The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the local environment and external APIs through shell commands using git and the pup CLI. It demonstrates secure practices by instructing the use of quoted heredocs (<<'EOF') when writing JSON payloads to temporary files, which prevents the shell from performing unintended variable expansion on the data.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted data in the form of test error messages (@error.message) and stack traces (@error.stack) retrieved from Datadog CI Visibility logs. This data is used to determine root causes and recommend fixes. There is a potential risk that an attacker who can influence test failure outputs could embed malicious instructions or deceptive information meant to influence the agent's logic.
Ingestion points: Data enters the context via search_datadog_test_events tool calls which fetch @error.message and @error.stack (SKILL.md, STEP 2).
Boundary markers: None explicitly defined to separate log content from agent instructions during analysis.
Capability inventory: The skill has access to shell execution (bash) for running git and pup commands, and the ability to update test states in Datadog via update_datadog_flaky_test_states.
Sanitization: While secure heredocs are used for file creation, there is no specific sanitization or escaping described for interpolating the ingested error messages into the prompt analysis for fix generation.

triage-flaky-test