v0.7.0: Proof first

Proof first. The trustbut-index pilot labels found that trustbut's facts were mostly true but rarely worth a reviewer's time; the useful signal was whether anything proves the change.

• New finding change-not-proven (warn), listed first in the receipt: code changed, and no test changed, or test files changed but no new or changed test that ran was found in them (only a test helper or type test changed, or a test file changed without a test-definition line), or none of the new or changed tests fails without the change. Comment-only and blank-line changes don't count as code; Python docstring edits do (the lexer reads one line at a time, and docstrings are strings). It is not emitted while a new or changed test fails at head (test-failing already says so; whether it proves the change is undecided until it passes), nor when the receipt already says there is no evidence (no-runner, base-run-failed, proofs-unmatched, tests-outside-root, suite-failing, no-test-results). It replaces no-tests-changed and the per-test test-passes-without-change, which trustbut no longer emits; tests that pass without the change are still listed in proofs.
• A receipt that warned only because a new or changed test passed without the change now PASSes when another new or changed test proves the change. This matters for --strict and the Action's strict input, which fail on WARN.
• The Markdown receipt opens with a proof line: Proof: N new/changed tests prove the change ✅ or Proof: nothing proves this change ⚠. In the hook's one-line summary, a FAIL lists change not proven among its parts; a WARN shows the finding title Nothing proves the change.
• Receipt spec 1.2: adds the core id change-not-proven. trustbut writes specVersion: "1.2"; trustbut validate rejects change-not-proven in a receipt that declares 1.1 or older.
• Moved tests: tests-removed and assertions-removed no longer count a removed line that another changed test file adds with the same text. A move cancels at most as many removals as were added elsewhere.
• Merged tests: tests-removed is not reported for a file that also adds a test definition and whose assertions did not drop on net (tests merged into one). Plain deletions of tests, and Rust inline tests (which have no assertion count), are still reported.
• pytest: base runs import workspace members installed in editable mode from the checkout (for example packages/lib/src) from the base worktree, so a test no longer passes at base on head's code.

v0.6.0: Less noise

Less noise from the trustbut-index pilot. Fixes for findings that were true only by accident, or not at all.

• test-regressed is now confirmed by a rerun. A test that fails at head and passed at base is rerun with the base run's files at head, then at base; it is reported only when it fails again at head and passes again at base. Otherwise trustbut adds the info finding trustbut/regression-not-reproduced with the rerun outcomes. This drops failures that happen only in the full suite or only once. "(passed at base)" in Markdown and agent feedback follows the confirmed findings.
• Type-test files are tests: test-d/, test-dts/, __typetests__/, *.test-d.ts, *.spec-d.ts, *.d-test.ts. Adding one no longer gives no-tests-changed, and its @ts-expect-error lines are no longer escape-added. They are never passed to a test runner.
• escape-added no longer reports a JS directive mentioned in comment prose (@ts-…, eslint-disable, coverage ignores count only right after a comment marker), any in a generic constraint (T extends (...args: any[]) => R), or anything in files .gitattributes marks linguist-generated. any in a type signature and suppressions with a rule code or a written reason are still reported.
• Tests: the runner parser tests resolve their fixtures with fileURLToPath, so npm test passes in a checkout whose path contains a space.

v0.5.0: Regressed tests

Regressed tests. A red suite no longer looks the same whether the change broke a test or the test was already failing.

• New finding test-regressed (fail): a test that passes at base and fails at head. The base run now also gets the files of tests failing at head (pytest module names map to their .py file), and runs even when no tests changed if something fails at head. Markdown and agent feedback mark these tests "(passed at base)", and the hook summary counts them.
• Receipt spec 1.1: adds the core id test-regressed. trustbut writes specVersion: "1.1". trustbut validate accepts 1.1, treats only minors newer than 1.1 leniently, and rejects test-regressed in a receipt that declares 1.0.
• New: verify --output-tail adds the end of each failed run's output (last 50 lines, at most 4,000 characters) under x-trustbut in JSON receipts. Off by default; receipts stay output-free without it.

v0.4.0: Agent receipts open spec v1

Agent receipts, spec v1. trustbut's receipt is now an open, versioned format that any tool can produce or read: spec/agent-receipt-v1.md, with a JSON Schema and examples.

• Breaking: verify --json and --receipt-json now write an agent receipt v1. version: 1 is replaced by specVersion: "1.0" and producer. base/head/headRef move under subject. The top-level scope moves to tests.scope, and tests is absent when tests didn't run. proofs[].id and proofs[].weak are dropped. trustbut-only detail (runner, languages, root, the after/before runs) moves under x-trustbut, and the raw runner output is no longer written to JSON. verdict stays at the top level, so the GitHub Action is unaffected.
• New: tests holds test counts (passed, failed, errored, skipped) taken from the head run.
• New: trustbut validate [file] checks any receipt against the spec (stdin for -; exit 0 valid, 1 invalid, 2 unreadable). --json prints the result as JSON, for unreadable input too. A receipt that declares a newer 1.x minor is accepted: fields and core ids unknown to 1.0 are warnings. A malformed finding id (Not_Valid: neither shaped like a core id nor a valid namespaced id) is an error whatever the version.
• Fixed: a new test was wrongly reported as passing without the change when the change added source under a test folder inside a source tree (src/spec/): that source was copied into the base run. It no longer is; test helpers elsewhere (spec/helpers/) and code fixtures under fixtures/ or testdata/ are still copied.
• Fixed: test helpers in a workspace package named like a source folder (packages/app/spec/, apps/app/test/) weren't copied into the base run, so new tests that use them couldn't prove the change there. A folder directly under packages/ or apps/ now counts as a package, not a source tree; source under app/models/spec/ or src/app/test/ stays out of the base run.
• Fixed: changed Go test-support packages (internal/test/, internal/testutil/, testhelpers/, testing/) are copied into the base run when, at head, only tests and other test-support packages import them and their own folder has no changed tests. A package that production code imports, or that is itself under test, still stays out.
• The npm package now ships spec/.

trustbut v0.3.1

First release published from CI through npm trusted publishing, with a provenance attestation. No changes to the CLI.

• Tests: the CLI tests start trustbut with node instead of npx, so npm's own config warnings on stderr no longer break the release workflow.

Install: npm i -D trustbut · npx trustbut verify

npm: https://www.npmjs.com/package/trustbut

trustbut v0.3.0

Agent receipts for AI-written code. trustbut reruns your tests in a clean checkout, proves new tests fail before and pass after the change, and flags test tampering and type/lint escapes, then prints a PASS / WARN / FAIL receipt. Local, deterministic, no LLM calls.

npx trustbut verify

First npm release. It includes everything built so far:

• v0.1, core engine: trustbut verify reruns your tests in clean temporary worktrees, proves new and changed tests fail before and pass after the change, checks test integrity and type/lint escapes, ranks risks and prints a PASS/WARN/FAIL receipt (Vitest, pytest).
• v0.2, every language: a runner registry and .trustbut.json config (package root, custom test command, setup command, env, extra args), with runners for Jest, node:test, Mocha, Go, Rust, Maven/Gradle (Java/Kotlin), .NET, RSpec, PHPUnit, Dart/Flutter, Swift and any command that writes JUnit XML, language-aware tamper and escape checks, and --working for uncommitted changes.
• v0.3, everywhere agents finish work:
  • trustbut hook <claude|codex|gemini|opencode>: a turn-end hook that verifies the agent's work and sends a FAIL back to the agent (up to hook.maxRetries times in a row), with the runner's failure messages and the test output tail in the feedback; trustbut's own errors and timeouts never block.
  • trustbut init <claude|codex|gemini|opencode> [--user] [--dry-run]: installs the hook into the agent's config without touching other hooks.
  • Claude Code plugin and marketplace (/plugin marketplace add isAlmogK/trustbut).
  • Reusable GitHub Action (isAlmogK/trustbut@v0) with a sticky PR comment and a job summary.
  • --receipt-json <file> / --receipt-md <file> to write the receipt to files.
  • npm release workflow with trusted publishing and provenance.

Install: npm i -D trustbut · Claude Code: /plugin marketplace add isAlmogK/trustbut · GitHub Action: uses: isAlmogK/trustbut@v0

npm: https://www.npmjs.com/package/trustbut