Tests

Stdlib unittest, no dependencies. Run the full (fast, offline) suite:

python -m unittest discover -s tests -v

Unit suite (runs on CI)

File	Covers
test_meter.py	token sums; dedupe by message.id (keep last); subagents inline + in a separate subagents/*.jsonl; cross-file dedupe; loop detection (is_looping, recent_tool_signatures)
test_pricing.py	per-model price resolution (incl. date-suffixed ids and unknown=$0); per-turn cost (opus is 5x haiku; cache rates); mixed-model session cost; subagent cost; the $0.10 and $1 boundaries
test_hook.py	budget parsing (dollars + tokens); resolution precedence (env vs arm file, with unit); the pure evaluate() verdict (halt/warn/allow in $ and tok, loop); per-event render() (Stop = continue:false; PreToolUse = permission deny)
test_cli.py	arm (tokens + dollars) sets baseline; reset re-baselines and keeps the cap; disarm removes the file; error paths. Real meter, temp projects dir, patched arm_path (never touches real ~/.claude)
test_dispatch.py	/budget sentinel parsing → CLI args + decision:block output (incl. default-to-status, fallback reason); /goal auto-arm of $50 when unarmed, and not clobbering an explicit budget. cli.main mocked, so offline
test_version.py	meter.VERSION equals the top CHANGELOG heading and .claude-plugin/plugin.json version

All of these are pure/offline and isolated via temp dirs and mocks. CI (.github/workflows/tests.yml) runs them on Python 3.9 / 3.11 / 3.12.

Live test (NOT on CI)

test_live_goal.py makes real claude -p calls (real cost) to prove what unit tests can't: that burnstop's Stop-hook continue:false actually halts a session and overrides a /goal-style block. It is @skipUnless gated:

BURNSTOP_LIVE=1 python -m unittest tests.test_live_goal -v

Without BURNSTOP_LIVE=1 (and a claude CLI on PATH) it skips, so GitHub never runs it.

What it does, and why it's safe. Hooks loaded via --settings or a project .claude/settings.json do not fire in claude -p (only user-scope settings are trusted headlessly — an empirical finding from building this). So the test temporarily injects hooks into ~/.claude/settings.json and restores it in tearDown. Every injected hook is a no-op unless BURNSTOP_TEST=1, which only the child claude -p run sets — so it can never trap your own session. The fixture tests/fixtures/_gated_block.py simulates an unmet /goal by returning decision: block; always_block.py is the un-gated variant kept for reference.

The assertion. It runs two headless sessions under the goal-block: one where burnstop can't trip ($999 cap) — the block forces the full --max-turns leash — and one with a $0.30 cap, which should stop before the leash. test_turns < control_turns proves continue:false overrode the block. (The control's > 1 guard also surfaces the case where headless hooks don't fire at all.)

Note: this is the live verification flagged in spikes.md §2b. Running it requires temporarily editing user settings, which an agent's safety classifier will (correctly) block — so run it yourself.