Agent workflow helpers
This directory contains deterministic checks used by the minimal agent harness. The
helpers use the Python standard library, except for pinned Ruff invoked through uvx.
They do not compile the product, load or run kernel modules, or evaluate model quality.
check.py provides:
snapshotfor a non-empty exact commit range, ancestry,HEAD, and tracked-cleanliness evidence.--require-no-untrackedalso rejects untracked files for an isolated deliverable worktree. Only initial validation of a newly prepared result worktree uses--allow-empty;instructionsfor the requiredscst-skill namespace, matching skill frontmatter and directory names, directly referenced Markdown files, requiredagents/openai.yamlinvocation policy, and standalone agent-role TOML consistency.
lint.py selects pinned Ruff for existing changed Python files. For changes to agent
instructions or configuration it also runs check.py instructions and the focused unit
tests in this directory. scst-lint-patch invokes it automatically for the committed
range. Run these examples from the repository root:
python3 .agents/workflow/lint.py \
--repository "$PWD" --base <full-base-sha> --head <full-head-sha>
Individual checks remain available for diagnosis:
python3 .agents/workflow/check.py instructions --repository "$PWD"
PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover \
-s .agents/workflow -p 'test_*.py'
These helpers validate observable repository facts. Base and result branches are created with ordinary Git commands because branch and worktree bootstrap must also work when the selected base snapshot does not contain these helper files. The helpers do not validate the semantic meaning of instructions, choose task scope, decide a design, grade a patch, or replace independent review. They provide no compilation or runtime evidence.