The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted review data through the source_batches input in SKILL.md.\n
Ingestion points: The review_fold_input parameter in SKILL.md accepts findings from external owners.\n
Boundary markers: The skill instructions in SKILL.md (e.g., under 'Counterexample admission') and agents/openai.yaml explicitly command the agent to establish obligations and validity independently of the reported claim's prose.\n
Capability inventory: The skill can read repository state and write to a local event log (.ledger/review-fold/counterexamples/events.jsonl) using the ledger CLI tool.\n
Sanitization: Data persisted to the ledger is governed by a strict JSON schema in definitions/ledger/counterexample-corpus.json, which includes regex validation, enums, and byte limits.\n- [DYNAMIC_EXECUTION]: The offline test script tests/test-counterexample-admission.mjs uses node:vm's runInNewContext to execute synthetic JavaScript code snippets. This is limited to testing synthetic fixtures and employs an isolated context (Object.create(null)) with a execution timeout.\n- [COMMAND_EXECUTION]: The skill requires the agent to execute ledger project and ledger transact commands to manage the counterexample corpus. These operations are scoped to a local repository ledger and are structured via a platform-provided definition.

review-fold