The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze session transcripts, which are untrusted data that may contain malicious instructions meant to influence the analysis subagents.
Ingestion points: The skill locates and reads JSONL transcript files from directories such as $PSTACK_TRANSCRIPTS_DIR (referenced in SKILL.md).
Boundary markers: Subagent prompts (divergent-reviewer.md, judgment-reviewer.md, tooling-reviewer.md, synthesizer.md) contain explicit warnings that transcripts may contain prompt-injection attempts and instruct the models to ignore any embedded directives.
Capability inventory: The skill has the capability to read local files, spawn subagent tasks, access MCP tools (like ticket trackers or observability tools), and modify skill files (SKILL.md).
Sanitization: The process enforces a human-in-the-loop requirement, where the agent must present all proposed edits to the user and wait for explicit approval before modifying any files. Reviewers are also instructed to maintain a read-only posture regarding MCP tools.
[COMMAND_EXECUTION]: The skill uses shell commands to identify and list transcript files for processing.
Evidence: SKILL.md contains a bash snippet using ls, head, and head to retrieve the most recent transcript files from the agent's history directory.

reflect