pawahara-harness

name	pawahara-harness
description	Use Pawahara Harness for difficult coding, CTF, competitive programming, debugging, optimization, or long-running tasks that benefit from diverse beam-search workers, scoring, resume, Karasu continuation, and the monitor UI.
metadata	{"short-description":"Run Pawahara Harness search"}

Pawahara Harness

Use this skill when the user asks to use Pawahara Harness, wants more diverse reasoning, asks for CTF or competitive-programming solving, needs multi-worker exploration, or the current agent is likely to stop too early.

Run Search

Prefer search for real work:

goal="$(cat <<'EOF'
<user task>
EOF
)"
pawahara-harness search \
  --goal "$goal" \
  --cwd "$(pwd)" \
  --beam-width 4 \
  --branch-factor 4 \
  --max-depth 2 \
  --max-workers 4

For a quick smoke test or a single direct agent, use:

pawahara-harness run --goal "$goal" --cwd "$(pwd)"

Resume

Continue an existing run instead of starting over:

pawahara-harness search \
  --resume-run <run_id_or_path> \
  --resume-message "$goal" \
  --cwd "$(pwd)"

Steering

Use Helm for forced context:

pawahara-harness search \
  --goal "$goal" \
  --cwd "$(pwd)" \
  --helm "worker:<extra worker instruction>" \
  --helm "manager:<extra manager instruction>"

Operating Notes

• Check command -v pawahara-harness before relying on the skill.
• Keep the user's original objective intact in --goal; do not summarize away constraints.
• Use --runs-dir when you need an isolated test run.
• Inspect .pawahara/runs/<run_id>/result.json and worker responses before reporting.
• Karasu is enabled by default; leave it on when the user wants persistence.
• Do not use the harness for tiny one-step edits unless the user explicitly asks.