The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources to construct chaos experiment manifests that can execute commands on infrastructure.
Ingestion points: Reads project/org data, infrastructure details, and workload names via harness_list and harness_get MCP tools. It also accepts free-text descriptions and custom tags from the user.
Boundary markers: Includes 'MANDATORY — Tool Result Verification' instructions across all files to ensure the agent verifies tool outputs rather than hallucinating or blindly trusting missing data.
Capability inventory: The skill uses harness_create and harness_update to save manifests that define system disruptions (faults), health checks (probes), and custom shell scripts/container actions.
Sanitization: Implements strict naming conventions (lowercase a-z, 0-9, and dashes) and YAML validation rules (no duplicate keys, no orphan references, no blank lines) before serialization.
[COMMAND_EXECUTION]: The skill facilitates the creation of ChaosExperiment manifests that include operational steps like 'Custom Script' actions and 'Command Probes'.
Evidence: Instructions in references/components.md guide the agent to configure shell commands and container commands within the actionRef and probeRef sections of the experiment YAML.
[CREDENTIALS_UNSAFE]: The skill provides procedures for integrating secrets into chaos experiments for fault authentication.
Evidence: references/components.md (Step 7F.2.S) describes fetching secrets via harness_list and constructing reference strings like secrets.getValue("account.<identifier>"). This is a platform-native safe practice for secret management, as the agent only handles the identifier rather than the raw secret value.
[SAFE]: The skill correctly uses MCP-provided schemas via harness_describe to ensure the generated requests match the expected API structure, reducing the risk of accidental misconfiguration.

chaos-experiment