The Agent Skills Directory

Dynamic Command Execution: The skill includes a mechanism to track and clean up resources by executing shell commands registered during the skill's lifecycle.
Evidence: In scripts/infra_ledger.sh, the sweep command retrieves stored teardown instructions from a ledger file (infra.tsv) and executes them using bash -c "$destroy". This pattern allows the skill to dynamically execute commands that were registered by earlier sub-processes or 'lanes'.
Background Process Persistence: The skill utilizes a long-running watchdog loop to monitor task progress and ensure the agent session remains active over extended periods.
Evidence: The scripts/highway_watchdog.sh script is designed to run in a detached tmux session for up to 12 hours. It periodically checks the state of active lanes and uses the platform's resume capability (amplifier run --resume) to wake the orchestrator session if a lane ends or the manager becomes inactive.
Indirect Prompt Injection Surface: The manager processes data and status markers generated by autonomous sub-agents to make merging and scheduling decisions.
Evidence: The skill's architecture involves a manager session that ingests results from 'lanes'.
Ingestion points: scripts/verify_lane.sh reads DONE.json and log files from sub-agent worktrees; scripts/infra_ledger.sh reads registered commands from infra.tsv.
Boundary markers: The skill uses structured JSON markers (DONE.json) and tab-separated ledger files to delimit data.
Capability inventory: The skill possesses the ability to execute arbitrary shell commands via bash -c, perform git merges, manage tmux sessions, and resume agent sessions.
Sanitization: scripts/infra_ledger.sh performs validation to ensure that registered infrastructure identifiers and commands do not contain tabs or newlines that could corrupt the ledger format.

ten-lane-highway