name: skill-evals-run description: "Run the OpenCode skill-loading eval suite for this repo. Use when asked to run skill evals, skill-loading evals, or the skill-evals-run command." allowed-tools: Bash(./evals/skill-loading/opencode_skill_eval_runner.sh) Read context: fork

Skill Evals Run

Run the local skill-loading eval suite with the shell runner.

Prerequisites

• opencode is installed and on PATH.
• Provider config is available under ~/.config/opencode/ (used even with --isolate-config).
• If using --disable-models-fetch, ~/.cache/opencode/models.json exists and includes the target model.
• Auth/credentials are present (typically ~/.local/share/opencode/auth.json).
• Network access is available for model calls.

Command

Run:

evals/skill-loading/opencode_skill_eval_runner.sh \
  --repo "$PWD" \
  --dataset evals/skill-loading/opencode_skill_loading_eval_dataset.jsonl \
  --matrix evals/skill-loading/opencode_skill_eval_matrix.json \
  --disable-models-fetch \
  --isolate-config \
  --parallel 3

Arguments

If the user provides any of the following flags, append them to the command:

• --filter-id <regex>
• --filter-category <substring>
• --parallel <n>

If --parallel is omitted, keep the default of 3.

After the run

• Summarize PASS/FAIL counts and list failed case IDs.
• If failures exist (PASS/FAIL, not ERROR), reference evals/skill-loading/docs/skill-optimization-steering.md and suggest the next remediation step.
• If any cases are ERROR, do not suggest optimization. Instead, inspect evals/skill-loading/.tmp/opencode-eval-results/<run>/results.json and any traces to identify the crash, then re-run the evals once the error is resolved.

Notes

• Run from the repo root so relative paths resolve.
• --isolate-config also disables project config, so no extra flag is required to avoid loading repo config/plugins during evals.
• With --disable-models-fetch, the runner falls back to ~/.cache/opencode/models.json when present. Use --models-url file://... if you need a different cache file.
• Include the exact command used in the response.