Repository navigation
Releases
v0.7.0
Minor Changes
-
What's new in 0.7.0
Verdict 0.7.0 is a breadth release. It adds a local code index, a read-only
servesurface with a browser
UI, and a backtest harness that mines past bug fixes from your own history, replays them against your
agent, and scores the result. Review output gains an auditable provenance receipt, diff-coverage
attestation, advisory risk scoring, and an optional validation pass that floors each finding on the
evidence behind it, while the CI gate now checks coverage by default. Posting grows a secret guard, honest
per-host capability reporting, and a summary-comment option, GitLab and Forgejo join GitHub as full publish
targets, and lens packs, declared recipes, and reported verification make the review pipeline itself data.
Thirteen pre-1.0 changes alter behavior — each names the migration step below — and the known limits close
the notes. -
New commands and flags
pr-wait— wait for a PR to merge, close, change checks, or change review state, with a bounded timeout and typed timeout data. Bounded foreground waitsingest— import SARIF 2.1.0 results into a bundle's findings store, sharing one dedupe and suppression path with model reviews. Static-analysis ingestionthreads resolve— close the review threads whose findings your store marks addressed, one receipt per run, with--dry-runand typed refusals for threads verdict did not author. Guarded submissionrisk— rank every local bundle into an advisory triage queue by band and 0–100 score, without a host call or an index build. Triage risk queuereport --html— export the six dashboard views as a static site, published through staging and backup siblings with rollback and symlink refusals. Static dashboard exportrun --agent— run your own agent CLI against an existing bundle, assert exactly one freshreview.json, then hand off to the unchanged record/post validation. Your agent, launchedpanel— plan independent verifier dispatches locally; keep-biased results enrich findings and summaries without voting or dropping anything. Panel verification- Recipes —
.verdict/recipes.jsondeclares a lens set, synthesizer hint, and budget as strict versioned data, read at the PR base and hashed into the bundle. Recipes skill show,skill status,skill uninstall— inspect one payload file byte-identically, gate on drift with an exit code, and remove only the files a receipt owns. Install routespost --summary comment— maintain one separate PR-level summary comment, updated in place across all three publish hosts; the defaultoffkeeps output byte-identical. Guarded submissionfix --all— the batch form of the bounded fix loop: every open inline finding at or above--min-severity, one pass per sig under the same ceilings, per-sig outcomes indata.batch, and afix.patchwritten beside the ledger. Fix loopsfix --selected— address exactly the sigs a human 👍 selected on the root verdict comment, read through one shared reaction reader; a selection file is written only when something was reacted to. Fix loopshandoff— one markdown projection of the open inline findingsfixwould take, selecting through the same selector, with--json,--format md, and repeatable--paths; read-only, with no ledger and no host call. Handoffrun --print --stage <n>— render one recipe stage at a time, so CI can dispatch the four stages under different models. Recipesbacktest reverts— record the reverts git already knows about into a sorted, byte-stable.verdict/reverts.jsonl, with.verdict/reverts.mdas the fenced reading surface. Revertsbacktest mine— mine SZZ candidates from a local clone with zero model calls and zero network, writing a timestamp-freemine.json. Backtestbacktest replay— replay mined candidates against your own agent in a throwaway worktree per candidate, bounded by review count, projected budget, and timeout. Replaybacktest score— grade a replay's findings against the candidate bug lines and report a catch rate, off-target findings per PR, and cost per review as a pure read. Scorelearn --incident— attach one revert or incident to a recorded bundle in a sorted, idempotent.verdict/incidents.jsonlledger. Knowledge base
-
Local code index
verdict index— build a local Go index with the standard library, stored by tree SHA so an empty commit reuses a tree and a rename parses nothing. Code index- Graph queries —
index tests-for,index co-changed, andindex ownersanswer candidate tests, co-change partners, and the CODEOWNERS rule at a tree. Code index - Three more languages — the index now reads TypeScript, JavaScript, and Python through two dependency-free lexical scanners, with no tree-sitter and no cgo. Code index
index blast-radius— one change's reach: callers to depth 2 plus candidate tests, co-change partners, and owners for the changed paths. Code index- Transitive tests —
tests-forwalks the reverse import graph up to--depthhops and labels tests it reached through imports. Code index fetch --index— hand an index you already built tofetchorfetch-localand the bundle gains animpactartifact and a fenced## impactsection. Impact hand-off- Missing-index hints —
index blast-radiusandfetchnow name the revision and tree an absent index is about, with the exact shell-quoted build step that answers for it. Code index - Scanner fixes — four corpus bugs are fixed: a transitive test no longer disappears, relative imports resolve, a Python attribute call is no longer misattributed, and a Go selector outside its package is counted unresolved. Code index
- One impact budget — the impact section's byte budget is shared across symbols, callers, and tests, each list carrying its own omitted count. Code index
-
Review quality and provenance
- Explicit context —
fetch --contextandfetch-local --contextattach named context files, and findings cite them. [Explicit context](https://github.com/stainless-code/verdict/blob/main/docs/m...
- Explicit context —
v0.6.1
Patch Changes
- The npm wrapper and each platform package now ship the repository README, so the registry page renders the project documentation instead of an empty README.
- The package page now names what the tool does rather than a claim that stopped being true: the description drops "zero install" (every route installs a binary) for the loop it actually runs — fetch, review, gate, post — plus cost rollups and one static binary, and the package carries the docs site as its
homepageinstead of letting npm fall back to the repository README. Keywords gaincli,mcp,go,llm,static-binary, and the agent vocabulary, and drop the self-referentialverdict. The docs site's own description moves to the same sentence, so the About box, npm, and the site stop disagreeing. - Public surfaces now match the binary. The docs site, its home pages, and the CI review example use
.verdict/state/; the operator manual documents deep validation, path-scoped profiles, the evidence ledger, posting guards, local review coverage, and the shipped CLI flags, and states thatverification.jsonis agent attestation rather than tool proof; the TypeScript client throws a typedVerdictGateErrorfor a gate-shaped exit 1 instead of a schema error; the Python smoke uses the real bundle path and runs in CI; the Go SDK exposes record dry-run/strict/cost-names and fetch/fetch-local head/worktree controls;config --explainhuman output shows privacy exclusions and plan group ids;spend-cappedcarries a hint; and the README says the Go SDK runs in-process.
v0.6.0
Minor Changes
-
Breaking (cost record):
cost.jsonaccepts onemodelsshape: an array of entries. A record whosemodelsis an object is no longer read by key order — that coercion dropped the keys, which are the only place the writer named a model, so the entry priced without anything able to attribute it. Such a record now reads as unreadable spend, named by the existingcost.json unreadable — spend unrecorded for this rollupwarning, and the run still completes. Every record the shippedprompts/review.mdtemplate writes already carries the array form, so no template and no documented recipe changes. An existing object-valued record migrates by rewriting it as"models": [{"name": "<panel slug>", "model": "<model id>", "input_tokens": N, "output_tokens": N}];record --cost-nameswill not convert it, because converting would restore the shape the reader just dropped. -
record --cost-names[=<names>]names themodels[]entries of a bundle's cost record and its per-chunk records, in place and without changing a count. Names come from the flag's own ordered list — synthesizer first, then lenses — or from a panel pinned in the bundle; with neither, the repair refuses rather than inventing a name from a slug.An entry that already names its seat keeps that name. A name that sits on the wrong seat, repeats, or comes from outside the panel refuses with a code naming the entry and the seat it should hold, and an entry the panel cannot explain refuses rather than being left half-attributed. A document holding several records gives each record the whole panel, never a share of it.
The repair plans every cost file before writing any, and it runs after the record verb's other checks, so no refusal rewrites a cost record. A cost file that changed after it was planned refuses rather than replacing the newer write, and a record that grew past the reader's 8 MiB cap refuses before anything is written. A write that fails mid-pass stops with the files already written named in the refusal, and re-running finishes the repair.
This verb is how an operator repairs the array that the same release made the only shape: a record whose
modelsis an object iscost-models-not-arrayhere rather than converted, because converting would restore the second shape the reader just dropped. Migrate by writing an array whose entries name their seat — the migration note sits with the array-shape change, and this verb is its repair path. -
Lens certification becomes version-scoped, and aggregation stops refusing records it can read. A cost record offered as
schema_version: "2"may carry only the top-level fields v2 defines: an extra field now refuses by name and version (carries top-level field "notes", which that version does not define — remove the field; no version this build certifies defines it) instead of a generic "unreadable" sentence, a record bound to the run at a version this build does not certify refuses asversion-unsupportedrather than being re-read as v2, and aschema_versionno reader can parse (null, a boolean, an array, an object, an empty string) ismalformedinstead of counting as absent — so an unreadable declaration can never let an older certified record stand in for the newest binding. The group estimate and the cost rollup take the other side of the split: they price the fields they know and ignore top-level fields a newer writer added, so one run's stray key no longer voids the history an estimate is built from — that closes abudget-unknownrefusal underrequire_known_estimatethat a single unregistered field could previously trigger. Nested shapes stay strict, and every fail-closed check still fails: foreign capture, duplicated key, unreadable entry, mismatched binding.
Patch Changes
report --costkeeps a chunk's spend when the bundle's owncost.jsonis unreadable. The broken top-level record used to be folded into the same rollup as the per-chunk records, so one bad file voided every chunk beside it and reported zero tokens for a run that had real measured spend. It is now dropped on its own — named once bycost.json unreadable — spend unrecorded for this rollup, the same way an unreadable chunk is named — and the chunk records still price. No other shape changes: a bundle with no chunk records reads its top-level record exactly as before.- The release body is asserted, not assumed.
scripts/release.shnow reads the published release back after GoReleaser publishes the assets, rewrites the body fromCHANGELOG.md's section for the tagged version when it differs, and fails the run when the rewrite does not stick or the release is incomplete. A body that differs only in the whitespace GitHub hands back is accepted rather than rewritten, and a release that is missing assets never gets a body at all. Release 0.5.0 shipped a one-byte body while the notes file held 15.9 kB and every publish step reported success; the newbodystage closes that gap, therelease-bodyeval holds it offline, and a re-run leaves an already-correct body untouched.
v0.5.0
Minor Changes
-
The bench runner learns a repair loop. Pass
-loop Nto let one sample try up to N rounds: the harness runsrecord --dry-run --json, feeds the problem list back, and records only the last valid round. Round evidence lives underruns/<slug>/round-NN, the report gainsloops,fix_rates, andloop_cost, and-loop 1keeps the old single-shot files byte-identical. -
Findings can now carry a closed category (
logic,edge-case,security,
performance,accidental-commit,style,docs,other): an optional
per-comment label that never changes the finding's signature.report --categoryfilters to one class,report --check --categorygates only that
class, andreport,calibrate, andtrendall publish per-category counts
(untagged for unlabelled findings). Unknown or non-string values warn once
and store null; they never refuse a review. -
Findings can now carry a repair brief: the affected sites, a one-sentence
outcome, an optional approach and open questions.recordvalidates it against
the fetched change and keeps it across re-records,show --jsonandfix
print it, and it never changes a finding's signature. A malformed brief warns
and is dropped; it never refuses a review. -
verdict gateis one composed CI gate over the three dimensionsreport
already computes: open findings at-or-above a tier (--max-severity), a
completed post for the fetched HEAD (--require-head), and live unresolved
threads (--live auto|strict|skip). One exit code carries the answer — 0 pass,
1 fail-or-unknown, 2 usage, 3 tool error — and the human tail names each
dimension, the tally, and onetry:hint covering every failing dimension.
--jsonaddsdata.passedplusdata.dimensions[]infindings, head, live_threadsorder withcountnull where a dimension has no tally; a failed
gate keepserror: nullbecause a verdict is not a refusal.gateis
read-only (no store writes, no post) and absent from MCP: a local pass is not
authority to merge. -
Breaking (experimental MCP): a host on
verdict mcp --read-onlyno longer
sees athreadstool, andreportthere drops its--output/--sarif
arguments and reportsreadOnlyHint: falseinstead of claiming to write
nothing. Seven tools remain:report,next,show,trend,ready,
doctor,check-tree. A host that needs the thread feed should refresh with
verdict threadsoutside the server or run it without--read-only; a host
pinningreportas a read-only tool should stop trusting that hint. -
Patch identity dedupes re-reviews across rebases.
fetchandfetch-localwrite.verdict/state/<bundle>/identity.json(git patch-id --stable, labeleddiff-sha256fallback),postrefusesalready-reviewed-patchon a same-patch sibling unless--force(recorded inpost.jsonforced[]),adopt --from automatches exactly one local twin, andreport --completenesspasses when the posted and fetched heads differ only by a rebase with the same patch. -
verdict recordcan now answer "would this record?" before writing anything.
record --dry-runruns the same checks and writes nothing, returning a
summary of the comments, sections, and anchors that would move to the body.
Both modes share one checker instead of stopping at the first problem: the
refusal carrieserror.problems[]in document order with a JSONPath, a
publishedcode, the level, the message, the exactfix, and an optional
example, pluserror.problems_truncatedpast the 50-problem cap; human mode
lists the same problems before its try/re-enter lines. The checker owns the
brief, anchor/side, kb-id, chunk-attestation, confidence, security, claim,
triage, section, and verification-input checks, not just the top-level shape.
Warnings (moved anchors, brief parts dropped without refusing, malformed kb
ids, unknown fields, an unset rigor) let the record proceed unless--strict
or"strict_input": trueincontrols.jsonpromotes them.
pkg/verdict.CheckReviewexposes the pure checker, the MCPrecordtool gains
dry_run(served on a read-only server; a non-dry record is still refused),
and the review skill now loops write -> dry-run -> fix for at most three
rounds, then reports any remaining problems instead of recording a partial
review. -
verdict recordnow repairs the mechanical shapes weak models emit instead
of refusing them. Before checking,NormalizeReviewunwraps one
whole-document code fence, drops prose around exactly one balanced JSON
object, wraps a scalar where a string array belongs, renames the
inBounds/rung/filealiases, folds rigor synonyms and enum case, and
removes a BOM, an explicit null on an optional key, or an exact-duplicate
comment. Every rewrite is a warning with its own code, listed in
record.json'srepairs[];--strict,REVIEW_STRICT_INPUT=1, or
"strict_input": trueincontrols.jsonturns each into an error.
record --dry-runreturns the exact stored object asdata.normalized
plusdata.repairs, and never rewritesreview.jsonon disk. The CLI also
owns the body's> Basis:line, replacing a disagreeing line (or inserting
a missing one) with the computed counts and reportingbasis-replaced.
Unset rigor records assaidwithrigor_source: default, and length and
numeric bounds now refuse asout-of-rangeinstead ofbad-enum.
pkg/verdict.NormalizeReviewexposes the pure, idempotent primitive. -
Breaking: the published schemas moved with the skill, so every schema's
$idbase path is nowskills/verdict/schemas/rather thanschemas/—
https://github.com/stainless-code/verdict/blob/main/skills/verdict/schemas/envelope.schema.json,
for example. If you resolve a schema by URL, or pin$idin a validator, update
the base path; the schema contents are unchanged. -
The Go SDK and the TypeScript client can read review-thread state in-process,
so the one review read an embedder had to shell out for is now a method.
verdict.Threadswraps thethreadsverb:ThreadsRequestcarries the repo,
PR, and--provideroverride, andThreadsResponsereturns the envelope plus
the refreshed feed tally (Dir,File,Provider,HeadSHA,FetchedAt,
Truncated,Threads,Open,Outdated,Verdict), withData
byte-identical tothreads --json. The TS client gainsthreads(...)
resolvingEnvelope<DataFor["threads"]>, and every review verb now has a typed
client method instead of falling back torun(). No wire or behavior change:
the verb, its data profile, and its MCP tool already existed. -
verdict skill install --detectadds the review skill to every coding agent
already set up on the machine, instead of making you look up one path per agent.
It names each target and its result, and--jsonlists each target with its
status, path, and files written. With no agent set up it refuses and points at
--agent <name>, and--detectis mutually exclusive with--dirand
--agent.install.shandinstall.ps1now close a successful install with
that exact command. -
The TypeScript client's verb methods now return their own
dataprofile
instead ofEnvelope<unknown>:readyresolvesReadyData,fetchresolves
FetchData,trendresolvesTrendData, and so on for every verb method.
DataForexposes the same verb → profile map, so code that passesdata
around can name the profile without restating it.Every profile is generated as a
<verb>Data$defsentry in
envelope.schema.json(with the nested$defsit references), so non-TypeScript
clients get the same shapes from the published schema. Additions only:data
stays an open object that allows additional properties, no existing property,
required key, or type changed, andclient.run()still returns
Envelope<unknown>for verbs without a profile. -
Add typed configuration in
.verdict/config.json, with path scopes, captured explanations, review profiles, grouped spend and diff limits, and recorded deep validation. Local run artifacts now live under.verdict/state/; retired.review/knob files are no longer read.Cost records use schema version 2 with grouped lens and cost evidence. The Go Commander interface now takes a command pointer and preserves its working directory.
Patch Changes
-
Corpus cases accept an optional
also_lineslist of alternate anchors for
the same defect. Each entry is matched with the same line tolerance and
defect-naming rule as the primary line, matching any anchor counts the bug
once, and the reported line stays the primary one. The generated
corpus.schema.jsonnow requires positive line numbers inalso_lines
items, so schema validation rejects0and negative entries just as
calibratealready refused them ascorpus-invalid. -
calibrateandtrendgrade a corpus finding on the defect it names, not the
line it picked. A finding used to count as a true positive whenever its path
matched and its line sat withinline_toleranceof a known bug, so on a
four-to-seven-line hunk every anchor was correct and the reported precision
measured line-guessing rather than review quality; the same line could hold a
comment about variable naming and score as a catch. A true positive now also
has to name the bug: one of itskeywords(new, optional in
corpus.schema.json), itskind, or a phrase from the published category
vocabulary. The default window drops from three lines to one, and a findingrecordmoved out of its hunk is
never a true positive, so the anchor is measured and not assumed.The human false-positive line says which rule failed, so a run distinguishes a
finding that landed off the bug from one that landed on it and described
something else. Corpus bundles the benchmark runner writes now carry the
reviewedfull.diffand ...
v0.4.0
Minor Changes
-
Measured review rates instead of impressions. A marked
corpus-*bundle kind
carries a checked-in case manifest with ground-truth bugs and provenance;
calibrategrades a model's findings against that ground truth offline and
reports precision, recall, and false-positive rate, andtrendaggregates the
same numbers per model and panel. Ordinarylocal-*reviews keep refusing the
live join. Ships a seeded starter corpus underbench/corpus/and
docs/benchmark.md, an operator runbook for real PRs, real agents, cost
estimates, and publishing measured rates. -
Experimental MCP server.
verdict mcpserves the review verbs to an MCP host
over stdio (newline-delimited JSON-RPC 2.0): no daemon, no HTTP, no new
dependency, sharing the host's credentials and exiting when stdin closes.
Tool schemas reflect the CLI request types, results are the--json
envelope'sdata, and typed refusals returnisErrorwith code, message, and
hint intact;--read-onlyexposes only the non-writing verbs, whilepostis
available in normal mode and in--operatormode.verdict mcp config --agent claude|cursor|codexprints the exact host entry, and--writemerges it
idempotently into the documented config (.mcp.json,.cursor/mcp.json, or
$CODEX_HOME/config.toml), backing the old file up to<path>.verdict.bak
and refusing unparseable or foreign entries instead of rewriting them.--json
prints one envelope, never prompts, and never starts the server;verdict --helpnow lists the MCP entry so the surface is discoverable. -
verdictis one static binary — no checkout, nojq, no shell script. Install
it with the verifiedinstall.sh(macOS/Linux) orinstall.ps1(Windows) into
~/.verdict/bin, the Homebrew tap, or the npm wrapper; both installers verify
the cosign signature and the archive SHA-256, smoke-test the installed binary
from its final path, and addPATHwith one marked line
(VERDICT_NO_MODIFY_PATH=1opts out). In CI the newsetup-verdictaction
installs a pinned, verified copy on Ubuntu, macOS, and Windows runners without
touching shell rc files.verdict upgradeupdates whichever channel installed
it — a verified self-replace for installer installs, the owning package
manager's global update for npm and Homebrew — andverdict uninstallremoves
only installer-owned files; both refuse, naming the exact command, when the
install belongs to something else.verdict skillprints or installs the
payload embedded in the binary (SKILL.md,prompts/,schemas/, and the
scripts/verdictshim), so the served skill can never skew from the binary.Upgrading from 0.3.x: the old script install is replaced, not patched.
scripts/review.shis gone — every verb isverdict <verb>with the same
flags and--jsonenvelopes. The old symlink-onlyinstall.shbecame
install-skill.sh(checkout-only), and the copy-basedinstallverb is gone,
so reinstall the skill withverdict skill installornpx skills add. Docs
and scripts that call the old script path must be updated. Start at the
rewritten quickstart, the new glossary of review terms, the CI recipe in
docs/ci-recipe.md, andverdict --help, which now names each verb's
provider scope. -
Posting is harder to get wrong and harder to do twice.
postrefuses a second
review on the same reviewed SHA before any network read, naming the exact
fetch --base-shaadvance; it also refuses duplicates found in the bundle's
own fetched review list, so a fresh checkout, a CI re-run, or another machine
cannot repost what the PR already has —--forcestill overrides, and the
refusal names the re-fetch needed when the earlier review was deleted. Points
outside the diff hunks post under "Unanchored observations" instead of as
inline anchors GitHub would reject. A body past the 65,536-character hosting
limit splits into follow-up COMMENT reviews — first review keeps the footer,
later parts carry a continuation header, and a failed follow-up refuses typed
naming the review that already landed — instead of being trimmed. A 403/429
carrying a rate-limit signal retries within bounds, honoringretry-afteror
the reset marker and never publishing twice; when the last permitted attempt
is still retryable, the refusal carriesretry_atin RFC 3339 so an agent can
wait for that instant.A stale-fetch downgrade note is now followed by a blank line, like the drift note, instead of a stray literal
\nin the posted review body. -
The PR conversation is a local, schema'd record instead of something each verb
re-reads from the host. Every fetch writes.review/<pr>/threads.json(schema
published) holding each thread's host id, diff anchor, resolved and outdated
state read from the host's own fields, root author kind, and the comments with
authors, timestamps, and capped bodies; GitHub comes from GraphQL
reviewThreads(REST carries no resolution), and GitLab, GitCafe, Forgejo,
Bitbucket, and Azure fill the same shape, paging rather than truncating.
verdict threads <owner/repo> <pr> [--provider …] [--json]re-reads just that
surface and rewrites only that file, so re-review, address-reviews, readiness,
and calibration can join prior feedback on any of the six hosts. Posted inline
comments now carry a hidden final-line<!-- verdict-finding sig=… -->marker,
so a new comment links back to the finding that produced it while older
comments keep reading as unlinked. -
Published machine contracts beside the CLI.
schemas/ships JSON Schemas for
the review input, findings store, cost and usage records, PR threads, and the
--jsonenvelope with per-verbdataprofiles, included in the npm tarball;
schemas/validate.shchecks documents against them offline (nojq; exit 0
valid, 1 rejected, 2 bad usage), and the drift rule is additive within schema
version 1. An experimental TypeScript client
(@stainless-code/verdict/client) wraps each verb with envelope types
generated from those schemas and refusal codes generated from the CLI, with no
runtime dependencies. The experimental Go packagepkg/verdictembeds the
same operations in-process through a context-first client with typed requests
and responses and a*RefusalErrorforerrors.As; its API may still change
before 1.0. A new "Build a client" guide documents the--jsonprotocol —
envelope keys, exit and refusal codes, versioning, binary resolution,
timeouts, output caps, and when MCP beats spawning a process — with a
dependency-free Python client generated from the schemas and an offline smoke
test. -
The readiness gate can say what it knows.
report --checkand
report --completenessgain a live-thread dimension: open threads the local
store does not know about fail the gate, and unreachable lookups warn and stay
neutral. New--live auto|strict|skippicks the posture —autokeeps that
fail-open default,strictfails closed when the host cannot be reached, and
skippays zero network and gates on the local dimensions only, saying so in
the output and the--jsonwarnings.postand the gate now name the same
advance for head drift, and a binding APPROVE on a known-stale fetch
downgrades to COMMENT instead of approving past the moved head. -
Chunked split-range reviews. Very large PRs are reviewable again: with
--chunk(or.review/chunk) an over-cap diff is sliced on file boundaries
into numbered chunks with a versioned manifest, the agent reviews one chunk at
a time, and the results synthesize into one verdict with a summed Basis line.
postandrecordrefuse coverage that claims to be complete but is attested
partial, and warn loudly when a chunked review is unattested.fetch-local
supports the same flow. -
verifyrefuses before it runs. Destructive, networked, credential-reading,
and repo-escaping commands are rejected with a typed error before execution
(over 40 regression shapes), so a hostile or carelessverifyCmdcannot
delete, exfiltrate, or escape the repository. The recorded tail also drops
credentials no environment value names — vendor-prefixed tokens,
AuthorizationandProxy-Authorizationheaders, bareBearervalues, and
URL userinfo — and strips ANSI escapes and other control characters, so a
command cannot rewrite the proof it lands in;.verify.outkeeps the raw
bytes.doctorwarns on excess token scope, and the manual documents an
opt-in hostile-PR sandbox recipe plus the residual risk a string scan cannot
remove.
Patch Changes
- Cost accounting reads what the operator actually knows.
cost.jsonentries
gain optionalcache_read_input_tokensandcache_creation_input_tokens
(subsets ofinput_tokens; absent ornullreads unknown), and
report --costrolls them into a top-levelcached_input_tokensscalar with
totals unchanged. The spend gate can now price comparable runs from usage
records the operator hands over withVERDICT_USAGE=<file|dir>, in a published
schema (schema, model, tokens, optional USD, source, window), with the
self-reportedcost.jsonas the fallback; a handed record joins by its window
bracketing the run, and a named path that is missing or malformed refuses typed
before the bundle is written. Automatic discovery of other tools' transcripts
is gone — the optional converters live outside the binary in
examples/usage-adapters/. With nothing handed over, output is byte-identical. - Isolated local reviews.
fetch-local --worktreechecks the review target into
a temporary detached worktree and builds the bundle from that checkout instead
of the live tree, so caller tracing and verification run against real, stable
code without disturbing the working tree. The checkout is always removed
before the verb exits (success, refusal, or cancel); when no checkout...
v0.3.1
Patch Changes
- Basis-line validation compares the agent's line against the review's own comments instead of the merged findings store, so re-reviews with carried-over findings no longer warn. Genuine miscounts still warn and keep the body.
- Same-head reposts omit the review link in the footer instead of pointing at the superseded review. Fresh posts keep the link; plain-text fallback unchanged.
v0.3.0
Minor Changes
- Binding APPROVE now rechecks live open threads before posting: unresolved live threads the local store doesn't know about refuse APPROVE with a typed error (triage or
--force); an unreachable lookup degrades to COMMENT with the reason stated instead of approving blind. COMMENT, dry-run, and local flows make no new network calls. - Attribution v2: the footer default is now a linked, leveled, logo-bearing line — a product-linked 18px logo (defaulting to the tracked
apps/docs/public/logo.svgraw asset) plusReviewed with [**Verdict {version}**](https://stainless-code.com/verdict)with the version inline space-separated, followed by the receipt fields.publish.attributionmigrates from the day-old P1 bool to anoff/product/fullenum (file bools refuse with the mapping named:"off"for false,"product"for true;fullappendsby [Stainless Code](https://stainless-code.com)). Newpublish.logo(default true) andpublish.logo_urlknobs (empty reads text-only, never a broken image); a<!-- verdict-footer -->machine divider separates content from footer with strip-and-replace re-rendering;attribution_textgains the{version}/{tier}/{product_url}inventory with unknown placeholders refused. - Opt-in behavioral diagram: agent-drawn mermaid under a ported contract (
prompts/diagrams.mdwith the 7-type selection map, honesty rule, and copy-exact skeletons), gated by a newpublish.diagramsnamespace (behavioraldefault off,structuraldefault on for the P3 topology half,expandeddefault collapsed), grammar-level warn-and-drop validation that never fails a review, and--diagram-behavioral/--diagram-structural/--diagram-expandedflag pairs. - Bounded verify-gated fix loops: new
fixverb runs one pass per invocation over local bundles (top open inline sig or--sig), printing finding + fullkb.txtcontext with patch → verification.json → record → re-run instructions. Hard-coded ceilings (max 3 passes per sig in.review/<bundle>/fix.json, scope-escape on out-of-anchor worktree touches), start gates (conflicted tree, closed bundle), terminal completion summary, local-only in v1. - Self-grading probe:
calibrate --probegrades recorded findings against triage dispositions as the script-owned baseline, and merges agent judgments (--judge/--judge-file, validated, atomic, sticky) that may diverge — the agreement/dissent rate over doubly-judged rows is the instrument reading. Additive grade file only; findings, joins, and calibration rows untouched. - Anchors validate against the fetch-time snapshot:
fetchpins the commentable set andpost/record/adoptreuse it (rebuilding only if missing, with a warning); dropped anchors ship as named body notes with reasons instead of failing or misplacing silently. Mid-run pushes read as drift context, not a moving validation target. - Structured local knowledge base:
learnverb (append/list/retire/restore scoped rules with provenance, rigor, hit counters),kb.txtrecall onto relevant diffs as context-never-orders, flip-time credit with false-hit-rate auto-retirement, and per-credit audit rows. - Knowledge base: re-posting a finding without
kbkeeps the prior cite while explicit"kb": []clears it (uncite); every.review/kb-joins.jsonlaudit row now also carriesbundleandreview_id. - Overlap audit:
trend --overlapjoins recall manifests × findings cites × the joins ledger into recall precision, recall coverage of cites, and the cited-outcome split, with joins corroboration and calibrate eligibility. Pure read, no network; missing manifests read as unmeasured, never zero. - Change overview:
report --outputrenders a script-computed collapsed file table plus aflowchart TD
change-topology diagram on large diffs (more than 8 files or 24KB,--overview/--no-overviewoverride),
with stable F-row ids, deterministic relations only, and a newpublish.overviewknob. - Permalink-first evidence: findings reference GitHub blob links pinned to the reviewed SHA (reference-style,
path:linelocators with definitions), with quotes over three lines kept inside collapsed detail blocks
only. Non-GitHub bundles degrade to plainpath:linetext. Link validation warns and keeps, never failing
a review over garnish. - Debug legibility:
fetch/fetch-local --show-promptsdumps the exact assembled prompt (system + mode + recall + coverage manifest) without running an agent;report --outputbodies lead with confidence plus safety and list up to seven checked files on clean reviews, with full reasoning staying in the machine envelope. - Publish footer: posted bodies and
report --outputend with a merged receipt block (tool version, tier, cost or unrecorded, files reviewed, review link when known) including a factual one-sentence attribution, plus a deterministic confidence-basis line (rigor counts, cited rules, verified fixes). Newpublishsettings namespace (footer/attribution/attribution_text/copyright) with flags beating file beating env; agent Basis lines validate warn-and-keep, never failing a review. - Recall manifest: the recall step emits
kb-recall.json(recalled rule ids in rank order with scores) besidekb.txt, rendered from the same single sort. The human note is byte-identical, silence on clean is preserved, and all readers tolerate a missing manifest. - Report gate answers "is HEAD reviewed?": new opt-in
--completenessflag onreport(alone or with--check <tier>) compares the fetched HEAD against the completed-post record, failing closed with a typed error naming the re-fetch command. Without the flag, output and exits are unchanged. - Onboarding pair:
review.sh initscaffolds the standard local control files (budget, panel, KB policy, KB starter) with chill/standard/strict presets and runs check-tree validation, offline with preview-then-confirm; an XDG-homed presentation-only settings file (output/layout/color, schema-hinted, credentials refused) ends env sprawl with flags winning over file over env and byte-identical behavior when absent. - Operator-owned template overrides: per-block fragments under
.review/overrides/(preamble, footer
order, walkthrough threshold, attribution) guarded by an enumerated locked-invariant list (marker line,
findings structure, guard satisfiability, metadata presence, divider shape), validated at check-tree time
and revalidated at post time with typed refusals.check-tree --jsonemits the standard envelope on all
paths. - Exactly one terminal write, enforced:
postrefuses a second completed post on the same reviewed SHA with a typed error (--forceoverrides, new commits re-arm automatically). Dry runs, attempts, and failures never arm the guard. The existing exact-payload replay guard stays as the fine layer underneath. - Copy-based skill installer: new
review.sh installverb detects the de-facto agent skill dirs (.agents/skills, .claude/skills, .codex/skills at the repo root and $HOME), previews the copy plan for the tool-owned verdict/ payload (SKILL.md + prompts/), and applies only on explicit confirm (same preview-then-confirm idiom as init, typed needs-confirm). Only installer-owned verdict/ dirs (with a .verdict-managed receipt) are ever written — foreign skills and unowned dirs stay byte-identical; re-runs are idempotent and receipt/tool version skew warns instead of failing. Offline, shellcheck-clean, operator docs updated.
Patch Changes
- Bird's-eye review follow-ups: indented fenced blocks (up to 3 columns) now validate, single-line
fences never open an end-of-body swallow, drop-notes insert before a carried footer divider so
strip-replace preserves them, the doc wrapper is a parseable multi-line block, and the manual
documents the full v1 placeholder inventory. - Pre-commit runs
shellcheck -S warningover staged shell files (missing binary skips with
a notice, same degrade-gracefully idiom as the formatter check), and the CI lint job covers
every tracked*.shinstead of three hardcoded paths. Pinned by a hermetic hook eval
(violate fails, clean passes, missing binary skips).
v0.2.0
Minor Changes
- Overridable fetch diff cap: REVIEW_MAX_BYTES env (or .review/max_bytes file) overrides the 512000-byte default for deliberate exceptions; refusal names the override last and the envelope carries cap provenance when set.
- Cost control plane: operator spend budget (env/file), pre-run estimate from comparable past runs, and consent flag for projected-over-budget runs with receipts in the cost rollup. Unset preserves today's behavior exactly.
- Record effort tiers (cheap/standard/deep) on cost runs with per-tier rollup, default delta re-reviews to cheap lenses, and emit a sensitive-path manifest plus no-quote rule on fetch (advisory warning, evidence untouched).
- Add a
readyreadiness-advisory verb (CI state, open threads, draft/conflict, description completeness — advice, never a gate) and per-review cost accounting (agent-reported run records plusreport --costrollup).
Patch Changes
- Robustness bundle: optional review-level
confidence(1–5) andsecurity_summaryrecorded onrecord.json/post.json(absent stays absent, out-of-range warns and drops); exact re-entry commands on post/record failures plus an echo-only interrupt trap on post;error:/try:failure-message pass with received values and valid ranges; fetch-local warnings for untracked/binary exclusions and empty paths filters; explicit fetch byte cap (512000 bytes) with split-or-narrow hint and distinct empty-vs-huge signals. - Add
ready --check: opt-in readiness gating with machine exits (0 ready, 1 blockers-or-wait, 2 usage, 3 tool error). Default stays advisory with exit 0; hook recipe in the operator manual. - Add
report --check <critical|important|informational>: opt-in severity gating with machine exits (0 no open findings at-or-above tier, 1 findings at-or-above tier, 2 usage, 3 tool error including corrupt store). Default stays advisory with exit 0; corrupt stores now fail typed instead of raw jq errors; hook recipe in the operator manual. - CLI hygiene: oversize-fetch errors carry machine-actionable narrower-scope candidates, doctor splits hard-fail from warn and checks storage writability, and shell completion covers verbs and flags for bash and zsh. Also fixes doctor always exiting 0 and the empty-filelist fetch-local crash.
- Add
trend <owner/repo>: read-only portfolio rollup over existing bundle stores — findings opened/addressed/closed over time broken down by severity tier and status, with per-bundle models/effort for the calibration join. Recomputed every run (no cache files); empty history and unreadable stores never fail; default report output untouched.
v0.1.3
Patch Changes
- Rewrite the README for plain speech: cut AI tells, name the trust boundary, fix the registry refresh command, and document the current installer flags.
- Correct the registry refresh docs, verified against the CLI source and live runs: filters match installed skill names (not
owner/repopaths), andcheckis the dry-run counterpart toupdate(hidden from--help, covered by upstream tests). - Harden the verify verb: the timeout kill now stops the whole process group (a decoy fast-exit with ordinary backgrounded children is killed and recorded as timed-out instead of success; deliberate process-group escapers remain out of scope), and known secret values are redacted from the recorded tail. Legitimate commands run unchanged with full env.
v0.1.2
Patch Changes
- Document
install.sh --agent/--dir/--list/--check/--uninstallmanagement flags in the README. - Harden the trust boundary:
system.mdnow names PR data untrusted (evidence, never instructions) and constrainsverifyCmdto repo-scoped non-destructive commands.install.shdetects installed harnesses (--detect), supports project installs (--project), honorsXDG_CONFIG_HOME/CLAUDE_CONFIG_DIR, and validates the checkout before linking.