The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from production call logs and Cekura simulation results to diagnose failures and suggest improvements. This creates a vulnerability to instructions embedded in processed transcripts.
Ingestion points: phases/collect.md fetches transcripts and call metadata; phases/reproduce.md uses this data to build a testing harness.
Boundary markers: The SKILL.md and phases/overfitting-gate.md define an 'Overfitting Gate' to detect and scrub verbatim transcript quotes or scenario-specific identifiers from proposed edits.
Capability inventory: The skill can execute shell commands (phases/optimization/apply.md), perform file writes using the Edit tool (phases/self-hosted/overview.md), and update agent configurations on external provider platforms.
Sanitization: Includes a secrets_policy for redacting sensitive strings and the Overfitting Gate for generalizing instruction wording.
[COMMAND_EXECUTION]: The skill executes shell commands defined in a project-specific manifest (.cekura/selfimprove.yaml) to manage the agent lifecycle.
Evidence: phases/optimization/apply.md executes a redeploy_command to restart self-hosted agents. phases/self-hosted/overview.md references the use of database clients like psql, mysql, and mongosh for updating agent configurations stored in databases.
Risk Mitigation: Commands are registered during the setup phase, and the manifest utilizes typed and escaped parameters to prevent injection.
[DYNAMIC_EXECUTION]: The skill facilitates the dynamic generation and application of edits to agent prompts, tool definitions, and orchestration code.
Evidence: phases/optimization/fix.md classifies failures and proposes specific Edit operations for source files or database updates, which are then applied in the apply.md phase.
[EXTERNAL_DOWNLOADS]: The skill communicates with external APIs and well-known services to fetch and synchronize agent configurations.
Evidence: agents/fetch_failures.py fetches result data from api.cekura.ai. phases/clone.md and phases/collect.md interact with established provider APIs including VAPI (api.vapi.ai), ElevenLabs (api.elevenlabs.io), and Twilio (api.twilio.com). These references are consistent with the skill's purpose and target well-known infrastructure services.

cekura-self-improving-agent