coordinator-mode
Warn
Audited by Gen Agent Trust Hub on Nov 11, 2026
Risk Level: MEDIUMPERSISTENCEPROMPT_INJECTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PERSISTENCE]: The skill provides detailed instructions for modifying the user's AI harness configuration (e.g.,
~/.claude/settings.json,~/.codex/hooks.json) to register shell command hooks. These hooks execute a local Python script (coordinator_mode.py) on session events likeSessionStartandPreCompact, allowing the skill to automatically re-inject its state and instructions into the AI context. - [PROMPT_INJECTION]: The skill includes instructions that attempt to override default agent behaviors and safety constraints, such as 'Nothing in this mode is forbidden to you' and 'Reshape internal architecture... without asking for routine approval.' It employs a 'reread' pattern to ensure the agent maintains this specific role and set of instructions even after context compaction or session resets.
- [COMMAND_EXECUTION]: The skill's functionality depends on the execution of a local Python script (
scripts/coordinator_mode.py). The instructions direct the agent to invoke this script to initialize the session state, manage journals, and provide the persistent hook logic. - [INDIRECT_PROMPT_INJECTION]: The skill utilizes a 'journal' (a Markdown file) to store project state and evidence from 'executors' (subagents). The coordinator is instructed to reread this journal in full after context resets, which creates a vulnerability surface for indirect prompt injection from worker outputs.
- Ingestion points: The journal file (
references/journal.md) serves as the primary location where external summaries and results are recorded and later reread. - Boundary markers: Absent. There are no clear delimiters or instructions to prevent the agent from following directions that might be embedded within the journal content.
- Capability inventory: The coordinator role has extensive capabilities, including executing shell commands, refactoring codebases, and managing subagents.
- Sanitization: Absent. The provided scripts and instructions do not include any mechanisms for sanitizing or escaping content before it is written to or read from the journal.
Audit Metadata