The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local CLI tools, specifically codex and claude, to perform engineering tasks and environment verification.\n
The harness.py script manages the lifecycle of these processes, including launching, tracking (via /proc), and terminating process groups.\n
A 'startup probe' phase is implemented in worker_policy.py which executes /bin/true through the codex app-server to verify security policies (e.g., sandbox mode, approval policy) before initiating any model inference.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data which could potentially contain malicious instructions intended to influence agent behavior.\n
Ingestion points: Untrusted data enters the agent context through user_messages (original requests and amendments) and named_evidence (files declared in the task contract, limited to 32 KiB each).\n
Boundary markers: The instructions explicitly direct the agent to separate user messages from agent summaries and evidence. Prompts include a specific boundary string emphasizing that evidence is not a user instruction.\n
Capability inventory: Depending on the authorized policy, the skill can perform file reads, writes, and shell command execution (Bash) via the underlying CLIs. The controller script harness.py also performs sub-process operations and filesystem management.\n
Sanitization: The skill uses strict JSON schema validation for both the task contract and the model response. It enforces hard ceilings on log sizes (8 MiB), prompt sizes (256 KiB), and file read buffers to prevent resource exhaustion or overflow attacks.

orchestrate-harnesses