The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted session transcripts to generate skill edits and backlog items.\n
Ingestion points: Reads JSONL transcript files located in the workspace's transcript store as identified in SKILL.md.\n
Boundary markers: Explicit instructions are included in the prompt templates for all subagents (divergent-reviewer.md, judgment-reviewer.md, tooling-reviewer.md, and synthesizer.md) to treat transcript data as untrusted and ignore any instructions found within it.\n
Capability inventory: The skill can execute shell commands, read and write files (editing SKILL.md files), and spawn subagents with access to MCP tools (e.g., ticket trackers, observability logs).\n
Sanitization: The skill relies on prompt-based instructions rather than automated data sanitization to handle untrusted input.\n- [COMMAND_EXECUTION]: The skill executes a shell command to locate and list the most recent transcript files for analysis.\n
Evidence: ls -t <agent-transcripts>/*.jsonl <agent-transcripts>/*/*.jsonl <agent-transcripts>/*/subagents/*.jsonl 2>/dev/null | head -10 in SKILL.md.\n- [DYNAMIC_EXECUTION]: The skill dynamically spawns three specialized reviewer subagents and a synthesizer agent using parallel Task calls to process the transcript data.\n
Evidence: Multiple Task calls in SKILL.md sections 2 and 3 using the pstack-harness skill for orchestration.

reflect