beton

BETON: Verifier Loop

A PostToolUse hook on Write, Edit, MultiEdit. Short-circuit chain:

1. Lint (≤4s)
2. Typecheck (≤5s, only if lint passed)
3. Affected tests (≤8s remaining, only if both passed)

Affected tests are discovered by naming convention: test_<basename>.<ext>, <basename>_test.<ext>, <basename>.test.<ext>, <basename>.spec.<ext>.

If everything passes, the hook is silent. If the file extension or language tool is not covered, the hook exits 0 and never blocks on something it cannot check.

Reading a block payload

BETON VERIFIER ERRORS:
LINT:
src/auth.py:3:1: F401 [*] `unused_module` imported but unused

BETON VERIFIER ERRORS:
TEST:
FAILED tests/test_auth.py::test_token_expiry - AssertionError: expected 401, got 200

Steps:

1. Read the error verbatim. Do not guess at the cause.
2. Fix the specific issue named.
3. Re-edit. The hook re-runs.

Do not bypass the verifier. Bypasses include: renaming files, adding # noqa / # type: ignore suppressions, deleting the failing test, or moving code to an unchecked path. Fix the root cause. The point is the loop.

Tuning

If the verifier times out repeatedly, suggest the user raise one of:

• BETON_VERIFY_TIMEOUT (total chain, default 8)
• BETON_TC_TIMEOUT (typecheck, default 5)
• BETON_TEST_TIMEOUT (tests, default 8)

Subagents

• repo-map: token-budgeted symbol index. Use before editing unfamiliar code.
• localize: ranked candidates from a stack trace or issue. Use at the start of a bug-fix task.

Native context primitives

Claude Code provides these; do not reinvent them.

• compact_20260112: /compact for server-side summarisation. Useful instructions: "Preserve edited paths, test status, failed assertions verbatim, active localization candidates. Drop search results, intermediate diffs."
• clear_tool_uses_20250919: auto-clears old tool results after ~5 tool-heavy turns.
• Memory tool / auto-memory at ~/.claude/projects/<project>/memory/: cross-session state.

Thinking budgets

Task	Opus 4.7 budget_tokens	Sonnet 4.6 effort	Haiku 4.5
Bug fix (beton-swebench)	~80K	xhigh	extended thinking via budget_tokens
Feature / refactor	~40K	xhigh	n/a (no adaptive thinking)
Test writing	~16K	medium	n/a
Explanation	~8K	medium	n/a

Sonnet 4.6 defaults to high effort (adaptive thinking on complex prompts). Haiku 4.5 has no adaptive thinking; configure budget_tokens explicitly at the API level. Changing effort/budget mid-loop invalidates the message cache.

Companion skill

beton-swebench for bug fixes, GitHub issues, failing tests.