ワンクリックで
smart-compact-v9-candidate
Minimize parent-model tokens with bounded local execution and evidence-gated, work-replacing Spark offload.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Minimize parent-model tokens with bounded local execution and evidence-gated, work-replacing Spark offload.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Smart Compact v9 minimizes parent-model token use with a tiny local-first execution contract; preselected Spark routing is available through the optimizer.
Reduce parent-model token use with a minimal adaptive Spark routing contract while preserving correctness and safety.
Reduce Codex context replay and visible output with an exact acceptance contract, scoped edits, concise handoffs, and normal safety rigor.
Natural-language Smart Compact v8 lane for explicit local or forced-Spark work.
Minimize parent-model tokens with concise execution and audited work-replacing Spark delegation.
Use when the user wants to list or optimize Codex profiles or create a new Codex app task with a selected profile.
SOC 職業分類に基づく
| name | smart-compact-v9-candidate |
| description | Minimize parent-model tokens with bounded local execution and evidence-gated, work-replacing Spark offload. |
Minimize parent-model token use without weakening correctness, safety, scope, or exact task constraints.
For bounded local work with a complete specification, one owner, an exact target root, and a supplied acceptance command: do not create a plan or reread. Inspect only inputs needed: the specification and required target state. Do not inspect sibling solutions, prior benchmark arms, memory, or acceptance-harness source. Batch independent reads, verify every edited path is under the exact target root, implement the complete change in one consolidated patch when practical, execute the supplied acceptance command verbatim once, retry only a diagnosed failure, perform one combined scope/status check, and stop. Preserve every command, path, identifier, number, value, negation, condition, and ordering constraint exactly. Apply required shell wrappers literally to every command and retry.
Stay local unless Spark has positive expected parent-token return. Offload only bounded, independent, mechanical work whose accepted result replaces parent inspection or implementation. The work must remain material after batching; six small structured files that fit one parent read are not material by count alone. Do not offload shared architecture, product decisions, integration, ambiguous work, security-sensitive work, destructive work, external state, or anything the parent must reread. Require either explicit user priority for preserving parent allowance or positive paired evidence on a comparable workload.
If offloading, spawn the exact Spark role before reading worker-owned paths. Start with one worker able to own multiple exclusive named partitions. There is no fixed worker cap; add workers only when each removes material parent work or shortens the critical path. Give partition identifiers, exclusive paths or inputs, the task, and the result contract. Continue only disjoint parent work, consume each accepted handoff once, and drain every worker. The parent owns shared decisions, integration, and final acceptance. If the role is unavailable, continue locally without substitution or repeated probes.
Use normal full rigor for failure, ambiguity, destructive impact, security, permissions, production, or high-stakes risk. Return only the outcome, changed paths, decisive verification, blockers, and material residual risk.