The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run, subprocess.Popen, and os.execve to execute local helper scripts and the official lark-cli tool for authentication and publishing.
Evidence: Execution patterns found in scripts/check_feishu_connection.py, scripts/run_feishu_auth.py, and scripts/run_feishu_config_init.py.
Mitigation: The skill uses scripts/feishu_process_environment.py to build a minimal execution environment, stripping sensitive ambient variables like proxy credentials or system tokens before launching child processes.
[EXTERNAL_DOWNLOADS]: The scripts/install_lark_cli.py script is designed to download the official Lark/Feishu CLI binary from GitHub.
Evidence: Downloads are fetched from github.com/larksuite/cli over HTTPS.
Safety: The script validates that URLs belong to official GitHub domains and requires a blocking user consent step as instructed in references/feishu-onboarding.md before any network activity.
[REMOTE_CODE_EXECUTION]: The skill executes the downloaded lark-cli binary to manage Feishu/Lark document creation.
Evidence: The binary is invoked by various Python helper scripts during the onboarding and publishing phases.
Safety: The implementation mandates SHA-256 hash verification of the binary before execution and employs hash-based identity pinning for the Feishu App ID and User Open ID to prevent unauthorized account switching.
[DATA_EXFILTRATION]: The skill's primary function is to transmit research data to the user's Feishu or Lark workspace.
Evidence: Document creation logic is managed by scripts/publication_checkpoint.py and documented in references/feishu-publishing.md.
Safety: Several scripts implement a sanitize_output function that uses regular expressions to redact credentials, including Bearer tokens, JWTs, and App Secrets, from process output and logs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from research papers, which creates a surface for indirect prompt injection.
Evidence: Data is ingested from sources like arXiv and OpenReview.
Safety: The references/evidence-policy.md and references/report-schema.md provide a structured validation process and require the agent to explicitly label and separate source facts from AI-generated interpretations.

feishu-paper-reading