The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external project configuration files (.codex/config.toml, .codex/agents/*.toml) and user-provided task descriptions to determine routing logic. This introduces an injection surface where external data could influence agent orchestration. The skill mitigates this through an explicit boundary verification step, a requirement to prioritize upper-level system rules, and a logic flow that defaults to human review for high-risk or ambiguous tasks.
Ingestion points: Project configuration files (.codex/config.toml) and runtime metadata referenced in SKILL.md and references/codex-project-onboarding.md.
Boundary markers: Explicitly mandates reading system and project rules first ("上層規則永遠優先") and verifying task boundaries in the "先確認邊界" section.
Capability inventory: File system writes (configuration generation) and sub-agent dispatching logic.
Sanitization: Enforces a "lowest competent tier" selection policy and requires explicit interactive user selection ("建立建議設定", "只顯示建議", "暫不設定") before any configuration modification.
[DYNAMIC_EXECUTION]: The skill generates agent configuration files (.toml) based on pre-defined templates in references/codex-project-onboarding.md. These files include system instructions (developer_instructions) for specialized sub-agents such as "Terra Worker" and "Sol Reviewer". The skill adheres to safe practices by requiring explicit user confirmation before writing these files and ensuring the generated configurations follow a minimal, least-privilege pattern (e.g., sandbox_mode = "read-only" and approval_policy = "never" for reviewers).
[COMMAND_EXECUTION]: The skill references the execution of the codex CLI tool to verify runtime capabilities and versions (e.g., codex --version, codex features list) as part of its environment verification process. These operations are descriptive "evidence" steps intended to ensure the environment supports required orchestration features and are not used to execute arbitrary, untrusted, or user-supplied shell commands.

custom-agent-router