The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides shell command templates for executing external CLI tools such as codex and claude. It explicitly instructs the agent to use the --permission-mode bypassPermissions flag when running the Claude CLI. This configuration disables standard security guardrails and user-confirmation prompts for tool execution, allowing a subagent to perform potentially destructive or unauthorized actions without oversight.
[DATA_EXFILTRATION]: The debug workflow involves capturing raw interaction traces from CLI stdout (in NDJSON or JSONL format) and saving them to a local repository directory (.heerogeneous-tracing/). These traces contain the full content of agent communications, which may inadvertently include sensitive source code, environment details, or credentials processed during the debugging session.
[PROMPT_INJECTION]: The skill defines a complex pipeline for ingesting untrusted data from external CLI agents (Codex and Claude Code) and mapping it into the application state. This creates a surface for indirect prompt injection, where malicious or unexpected output from the external CLI could influence the behavior of the main agent.
Ingestion points: The HeterogeneousAgentCtr module in the Electron main process captures raw stdout/JSONL streams from spawned CLI instances and broadcasts them via IPC (heteroAgentRawLine).
Boundary markers: The skill documents the use of assistant message.id changes and turn-boundary signals (turn.started, turn.completed) to delimit steps, which helps regulate turn transitions but does not sanitize content.
Capability inventory: The pipeline supports spawning subprocesses, persisting assistant/tool messages into the chat store (persistToolBatch), and forwarding events to the UI for hydration.
Sanitization: The skill employs specific adapters (e.g., packages/heterogeneous-agents/src/adapters/claudeCode.ts) to map raw provider events into a structured HeterogeneousAgentEvent schema, providing a layer of validation against malformed event structures.

heterogeneous-agent