⏩ thunderkit

Opinionated multi-model delegation for very large repos.

skill

tk-test

Use to prove the fleet configured by tk-router is actually reachable: pings every model in .thunderkit/config.json through its real harness CLI with a one-word probe and reports reachable/unreachable per model and per class before any real work starts.

Install: npx skills add thunderock/thunderkit -s tk-test -g

---

tk-test — does the configured fleet actually answer?

A smoke test for the model classes tk-router chose. Before committing a big run to a fleet, tk-test proves each configured model is actually reachable through its harness on this machine — not assumed from config. It sends the smallest possible tk-ask probe ("Reply with exactly one word: pong") to every model and reports what came back, with the resumable id.

Run it right after tk-router writes .thunderkit/config.json, on a fresh machine, or whenever a run mysteriously stalls — a stalled lane is usually an unreachable or unauthed model, and this finds that in seconds instead of minutes of silence.

What it does

scripts/tk-test.py reads .thunderkit/config.json, resolves each short name to a harness + provider id via the roster, and dispatches the probe through the real CLI for each:

Harness	Probe command	Reads
claude	claude -p "<probe>" --model <id> --output-format json	.result == pong, .session_id
codex	codex exec --json --skip-git-repo-check -m <id> "<probe>"	item.completed text, thread.started.thread_id
hermes	hermes chat -q "<probe>" --oneshot -Q --provider <p> -m <id> -t ""	last line == pong, session_id

Each model is reported reachable (answered "pong"), unreachable (CLI ran but wrong/failed answer — auth, quota, bad id), or not-installed (harness not on PATH). It also checks the cross-family invariant: how many distinct model *families* are reachable among the reviewers, against review_families_min — because if only one family answers, tk-review can't do a real cross-family review and tk-ship will block.

Run it

python3 skills/tk-test/scripts/tk-test.py                 # human table, exit 0 iff all reachable
python3 skills/tk-test/scripts/tk-test.py --json          # machine-readable
python3 skills/tk-test/scripts/tk-test.py --timeout 150   # per-probe timeout (default 120s)
python3 skills/tk-test/scripts/tk-test.py --config path/to/config.json

(When the pack is installed via npx skills, the script lives at ~/.agents/skills/tk-test/scripts/tk-test.py.)

Output — reachability report

Per-model status + timing + resumable id, then a per-class roll-up and the family check:

✓ opus48  claude   reachable   8.9  pong
✓ opus5   hermes   reachable  28.1  pong
✓ fable51 hermes   reachable  14.8  pong
✗ sol     codex    unreachable      rc=1 quota exceeded

  reviewer families reachable: 1 (anthropic); required ≥ 2  ← cross-family review NOT possible

Exit 0 only when every configured model answered; non-zero otherwise, so it drops straight into a Makefile target or CI preflight.

Discipline

• Never fakes a result. A model that doesn't answer is unreachable/not-installed, never a

silent pass. The probe asserts the literal word pong came back, not just that the CLI exited 0.

• Degrade honestly. If a class loses a model, tk-test says which class and whether the

cross-family invariant still holds — the same honesty rule the rest of the pack follows.

• Cheap and bounded. One tiny turn per model, each under a timeout, so a hung harness can't

stall the preflight.