superscalar

{
  "name": "superscalar",
  "version": "0.5.2",
  "description": "Execution-scheduling discipline for parallel sub-agent dispatch. v0.5.0 / Superscalar v0.14.0 adds the /speculation toggle (§4 speculation authorization) — four modes plus a read-back (off default / auto asks per speculation with the four-element trade-off / on pre-authorizes low-downstream-sensitivity lanes while high-sensitivity ones still ask / always pre-authorizes every lane inside the §3 barrier / status reports the mode and where it would next apply). The modes graduate along downstream sensitivity, an axis §4 already named, so this is a threshold on existing vocabulary rather than new vocabulary. One constraint governs all four and cannot be read past: the mode changes WHO authorizes a lane, never WHAT a lane may touch — read the other way, there is no mode in which speculation reaches a write, deploy, send, or delete, so no operator setting turns speculation into an outward-facing action; widening scope is a separate and narrower decision than changing who signs off. Andon transparency (visual signal, /stop-spec pull-cord, misprediction logging) is unconditional at every mode including always, because pre-authorization removes the question and not the visibility — dropping the announcement at always would make the loudest mode the quietest one. State is one marker (.agent/speculation.json, absent ⇒ off) read at dispatch time and never mirrored, and a mode change is a declaration event where a board is present. v0.4.0 / Superscalar v0.12.0 adds the /context-caching toggle (§5.4 memory-hierarchy borrowing) — three modes (off/self/agent): pin model+effort at session head (each is a cache key), compact at task boundaries with rewind-over-compact for abandoned paths, prefetch context right after boundaries, account for cache-cold subagents in fan-out pricing, consult the per-harness flush list before mid-session mutations, and measure with the cache counters instead of declaring; volatile vendor numbers (multipliers, TTLs, minimums, storage rents) live in cache-registry.json under the dated/confirmedBy/revisit contract with singleSource marks. v0.3.0 / Superscalar v0.11.0 adds the /ooo question-scoreboard toggle (§5.3) — out-of-order execution for the main lane itself: a mid-task question registers as a one-line scoreboard entry instead of stalling the queue (self-resolve by measurement → bypass independents → assumption-run only under §4 speculation gates → park), late answers wake only their dependents, and loss/external-publish gates are never bypassed — the toggle moves the waiting, never the gates. v0.2.0 / Superscalar v0.6.0 adds the /subscaler toggle (§5.1 tiered model composition): when the main agent runs a frontier reasoning model, delegate execution-shaped work (code writing/editing) to the strongest execution-tier model as subagents — Fable-class→Opus-class / Sol-class→Terra-class, effort high — with a delegation rubric (spec-complete work delegates, architecture/judgment/review stays), per-harness application surfaces (Claude Code Agent-tool model param + Workflow opts; Codex .codex/agents TOML; Kimi per-invocation), single-marker-file toggle state, and default OFF (recommended ON for fan-outs, where the shared prompt cache is already forfeited). v0.2.4 / Superscalar v0.10.0: five-tier vocabulary — T1.5 frontier-execution band (the vendor's start-here flagship at ~half T1 price, split from T1 on temperament: it converts thinking into action sooner, so T1.5 lanes always pair with verification gates) + per-account entitlement check for invitation-gated models. Wraps Superscalar.md as model-invoked skills the agent consults before fan-out decisions: issue_width formula + worktree-isolated reorder buffer + cost-benefit gate + opt-in speculation + adoption thresholds + n=8 dogfood ledger (Entry 05 absorption sync to bundle 005 — case 8 first-worktree-write data point + §2b nested-repo limitation + R3 + secondary verification-depth lesson) + Entry 06 controlled A/B measurement showing 'parallelism is not sufficient; orchestration discipline produces consistency' + v0.4.1 §3.1 Hyperbrief decision-delegation interlock (write/deploy/send lanes pass through both cost-benefit gate AND hyperbrief-trigger-check; read-only lanes exempt).",
  "author": {
    "name": "SoliEstre",
    "url": "https://github.com/SoliEstre/EstreGenesis"
  },
  "license": "Apache-2.0",
  "homepage": "https://github.com/SoliEstre/EstreGenesis/blob/main/Superscalar.md",
  "repository": "https://github.com/SoliEstre/EstreGenesis"
}