The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as user-provided error logs, stack traces, and code diffs which could contain malicious instructions.
Ingestion points: Symptoms and reproduction context are collected in Phase 0b; stack traces and log excerpts are captured in Phase 1a; git history, lock files, and config diffs are recorded in Phase 1b.
Boundary markers: The skill does not employ explicit delimiters or escaping for the ingested user data, though it uses structured markers for its own internal agent outputs.
Capability inventory: The skill utilizes Task to spawn sub-agents, Bash to run commands and aggregation scripts, and the Edit tool to modify source code.
Sanitization: No sanitization or filtering is applied to external data before it is interpolated into prompts for hypothesis generation or judging.
[DYNAMIC_EXECUTION]: The skill generates and executes "discriminating probes," which are shell commands or scripts created at runtime to test debugging hypotheses.
The coordinator or an 'evidence-gatherer' sub-agent executes these probes via Bash in Phase 4.
This represents dynamic code generation where the exact command is determined by the agent's logic during the debugging session.
Mitigation: A mandatory user confirmation gate is required for any probe that is irreversible or touches production data, reducing the risk of autonomous damage.
[DATA_EXFILTRATION]: The skill automatically collects system and project metadata that may contain sensitive information.
It captures git log output, environment variables (env | sort), and configuration file diffs during Phase 1.
This information is stored in local project files (e.g., evidence.md), which could lead to accidental exposure of credentials or intellectual property.
[EXTERNAL_DOWNLOADS]: The skill leverages an ensemble panel that sends hypothesis and evidence data to external AI providers (OpenAI and Google) for evaluation.
Specifically, it utilizes the mcp__pal__chat tool to call models gpt-5.4 and gemini-2.5-pro.
These are well-known technology services, and their use is integral to the skill's intended purpose of multi-provider judging.

deep-debug-ensemble-v1