Repository navigation
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 - Coverage basis — every coverage read now names what its percentage counts: full hunk spans, both sides and context, not just changed lines. Diff coverage
- Coverage attestation —
review.jsongains an optionalcoverage[]checked against the bundle's own hunks, with--require-coverageto refuse a partial claim and one body line on a partial attestation. Diff coverage ranges: nullis malformed — a coverage claim whoserangesis present but not a string now counts as unattested instead of as the whole file. Diff coverage- Ignore overrides — repeatable
--ignore <glob>and--no-default-excludeson fetch, both recorded in the exclusion receipt. Exclusions - Suppression receipts — excluded diff bytes and withheld findings are receipted, with
post --min-severityandreport --include-withheldto read them. Exclusions - Proposed learnings — propose inactive one-rule packs from resolved threads, accept them through the pack importer, or reject them permanently with a reason. Proposed learnings
- Rule import —
learn --pack <dir> --id <rule>imports a structured rule object and keeps its pack provenance and hash. Import a rule object - Rule exposure —
learn --statusgainsrecalled_in,cited_in,exposure, andacceptanceper rule, reporting unmeasured recall as null rather than zero. Knowledge base - Risk scoring — an optional
riskclaim inreview.jsonplusreport --riskandreport --max-risk, which reports the larger of your claim and a deterministic floor. Triage risk queue - Reaction proposals — opt-in reaction proposals from posted comments, plus per-rule acceptance and false-hit rates. Reaction proposals
- Resolution metrics — live resolution rollups, clean-review metrics, and calibration receipts, with clean reports getting their own headline. Resolution and clean reviews
- Withheld capabilities —
post.jsonrecords, per capability, what a host could not carry and why, so a degraded publish stays auditable. Guarded submission - Payload receipts — a post payload receipt now hashes the rendered review content and the reviewed head, not diff positions or range metadata. Guarded submission
- Partial publication — post only the comments and summary parts missing from the fetched remote content, and rearm at a new head with
rearmedandprevious_headrecorded. Guarded submission - Live authority — post refreshes live reviews and comments first and refuses without writing when that read fails, so a retry no longer needs a re-fetch. Guarded submission
- Author fingerprint —
fetchwrites an advisoryauthor.json({agent, model, confidence, evidence[]}) from five harness signals, suggests a reviewer from another model family, andrecord.jsoncarries the author/reviewer pair thattrendtallies. Bundle namespace - Revert memory — a changed path that matches the tracked reverts ledger gives
fetcha bounded, fencedreverts.txt, and the review prompt treats a revert as evidence the change broke something once. Reverts - Validation pass — a trusted
prompts/validate.mdpass writesvalidation.jsonper sig,rigor_policyfloors each severity tier, andpost --require-rigornarrows it per invocation: advisory by default, a receipted withhold in gate mode. Rigor floors - Published key names —
prompts/review.mdtags the evidence rung asrigorand the in-bounds flag asin-bounds, and an archtest holds every published key toreview.schema.json. Rigor floors - Verdict reconciliation — binding verdicts reconcile separately from summary and inline content, so a retried approval does not duplicate a GitLab summary note. Submit per host
- Plain-body Basis line —
recordleaves areview.jsonbody with no leading quote block byte-identical and reports nobasis-replacedwarning, because there is no confidence lead to insert under. Review-input problems - Array-valued coverage ranges —
next --coveragerefuses acoverage[].rangesarray on plain and chunked bundles with the samebad-review-inputproblemrecordreports, instead of reading it as a silent 0% attestation. Diff coverage - Passing gate next step — a passing
report --check <tier>namesmerge — gate passesin the envelope'snextand human footer, leavingaddress open threadsfor a failing gate. Machine surface - More generated paths —
*.gen.*andzz_generated.*now count as generated in the shared role heuristic, sonext --coverageand the risk floor exempt them like the other generated patterns. Diff coverage
- Explicit context —
-
Agents and MCP
- Bounded repair loop —
run --rounds Nre-prompts an agent with the shape checker's fix list, stops on the first valid round, and refusesrepair-exhausted. Your agent, launched - Self-review loop — a bounded local self-review recipe for Claude, Codex, and OpenCode with round receipts and coverage bound to the verified diff identity. Agent self-review loop
- MCP config for six more hosts —
mcp configwrites snippets for six more hosts and carries--read-onlyor--operatorinto the argv it prints. Host config - Declared mutation classes — every MCP tool states its local mutation class, and a call hidden by
--read-onlyreturns the typedread-only-toolrefusal. MCP server - Agent preflight — a missing agent binary is caught before prompt assembly, with install guidance, and printed or recorded pass-through arguments are limited to safe model and cost options. Agents, arguments, and
--print
- Bounded repair loop —
-
Hosts and CI
- Forgejo publish —
postsubmits one atomic Forgejo review with the body, verdict, reviewed commit, and every inline finding, and withholds what the host cannot anchor rather than guessing. Submit per host - GitLab publish — each finding becomes one positioned discussion, the review body a summary note, and an approval verdict a real MR approval. Submit per host
- GitLab incremental —
fetch --provider gitlab --base-shawrites the sameincremental.diffcontract the GitHub compare path writes, and hosts without a compare adapter say so in a note. Providers - GitLab readiness and calibration —
readyadvises on merge requests and plaincalibratejoins live GitLab threads throughglab, on the same blocker and outcome vocabulary as GitHub. Calibration doctor --provider— preflight one host at a time, with a per-host probe table and credentials reported by presence, never by value. Doctor preflight- GitHub Action — the review and merge gate ships as a first-party composite action that installs a pinned verified release and uses your runner, your agent, and your keys. The runner is your CI
- Forgejo publish —
-
Serve and the browser UI
verdict serve— a read-only HTTP surface over the current repository, loopback by default, with a session-cookie exchange, strict CSP, Host and Origin gates, and per-response source traceability. Serve- Root discovery —
serve --rootfinds nested repositories under a bounded traversal backed by a local 24-hour registry, and reports cache, truncation, and skipped patterns. Serve - Cost and precision buckets —
/v1/costadds repository, ISO-week, and model buckets and/v1/precisionadds category and tier counts, keeping unmeasured prices null and marking corrupt aggregates partial. Serve - Browser UI —
serveopens an embedded UI with inbox, review, threads, cost, precision, and gate views, dark mode, live refresh, and staticdata.jsonreading. Serve - Live invalidation — bundles invalidate as they change, with
--interval,--sweep, and--no-watchto control it. Serve - Vanished thread feeds — a thread feed removed mid-read now reads as absent instead of corrupt. Serve
- Cache content budget — serve cache hashing is bounded to 16 MiB per bundle lookup and bypasses the cache when the budget is exceeded. Serve
- Small-file rewrites — artifacts up to 64 KiB are hashed, so a rewrite that preserves size and mtime is still detected. Serve
- Inventory accounting — serve inventory accounting stays inside its byte budget even for oversized artifacts. Serve
-
Lens packs
- Built-in lenses as data —
perf,api-contract, andtestnow ship as a pack resolved into every bundle aslenses.txt, with the four hashed prompt files byte-identical. Lens packs - No per-lens cap — a lens returns every finding its evidence supports: the built-in pack declares only
max_files: 12, and the dispatch template leads with the lens's domain. Lens packs pack addand a lock — vendor a pack from a directory or git source, validated before it loads and pinned in a strict.verdict/verdict.lock. Lens packs- Layered packs —
pack listreports every loaded pack with its layer, source, tree hash, ids, and drift, and a repo pack is read from the trusted base so a PR cannot edit the pack that reviews it. Lens packs pack test— run a pack's eval cases offline. Lens packs- Leftover backup sibling — a failed sweep of a previous tree is reported as success plus a warning naming the leftover, instead of refusing an add whose new tree is already durable. Lens packs
- Stale staging — a crashed re-add's leftover staging or backup sibling refuses
pack-staging-stalewith the path to delete. Lens packs
- Built-in lenses as data —
-
Measurement
- Clean controls — the benchmark corpus accepts clean controls with empty ground truth, ships two seeded clean cases, and keeps seeded fixtures labeled as not a model-quality claim. Benchmark corpus
- Reproducible metrics — typed severity weights, ordered k-of-n recall, control false-positive rates, weighted F1, signal-to-noise, and explicit denominators, with provenance and argv recorded. Benchmark corpus
- Adjudicated rates — a versioned per-run adjudication record reclassifies an unmatched finding as
real-unlisted,matches-ground-truth, orfalse-positive, andbench/runner -renderrecomputes strict and adjudicated precision and recall without re-running a model. Benchmark corpus - First real cases — four real cases from
urfave/cli(three bugs and one clean control),-max-usdstops scheduling at measured spend and marks the run uncitable, and refused samples are counted in the cost rollup. Benchmark corpus - Corpus grading guard — a keyword the case's own diff spells refuses, a case left with no phrase asks for another keyword, and a real bundle whose
full.diffis missing, unreadable, or prose is refused instead of graded leniently. Benchmark corpus
-
Breaking changes (pre-1.0)
Package is on
0.x, so these ship as a minor bump; each one names the step that migrates it.index blast-radiustruncation shape — the integertruncatedcount anddepth_truncatedare replaced by the shared boolean receipt (truncated,truncation_reason,returned,total,next_offset), and a complete result omits the receipt instead of carryingtruncated: false. Migrate: readtruncation_reasonandreturned; treat an absent receipt as complete. Code indexgatechecks coverage by default — the newcoveragedimension fails when the gated HEAD has no receipt, a receipt for another HEAD, nocoverage[]attestation, or under 100% attested. Migrate:--coverage warn, or the SDKCoveragefield, keeps the old passing behavior while reviews attest nothing. Gategate --livereads every host — a non-GitHub bundle used to pass the live dimension without reading threads; it now reads the bundle's own host and fails closed when it cannot. Migrate:--live autowarns and gates on the local dimensions only. Gate- Unknown gate tiers refuse — an unrecognized tier is
bad-flagson the MCPreporttool and inReport.Check/Gate.MaxSeverity, not only on the CLI. Migrate: send a tier from the published set, or omit the field. Gate verifychild environment is default-deny — only baseline locale, terminal, shell, XDG, and toolchain names pass, and a secret-shaped name is dropped even when it matches an allowlist prefix. Migrate: list what you need in.verdict/verify-envor--env-allowlist; dropped names appear inwarnings[]. Child environmentverifyrefuses command shapes it used to run —wget --remote-name/-i,curl -K,busybox wget,pip3 download <url>,telnet,git archive --remote, and the rest refuseverify-networkedin attached and clustered spellings, and a command word built from a variable (E=rm; $E -rf ./target, including the braced, quoted,exported, loop, and wrapper spellings) refusesverify-indirection. Migrate: write the tool word plainly and rewrite the command to a plain fetch your recipe can run; every refusal keeps itstry:hint. Known residuals.verify.outis owner-only — the verify stream is written0600on Unix where it was0644, because a command that prints a credential leaves the value in it. Migrate: read it as the user that ranverify, or copy the bytes inside that same step; on Windows the file keeps the directory's inherited ACL. Enforced sandbox- Bundle paths are guarded — a bundle name that is not one path element (
a/b,..,CON,x:) refusesbad-flags, and a symlinked bundle root or component refuses before any write, including a link that points inside the repository. Migrate: one plain name per bundle, and a real directory where a link used to work. Bundle namespace - Provider is inferred from origin — with no
--providerflag, an unknown or conflicting origin now refusesprovider-unknowninstead of defaulting to GitHub. Migrate: pass--providerexplicitly, or keep a plain GitHub origin. Providers capabilitiessplits,truncatedwidens — the binding verdict becomesapproveandrequest_changeswithreview_verdictskept as the compatibility sum,unsupportednames each absent capability, anddata.truncatedwidens frombooleantoboolean | row[]. Migrate: branch on the two new flags, and narrowtruncatedbefore you read it as a boolean. Published schemaspack addsource allowlist — onlyhttps://host/pathwith no userinfo,ssh://user@host/path, anduser@host:pathare accepted; anything else refuses before a git call, and a lock row carrying one refuses atpack list,fetch,post, andrecord. Migrate: rewrite the source in one of the three forms. Lens packs- Benchmark cases name their commits — a real corpus case now requires
provenance.base_shaandprovenance.head_sha(40-hex), the two commits the reviewed diff spans, and a bundle stamps the same pair. Migrate: add both commits to every realcase.json; a case without them fails validation. Benchmark corpus - Rebuild indexes — indexes written by the previous scanner over-report callers, so rebuild them. Migrate:
verdict index build --at HEAD. Code index
Patch Changes
-
Safety and trust
- Tokens off the argv — Forgejo and Bitbucket reads hand the credential to curl on stdin or a 0600 file, so it is gone from
psand/proc/<pid>/cmdline. Providers - Secret guard —
postrefuses a secret-shaped value before the first host call, naming the shape and never the value, with--allow-secret-valuesas the explicit exception. Guarded submission - Mentions neutralized —
recordescapes@-mentions once, where the body is built, so recorded, posted, dry-run, and digested bytes cannot diverge. Mentions - Punctuated mentions —
(@alice),**@alice**, and_@alice_are escaped while real addresses stay byte-identical. Mentions - Strikethrough mentions —
~~@alice~~is escaped too, while a lone tilde in prose stays as written. Mentions - Learning evidence — cited paths and comment identifiers must stay inside quoted learning evidence, and unsafe ones refuse before a learning is proposed or accepted. Learnings
- Unicode line breaks — U+2028 and U+2029 now refuse a serve URL, refuse a learning path, and flatten in a rendered
pack testline. Serve - Store locking — every
.verdict/kb.jsonlwrite takes one advisory lock; a second writer refusesstore-busyinstead of dropping a row. Findings store - Enforced sandbox —
verify --sandboxruns the verify command under bubblewrap on Linux with no network, a private PID namespace, empty~/.ssh/~/.aws/~/.docker, and no silent fallback. Enforced sandbox - macOS sandbox —
verify --sandboxnow denies everybind()andconnect()on macOS through the built-insandbox-exec, loopback and unix sockets included, so filesystem isolation stays Linux-only. Enforced sandbox - Variable command words refuse —
E=rm; $E -rf ./targetand the braced, quoted,exported, loop, concatenated, and own-line spellings now refuseverify-indirection, including afterenv/timeout/xargswrappers and insidesh -c. Known residuals .verify.outis owner-only — the verify stream is written0600on Unix, like.run.out, because a command that prints a credential leaves the value in it. Enforced sandbox- Launch-record redaction — a secret-shaped
run --agentoperand no longer reaches the recordedoutput_tailor the 0644 child log. Whatrunasserts - setup-verdict hardening — the action passes its inputs through environment variables and verifies a cached archive before reuse, failing closed on a tampered one. The runner is your CI
- Install preflight —
skill installprobes exactly the directories it will write before the first byte, so a read-only target leaves nothing half-installed. Install routes
- Tokens off the argv — Forgejo and Bitbucket reads hand the credential to curl on stdin or a 0600 file, so it is gone from
-
Fixes
- Read-verb refusals —
report,next, andshowrefusals carry a shell-quotedre-enter:line naming the invocation you typed. Machine surface next --coveragewith no positionals — leads with oneerror:line naming what is missing, and a plainBasis:line is accepted without a warning. Machine surface- Unnamed cost models —
recordwarns once percost.jsonmodels[]entry with noname, and still accepts and prices the record. Cost control - Over-cap and missing-index refusals —
diff-too-largeandmissing-indexnow carry anext:line naming the exclusions that fit, and the exact build command. Over-cap diffs - Five small repairs — an agent that fails just before its budget reads
agent-failed, MCP refusals carry their warnings,index-cappednames the bytes it measured,index statuspoints at the query, and a crashed agent's partial review can no longer block on a FIFO. Machine surface - Windows watchdog — a command that failed to start and survived the cleanup kill now returns at once, naming both errors, instead of hanging forever. Bounded foreground waits
- Quoted upgrade hints —
verdict upgradequotes its project-localcdhint so a path with a space or quote pastes resolvably, and the uninstall sibling hint reads as a sentence. Machine surface - Release asset check — the release script's completeness check no longer reports a shipped asset as missing when the asset list is large. Release train
pr-waitinconsistent reads — a head that is not yet the same on both APIs is retried instead of failing the bounded wait, andinconsistent_observationskeeps the retry visible. Bounded foreground waits- Windows watchdog resume — a command that failed before starting under load retries the thread lookup instead of refusing
resume suspended verify process. Bounded foreground waits backtest minein a partial clone — a--filter=blob:noneclone refusespartial-clonewith the step that fetches the missing objects, instead of stalling. Backtest- Quiet local live skip — a local bundle's live recheck writes one
info:line to the human stream instead of a warning or a--jsonwarnings[]entry; failed remote reads still warn. Machine surface - Optional review schema version —
recordaccepts areview.jsonwithoutschema_versionand reads it as version 1, while a present value other than"1"still refuses. Review-input problems
- Read-verb refusals —
-
Docs
- Refusal code reference —
docs/troubleshooting.mdlists every refusal code with its cause and fix, and a drift test fails the build when one is missing. Refusal codes - Docs site structure — grouped sections with short sidebar labels, a Markdown twin for every page, and a glossary route from
verdict --help. docs/ - Defaults at a glance — a generated table names every default a bare run takes and the verb it shapes, so it cannot drift from the shipped binary. Defaults
- Refusal code reference —
-
Known limits
- The macOS sandbox covers the network half only —
verify --sandboxdeniesbind()andconnect()on macOS but has no filesystem namespace, so credential paths,/proc,/run, and a private/tmpare hidden only on Linux underbubblewrap; filesystem isolation on macOS needs a Linux VM or a container running that recipe. Enforced sandbox - The string gate is a tripwire, not a boundary —
sh "$V", a script named by a variable, still passes the pre-exec scan, which is whyverify --sandboxis the boundary and a spelling outside the gate's reach is a gate gap to arm rather than a vulnerability. Known residuals - A backslash before a CRLF line ending is not joined as a continuation — the gate refuses that shape rather than passing it, so a CRLF continuation fails safe. Known residuals
- The macOS sandbox covers the network half only —