The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary function involves executing shell commands provided by the user for running experiments and measuring metrics. It also makes extensive use of git commands (such as git commit, git reset --hard, and git checkout) to manage the state of the codebase during experimentation.
[INDIRECT_PROMPT_INJECTION]: During its 'THINK' phase, the agent reads data from project source files and its own experimental logs (.lab/log.md, .lab/results.tsv). Because these files are ingested without explicit boundary markers or sanitization, there is a risk that instructions embedded in the project data could influence the agent's autonomous decisions.
[PROMPT_INJECTION]: The instructions utilize strong directives to enforce an autonomous persona, telling the agent it has 'complete freedom' and should 'Consult the user ONLY when' specific dead ends are reached. These patterns attempt to reduce user oversight and override the agent's typical interactive behavior.

autoresearch