code-claw

Manage agent runtime lifecycles from spawn through context fork, permission shaping, transcript recording, progress streaming, and deterministic cleanup. Use when Codex needs to implement or audit agent spawning, resume support, background workers, or subagent runtime scaffolding.

Encode desired agent behavior into durable prompt policy using concrete do and dont rules, escalation triggers, communication constraints, and reporting standards. Use when Codex needs to design or audit system prompts, agent constitutions, collaboration rules, or behavior control layers.

Design or audit permission systems that combine rule-based allow ask deny logic, safety checks, auto-mode classifiers, dangerous-rule stripping, and blast-radius-aware user confirmation. Use when Codex needs to build or review tool permission layers, approval workflows, or risk-scoped execution policies.

Keep long-running agent conversations healthy with attachment filtering, tool-result budgeting, snip and microcompact passes, context collapse, autocompact, reactive recovery, and queued-command hygiene. Use when Codex needs to build or audit long-context runtime loops, compaction systems, or context management pipelines.

Design, review, or extend hook-based governance layers for tool runtimes with pre and post hooks, permission mediation, continuation control, additional context injection, and non-fatal hook errors. Use when Codex needs to build or audit hook systems, policy middleware, or execution governance around tools.

Design, review, or debug MCP integration layers with multi-transport connectivity, auth lifecycle handling, tool and resource discovery, output persistence, session-expiry recovery, and elicitation support. Use when Codex needs to build or audit MCP clients, connector planes, or external tool integration systems.

Design, review, or operate multi-agent systems with role-specialized workers, scoped permissions, forked context, skill preloading, additive MCP connections, and explicit cleanup. Use when Codex needs to build or audit agent swarms, subagent delegation, worker roles, or cooperative task execution pipelines.

Design, review, or refactor prompt assembly systems with static and dynamic prompt boundaries, section registries, cache-aware composition, feature-gated overlays, and session-specific instruction injection. Use when Codex needs to build or audit system prompt pipelines, prompt builders, prompt layering, or instruction composition code.

Optimize prompt systems for cache efficiency, stable prefixes, fork-safe reuse, compact summarization, and instruction budgets. Use when Codex needs to design or audit prompt caching, prompt reuse, compaction prompts, skill listing budgets, or cache-safe agent forks.

Package reusable workflows as discoverable skills with command metadata, invocation contracts, char-budgeted listings, forked execution, and MCP-sourced skill merging. Use when Codex needs to create or audit skill systems, slash-command skills, or workflow packaging for reusable agent capabilities.

Design, review, or debug tool execution runtimes with input validation, hook integration, permission resolution, telemetry, structured results, and failure handling. Use when Codex needs to build or audit tool call pipelines, tool runners, tool middleware, or execution orchestration code.

Design or operate adversarial verification agents that independently validate implementation work with command-backed evidence, strict read-only boundaries, and explicit PASS FAIL PARTIAL verdicts. Use when Codex needs a post-implementation verifier, QA subagent, or evidence-first acceptance gate for non-trivial changes.