The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several local commands to gather diagnostic information, including git, node, npm, ha (the Harness Anything CLI), and gh (GitHub CLI).
[COMMAND_EXECUTION]: It instructs the agent to execute reproduction commands derived from the user's "finding" to capture logs and exit statuses. While this presents a surface for executing unsafe commands, the skill explicitly mandates that these commands must be "read-only" and "safe in the current checkout."
[DATA_EXFILTRATION]: The skill automates the creation of public GitHub issues, which involves sending local system data (versions, logs, environment details) to a remote repository. This is mitigated by Step 4, which requires a scan and redaction of secrets (like ghp_ tokens, sk- keys, and Authorization headers) and absolute local paths. The destination repository is a vendor-owned resource.
[INDIRECT_PROMPT_INJECTION]: The skill processes external "findings" (bugs, feedback) provided by users to generate the issue body. This creates an indirect prompt injection surface where a malicious finding could attempt to influence the agent's behavior or cause the execution of unintended commands during the reproduction phase.
Ingestion points: User-provided findings or bug reports (SKILL.md).
Boundary markers: The skill uses a fixed markdown template (references/issue-body-template.md) and a checklist to constrain the output format.
Capability inventory: Shell execution (bash), file reading/writing (temporary files), and network activity via gh issue create.
Sanitization: Explicit redaction step using grep to identify and remove secrets and local directory paths before publishing.

harness-issue