The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to trigger research waves via the codex CLI. Technical details in references/codex/codex-exec-contract.md specify the invocation of codex exec with the flag --dangerously-bypass-approvals-and-sandbox. This flag is explicitly used to circumvent sandboxing and approval prompts, which are fundamental security controls for sub-processes.
[REMOTE_CODE_EXECUTION]: Research tasks are dispatched to external executors that process untrusted data from the internet. Executing these tasks in un-sandboxed environments, as encouraged by the skill's technical documentation, represents a high-severity security risk.
[DATA_EXFILTRATION]: The skill's architecture uses the local filesystem as a context channel, reading files from the workspace and passing them as input to research jobs that have network access. This creates potential pathways for sensitive data to be exfiltrated to external domains.
[CREDENTIALS_UNSAFE]: Technical references in references/codex/codex-exec-contract.md describe accessing sensitive authentication paths, including ~/.codex/auth.json and ~/.codex/config.toml, to verify tool authorization status.
[PROMPT_INJECTION]: The skill ingests untrusted data from the web, creating a surface for indirect prompt injection attacks. 1. Ingestion points: Scraped content from the web and Reddit processed during research waves (references/subagent-briefs.md). 2. Boundary markers: Briefs define explicit input/output paths and use a seven-section prompt skeleton (references/codex/codex-prompt-skeleton.md). 3. Capability inventory: The skill possesses the ability to spawn subagents, execute shell commands, and run codex exec subprocesses with network and filesystem access. 4. Sanitization: The skill relies on LLM-driven citation discipline and synthesis protocols (references/synthesis.md) rather than automated sanitization or filtering.
[COMMAND_EXECUTION]: The skill includes a shell script scripts/init-corpus.sh and a Ruby snippet in references/industry/industry-architecture.md intended for execution by the agent to perform environment scaffolding and link validation tasks.

run-deep-research