-score escalation + 5 MUST-trigger); on FULL_HYPERBRIEF pause the lane and emit Constellation `DECISION_REQUEST + HyperbriefCard`, await `ack_tier='decided'`; read-only lanes are exempt by construction. Especially relevant when the work shape is "audit / cross-dimension consistency / handover-grade output" — where Entry 06's A/B measurement showed Arm B (discipline) catches the contradictions Arm A (naïve max-parallel) leaves silently unresolved.
SoliEstre/EstreGenesis · ★ 5 · AI & Automation · score 77

Install: claude install-skill SoliEstre/EstreGenesis

# Superscalar — execution-scheduling discipline You are about to dispatch parallel work. Before you do, run through this checklist. ## 1. issue_width formula (the concurrency cap) Lane class matters. Read-only and write/worktree lanes have DIFFERENT caps: ``` issue_width_read = min(effort_band, runtime_concurrency_ceiling, autonomy_available_workers) issue_width_write = min(effort_band, pace_mode_cap, Little's-Law-review-throughput, Kanban-WIP, autonomy_available_workers) ``` - **effort_band** — Anthropic effort tier (Cautious 2-4× / Proactive 5-6× / Burst 6-8× / Sprint 9-10×) - **pace_mode_cap** — write lanes only (Burst = 6, Sprint = unbounded modulo runtime ceiling) - **Little's-Law** — write only (review throughput ÷ task duration; binds retire-merge) - **Kanban-WIP** — write only (≈ team_size + 1; binds WAW/merge contention) - **autonomy_available_workers** — both classes (Haiku-tier worker = 0 for autonomous lanes per dogfood telemetry) - **runtime_concurrency_ceiling** — `min(16, cores - 2)` per workflow; `max parallel tool calls` per agent-tool fan-out `effective_concurrent_lanes = min(policy_cap, runtime_ceiling)`. `issue_width_write` is a hard policy bound; `issue_width_read` is a soft preference under the runtime ceiling. ## 2. Cost-benefit gate (spawn vs inline) Open a lane only when `estimated isolation + merge overhead < expected parallel/early-start benefit`. Inputs: worktree setup cost · task size · dependency fan-out · conflict likelihood · pace mode