The Agent Skills Directory

[COMMAND_EXECUTION]: The launch.sh script dynamically generates a bash script (orchestrator.sh) at runtime and saves it to the fleet root before execution.
[COMMAND_EXECUTION]: The orchestrator script constructs and executes shell commands for AI agents using bash -c in the main loop of the orchestrator.sh script.
[COMMAND_EXECUTION]: The skill executes a user-provided eval_command (defined in fleet.json) within a loop to measure project metrics, which can include arbitrary shell commands.
[REMOTE_CODE_EXECUTION]: The agent is instructed to perform an autonomous 'research' loop which involves modifying local source code and executing it via the eval_command to optimize performance metrics.
[COMMAND_EXECUTION]: The skill manages global state by reading and writing to ~/.claude/fleet-registry.json, accessing the user's home directory via shell operations in lib/registry.sh.
[PROMPT_INJECTION]: The agent's instructions include reading results.tsv, which contains output and descriptions from previous iterations of the agent. This creates an indirect prompt injection surface where adversarial content in one iteration could manipulate the agent's behavior in subsequent iterations.
Ingestion points: results.tsv (read by both the orchestrator prompt builder and the agent).
Boundary markers: Absent; the agent is not instructed to treat previous results as untrusted content.
Capability inventory: The agent is provided with Bash, Read, Write, and Glob tools, and nested agents are granted shell access.
Sanitization: Absent; descriptions generated by the agent are recorded in the TSV and read back without validation.

autoresearch-fleet