Troubleshooting

Start with the symptom, then return to the owning page. Keep setup problems separate from model-quality evidence.

Setup checks fail

Run:

gh auth status
stet --version
stet auth status
npx skills list
docker info
command -v uv

The quickstart explains which checks are required for the default Docker-backed path. If the binary is old, reinstall from the public installer and verify stet --version again:

curl -fsSL https://raw.githubusercontent.com/Stet-AI/stet-cli/main/install.sh | sh

The installer selects the latest stable CLI by default; verify the result with stet --version.

GitHub or Stet authentication fails

The public installer and skill do not require GitHub login. GitHub CLI auth is needed for private release overrides or PR-backed discovery. Commercial eval execution requires:

stet auth login
stet auth status

If a private override fails, confirm repository access before debugging Stet.

Docker or `uv` is unavailable

Start Docker Desktop or the Linux Docker service, then rerun docker info. If Docker is intentionally unavailable and the surface supports it, ask your agent to use the explicit worktree backend:

Use the Stet skill. Run the supported local check with
--harbor-backend worktree, then read status/report and inspect worktree
integrity artifacts only if diagnosis is needed.

Install uv with the public installer when command -v uv returns no path:

curl -LsSf https://astral.sh/uv/install.sh | sh

Worktree-backed evidence is inspect-only by default. It is not Docker-equivalent isolation and should support a decision only when the Trial Result explicitly says that the evidence is decision-grade; use Docker-backed replay for gateable decisions otherwise.

Claude Code authentication fails

Authenticate before a Stet-managed Claude run:

claude setup-token
mkdir -p ~/.config/stet
chmod 700 ~/.config/stet
printf '%s\n' '<printed token>' > ~/.config/stet/claude-oauth-token
chmod 600 ~/.config/stet/claude-oauth-token
unset CLAUDE_CODE_OAUTH_TOKEN

Do not commit the token or put it in repository .env files. If the skill is missing, reinstall it with the command in the quickstart.

A run is slow, stalled, or incomplete

Do not restart blindly. Ask the agent to inspect status first:

Use the Stet skill. This Stet rules run looks incomplete. Check status first
and identify whether the launcher is still active. Use rules repair only if the
run has exited or stalled. Do not restart unless I explicitly approve
discarding evidence.

Then use Read a Trial Result to distinguish an infrastructure state from a model-quality result.

Replay validity is missing

Errors such as no_gold_pass_commands, all_commands_ignored_gold_failure_mode_unset, or replay-validity failures mean the selected slice is not decision-grade yet. Fix the test or harness setup, choose a better task slice, or stop with an inspect-only diagnosis before relaunching.

Grader coverage is missing

Tests may still have run, but missing declared grader coverage limits the claim. Use the repair or regrade command emitted by stet eval status or stet eval report, then reread the complete Trial Result.

The result says `inspect`

inspect is useful evidence for diagnosis or iteration, not a generic failure. It can reflect a small sample, stale evidence, missing graders, replay validity, infrastructure failure, or mixed results. Ask the agent to name the concrete blocker and propose one bounded next action; do not infer promotion from prose.

​Setup checks fail

​GitHub or Stet authentication fails

​Docker or uv is unavailable

​Claude Code authentication fails

​A run is slow, stalled, or incomplete

​Replay validity is missing

​Grader coverage is missing

​The result says inspect