flyingnobita-deliver-retro
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill provides explicit instructions to bypass potential platform-level tool invocation restrictions. If the system harness refuses to load a dependent skill because it is user-invoked, the agent is instructed to manually read the markdown source file of that skill as a workaround.
- [DATA_EXFILTRATION]: The skill accesses sensitive local files to gather evidence for retrospectives. This includes reading Claude Code session logs (
~/.claude/projects/), Codex session transcripts (~/.codex/sessions/), and delivery run state records that contain environment variable metadata. It also facilitates sending data to external project trackers. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data by ingesting raw session transcripts and worker reports which may contain malicious instructions.
- Ingestion points: Orchestrator transcripts, worker transcripts, and delivery reports located in the local filesystem.
- Boundary markers: No specific delimiters or safety instructions are defined to isolate analyzed transcript content from the agent's core instruction set.
- Capability inventory: Writing to external task trackers, creating or modifying local skill code, and committing changes to git repositories.
- Sanitization: The skill advises against printing secret values from environment variables but lacks sanitization or escaping for executable content within the analyzed transcripts.
- [DYNAMIC_EXECUTION]: The skill manages the modification of local source code and skill definitions. It guides the agent to implement proposed fixes in the
skills/ortests/directories and commit those changes to the repository based on user approval.
Audit Metadata