The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes r.content fetched from the released-db database, which contains externally sourced, untrusted release notes and changelogs.
Ingestion points: Data enters the agent context via the output of wrangler d1 execute as described in SKILL.md.
Boundary markers: The skill utilizes buildReleaseBlock for structural formatting and provides specific instructions to sub-agents to parse and discard certain tags, although it lacks comprehensive sanitization against malicious content within the release bodies.
Capability inventory: The skill environment provides access to database queries (wrangler), local CLI modifications (releases admin), and network API communication (curl).
Sanitization: Processing relies on XML tag parsing logic and boilerplate detection routines (isEmptyContent) defined in the local packages/ai/src/release-content.ts file.
[COMMAND_EXECUTION]: The skill instructions involve executing multiple shell commands to manage and update release data.
Evidence: Uses wrangler d1 execute to query a remote database and releases admin release update to modify local release records.
Evidence: Invokes bun scripts/generate-release-content.ts to delegate large-scale backfill tasks to a local script.
[DATA_EXFILTRATION]: The skill performs network operations to transmit release data and composition metadata to an API.
Evidence: Employs curl to send PATCH requests to $RELEASES_API_URL to update release composition information.
Evidence: These requests include an Authorization: Bearer header utilizing the $RELEASES_API_KEY environment variable.
[DYNAMIC_EXECUTION]: The skill implements a multi-agent workflow to parallelize data processing.
Evidence: Programmatically dispatches sub-agents using the Agent() function with dynamically constructed prompts and structured JSON output contracts.

generating-release-content