Repository navigation
chore: version packages - #173
Open
github-actions[bot] wants to merge 1 commit into
Open
github-actions[bot] wants to merge 1 commit into
github-actions[bot] wants to merge 1 commit into
Conversation
|
Important Review skippedBot user detected. Ask an organization administrator to verify that this bot has an active assigned CodeRabbit review seat. See the seat-assignment guide. After the seat is confirmed, request a review again.
|
github-actions
Bot
force-pushed
the
changeset-release/main
branch
21 times, most recently
from
November 19, 2026 22:12
fd84953 to
92192c0
Compare
github-actions
Bot
force-pushed
the
changeset-release/main
branch
from
November 19, 2026 23:48
92192c0 to
0463198
Compare
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
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@0.8.0
Minor Changes
ecc6b89:
drainis a hidden CLI verb that refreshes threads, plans the review threads waiting on a verdict reply, and pins the plan with a digest.drain --agent <name>launches one agent through the same launcherrunuses and validates itsdrain.replies.jsonagainst the plan. No host writes or reply posting ship in this slice.6923d0e:
drain --postanswers the waiting review threads from a validateddrain.replies.json. GitHub and GitLab post one marked reply per draft and appendreplies.jsonunder a lock after each write;withdrawalso resolves its thread. Secret-shaped bodies refuse before any write, stale threads are skipped, and a failed write recordsfailed[]and refusesdrain-incompleteso the same command can finish the remaining replies.9c68ea2:
run --agentnow supports fourteen verified official agent CLIs, includingagy,amp,copilot,crush,droid,gemini,goose,kiro,pi, andqwen, with prompt flags placed immediately before the directive for agents that take it as a flag value. Operators can define more agents under the newagentskey of their own settings file; the shape is prefs-only, closed, and refuses name collisions and invalid binaries asbad-settings.bc6f739: A bad
--contextpath now refusescontext-invalidwith a path-specific hint instead offetch-failed.Breaking changes
A bad
--contextinput — missing, symlinked, too deep or too large, or the wrong type — now refusescontext-invalidinstead offetch-failed. Scripts that matchedfetch-failedfor these cases must matchcontext-invalid. Bundle-write failures still refusefetch-failed. The refusal names the path as typed and, for a relative path, its absolute resolution, and its hint tells you to check each--contextpath.98d8bec:
run --agentnow offers 32 more cited headless agents, including Cline and Junie after CLI-reference checks. Rows whose binary shares a name with another tool run a harmless identity probe before launch; a non-zero exit, a five-second timeout, or a missing match refuses the newagent-mismatchcode before any agent starts, andlauncher.jsonrecordsprobe: {args, ok}for a passing check. Operator agents can setprobe_argsandprobe_matchunder settings.jsonagents, withprobe_matchrequired whenprobe_argsis present.770b32f:
drain --post --dry-runruns every post gate and live read, then writesdrain.dryrun.jsonwith the exact marked replies and typed skips a real post would produce. It makes zero host writes and leavesreplies.json,drain.json, andthreads.jsonbyte-identical;--agent --post --dry-runplans, drafts, then previews. The flag requires--postand secret-shaped bodies still refuse.5712cec: Breaking changes
The installed skill is now a small Bootstrap skill: it only says when to use Verdict and how to reach the binary. The binary serves the full Procedure, so the instructions always match the version that checks your work.
Reinstall the skill on every machine. For harness installs, for each installed agent, run
verdict skill status --agent <x>and delete the oldverdictskill directory at the path printed after->; then runverdict skill install --agent <x>.verdict skill install --detectis fine instead of--agent.verdict skill uninstallremoves only the new Bootstrap skill files, so it would leave the old Procedure files beside the stub.For registry installs, delete the old
verdictskill directory, then re-runnpx skills add stainless-code/verdict.Copies went stale between releases: the Procedure lived inside the installed skill, so an older copy could hand an agent instructions that disagreed with the binary grading its work. Keeping it in the binary removes that skew.
66342c5: Every fetched bundle now carries
procedure.md, the review Procedure assembled by the binary from its embedded payload. Drain plans carrydrain.procedure.mdfor reply drafting. Therun, stage, and drain directives name those files as authoritative, so an installed skill that is older can no longer displace them. The fetchnext:hints and--show-promptsoutput point at the bundle Procedure, and provenance receipts record its sha256 as a bundle input.PromptsHashnow reads the embedded payload, so an installed user gets the real prompt-set hash without aprocedure/checkout.85b2e97: MCP hosts can now load the Procedure straight from the binary.
prompts/listandprompts/getserve five task prompts (address-reviews,fix-ci,incremental-review,review,validate), andresources/listandresources/readserve every payload file atverdict://skill/<path>, byte-identical toverdict skill show. Prompts and resources are read-only and available in every server mode, including--read-onlyand--operator.initializeadvertises both capabilities and points hosts at thereviewprompt orverdict://skill/PROCEDURE.md.31efb1b: Breaking changes
Published JSON schemas moved from
skills/verdict/schemas/to/schemas/; every$idchanged accordingly.9de3880: Verdict now writes an append-only per-bundle invocation journal under
<state>/verdict/journal/<repo-key>/<pr>.jsonl($XDG_STATE_HOME, falling back to~/.local/stateon Unix or%LOCALAPPDATA%on Windows) for every bundle-writing CLI and MCP call, and adds the read-onlyverdict journal <owner/repo> <pr|local-name> [--json]verb (plus its MCP tool, Go facade method, TypeScript client method, shell completion, andschemas/journal.schema.json). Each call records pairedstartandendlines carrying the canonicalreporoot, redactedargv, content-digestbundle.before/bundle.aftertamper-evidence hashes, exactwrote/removedpaths with SHA-256 digests, exit code, refusal code, warnings count, andnexthint, bounded at 1 MiB with atomic whole-line truncation and non-fatal append warnings. The foldedverdict journalsummary pairs each call under a uniquestepkey, reports the size-cap notice undertruncation, and publishesbundle.unobservedwith no file diff when a content sweep fails instead of claiming every file changed. Because the journal lives outside the repository and.verdict/, dry runs, refusals, and bundle receipts stay completely untouched.938bc6d:
review.jsonis now a findings-only handoff: each finding carries a path, strict integer line, severity, claim, quoted code, and finder confidence; optional spec checks and receipts reserve the evidence S4 verifies. Verdict renders the body and derives approval from the stored findings, with a documented default threshold and the existing block-clause policy. The document keeps"schema_version": "1"— this is a breaking change inside that version, shipped with no shim.Breaking changes
review.jsonmust carryfindings[]and may no longer carryapproved,body,comments[],sections[]or a top-levelconfidence; the document is closed, so any of those keys refuses as an unknown field and there is no repair hint for the old shape. To rewrite a stored v1 file, move eachcomments[].bodyintofindings[].claim, addquote,severityandconfidenceto every finding, convertlineto an integer, move the oldbodyprose tosummary(at most 2000 characters), dropapproved,sections[]and top-levelconfidence, and keep the claim arrays (resolved,triage,coverage,chunks,suppressed,risk,security_summary) unchanged.One problem code that policed the old contract is gone:
basis-replaced(the agent writes nobody, so there is no line for the CLI to repair). The closed document refuses the removed keys
as
unknown-field, and the review prompt dropsin-boundswith the rest of the body grammar.The dry run's
data.checknow reportsfindingsandanchors_moved_to_bodyin place of thecommentsandsectionscounters that described the old body grammar.Approval is no longer an agent claim. With no repository policy, any open, visible finding whose quote standing is not
unverifiedatimportantor above withholds approval, as does any violatedspec_checks[]entry;block_clausesandreport --checkrefine that default when configured.recorddecides once and stores the decision inrecord.jsonasapproval, beside thereview_sha256it decided on;postpublishes that stored decision, renders its body from the same decision, and refusesrecord-stalewhen the review, diff, or head moved after the record, rather than deciding again over a review nobody recorded.Do not release between this slice and S4. S1 stores
spec_checks[]andreceipts[]but does not verify quotes or require receipt coverage; S4 adds the insufficient-review refusal and closes the gap where a findings-free review would otherwise derive approval. The release hold is tracked in Linear.1dfcc29:
backtest fixturesgrades converted fixture recordings against hand-known defects in a new fixture corpus. It publishes recall, bonus-union, false-green, hallucination (with explicitunverifiablecounts) andreview-insufficientmetrics, reproduces seven hand scores as calibration, and writes one report under.verdict/backtest/fixtures/report.json. The mode never launches an agent or calls a host.4ab8691: The review handoff is now lean and layered.
<bundle>/procedure.mdcarries a pushed core — role, two method lines, the findings-only contract, and a capability map — capped at 3 KB by an archtest. Sections for chunked diffs, prior threads, knowledge-base rules, lenses, and handed context are appended only when the bundle carries that fact. A sliced review checklist and the remaining reference material are written to<bundle>/refs/and pulled on demand; the full operator workflow and body-format rules stay withverdict skill show.--show-promptsprints the assembled handoff and one line naming the refs, never their bodies. Drain and the MCP task prompts use the same core where it applies.The directive now says to write
review.jsonfirst with empty findings and update it as the review proceeds; after a clean exit only (no crash, timeout or non-zero exit),runsalvages the last fencedjsonblock on stdout that decodes to an object with afindingsarray, addingschema_version"1" if missing, recordssalvaged: trueinlauncher.json, and warns. It replaces only an absent artifact or the exact empty placeholder; a salvaged or written review that fails the shared checker is refusedbad-review-inputwith its problems, which feed the next round under--rounds, while only stdout with no such block keepsno-review-artifact.A finished review with no findings must carry a non-empty
summary; the empty placeholder alone is never a clean review, andrunchecks every artifact with the same checkerrecorduses before reporting success.prompts_hashand the provenance receipt now cover every file that feeds the assembled handoff and its refs, so an edit to any of those bytes is visible in the bundle's recorded prompt set.8b159e9:
fetchandfetch-localnow accept a repeatable--spec <file>flag to handspecification documents into the bundle separately from plain
--contextreference material.
--context(size, type, symlink, and depth caps).context/directory and fencedcontext.txt, whilecontext.jsonrecords"role": "spec"or"role": "context"per entry in place andtoc.txt's## contextsectionmarks spec rows with
[spec].the cap refuses
spec-too-largewith no bundle files written.5e1ab56: Verdict now checks the evidence a review states before it decides anything: each finding's
quoteis matched against the pinned diff at itspath:line, a review whose findings are all below the approval tier must receipt every non-exempt changed hunk, and--specsources get a per-item checklist. The verdict is stored as an evidence receipt inrecord.jsonand read by every verb:record --dry-runnames each gap,runrepairs them across its rounds and refusesreview-insufficientwhen they survive,report --checkandgatefail an always-onevidencedimension, andpostpublishes the findings with an insufficient lead and never approves.Breaking changes
A review that states no evidence is now insufficient rather than approved.
review.jsonmust quote the exact line at each finding'spath:line(start_lineplus one quoted line per range line for a multi-line quote), and a review with no gating finding must carryreceipts[]— one{"hunk":"<the hunks.tsv row>","quote":"<a line that hunk carries>"}per non-exempt hunk;verdict record --dry-runnames every gap and now ends a clean review with a warning instead of silence.run --roundsdefaults to 2 instead of 1, so an invalid review is re-prompted once unless you pass--rounds 1;runalso refusesreview-insufficientwhere it used to hand off, and a plainrunenvelope now carriesdata.repairon that default.report --checkandgatealways grade a seventh dimension,evidence, so a bundle that never recorded a receipt for its head readsunknownand fails instead of passing on the other six.fetch --spec, and thedefaults.spec_coverage_mindefault of 80 it feeds, become part of the review contract: an enumerated spec item with no check, or with test-output evidence (logs, goldens, snapshots,testdata; test source counts as code), withholds approval. There is no compatibility shim for any of this — v0.7 is not yet at 1.0, and the previous behaviour approved reviews nothing had checked.abb3389:
ready --filternow gates the review on the closed trigger predicates in the trusted base configuration: matching PRs run and record the decision, non-matching PRs refusefiltered-outbefore any host call, and unusable policy or facts refusetrigger-failed.Fetch now requests and records PR labels in
PR.jsonfor every host, so thelabelstrigger predicate can read them from the bundle.Patch Changes
--specchecklist cap is now 150 items. The 60-item ceiling refused real specification documents mid-PR;fetchandfetch-localstill refusespec-too-largebefore writing any bundle file once the combined--specsources enumerate more than 150 items, and the refusal message and hint name the new limit.report --checkandgatenow fail on a violated spec check. The new always-onspecdimension reads the spec checks the stored record graded: aviolatedone fails the gate withviolated spec check: <item>and a tally of every violated item, whilemet,n/a, and a review with no spec checks pass. Before this, a bundle whose only defect was a violated spec check passed the gate while the derived approval withheld withviolated spec check, so the gate certified a review the decision refused.--contextspec.build,ci, andgradeas task prompts, one per task prompt theProcedure's mode router names, so
prompts/listoffers eight instead of five. Each body is thesame assembly the other task prompts use —
PROCEDURE.md,prompts/system.md, and that task'sown prompt — and an unknown name still refuses
-32602and lists the valid tasks.