Optimize Codex Usage

Optimize Codex Usage

Reduce total Codex consumption while preserving the requested outcome, authorization boundaries, safety checks, and completion evidence.

Load Policy

• For a conceptual explanation or a narrow current-product fact, use this file and current official OpenAI documentation only. Do not inspect a repository or load the audit reference unless the request needs it.
• For an audit, design change, or implementation involving Skills, AGENTS.md, MCP, context, history, tools, validation, or reporting, read references/context-audit.md before changing files.
• Add another Axiom skill only when the requested implementation crosses into that skill's distinct owner: an AGENTS instruction system, traceable Git state, or a persistent system change. Usage analysis alone needs only this skill.

Measurement Boundary

• Use exact token, credit, cache, reasoning, or tool metrics only when the host exposes them for the scoped run. Axiom cannot recover hidden usage data.
• Otherwise label UTF-8 bytes, words, lines, reference counts, tool calls, and route-chain sizes as proxies or estimates.
• Do not claim a percentage saving without a repeatable before/after workload and equivalent quality acceptance.
• Fetch current official OpenAI documentation only when a volatile Codex fact affects the decision. Never place current prices, plan limits, or model quotas in always-loaded instructions.

Non-Negotiable Boundaries

• Define the required quality, safety, authorization, rollback, and evidence bar before optimizing. Lower resource use counts only when that bar still passes.
• Do not remove required tests, rollback checks, secret handling, destructive safeguards, or outcome-owning evidence to reduce context or calls.
• Do not automatically change the model, reasoning effort, execution mode, or host configuration. Offer a bounded comparison only when the user requests usage optimization and the host supports the choice.
• Preserve unrelated work. Do not reset, stash, clean, stage, commit, push, deploy, install tools, change user Skills, or mutate remote state without separate authorization.
• Do not add telemetry, background work, automatic updates, or network checks.

Optimization Order

1. Bound the target workflows and representative requests.
2. Measure always-loaded and common-path context before rare worst cases.
3. Inventory metadata and route edges before reading bodies.
4. Read only candidate owners, then remove duplication or route detail behind one directly discoverable on-demand reference.
5. Reduce repeated tool calls, checks, updates, and reports only where the next decision does not depend on each intermediate result.
6. Re-run the same routing, authorization, safety, and outcome scenarios.
7. Stop when further reduction would trade away required quality or evidence.

Prefer precise trigger descriptions, one canonical owner per rule, phase-specific references, bounded retries, conclusive stopping conditions, and concise evidence-led reporting. Keep ordinary no-match work outside Axiom.

Report

Lead with the implemented or recommended outcome. Include affected files or surfaces, comparable before/after measurements, validation results, and real unverified gaps. Omit repeated rationale, routine successful details, and exact usage claims the host did not expose.