The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The orchestration script run-codex-handoff.sh executes the codex CLI with the --dangerously-bypass-approvals-and-sandbox flag for implementation agents. This grants delegated agents broad permissions to read, modify, or delete files accessible to the host account without interactive user approval for each step.
[COMMAND_EXECUTION]: The skill relies on executing the local codex binary and shell scripts (run-codex-handoff.sh, watch-codex-wave.sh) for agent lifecycle management. It also utilizes git for worktree identification and python3 for monitoring event streams.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input and sub-agent results, propagating them through the orchestration pipeline.
Ingestion points: User-provided task arguments and sub-agent result artifacts (JSON/stderr).
Boundary markers: Implementation prompts use structured blocks for outcome summaries, briefs, and constraints to isolate instructions.
Capability inventory: Broad file system access, command execution via codex, and coordination claim management via ai-coord.
Sanitization: Results are validated against strict JSON schemas (result.schema.json and research-result.schema.json).
[DYNAMIC_EXECUTION]: The skill dynamically generates implementation prompts for sub-agents at runtime based on the approved plan. Additionally, the watch-codex-wave.sh utility executes an embedded multi-line Python script to process agent event logs.

codex-handoff