The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes findings files from .context/findings/*.json which contain free-text fields like title, repro, and suggestedFix. These fields are used to drive automated code modifications and are interpolated into prompts for external agents (Conductor) and issue trackers (Linear).
Ingestion points: Findings data is loaded from the .context/findings/ directory in SKILL.md Step 1.
Boundary markers: There are no explicit delimiters or instructions provided to the LLM to ignore potentially malicious content embedded within the findings data fields.
Capability inventory: The skill has high-privilege access, including the ability to write to the local file system, execute git commands, run a user-defined testCommand, and launch external applications via deep links.
Sanitization: The skill enforces a rule to redact credentials from the evidence field before transmission, but it does not perform sanitization of the instructional text fields.
[DYNAMIC_EXECUTION]: The skill dynamically assembles an agent prompt, encodes it, and embeds it into a conductor:// deep link. This link is then executed using the system open command to initiate activity in a separate workspace (Step 8b).
[COMMAND_EXECUTION]: The skill performs several command-line operations, including git branch renaming, the system open command for deep links, and the execution of a configurable testCommand (e.g., npm test) to verify fixes.
[DATA_EXFILTRATION]: The skill transmits QA finding metadata and repro steps to the Linear issue tracker via an MCP tool and to the Conductor platform via deep-link parameters.

candid-chrome-qa-fix