Skip to content

chore: version packages - #58

Merged
SutuSebastian merged 1 commit into
mainfrom
changeset-release/main
Nov 16, 2026
Merged

SutuSebastian merged 1 commit into
mainfrom
changeset-release/main

Conversation

@github-actions

@github-actions github-actions Bot commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@stainless-code/verdict@0.7.0

Minor Changes

  • cc6bfbf: What's new in 0.7.0

    Verdict 0.7.0 is a breadth release. It adds a local code index, a read-only serve surface 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.

  • cc6bfbf: 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 waits
    • ingest — import SARIF 2.1.0 results into a bundle's findings store, sharing one dedupe and suppression path with model reviews. Static-analysis ingestion
    • threads resolve — close the review threads whose findings your store marks addressed, one receipt per run, with --dry-run and typed refusals for threads verdict did not author. Guarded submission
    • risk — 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 queue
    • report --html — export the six dashboard views as a static site, published through staging and backup siblings with rollback and symlink refusals. Static dashboard export
    • run --agent — run your own agent CLI against an existing bundle, assert exactly one fresh review.json, then hand off to the unchanged record/post validation. Your agent, launched
    • panel — plan independent verifier dispatches locally; keep-biased results enrich findings and summaries without voting or dropping anything. Panel verification
    • Recipes — .verdict/recipes.json declares 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 routes
    • post --summary comment — maintain one separate PR-level summary comment, updated in place across all three publish hosts; the default off keeps output byte-identical. Guarded submission
    • fix --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 in data.batch, and a fix.patch written beside the ledger. Fix loops
    • fix --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 loops
    • handoff — one markdown projection of the open inline findings fix would take, selecting through the same selector, with --json, --format md, and repeatable --paths; read-only, with no ledger and no host call. Handoff
    • run --print --stage <n> — render one recipe stage at a time, so CI can dispatch the four stages under different models. Recipes
    • backtest reverts — record the reverts git already knows about into a sorted, byte-stable .verdict/reverts.jsonl, with .verdict/reverts.md as the fenced reading surface. Reverts
    • backtest mine — mine SZZ candidates from a local clone with zero model calls and zero network, writing a timestamp-free mine.json. Backtest
    • backtest replay — replay mined candidates against your own agent in a throwaway worktree per candidate, bounded by review count, projected budget, and timeout. Replay
    • backtest 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. Score
    • learn --incident — attach one revert or incident to a recorded bundle in a sorted, idempotent .verdict/incidents.jsonl ledger. Knowledge base
  • cc6bfbf: 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, and index owners answer 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-for walks the reverse import graph up to --depth hops and labels tests it reached through imports. Code index
    • fetch --index — hand an index you already built to fetch or fetch-local and the bundle gains an impact artifact and a fenced ## impact section. Impact hand-off
    • Missing-index hints — index blast-radius and fetch now 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
  • cc6bfbf: Review quality and provenance

    • Explicit context — fetch --context and fetch-local --context attach 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.json gains an optional coverage[] checked against the bundle's own hunks, with --require-coverage to refuse a partial claim and one body line on a partial attestation. Diff coverage
    • ranges: null is malformed — a coverage claim whose ranges is 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-excludes on fetch, both recorded in the exclusion receipt. Exclusions
    • Suppression receipts — excluded diff bytes and withheld findings are receipted, with post --min-severity and report --include-withheld to 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 --status gains recalled_in, cited_in, exposure, and acceptance per rule, reporting unmeasured recall as null rather than zero. Knowledge base
    • Risk scoring — an optional risk claim in review.json plus report --risk and report --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.json records, 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 rearmed and previous_head recorded. 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 — fetch writes an advisory author.json ({agent, model, confidence, evidence[]}) from five harness signals, suggests a reviewer from another model family, and record.json carries the author/reviewer pair that trend tallies. Bundle namespace
    • Revert memory — a changed path that matches the tracked reverts ledger gives fetch a bounded, fenced reverts.txt, and the review prompt treats a revert as evidence the change broke something once. Reverts
    • Validation pass — a trusted prompts/validate.md pass writes validation.json per sig, rigor_policy floors each severity tier, and post --require-rigor narrows it per invocation: advisory by default, a receipted withhold in gate mode. Rigor floors
    • Published key names — prompts/review.md tags the evidence rung as rigor and the in-bounds flag as in-bounds, and an archtest holds every published key to review.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 — record leaves a review.json body with no leading quote block byte-identical and reports no basis-replaced warning, because there is no confidence lead to insert under. Review-input problems
    • Array-valued coverage ranges — next --coverage refuses a coverage[].ranges array on plain and chunked bundles with the same bad-review-input problem record reports, instead of reading it as a silent 0% attestation. Diff coverage
    • Passing gate next step — a passing report --check <tier> names merge — gate passes in the envelope's next and human footer, leaving address open threads for a failing gate. Machine surface
    • More generated paths — *.gen.* and zz_generated.* now count as generated in the shared role heuristic, so next --coverage and the risk floor exempt them like the other generated patterns. Diff coverage
  • cc6bfbf: Agents and MCP

    • Bounded repair loop — run --rounds N re-prompts an agent with the shape checker's fix list, stops on the first valid round, and refuses repair-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 config writes snippets for six more hosts and carries --read-only or --operator into the argv it prints. Host config
    • Declared mutation classes — every MCP tool states its local mutation class, and a call hidden by --read-only returns the typed read-only-tool refusal. 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
  • cc6bfbf: Hosts and CI

    • Forgejo publish — post submits 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-sha writes the same incremental.diff contract the GitHub compare path writes, and hosts without a compare adapter say so in a note. Providers
    • GitLab readiness and calibration — ready advises on merge requests and plain calibrate joins live GitLab threads through glab, 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
  • cc6bfbf: 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 --root finds 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/cost adds repository, ISO-week, and model buckets and /v1/precision adds category and tier counts, keeping unmeasured prices null and marking corrupt aggregates partial. Serve
    • Browser UI — serve opens an embedded UI with inbox, review, threads, cost, precision, and gate views, dark mode, live refresh, and static data.json reading. Serve
    • Live invalidation — bundles invalidate as they change, with --interval, --sweep, and --no-watch to 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
  • cc6bfbf: Lens packs

    • Built-in lenses as data — perf, api-contract, and test now ship as a pack resolved into every bundle as lenses.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 add and 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 list reports 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-stale with the path to delete. Lens packs
  • cc6bfbf: 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, or false-positive, and bench/runner -render recomputes 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-usd stops 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.diff is missing, unreadable, or prose is refused instead of graded leniently. Benchmark corpus
  • cc6bfbf: 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-radius truncation shape — the integer truncated count and depth_truncated are replaced by the shared boolean receipt (truncated, truncation_reason, returned, total, next_offset), and a complete result omits the receipt instead of carrying truncated: false. Migrate: read truncation_reason and returned; treat an absent receipt as complete. Code index
    • gate checks coverage by default — the new coverage dimension fails when the gated HEAD has no receipt, a receipt for another HEAD, no coverage[] attestation, or under 100% attested. Migrate: --coverage warn, or the SDK Coverage field, keeps the old passing behavior while reviews attest nothing. Gate
    • gate --live reads 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 auto warns and gates on the local dimensions only. Gate
    • Unknown gate tiers refuse — an unrecognized tier is bad-flags on the MCP report tool and in Report.Check/Gate.MaxSeverity, not only on the CLI. Migrate: send a tier from the published set, or omit the field. Gate
    • verify child 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-env or --env-allowlist; dropped names appear in warnings[]. Child environment
    • verify refuses command shapes it used to run — wget --remote-name/-i, curl -K, busybox wget, pip3 download <url>, telnet, git archive --remote, and the rest refuse verify-networked in 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) refuses verify-indirection. Migrate: write the tool word plainly and rewrite the command to a plain fetch your recipe can run; every refusal keeps its try: hint. Known residuals
    • .verify.out is owner-only — the verify stream is written 0600 on Unix where it was 0644, because a command that prints a credential leaves the value in it. Migrate: read it as the user that ran verify, 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:) refuses bad-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 --provider flag, an unknown or conflicting origin now refuses provider-unknown instead of defaulting to GitHub. Migrate: pass --provider explicitly, or keep a plain GitHub origin. Providers
    • capabilities splits, truncated widens — the binding verdict becomes approve and request_changes with review_verdicts kept as the compatibility sum, unsupported names each absent capability, and data.truncated widens from boolean to boolean | row[]. Migrate: branch on the two new flags, and narrow truncated before you read it as a boolean. Published schemas
    • pack add source allowlist — only https://host/path with no userinfo, ssh://user@host/path, and user@host:path are accepted; anything else refuses before a git call, and a lock row carrying one refuses at pack list, fetch, post, and record. 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_sha and provenance.head_sha (40-hex), the two commits the reviewed diff spans, and a bundle stamps the same pair. Migrate: add both commits to every real case.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

  • cc6bfbf: 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 ps and /proc/<pid>/cmdline. Providers
    • Secret guard — post refuses a secret-shaped value before the first host call, naming the shape and never the value, with --allow-secret-values as the explicit exception. Guarded submission
    • Mentions neutralized — record escapes @-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 test line. Serve
    • Store locking — every .verdict/kb.jsonl write takes one advisory lock; a second writer refuses store-busy instead of dropping a row. Findings store
    • Enforced sandbox — verify --sandbox runs 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 --sandbox now denies every bind() and connect() on macOS through the built-in sandbox-exec, loopback and unix sockets included, so filesystem isolation stays Linux-only. Enforced sandbox
    • Variable command words refuse — E=rm; $E -rf ./target and the braced, quoted, exported, loop, concatenated, and own-line spellings now refuse verify-indirection, including after env/timeout/xargs wrappers and inside sh -c. Known residuals
    • .verify.out is owner-only — the verify stream is written 0600 on 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 --agent operand no longer reaches the recorded output_tail or the 0644 child log. What run asserts
    • 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 install probes exactly the directories it will write before the first byte, so a read-only target leaves nothing half-installed. Install routes
  • cc6bfbf: Fixes

    • Read-verb refusals — report, next, and show refusals carry a shell-quoted re-enter: line naming the invocation you typed. Machine surface
    • next --coverage with no positionals — leads with one error: line naming what is missing, and a plain Basis: line is accepted without a warning. Machine surface
    • Unnamed cost models — record warns once per cost.json models[] entry with no name, and still accepts and prices the record. Cost control
    • Over-cap and missing-index refusals — diff-too-large and missing-index now carry a next: 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-capped names the bytes it measured, index status points 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 upgrade quotes its project-local cd hint 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-wait inconsistent reads — a head that is not yet the same on both APIs is retried instead of failing the bounded wait, and inconsistent_observations keeps 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 mine in a partial clone — a --filter=blob:none clone refuses partial-clone with 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 --json warnings[] entry; failed remote reads still warn. Machine surface
    • Optional review schema version — record accepts a review.json without schema_version and reads it as version 1, while a present value other than "1" still refuses. Review-input problems
  • cc6bfbf: Docs

    • Refusal code reference — docs/troubleshooting.md lists 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
  • cc6bfbf: Known limits

    • The macOS sandbox covers the network half only — verify --sandbox denies bind() and connect() on macOS but has no filesystem namespace, so credential paths, /proc, /run, and a private /tmp are hidden only on Linux under bubblewrap; 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 why verify --sandbox is 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

@coderabbitai

coderabbitai Bot commented Nov 9, 2026 •

Copy link
Copy Markdown

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

Run configuration

  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: 5179dd27-aa34-49c6-bf3d-d11b828f06b7

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • Trigger review
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

@github-actions
github-actions Bot force-pushed the changeset-release/main branch 28 times, most recently from 1cb31ab to 84c9ae0 Compare November 12, 2026 10:27
@github-actions
github-actions Bot force-pushed the changeset-release/main branch 28 times, most recently from 1cf3bdc to 306c1ce Compare November 14, 2026 13:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant