The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill performs remote updates of the third-party tool via the "opencode upgrade" command. It also provides a shell command (curl piped to bash) for the user to manually install the tool if it is missing from the system.
[COMMAND_EXECUTION]: The skill executes various shell commands to synchronize the local repository (git pull --rebase), manage background processes (kill, pkill), and handle temporary log files in /tmp.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted output from the opencode tool which could contain malicious instructions. -- Ingestion points: The agent reads status lines and a final 40-line task summary from a local log file populated by the tool's output. -- Boundary markers: No specific delimiters are used to wrap the ingested log content in the agent's context. -- Capability inventory: The skill possesses capabilities to execute shell commands like git, kill, and rm. -- Sanitization: While status polling uses "printf %q" for basic shell escaping, the final task summary is ingested into the context without explicit sanitization or filtering.

opencode-runner