The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests and processes untrusted data from external financial news and search APIs, creating an attack surface for indirect prompt injection.
Ingestion points: scripts/pit_fetch.py retrieves content bodies from Benzinga news and search results from Perplexity.
Boundary markers: There are no explicit delimiters or system instructions used when interpolating this external data into the agent's context to prevent the agent from following instructions embedded in the data.
Capability inventory: The skill is configured with broad tool access in SKILL.md, including Bash, Write, and Edit capabilities.
Sanitization: The scripts perform basic text cleaning for formatting but do not implement security-focused sanitization or instruction filtering on external content.
[COMMAND_EXECUTION]: Multiple components of the skill perform shell command execution and access local system files.
Subprocess usage: scripts/get_quarterly_filings.py runs the column utility. The test suite (scripts/test_pit_fetch.py) uses subprocess.run to execute project scripts for verification purposes.
Transcript access: scripts/build-thinking-index.py programmatically reads agent session transcripts from the .claude/projects directory to build a research index. While this is an intended feature for indexing past work, it involves automated processing of sensitive interaction history.
Hooks: SKILL.md utilizes a PostToolUse hook to automatically trigger a Python script (build_orchestrator_event_json.py) after Bash commands are executed.

earnings-orchestrator