The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to analyze "full transcripts" and "task statements" from other agent trials (found in jobs/analysis-briefs/). This presents an attack surface for indirect prompt injection where malicious instructions embedded in a trial transcript could influence the analyzer sub-agents.
Ingestion points: Files located at /w/seungpil.lee/deepswe/jobs/analysis-briefs/<id>.md and .win.md containing full agent message histories.
Boundary markers: The workflow prompts do not use explicit delimiters or "ignore previous instructions" warnings when presenting the transcripts to the judge agents.
Capability inventory: The skill possesses file read/write access, the ability to execute shell commands via sub-agents, and the ability to spawn new agent instances.
Sanitization: There is no evidence of sanitization or filtering of the transcript content before it is processed by the Sonnet/Opus judge models.
[COMMAND_EXECUTION]: The JavaScript workflows in references/workflows/ frequently use shell commands to manage the analysis process.
Evidence: Agents are instructed to use sed to read specific trial IDs, ls, sort, and awk for sharding file lists, and mkdir for directory creation.
Evidence: The Site phase in multiple workflows starts a local web server (python3 -m http.server) and executes chromium-browser --headless with the --no-sandbox flag to take verification screenshots.
[DYNAMIC_EXECUTION]: The skill generates and executes code dynamically at runtime.
Evidence: references/workflows/3-winners-contrast.js constructs a multi-line Python script that is piped directly into the interpreter (`python
<<'PY') to process trials.json` and generate task pairs.
Evidence: Sub-agents are tasked with modifying Python source code (deepswe/collect.py) and HTML (site/index.html) to integrate new analysis fields.

deepswe-analyze