The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core logic is driven by external artifacts (design.md and plan.json) and conversation context. If these sources contain malicious instructions, they could influence the agent's behavior during implementation.
Ingestion points: The skill reads docs/specs/*/design.md and docs/specs/*/plan.json in the 'Prerequisites' and 'Step 1' sections.
Boundary markers: The skill includes a validation step where plan.json is checked against a schema defined in spec-plan.
Capability inventory: The agent can write files, execute arbitrary tests/linters, perform git operations via code-commit, and dispatch tasks to code-subagents.
Sanitization: No explicit prompt sanitization or escaping mechanisms for the interpolated plan content are described.
[COMMAND_EXECUTION]: The skill performs various system-level operations including running type checkers, linters, and language-specific test suites as part of the TDD workflow. It also manages git branches and worktrees.
The execution of tests involves running potentially untrusted code generated by the agent based on the plan.
The 'Subagent Mode' dispatches work to code-subagents, extending the scope of execution.

spec-implement