The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from external sources during its history mining phase.
Ingestion points: The mine.mjs script (via lib/opencode.mjs and lib/github.mjs) reads session history, including user messages and turns, from the local OpenCode database (~/.local/share/opencode/opencode.db) and retrieves pull request comments and commit data via the GitHub CLI.
Boundary markers: The script is documented to operate in a read-only Phase 0 mode, and output quotes extracted from external data are truncated to a maximum of 240 characters.
Capability inventory: Across its various modules, the skill has extensive filesystem access, can execute shell commands (via gh and playwright-cli), and manages Git and Graphite repository states.
Sanitization: A redaction mechanism implemented in lib/classify.mjs uses regular expressions to identify and mask sensitive information such as GitHub tokens, AWS keys, and Bearer tokens before data is written to the output failure-mode logs.
[COMMAND_EXECUTION]: The skill utilizes external command-line tools to interact with repository services and perform automation.
Evidence: The lib/github.mjs script uses node:child_process.spawn to execute the gh (GitHub CLI) tool for searching pull requests and fetching API data. Additional documentation in the skill references the use of playwright-cli and Executor for browser automation and performance analysis.
Security Controls: The script constrains command arguments (such as issue IDs and pull request numbers) using regular expressions and type-casting to mitigate the potential for command injection.

computa-please