Repository navigation
Releases
Releases · NCodeGroup/code-review
v1.4.2
Immutable
release. Only release title and notes can be modified.
Added
- The repository is licensed under the Apache License 2.0, as the other NCodeGroup repositories are:
LICENSE.txtholds the license andCOPYRIGHT.txtthe copyright notice. .git-blame-ignore-revslists the commits of the one-time Biome reformat, so GitHub's blame view andgit blame(withgit config blame.ignoreRevsFile .git-blame-ignore-revs) show who last changed a line instead of the reformat (#212, ADR-1010).
Changed
GOVERNANCE.md,ROADMAP.md,docs/architecture.mdanddocs/assurance-case.mddescribe code-review itself: its maintainer, its plans for the next twelve months, its components and design rules, and its security requirements, threat model and the weaknesses it counters (#206).README.mdshows the OpenSSF Best Practices badge (project 16752, passing) and says how to verify a release withgh release verify, anddocs/best-practices-badge.mdrecords where the silver level stands.- CI runs
npm run lint, which fails on TypeScript that Biome's formatter or linter doesn't accept asbiome.jsonsets them, andnpm run formatapplies them; the code is reformatted once to match, andAGENTS.md,CONTRIBUTING.mdanddocs/best-practices-badge.mdsay so (#210, ADR-1010). - The repository follows project-template v0.99.0 (from v0.96.0): it gains the
/best-practicescommand, which checks the OpenSSF Best Practices answers against the repository, anddocs/best-practices-criteria.md, which maps every criterion to its evidence;tag-release.ymlruns CI on a release's commit before it tags it and signs the tag with gitsign, whichSECURITY.mdandREADME.mdsay how to check;CONTRIBUTING.mdhas a Review Requirements section,SECURITY.mdpromises an acknowledgement within 14 days, anddocs/assurance-case.mdhas a Security Reviews section.
Fixed
- The action loads again:
action.yml, from 1.4.0 on, wrote theapproveinput's description as a plain YAML scalar holding:, which GitHub's runner can't parse, so every job that used the action failed in "Set up job". A test parsesaction.ymlas YAML (#221).
v1.4.1
Immutable
release. Only release title and notes can be modified.
Changed
- The repository follows project-template v0.96.0 (from v0.86.1): it gains
CODE_OF_CONDUCT.md,GOVERNANCE.md,ROADMAP.md,docs/architecture.md,docs/assurance-case.mdanddocs/best-practices-badge.md, the documents the OpenSSF Best Practices silver level asks for, an OpenSSF Scorecard workflow (.github/workflows/scorecard.yml), read-only top-level permissions in.github/workflows/copilot-setup-steps.yml, and the template's newer review commands, which judge every NCode Review finding whatever its severity;AGENTS.md,CONTRIBUTING.md,README.mdanddocs/README.mddescribe them.
v1.4.0
Immutable
release. Only release title and notes can be modified.
Added
- The
approveinput (off by default) submits a review whose verdict is Approve, with the head unchanged and no earlier should-fix thread open, as an approving review, so a ruleset can require one approval and Scorecard's Code-Review check counts it. A review with only nits approves too, when every reviewer ran in full and verification lowered or refuted nothing. Any other review stays a comment and dismisses the bot's earlier approval, and the summary says "Approved" or "Not approved" with the reason (#203, ADR-1009).
v1.3.1
Immutable
release. Only release title and notes can be modified.
Fixed
- A review whose check run would list more than GitHub's 65,535 characters publishes: the check run's text shows as many findings, already raised, tracked, upstream, refuted and dropped entries as fit and says how many more each section has, or that all of a section's entries are in the run's trace when none of them fits, and its summary is cut to fit, so the review's threads and body are posted instead of the whole publish failing with HTTP 422 (#201).
v1.3.0
Immutable
release. Only release title and notes can be modified.
Added
- Synthesis merges one problem found in several files, such as one rule broken the same way in three of them, into one finding: its one thread, the review's body and the check run list the other places under "Also at", so the author answers it once. A reviewer's finding in another file counts as merged only when the finding names its place, within three lines; any other is posted on its own (#193, ADR-1008).
Fixed
- A re-review whose changed hunks are more than one part of a first review holds (160,000 characters) is a full review, in parts and under the run's cost ceiling, and the check run names the files it couldn't fit, instead of one reviewer call that stops at its $1 budget and reviews nothing (#194, ADR-1007).
- A re-review that raises an earlier finding again in new words lists it under "Already raised", with a link to the earlier thread, instead of posting it again: its reviewer names the earlier finding it repeats, and the re-review leaves it out of its verdict (#192, ADR-1008).
- A re-review whose verdict is Comment, because an earlier finding is still waiting or the re-review was incomplete, keeps it when the publish step lists a finding that repeats a dismissed thread as already raised, instead of turning to Approve.
v1.2.0
Immutable
release. Only release title and notes can be modified.
Added
- A
/reviewcomment from someone with write access that starts with/reviewand then a space or the end of the comment but isn't a command, such as/review please, gets a 😕 reaction within seconds, so its author can tell that no review started; the run's log lists the commands. The reaction uses a token of its own that holds only the pull requests permission, minted before setup-node and revoked before the review step, and never in the review step's environment (#174, ADR-1004). npm run trainlearns from the review bot's own findings that the maintainer answered, one source per pull request with the replies and whether each finding was fixed or declined: a decline with a lasting reason teaches reviewers what not to report, and a fix the rule it enforced. Only the maintainer's replies count (ADR-1005).
Changed
- The repository follows project-template v0.68.4 (from v0.59.0): the
Commit attributioncheck and thecommit-msghook catch attribution from any AI agent or model, and the check's remedy for an attribution line in a commit message runs the hook, asCONTRIBUTING.mdnow says; the workflow commands remove an agent footer a session appends to what they post, except on a/reviewcomment, which/run-issuesand/run-feedbackleave as posted because NCode Review reads the command from its first line; the commands that callghuse its REST forms in a session that refuses GitHub's GraphQL API, such as a Claude Code cloud session;/address-feedbackand/run-feedbackalso run in Copilot Chat and Copilot CLI and finish a run that stopped, also one the other tool started;.template/Update-Project.ps1stages only the files an update changes; andAGENTS.md's Landing section setspre-release: tracked, the template's new key, as it is for a project with releases. - The repository follows project-template v0.75.1 (from v0.68.4): Claude Code sessions run Bash and PowerShell commands without asking, except
git pushand a command that mentions.envorsecrets, and a Claude Code cloud session installspwshwhen its container lacks it;/ship-plancan mark follow-ups to act on next, which/finish-planthen names as/plan-issueruns; the Copilot setup workflow runs again when.node-versionchanges; and.template/Initialize-Nvs.ps1sets nvs up once per PC, asCONTRIBUTING.mdsays. The template'stag-release.ymlpublishes nothing here, becausereleasestaysnone: releases stay manual throughrelease.yml, theRelease (manual)workflow in the Actions tab. - The repository follows project-template v0.83.0 (from v0.75.1): nothing checks commits or pull requests for AI agent attribution, so the
Commit attributionCI job,.template/Test-CommitAttribution.ps1and the.githooks/commit-msghook are removed, andCONTRIBUTING.mdasks contributors to avoid it (the template's ADR-0018); CI doesn't run again when a pull request's title or description is edited; a Claude Code cloud session runs each.claude/hooks/ensure-*.shat start, which installspwshand the.node-versionNode.js;/run-issuesand/ship-plancan fix a plan's follow-ups in its pull request before merge; and plans can land through a branch a session's host made for it. The prelude is the template's.template/Use-Node.ps1, which accepts the.node-versionNode.js already onPATHwhere nvs isn't installed, andtools/Use-Node.ps1is removed (ADR-1003). - The repository follows project-template v0.84.1 (from v0.83.0):
.template/Update-Project.ps1stages every change outside the conflicting files, deletions included, when it stops on a conflict, and leaves a catch-all text rule such as* text eol=lfout of the.gitattributeslines an adopted project kept;/plan-template-updateleaves a deletion an older target's script left unstaged unmarked, so its commit note stages it (#181). - The repository follows project-template v0.84.3 (from v0.84.1): the workflow commands record a Claude Code cloud session's push branch without remarking on it, and
/ship-plannames the branch it pushes to in the summary it shows before it pushes; Claude Code's planning commands enable.githooksand record the session's push branch from the main checkout before they enter the plan's worktree, and/run-issuesand/address-feedbackcheck the hooks before they create theirs; and theaskrules in.claude/settings.jsonmatch a command that mentions the rootsecrets/folder, not any text containingsecrets, so a source folder namedSecretsno longer asks for approval. - The repository follows project-template v0.86.1 (from v0.84.3): Copilot Chat, Copilot CLI and the VS Code agents keep each plan in its own git worktree under
.claude/worktrees/, as Claude Code does, asCONTRIBUTING.mdsays;/run-issuesand/address-feedbackin Copilot work in worktrees of their own and never switch the main checkout, and/run-issuesin every tool runs the next entry ahead in its own worktree while a draft pull request waits, then lands the entries one at a time, in queue order;/analyze-architectureand/analyze-decisionwork in a worktree of their own, whereANALYSIS.mdstays untracked and a decision record lands with the plan made there, asdocs/decisions/README.mdsays, and/continue-planlists and/finish-planremoves such worktrees; and/finish-planfast-forwards the default branch without switching a checkout that holds other work. - The reviewers settle in the checkout what they would otherwise ask, and don't report a production rule in test code or a C# rule in a script, a rule a line broke before a change that only renames or renumbers on it, compatibility or migration notes in a repository whose conventions say it is pre-release, another repository's process rules, duplication between places that can't share code, a parameter the conventions require as unused, or a date against today's. The rule check reports only what the change writes, synthesis merges one problem raised at several places in one file into one finding,
testsnames the search behind a missing test, anddata-modelandapi-compatexpect no earlier data or callers where nothing is released (ADR-1005). - The learned guidance is regenerated from NCodeGroup/identity, identity-platform and project-template: their decision records, instruction files and review prompts, and the bot's findings the maintainer answered on their pull requests. It holds 1,234 rules, 157 of them on what not to report; production and test rules name their scope; and the rules of decision records that no longer exist are gone (#190).
- A release is made by merging: a plan whose change reaches consumers carries a
Release:line, its last step dates the changelog section, and merging its pull request tags and publishes the release, so nobody runsrelease.yml, which stays as the fallback. This repository's own pin of NCode Review is set to follow a release within a weekday: itsgithub-actionsDependabot entry runs daily and its cooldown excludes the action, which Dependabot's documentation doesn't say applies togithub-actions, so without it the bump waits the 3 days (#196, ADR-1006).
Fixed
- A training run deletes the rules it extracted from a source the repository no longer has, such as a renumbered decision record, so they no longer reach the guidance;
--repofinds the repository's extraction folder ignoring case; and a rule learned from an instruction file for production or test code says which, so reviewers no longer apply production rules to test projects (ADR-1005). - Training reads a repository's pull request comments one page at a time, by number, so it runs behind a proxy that refuses GitHub's next-page links, such as a Claude Code cloud session's.
v1.1.2
Immutable
release. Only release title and notes can be modified.
Fixed
- A
/reviewcomment followed by more lines, such as the---and_Generated by …_footer an agent adds, reviews the pull request instead of cancelling the review in progress and then doing nothing: a command is the comment's first line, in any case, and/review explainand/review teachdrop everything from a---line on, so a footer never reaches a question or a taught rule. The caller's concurrency group puts only a comment whose first line is exactly/reviewin the pull request's group, so/review pleaseor any other comment the engine won't act on can't cancel a review; copy thegroup:line from the README's caller into your workflow (#172, ADR-1001).
v1.1.1
Immutable
release. Only release title and notes can be modified.
Changed
- The repository follows project-template v0.59.0 (from v0.53.0): agents get the
/run-feedbackcommand, which runs/address-feedbackround after round until a pull request merges or closes;/address-feedback,/run-issuesand/run-feedbackreplyAutopilot:to every unresolved review thread they read, as the template's new ADR-0014 records; and/run-issuescomments/reviewafter it pushes fixes, so NCode Review reviews them (#166).
Fixed
- A closed pull request's outcomes count a writer's
Autopilot: nothing to changereply, which project-template v0.57.0's commands post on a finding they judge needs no change, as dismissed, asAutopilot: not changedis, instead of acknowledged, so the monthly report's dismissal counts stay right (#168). - A fetch of the pull request's head that reaches its time limit no longer leaves git's transport child,
git-remote-http, waiting on a remote that never answers: every fetch with a limit sets git's low-speed limit to the same time, so the child gives up on its own about when git is killed. The cleanup after a failed fetch retries a held folder for up to 30 s instead of about 2 s, and throws any other error at once, so on Windows, where that child holds the folder until it exits, the folder is removed and the next attempt starts clean (#165).
v1.1.0
Immutable
release. Only release title and notes can be modified.
Changed
- When the review, publish or finish step stops on an unexpected error, the run's log lists the error's stack frames, one line each, below its
::error::line, so a crash can be traced from the log without reproducing it; a thrown value that isn't an error is logged as text instead ofundefined. - In a pull request that adopts or updates project-template, the engine reads which files the template owns from the template itself, at the
_committhe head's answers name: every file itstemplate/folder renders except those itscopier.ymlwrites once (_skip_if_exists) or merges (_merged_files) and thoseproject_owned_fileslists. A file a new template release adds moves its findings upstream in the pull request that updates to that release, without waiting for an engine release that lists it. A template file the pull request deletes, or one the template no longer renders, is the project's, and whencopier.ymlcan't be read or lists a pattern instead of a path, every file is the project's and the run's log says why (#152, ADR-1000).
Fixed
- A re-review's summary comment links each earlier finding under "Earlier findings" to its thread again, instead of showing the link's markdown with its address broken: the comment keeps the links the engine writes to the bot's own threads as they are, and still neutralizes the model's text around them. An earlier finding whose thread address isn't an https link is named without a link (#154).
v1.0.1
Immutable
release. Only release title and notes can be modified.
Added
CONTRIBUTING.md,SECURITY.mdanddocs/, with the project template's decision records indocs/decisions/project-template/; the README links them.
Changed
- This repository's own workflow and the README's examples pin v1.0.0.
- The caller runs a plain
/reviewonly on an open pull request, so a/reviewon a pull request that just closed can't cancel or displace the outcomes job in their shared concurrency group;/review explainand/review teachstill answer on a closed one. - The README's
sensitive:example shows NCodeGroup/Identity's real project paths and the set it uses: tokens, keys and secrets, authentication, client and server. - A finding on a repository's
/audit-conventionscommand, prompt file or skill, which project-template generates from v0.45.0, moves upstream to NCodeGroup/project-template like a finding on the template's other generated files. - The repository follows project-template v0.53.0 (from v0.41.1): CI also checks the decision index and the pull request's title and description for agent attribution, and cancels a run a newer push supersedes; Dependabot waits three days before offering a new version and gives security updates and NCode Review bumps pull requests of their own; a workflow auto-merges a Dependabot bump of this repository's own NCode Review pin to a minor or patch release; the Copilot cloud agent's runner is set up as CI is; and agents get the
/address-feedbackand/audit-conventionscommands.
Fixed
- The context, selection and synthesis calls no longer cache their prompts: each makes one request, so caching paid 25% more to write about 150,000 tokens a review that nothing read. Reviewer followers keep reading their leader's cache, as the traces confirmed (#133).
- In a pull request that adopts or updates project-template, a finding on a file the template added through v0.54.0 and replaces on every update moves upstream to NCodeGroup/project-template like a finding on its other files, instead of being posted as the project's: its decision records in
docs/decisions/project-template/, itstemplate-update.ymlandcode-review-auto-merge.ymlworkflows, its C# test rules (.github/instructions/csharp-tests.instructions.mdand.claude/rules/csharp-tests.md) and its .NETbuild/files (ProjectTemplate.globalconfig,ProjectTemplate.Tests.globalconfig,BannedSymbols.txt,dod.ps1andUpdate-PublicApi.ps1) (#148). - A finding that repeats one of the bot's earlier threads on the pull request, posted or dismissed, is listed apart under "Already raised" in the summary comment and the check run, naming the earlier finding, linking its thread and saying "dismissed" when the maintainer dismissed it, and is left out of the finding counts and the annotations; a dismissed one is left out of the verdict too. Before, when every finding repeated a thread, no review was posted, yet the summary comment and the check run still reported them as new findings (#151).
- A pull request that deletes a C# file in a packed project no longer crashes the review with
Cannot read properties of undefined (reading 'trim')when a removed line is an attribute: the public-surface check looked for the declaration below the attribute before the deleted file's first line (#155). - The check run no longer says it's in shadow mode: its title drops "(shadow mode)" and its summary says, as the summary comment does, that findings are posted as review threads and in the check run; the action's description and the README's opening say the same (#151).