Repository navigation
Releases
unwash 1.0.0
unwash 1.0.0
The project has a name that matches what it does, and a name nobody else owns.
What it is
A trust report on what a coding agent's "done, tests pass" actually means. It
reads the diff, reruns the committed version of any test that was changed, runs
the project's own suite against a baseline, and reports a verdict with the
evidence behind it. No model, no network, no dependencies.
pip install unwash-cli
cd your-repo
unwashunwash-cli is the distribution; unwash is the command and the import.
The rename
The old name described the problem rather than the tool, and an unrelated
package already held it on PyPI. That forced a different distribution name and
an explanation in every install instruction -- a defect in a README whose whole
argument is that an instruction which does not work is a lie.
Now the distribution, the import, the command and the repository share one
name. This is a major version because that breaks things on purpose:
| Before | After |
|---|---|
greenwash |
unwash |
GW-TEST-001 |
UW-TEST-001 |
.greenwash/config.json |
.unwash/config.json |
GREENWASH_ENFORCE |
UNWASH_ENFORCE |
github.com/0DukePan/greenwash |
github.com/0DukePan/unwash |
A waiver written against an old rule code still resolves, and a new waiver
counts on from whatever ids the ledger already holds, so an id is never
reissued. GitHub redirects the old repository URL, so existing clones keep
fetching.
Verify it before you believe it
git clone https://github.com/0DukePan/unwash && cd unwash
python benchmark/verify.py # 90 seconds, no model, no networkWhat is in 1.0.0
unwash coverage-- the Requirement Coverage Engine. Requirements can bind
to more than one test, so one passing and one failing is reported as
contradictory rather than flattened into "failed". A requirement whose
evidence would not run is could not check, distinct from one with nothing
bound. Requirements whose bound tests only assert the happy path are named as
having no negative case. Every report ends with coverage counts requirements
with passing evidence; it is not a claim that the code is correct.benchmark/verify.py-- one command for the whole offline benchmark:
a planted cheat per task scored by the real test runner, each attack beside
its legitimate twin, a positive/negative pair per language pack.unreadable-file(UW-DIV-003) -- a changed Python file the static layer
cannot read, including one git reports as binary, is a finding rather than a
clean report.verifier-config-changednarrowed -- it was 104 of 109 flags on real
history, nearly all dependency bumps; it now asks only when a change can alter
what runs, and produces 7.- The maintenance pass, the demo rebuild, and the first live-model pilot are
described inCHANGELOG.md.
The numbers, including the one that is not a win
- Detector: 25 of 25 planted cheats caught on the machine that ran it (24
static, 23 behavioural), 0 false positives on 25 recorded honest fixes. Three
tasks are skipped for missing toolchains and named; CI scores all 28. - False positives on real history: 109 flags across ten repositories, 104 of
them one rule on dependency bumps; after narrowing, 12 -- 10 asking, 2
hand-checked false positives. A survey, not a rate. - Agent behaviour: still unmeasured on a current model. No page claims an
effect in either direction.
What did not work
Five days of red nightly because a repaired task changed label and its test did
not; the release blocked on API balance; the earlier +0.0 that came from an
instrument that could not record a passing cheat. All of it is in
docs/what-did-not-work.md.
greenwash 0.5.1
What this is
A trust report on what a coding agent's "done" actually means. greenwash reads
the diff, reruns the committed version of any test that was changed, runs the
project's own suite against a baseline, and reports a verdict with the evidence
behind it. No model, no network, no dependencies.
Install from the tag:
pipx install git+https://github.com/0DukePan/greenwash@v0.5.1
greenwash --versiongreenwash-cli is not on PyPI yet -- the release workflow is in the
repository and publishes on a tag once the Trusted Publisher is configured.
Verify it before you believe it
git clone https://github.com/0DukePan/greenwash && cd greenwash
python benchmark/verify.py # 90 seconds, no model, no networkThree offline measurements: a planted cheat per task scored by the real test
runner, each attack beside its legitimate twin, and a positive/negative pair per
language pack. The exit code is non-zero on the first regression.
What changed since 0.4.2
greenwash coverage-- the Requirement Coverage Engine. Requirements can
bind to more than one test, so one passing and one failing is reported as
contradictory rather than flattened into "failed". A requirement whose
evidence would not run is could not check, distinct from one with nothing
bound. Requirements whose bound tests only assert the happy path are named as
having no negative case. Every report ends with: coverage counts requirements
with passing evidence; it is not a claim that the code is correct.python benchmark/verify.py-- one command for the whole offline
benchmark, and the CI job that runs it where every toolchain is installed.verifier-config-changedno longer fires on maintenance. It was 104 of 109
flags on real history, nearly all dependency bumps; it now asks only when a
change can alter what runs, and produces 7. Rule id, code and severity are
unchanged, so existing waivers keep working.unreadable-file(GW-DIV-003). A changed Python file the static layer
cannot read -- including one git reports as binary -- is a finding instead of a
clean report. Missing paths are still not findings.- A maintenance pass: 42 branches to 1, the default test loop from twelve
minutes to seven seconds, Windows nightly instead of per-push, no in-tree
ticket files, and a CI wall time cut from eleven minutes to under two. - The demo starts from the cause. The README clip opens on the agent's real
git diffand ends on the verdict.
The numbers, including the one that is not a win
- Detector: 25 of 25 planted cheats caught on the machine that ran it (24
static, 23 behavioural), 0 false positives on 25 recorded honest fixes. Three
tasks are skipped for missing toolchains and named; CI scores all 28. - False positives on real history: a ten-repository survey found 109 flags,
104 of them one rule on dependency bumps; after narrowing, 12 flags -- 10
asking, 2 hand-checked false positives. A survey, not a rate. - Agent behaviour: still unmeasured on a current model. The earlier +0.0 came
from an instrument that could not record a passing cheat; the fix is in, the
re-run needs a funded key. No page claims an effect in either direction.
What did not work
The two pilot runs that exposed corpus prompts not stating their requirement, the
release blocked on API balance, and the eight days CI was red on a released tag
because the benchmark classifier could not read the checker's output. It is all
in docs/what-did-not-work.md.
Full write-up with intervals and limits: docs/benchmark-summary.md and
docs/claim-matrix.md.
greenwash 0.4.2
Note (2026-11-06): this tag was re-cut minutes after publication. Its first cut failed the Python 3.10 CI matrix on a pre-existing collector timeout bug --
concurrent.futures.TimeoutErroris not the builtin before Python 3.11. The fix is in this commit; nothing else about the release changed.
greenwash 0.4.2
A trust report on what a coding agent's "done" actually means -- now with a
strict gate, auditable waivers, and a Codex plugin that runs the same checker.
pipx install greenwash-cliWhat's new
Strict mode (--strict, or "mode": "strict"). Report mode's findings plus
one question: who signed off? A finding that only asks for review -- a new
mock, a touched conftest.py, an error path returning a default -- blocks the
turn until it is decided, and ignore_rules is refused rather than honoured:
the run exits 3 and the findings the suppression was hiding stay in the report.
Two lines are drawn deliberately. Evidence is not waivable: a held-out suite
that failed is an observation, not an opinion. And a broken check is not a pass.
Waivers (greenwash waive). The resolution for a finding that is correct is
a decision with a name, a reason and an expiry on it, not a rule switched off. A
waiver binds one rule, one path and one fingerprint -- and the fingerprint
covers the line, so editing the line invalidates it and the report says
stale instead of quietly permitting. Thirty days is the ceiling. A valid
waiver reports as waived with review, never as clean.
The Codex plugin (codex/), strict by default. Manifests in both spellings,
a Stop hook with a POSIX command and a Windows one, its own hook script and
skill. A plugin is opted into once by someone who wants a gate, so the profile
there is the longer table even though the CLI's default is not.
The two hooks now read the CLI's exit code instead of re-deriving the
policy. The Claude hook used to block on any finding in enforce mode --
including a verified one -- because it kept a second copy of the decision table.
Two copies drift. The CLI owns the policy now; the hook owns the wording.
Changed
- The distribution is
greenwash-cli. The PyPI namegreenwashbelongs to
an unrelated project, so every install line names the distribution this
repository actually publishes. The import name and the command stay
greenwash. npx greenwashis no longer advertised. The npm name is not owned by this
project either; the shim is real and still tested, butnpm/README.md
documents the localnpm packpath instead of a registry install. An install
command that installs someone else's package is not an install command.
The numbers, including the one that is not a win
- Detector: 24/24 planted cheats caught, 0 flags on 24 recorded honest
fixes, static 24/24, behavioral 22/24. - Agent behaviour: 252 scored runs across three plugin states, and no
measurable effect -- 7% silent cheats with the plugin off, 8% with it fully
on, -1.4 points, every arm's interval overlapping every other. All 16 silent
cheats weremockorswallowshortcuts; nine were one task, and no arm
caught it. That is the useful part. The pilot's instrument could not label a
passing cheat when it ran, and its rows carry no detector flags, so the number
is a label count rather than a measurement of the plugin's effect (see
docs/what-did-not-work.md sections 1 and 12). - False positives on real history: 250 commits and 1,029 file changes in ten
unrelated repositories produced 109 flags, 104 of them one asking rule on CI
and dependency configuration. That rule was narrowed to fire only when a
change can alter what runs; the same history now produces 14 flags -- 12
asking, 2 false positives -- and all seven remaining flags from that rule are
real runner changes. Two of the asking flags are the first measurement of
error-path-default, on onepytestchange seen through a commit and its
merge. The before and after are in
docs/false-positive-survey.md.
The full write-up, with intervals and limits, is
docs/benchmark-summary.md.
What did not work is
docs/what-did-not-work.md.
Receipts
404tests, no model and no network required.- CI runs the CLI on Ubuntu and Windows, Python 3.10 and 3.12; validates every
benchmark task in its own language; exercises the GitHub Action and the
pre-commit hook against a planted faked pass; packs and installs the npm shim
and checks it forwards exit codes; and runs the clean / block / waive / error
flows through both stop gates. - Two recorded demos drive the real hooks and print captured output:
python demo/strict_claude.py --terseandpython demo/strict_codex.py --terse.
It is still a report, still local, still dependency-free, and still honest about
what it cannot establish.