The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes external implementation plans and specification files that serve as the primary source of instructions for the controller and its subagents.
Ingestion points: Plan files located in the repository (e.g., docs/plans/*.md) and specification files defined in the ledger.
Boundary markers: While the protocol uses dedicated file paths for briefs and reports, the subagent prompts in implementer-prompt.md and reviewer-prompt.md lack explicit instructions to ignore potentially malicious embedded directives within the plan content.
Capability inventory: The controller executes shell commands (git, awk, mkdir, rm), writes to the file system (ledger and workspace files), and dispatches subagents that write and execute code and tests.
Sanitization: No sanitization or filtering was observed for the plan content before it is extracted and passed to the implementer subagents.
[COMMAND_EXECUTION]: The skill relies on local shell command execution for its operational protocol as defined in references/ledger-and-workspace.md.
Workspace management: Commands including mkdir -p and printf are used to initialize workspace directories and configuration files.
Git operations: The skill uses git rev-parse, git merge-base, git log, and git diff to manage versions and generate review packages.
File manipulation: An awk script is used to extract task sections from the plan file, and rm -rf is used for workspace cleanup.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes shell commands at runtime based on the state of the plan.
Task extraction is performed using an awk script that processes variables such as the task number.
Review diffs are generated based on dynamic commit ranges computed during the execution loop.

delegated-execution