- Host-Agnostic Correctness Review (
correctness-review):ag-review-implementationadds a read-only Correctness Scan (step 7b,templates/correctness_review.md) that hunts logic defects in the changed code and its callers using file reading only, so it behaves the same on Claude Code, Codex, Antigravity (agy), and OpenCode. Each finding carries a concrete failure scenario, the smallest fix, and the test that proves it, is verified asCONFIRMEDorPLAUSIBLEbefore it is reported, is labelledStatic analysis; not executed, and never blocks by itself.ag-review-pr,ag-pr-feedback,ag-governed-implement, andag-verifyinherit it through delegation. Depth isoff,quick(default, inline, at most 5CONFIRMEDfindings),standard, ordeep(read-only parallel reviewers and a fresh verifier when the host can dispatch them), set per run withcorrectness=<depth>or per project withreview.correctness.depthin.architecture-guard/config.yml; setoffto skip the extra review cost.
- Add unified team-role setup and synchronization across Claude Code, Codex, Antigravity (
agy), and OpenCode (.opencode/agents/agteam-<role>.md). - Add optional
implementor_e2e([E2E]) role across bundled role templates with read-only reviewer permission guarantees. - Support interactive
architecture-guard initteam-role auto-activation and configureagent_teams(enabled_hosts,activation: auto | ask | manual,isolation: auto | worktree | shared) in.architecture-guard/config.yml. - Add tiered workspace isolation for parallel implementor subagents: host-native workspace sharing (
agyshare) -> per-role Git worktrees (.git/ag-worktrees/agteam-<role>) managed via CLI helpers in the owning repository (with manual unmanaged fallback outside.gitfor read-only sandboxes) -> disjoint in-place writes. - Share execution modes, delegation thresholds, complete handoffs, ownership, and independent-review policy across canonical skills.
- Orchestrate parallel lifecycle phases across spec review, implementation, code review, security review (
sr-branch), and E2E verification. - Distinguish native feature reductions from behavioral guidance; parent sandbox, approval, and MCP settings are not per-role overrides. Read-only reviewer permissions are declared natively via tools in OpenCode (
tools: { write: false }), while Codex (disablingshell_tool) and Antigravity CLI govern reviewer roles via prompt contracts and human approval gates. - Protect resource and archive paths against path traversal escapes (CWE-22), while permitting readable external symlinks specifically for .architecture-guard/config.yml.
- During interactive or scripted
init, role installation failures log warnings and omit the failed host fromagent_teams.enabled_hostswhile allowing the rest of the workspace install (framework, presets, commands) to finish with exit code 0. Explicitsync-team-rolescontinues to exit with status 1 on role installation errors. - Runtime Codex/agy/OpenCode role discovery/enforcement and independent Security Review require separate evidence; fixture checks alone do not establish them.
- Per-Item User Story Decomposition & Tracker Sync (
split-user-stories-per-item):ag-governed-delivery-teamdecomposes multi-story features into a parent overview (user-stories/<feature>.md) and per-item child stories (user-stories/<feature>/<story-id>.md), while single-story features remain a singleuser-stories/<feature>.mdfile. Tracker sync maps parent + child stories to Jira Epics + linked Stories (or standalone Stories with copied context whenjira.epic: falsein.architecture-guard/sync.yml) and to GitHub parent + child Issues, recording per-item status (<feature>and<feature>#<story-id>) withrolemetadata in.architecture-guard/sync_status.json. - Expanded SonarLint Code Quality Bundle (
expand-sonar-rules): Expanded bundledsonar-rules/sonarlint-rules.jsonfrom 24 to 37 architecture-relevant rules acrossjava,javascript,php,python, andtypescript— adding cross-language parity rules (java:S1192,java:S1186,java:S112,python:S1192,python:S1066,python:S1186), TypeScript boundary type-safety (typescript:S4325), and a 6-rulephpbundle (php:S3776,php:S138,php:S107,php:S1192,php:S1066,php:S1186) supporting Laravel and CodeIgniter presets. - Light-First Landing Page & Flagship Skills Showcase (
revamp-landing-page-light-theme): Modernized public landing page with a default light theme, an interactive theme toggle withlocalStoragepersistence, hero spotlight cards for/ag-workflowand/ag-governed-backlog, native multi-host agent team optimization details (Claude Code, OpenCode, Antigravity, Codex), actionable usage paths, and a dedicated on-page Ecosystem & Credits section honoring OpenSpec, Spec Kit, Ponytail, Flash-Mem, Security Review, SonarSource (37 rules), Beads, and Playwright.
- Every Open Question comes with a recommendation: a new Open Questions rule in the Ponytail Core contract applies to every skill. A question left for someone else to decide now states the answer the agent would choose, why, and its confidence, even when the evidence is thin, and a section with no questions says
None. "TBD" and "needs discussion" no longer count. The recommendation never decides the question, grants approval, or widens scope. The Discovery Summary, User Story, specification clarification,ag-governed-backlogandag-pr-queuedecisions,ag-pr-feedbackDiscuss replies, andag-e2e-testingspec gaps follow it, as do the clarification loops in the built-in adapters and the legacy Spec Kit commands. Requests for facts only you hold and confirmations of a specific action are still asked directly. - Fixed —
validate-setuphonorsgovernance:: a readable document declared asconstitution,architecture, orsecuritynow stands in for the adapter's default path. Previously a project keeping its constitution elsewhere gotGovernance constitution missingand a failing exit code, with advice to runag-init, which would create a second copy. A declared document that is missing or gitignored replaces nothing. ag-pr-queueis one command for your pull-request work: a new skill scans a repository's open pull requests (with no arguments, or from a repository or/pullsURL), or takes numbers or URLs, and runsag-pr-feedbackorag-review-prfor each one. You can name the action per pull request; otherwise your own pull requests get feedback and everyone else's get a review, and it asks when your identity cannot be read. A scan sets aside your approved pull requests, your own with nothing open, other people's drafts, and pull requests you already reviewed at the current head, and lists them at the queue gate so you can add any back. After you approve the queue, it prepares every review draft and feedback triage without further questions (up to 4 in parallel on a host that can run subagents; fixes, tests, commits, pushes, and posts stay one at a time), then asks for the writes at batched gates: decide drafts and fixes (D), commits and pushes (P), and replies and thread resolution (R). A group answer can post drafts you were shown, but Approve with comments must always name its pull request. Each feedback entry works in its own worktree, your checkout is never switched, and rules always come from your checkout. A pull request that moved goes back to Gate D with an updated draft, and a failed entry is parked while the run continues. State lives in.architecture-guard/pr-queue.yml, so a later run resumes; approvals never carry over.ag-review-prandag-pr-feedbackgain a Queue Mode section that accepts these gates; run directly, they behave as before. Theprcommand alias now includespr-queueandreview-implementation, whichag-review-prandag-pr-feedbackdelegate to and which the alias previously left out. New optional forge keyscurrent-userandpr-list, defined for every built-in forge; a custom forge without them makes the skill ask you.
- Idea-to-backlog decomposition:
ag-governed-discovercan break an idea into a proposed backlog without writing files.ag-governed-delivery-teamapproves and persists that breakdown, then stops at Phase 3 in breakdown-only mode so planning and implementation do not start before the backlog is accepted. ag-review-artifactsfinds what the artifacts leave out, not only what they get wrong: a new read-only Consumer Inventory step takes each privileged or irreversible state the plan guards (a role grant, a reactivation, a deletion, an external send) and searches the base branch for the files that reference the table, model, field, route, or writer the artifacts name for it, reading only the matching line and its enclosing declaration. A reader or writer the spec, plan, or tasks do not name becomes a Path Coverage Gap finding (Unlisted Consumer). In a workspace with several repositories it searches onlycoderepositories, each at its own base ref, and reports a repository whose ref is unavailable asNot inventoriedinstead of searching it at another's. It is capped (10 names, 50 hits per name), reportsNot runwhen git is unavailable, and never reports an inventory that did not run as empty. Premise Grounding now also grounds a method or trait the plan calls on an existing type without defining it (for example anauthorize()the base controller must provide).- Two more artifact checks: Artifact Self-Consistency reads the whole artifact set after each repair, not only the repaired section, and reports a cross-reference or count that no longer matches the section it summarizes, an interface rule that disagrees with its server-side rule for the same action, two sections that order or resolve the same operation differently, a rule with no test row (or a test row for a rule that no longer exists), a field one artifact reads that no contract defines, and citation baselines that differ between sections. Convention Breach holds the plan to the host's own instruction files: the review now reads
AGENTS.md,CLAUDE.md,GEMINI.md, Copilot instructions, always-on Cursor and Claude rules, every file they import, and a style guide beside the specification, as conventions (MEDIUM for a MUST, otherwise LOW; non-blocking unless the same rule is in a declared governance document). In a multi-repository workspace a code repository's instruction files are read at its base ref, never from a working tree or pull-request copy, and rules still come only from the workspace. The report gains a Consumer inventory line and a Convention Sources section. No adapter key or configuration is added, so custom adapters and existing.architecture-guard/config.ymlfiles keep working unchanged, and the legacy Spec Kit extension command carries the same checks. ag-governed-deliveryandag-governed-delivery-teamnever ship an un-reviewed repair:No progressandPass limitnow stop before repairing that pass's findings. When you choose to apply the remaining repairs afterNot converged, the loop runs exactly one more review pass with no repair after it (a confirmation pass, outside the pass limit) and reports its findings next to the stop reason, so the summary never describes an applied but un-re-reviewed repair as reviewed. Previously a delivery could end with a repair batch applied after the last pass, and only a later pull-request review found the contradictions it introduced.
ag-review-artifactschecks what the artifacts claim about existing code: a new read-only Premise Grounding step reads only the base-branch locations the spec, plan, and tasks cite or name (paths,path:linecitations, named symbols, and "today the code does X" statements) and confirms or refutes each one. In a workspace with several repositories, it resolves the base branch and commit SHA for each owning repository separately (callers pass one base ref per repository), reports them per repository, and marks a repository's grounding unavailable rather than reading it at another repository's branch when its destination cannot be established. Paths the plan proposes to create are not premises: they are only checked for an existing file with the same name. It also checks that names the plan introduces do not collide with names already in that file or namespace, and reads the state of referenced pull requests or tickets through the forge'spr-viewaction or the tracker when available. It never checks out, fetches, writes, or posts, caps the inventory at 40 premises, and reportsUnavailablerather than passed when git or a base ref is missing. Contradicted, stale, or unresolvable premises become Premise Mismatch findings, and the report gains a Premise Grounding section. The "No Implementation Code" rule is narrowed to "No Implementation Code Review": the skill still never reviews code quality or design.- Two more artifact checks: Root-Cause Miss widens to Path Coverage Gap, which lists every entry path into a privileged or irreversible state and checks that each carries the same guard, confirmation, and audit, and that ordered steps still reach every "MUST refuse". The new Unqualified Guarantee check flags absolute concurrency or consistency claims ("cannot deadlock", "exactly once") that name no mechanism, conditions, or verification task.
ag-governed-delivery,ag-governed-delivery-team, andag-workflowpass the destination branch to the artifact review as its Base ref when they know it. The changes ship in every generated host variant (Claude Code, Codex, Antigravity, Cursor, Copilot, Gemini, and the rest) and in the legacy Spec Kit extension.ag-governed-backlogdelivers a prioritized backlog as a queue: a new skill reads work items from Jira (a JQL query), Beads (bd, when the repository has.beads/), or a local.architecture-guard/backlog.yml, and normalizes them to one shape (newbacklog_itemtemplate). It groups related items (Jira throughstory_scopePlanning Mode, Beads by parent epic, the local file bygroup), orders groups by dependencies, priority, and size, and runs each group through a lane matched to its type: full for stories and features, fix for bugs (a failing test before the fix), light for tasks and chores, and research for spikes. Each group gets its own worktree, branch, change, and pull request. A run has four gates: approve the queue, approve all new specifications together, approve every group's implementation scope (the exact files and operations) together, and approve each group's push and pull request. A confirmation a phase needs (for example continuing without security rules) is never taken for you; the group is parked until you answer at the next gate. Everything between them runs without stopping, and a group that does not converge, has a blocking security finding, needs your decision, or waits on an unmerged dependency is parked while the run continues. State lives in.architecture-guard/queue.yml, so a later run resumes. Writing back to Beads (in progress, pull request note, close after merge; research beads stay open for you to close) is opt-in; the queue never writes to Jira, never merges, and never force-pushes. Flash-Mem stays the context store and never holds backlog items. Newarchitecture-guard validate-backlogcommand (it reads only a file inside the project root, after resolving symlinks, and rejects dependency cycles between items and between groups), andsync.ymlaccepts abacklogblock (jql,lanes,max_groups).ag-governed-discoveraccepts a story set passed by its caller,ag-governed-delivery,ag-governed-tasks, andag-governed-implementaccept a caller approval whose preview showed their exact scope,ag-governed-tasksandag-verifyaccept a change with no specification delta (its stories and acceptance criteria stand in for the specification, a minimal plan is still required, and verification records where the criteria came from; a fix or light item with no acceptance criteria needs criteria you approve at the queue gate before its group starts), and the OpenSpec adapter'screate-changesetsskip_specs: truefor a change with no specification delta.
ag-pr-feedbackreads the whole review, not only threads: it now also reads each review's summary body and verdict and the general conversation comments, through new optional forge keyspr-reviews,pr-comments, andpr-comment(defined for GitHub, GitLab, Bitbucket, Azure DevOps, and plain git). Summary points that restate a thread are merged with it; points with no thread become their own items and are answered in one general comment per reviewer. A changes-requested verdict keeps the pass open until that reviewer reviews again, even when every thread is resolved. Previously a pass with no unresolved threads and no failing checks stopped and reported no feedback, missing findings stated only in a review summary. A custom forge without the new keys makes the skill ask you to paste them and reportDegraded.ag-review-prreads the same sources and now de-duplicates against all of them, not only threads.ag-review-prhandles re-reviews: when you have reviewed a pull request before, each earlier finding is classified Addressed (fixed since, never reposted), Answered (the author replied with evidence: re-checked, dropped when the author is right, otherwise answered in the same thread), or Open (kept, linked to the existing comment). Replies to Answered inline findings are part of the draft and are posted withthread-reply: after the review for Post as is; for Approve with comments, first, then the head is rechecked, then the approval, so a failed reply never leaves the pull request approved. Before posting it also re-reads feedback, not only the head commit. On a custom forge withoutpr-comment,ag-pr-feedbackwrites general replies to the replies file for you to post and reports posting asDegraded.
-
Discovery plans across related stories: with Jira story scope configured,
ag-governed-discoverapplies thestory_scopetemplate in a new Planning Mode. It reads the main story, its related tickets, and open tickets touching the same areas, and sorts them into Bundle now (strict: same files and boundaries, a few acceptance criteria, no new boundary or security-sensitive area, unassigned or yours), Do next, Conflicts, and Owned elsewhere. You accept the story set, andag-governed-specrecords it in the change's## Storiessection with requirements covering every story's acceptance criteria. The story scope check in reviews then reads that set, so bundled work is in scope and anything else is still flagged. -
Discovery suggests governance changes, with evidence: a new Governance Suggestions step drafts Constitution Update Proposals or workflow suggestions when the discussion or Flash-Mem shows a rule repeatedly in the way, citing how many independent cases support it. Discovery never edits rules or Jira; the user decides.
-
Archived changes stay archived during pull-request feedback: the usual order is to finish locally, run
ag-governed-archive(verify, archive, then open the pull request), and address feedback afterwards.ag-pr-feedbackandag-review-prnow treat an archived change as read-only reference, reading its full specification text from merged current specifications, a moved archive, or artifacts retained in place. A specification index only locates that text. They never move, edit, restore, or recreate the archived change. Feedback that changes a specified requirement is triaged as Discuss, with the choice of keeping the specification or opening a follow-up change, instead of a silent code fix that would leave the archived specification out of date. -
Story scope check against Jira: the shared change checks gain a read-only Story scope check (
story_scopetemplate), soag-review-implementation,ag-verify,ag-review-pr, andag-pr-feedbackcompare a change with its Jira story's acceptance criteria and related tickets (subtasks, links, open epic siblings). With no story key in the branch, title, or commits, it asks for one, then searches open tickets for overlap and asks you to confirm a candidate. It reports out-of-scope changes, work that belongs to another ticket, undelivered acceptance criteria, and untracked changes, advisory unless the Constitution makes them P0.ag-pr-feedbackuses it to classify Out of scope threads and name the covering ticket. Configure it withscope_checkin.architecture-guard/sync.yml(enabled,candidate_jql,max_candidates);validate-sync-configrejects it for providers other than Jira. -
One read-only Security Review operation for code: every adapter now maps
security-review-implementationtosr-branchagainst the base branch, soag-governed-implementand its review loop use the same operation asag-verifyandag-review-pr, and neversr-verify, which can execute code. Standaloneag-review-artifactsnow runs Security Review on the plan and tasks when it is installed (read-only, reportedUnavailableotherwise); insideag-governed-deliverythe delivery skill still owns that dispatch, so it runs once per pass. -
Workflow final review:
ag-workflowadds a specification approval gate, reports the artifact and implementation review convergence loops it delegates, confirms the final review ran against the destination branch, and suggests archival only after a final-review approval gate. In the SpecKit extension channel the final review reports Change Checks asUnavailable, because that channel'sag-verifydoes not provide them. Read-only reviews may run in subagents when the host provides them.
- Claude Code teammate roles: Agent Teams setup generates
.claude/agents/agteam-<role>.mdfiles from the configured roster, with tools restricted to each role. Existing permissions and hooks are preserved when Claude settings need manual repair. - Roster interview and synchronization:
ag-initrecommends roles from repository evidence and lets users keep, drop, or add implementors. The newsync-team-rolescommand refreshes generated roles and removes obsolete generated files while preserving user-authored files. Task tagging and dispatch follow the configured roster. - Safer Agent Teams setup: setup is offered only in Claude Code. Role profiles are validated before any files are written or removed; invalid tools, malformed profiles, and conflicting role names fail without changing role files.
-
Change checks run locally, not only on the pull request: the checks
ag-review-pradds on top of the architecture review (governance file changes, intent versus the stated purpose, missing or weakened tests, secrets and trust-boundary changes, security-sensitive files, CI) now live in a sharedchange_checkstemplate, andag-review-implementationandag-verifyapply it to your own branch against its target (/ag-verify base=main). Locally the intent source is the active change's spec, plan, and tasks plus the commit messages, CI isNot Applicable, and each finding becomes a refactor task forag-refactor-generatororag-apply.ag-verifyalways dispatches Security Review'ssr-branchon the whole branch diff against the base (base=<ref>, or the target branch it detects), the same scope the pull-request review uses, and never falls back tosr-verify; withoutsr-branchit reports Security ReviewUnavailable. Security Review runs once per run:ag-review-pr,ag-governed-implement, andag-pr-feedbacktell the review they own it. A branch that is clean locally should reach the pull-request review with few findings. -
ag-review-prcan approve with comments: when no finding is blocking, Security Review ran (or no security-sensitive file changed), no CI check is failing, coverage is not partial, and you are not the author, the review offers a fourth choice, Approve with comments. It posts the comments, approves at the recorded head SHA, and asks the author to address the findings before merging. Comments are posted first, and nothing is approved if any comment fails. Forges gain an optionalreview-approveaction (GitHubAPPROVEreview,glab mr approve --sha, Bitbucket approve, Azure DevOps vote approved with suggestions; unsupported on plain git). Custom forges without it keep working, and the option is simply not offered. -
ag-review-prruns Security Review: when the separate Security Review tool is installed with a branch operation (sr-branch),ag-review-prnow runs it on every pull request, after confirming it is registered, read-only, and does not run the pull request's code. Its findings join the draft as Security comments with their own severity. When it is not installed, the review saysSecurity Review: Unavailableand never reports it as passed.ag-review-implementationno longer starts its own Security Review when its caller runs one. -
ag-e2e-testing: a new skill that decides from the specification whether a feature needs end-to-end tests: UI E2E, API E2E, both, or Skip when nothing in scope is end-to-end. Each scenario traces to an acceptance criterion. Tests are written with Playwright: browser tests with thepagefixture for UI journeys, and the browser-freerequestfixture for API-only scope. Setup reuses an existing Playwright config, or installs Playwright only after approval. Tests run only against local or test environments, and a failing criterion is reported as a product defect instead of being weakened. -
Workspaces with separate application repositories: the governance and specifications repository (the workspace, holding
.architecture-guard/) can wrap application repositories with their own git history andorigin, such as a gitignoredsrc/. A new optionalrepositories:list in.architecture-guard/config.yml(name,path,roles: [code, specs], optionalforge) describes them;initdetects standalone nested repositories (a.gitdirectory; submodules and worktrees are left out), records them unless the key is already set (repositories: []opts out), and detects the forge from the first code repository.detect-changed-fileslists every repository's changes as workspace-relative paths, takes--repo <name>, applies workspace-relative and per-repositorysecurity_sensitive_paths, and still uses the workspace's settings when run inside a configured repository (previously it silently reported nothing sensitive there).hygienescans configured repositories, andvalidate-setupchecks each one. A newrepository_contexttemplate, referenced by every skill that uses git, tests or the code host, runs those steps in the repository that owns the files, checks the branch per repository, and proposes one commit and one push approval per repository. Projects withoutrepositoriesare unchanged.initrecords only nested repositories the workspace ignores (or any, when the workspace is not a git repository), never hidden, vendored, example or test checkouts. The host Security Review is dispatched once per owning repository with its changed files, or markedDegraded;validate-setupwarns about unlisted ignored repositories and workspace security globs that miss every repository;detect-changed-fileswarns when run inside a nested repository its workspace does not list, requires--repowith--base, and fails on a--repothat is not a git repository. -
architecture-guard hygiene --policy-root <dir>: takes the policy and local rules from that directory's.architecture-guard/instead of the scanned target's, so a review of someone else's checkout cannot change its own rules;ag-review-implementationpasses the workspace. -
Legacy Spec Kit extension safety:
commands/governed-implement.mdnow refuses to implement on a protected branch, andcommands/governed-archive.mdrefuses to commit or push on one, asks separately before each push, and never force-pushes. -
Governance documents reviewers can read: a new optional
governance:block in.architecture-guard/config.ymldeclares the constitution, architecture, security and layout documents; every built-in adapter's constitution set includes them.validate-setupwarns when a declared or mapped constitution is gitignored, outside the repository, or missing, and a new hygiene rule,unresolved-governance-citations(Warning), reports cited*constitution*.mdpaths in code, comments or specs that reviewers cannot open. Skills still read gitignored governance files but now report them as a hygiene warning. A rule only the author can read was cited in shipped code without anyone noticing. -
Security re-review on security-sensitive changes: a new optional
security_sensitive_paths:list of globs in.architecture-guard/config.ymlandarchitecture-guard detect-changed-files --security-sensitive(lists only the changed files that match; the default output is unchanged).ag-pr-feedbacknow dispatches the host Security Review before committing a fix that touches one,ag-governed-implementre-runs it after every later fix that does and recordssecurity re-review requiredas an open blocking item when Security Review is unavailable, andag-review-prsays a security review is required, matching only against the base checkout's list. Previously a review-thread fix that changed an authentication rule was re-checked for architecture only. -
Accepted risks must be complete: the constitution and plan templates require an owner, rationale, compensating control, trigger or expiry, and gate for every accepted risk or deviation.
ag-review-artifactsandag-verifyreport missing fields (LOW) and a security-relevant risk whose only mitigation is to revisit it later (MEDIUM, non-blocking unless the security policy says otherwise);ag-verifyalso reports a risk whose trigger has been reached as open. -
Safer skill gates:
ag-governed-specno longer creates a change before the branch gate (it asks separately for the change and capability names, and creates the change only after approval); auto-fix loops inag-governed-specandag-governed-tasksre-run validation (at most 2 attempts) and re-evaluate the blocking gate, andag-governed-tasksevaluates its gate before writing the summary;ag-consolidate-specsbuilds the index in memory and writes only after confirmation, and other workflows recommend it instead of regenerating the index themselves;ag-governed-archiverefuses to commit or push on protected branches, never force-pushes, and asks separately before pushing;ag-applyadds the protected-branch preflight and asks separately before any Flash-Mem write;ag-initpreviews before saving a paused draft and only proposes memory captures;ag-governed-delivery-teamvalidates sync config only with an installedarchitecture-guardCLI (nonpxfetch) and skips sync asUnavailableotherwise. -
Consistent skill behavior: every skill resolves the adapter through the same chain (override,
selected-adapter,config.yml, markers);ag-governed-implementuses the hygiene runner'sblockingflag and aligns its guardrail with the P0-only architecture blocking rule;ag-workflowpicksag-governed-delivery-teamwhenever User Story approval is required and no longer runs verification twice;ag-verifyrates boundary bypass HIGH unless policy says otherwise;ag-review-artifactslimits hygiene to runner findings on planning artifacts; constitution loading is stated once per skill. -
OpenSpec adapter: change discovery now uses
openspec list --json(changes) instead ofopenspec list --specs --json(specs), andopenspec new changeruns only after explicit approval. Generic adapter archive requires approval for botharchitecture-guard archiveand the manual move. -
Adapter Path Map
Notescolumn: annotations such as "optional" or "or a user-provided path" moved out of the path value into a thirdNotescolumn, so installed prompts receive clean paths. Custom adapters keep working; move annotations toNotesto get clean materialized paths. -
debug-artifactshygiene rule now defaults to Warning (was Critical), soconsole.log/console.errorat an approved Entry boundary no longer blocks by default. To keep blocking on it, adddebug-artifacts(orwarning) torepository_hygiene.fail_on. -
ag-pr-feedback: work through review feedback on your own pull request on any code host. It collects unresolved threads and failing checks, triages each (fix, discuss, disagree, already addressed, out of scope) against your constitutions, applies only approved minimal fixes, runs tests and a scoped architecture review, commits and pushes with separate approvals (never force-pushes), and drafts replies. You then choose Reply only, Reply and resolve (only fixed or already-addressed threads are eligible), Adjust first, or Don't post. Theprcommand alias now installs both pull-request skills. -
ag-review-pr: review someone else's pull request on any code host. It fetches the pull request through the selected forge, reads its full files from a temporary worktree (after approval), reviews the change against the constitutions and any linked SDD artifacts by reusingag-review-implementation, adds intent, test, secret, and CI checks, and drafts a comment-only review. You then choose Post as is, Adjust first, or Don't post; it re-checks the head commit before posting and never approves, requests changes, or merges. Install with--commands pr. -
Forge adapters: a new adapter layer for code hosts — GitHub, GitLab, Bitbucket Cloud, Azure DevOps, and plain git (
generic) — independent of the SDD adapter.architecture-guard initauto-detects the forge from theoriginremote (override with--forge <name>), installs forge adapters to.architecture-guard/forges/, and records.architecture-guard/selected-forge. Each action uses a registered MCP server for the host, then the host CLI (gh,glab,az), then a stated fallback. Custom forges work like custom SDD adapters; seeFORGES.md.validate-setupchecks the selected forge. -
architecture-guard hygiene --files <paths>: limits findings to the listed files (comma-separated, relative to--target) and records the scope in the JSON report (scope.mode,scope.files,scope.unmatched). Duplicate-detection rules still compare the listed files with the rest of the repository.ag-review-implementationuses it, so pull-request reviews no longer report pre-existing findings from unchanged files. -
ag-review-prhardening: the review worktree is created in the system temp directory, outside the repository, and removed at the end whatever the outcome; rules, constitutions, and configuration are read only from your base checkout, never from the pull request; changes to governance files are reported as security findings; the pull request's code, tests, and scripts are never run. -
ag-review-implementation: accepts a review root and base ref from calling skills, runs the hygiene runner on the changed files and reports its severity and blocking flags verbatim (manual rules are listed as degraded), adds aTargetcolumn soag-applyhandles artifact findings andag-refactor-generatorhandles code findings, and passes read-only and untrusted-content constraints to sub-agents. -
ag-pr-feedback: re-reads threads right before posting and asks again if reviewers changed anything, stops on a failed resolve, and never cites unpushed commits in replies. -
Forge adapters: review and reply text is passed to
gh,glab,az, andcurlthrough temporary files outside the repository instead of shell arguments. GitLab reviews publish only the drafts the run created (and stop if other pending drafts exist) instead ofbulk_publish, and never approve or set a review state. Azure DevOps diffs come fromgit fetchplusgit diffwith no checkout, REST calls pass the Azure DevOps resource id toaz rest(or--area/--resourcetoaz devops invoke), and threads are posted only if the source commit has not moved. GitHub review worktrees usegh pr checkout --detach, andgh pr checkshas a fallback for gh releases without--json. The generic forge slugifies ref names in file names and never runs tests for someone else's change. Installed skills list what each placeholder means. -
Forge detection:
github.*and*.ghe.comhosts are detected as GitHub.--forge bitbuckettargets Bitbucket Cloud only; Bitbucket Server and Data Center needgenericor a custom forge. -
Fixed — lean
initno longer deletes local overrides: files under.architecture-guard/templates,presets, andsonar-rulesare local overrides. A leaninitnow removes only byte-identical copies left by an older--vendorinstall, keeps modified or extra files (and lists them), and never prunes with--overwrite skiporkeep-both. -
Fixed — valid TOML for Gemini and Tabnine: prompts containing backslashes (for example
$\ge 15$) produced unparseable.tomlcommands. Prompts are now written as TOML literal strings, or as fully escaped basic strings when needed. -
Fixed —
--overwriteis validated: a mistyped mode (for examplekeepboth) now fails before anything is written instead of silently replacing files.keep-bothskill copies are named after their directory (ag-init-2), and Antigravityagx-*.mdworkflows follow the overwrite mode. -
Fixed — rerunning
initkeeps the selected forge: without--forge, a previously selected forge that still validates (built-in or custom) is kept instead of being replaced by auto-detection. -
Fixed —
archivesafety: change names must be a single directory name (no/,\,.., or absolute paths), soarchive ../../srccan no longer move project files. Changes are moved with a rename (copy and remove only across devices). Errors, including an unknown--framework, print oneError:line and exit non-zero without a stack trace. -
Fixed — SpecKit finalize: re-finalizing no longer erases hand-written text after an entry in
specs/system_context.mdor truncates a multi-line Purpose, the spec link is relative to the index (feat/spec.md), and a missingspec.mdor an unwritable index exits non-zero with{"status": "skipped" | "failed", "reason": ...}under--json. -
Fixed — smaller fixes:
init > logwith an interactive terminal no longer waits for stdin;validate-setuprejectsselected-adapter/selected-forgevalues that are not plain names; the hygiene file walk reports POSIX paths on Windows outside a git repository; the build cleansdist/and the npm package no longer ships compiled tests or stale removed commands. -
Legacy Spec Kit extension fixes (
commands/, safety and breakage only):- Delegation now resolves capabilities to the registered
speckit.architecture-guard.ag-<capability>command file or.specify/extensions/architecture-guard/commands/<capability>.md, so delegated commands no longer fail to resolve on Spec Kit installs. - The repository hygiene policy is written to and read from
.specify/config/repository_hygiene.yml(a top-levelrepository_hygiene:block), whicharchitecture-guard hygieneactually reads; the config template comment now names both locations. - Delivery commands no longer create a separate
feature/*branch before/speckit.specifycreates its own numbered feature branch; branch creation stays approval-gated and work never continues on a protected branch. ag-applyproposes Flash-Mem entries and writes them only after a separate[y/N]approval;ag-initasks before saving or archiving a constitution draft.ag-governed-delivery-teamrunsvalidate-sync-configonly when the CLI is on PATH (never throughnpx) and skips User Story sync, reported asUnavailable, otherwise; an invalid config means no remote write.- The budgeted context contract names the registered
/speckit.architecture-guard.ag-consolidate-specscommand; delivery repair loops stop asBlockedafter two cycles; review commands runarchitecture-guard hygiene --jsonwhen available (artifact review stays artifact-only);SPECKIT-INTEGRATION.mdstates that adapter overrides apply to standalone installs only.
- Delegation now resolves capabilities to the registered
Architecture Guard's standalone orchestrator now supports any SDD tool through custom adapters, and its skills are SDD-tool-agnostic. This release also closes several safety-gate gaps, repairs broken CLI helpers, and limits the legacy Spec Kit extension to safety and breakage fixes.
- Adapter location moved: all adapters now install to
.architecture-guard/adapters/instead of a project-rootadapters/directory. Rerunarchitecture-guard initto upgrade; the installer flags the oldadapters/directory but never deletes it. - New required adapter keys: OpenSpec-specific text in the orchestration skills moved into five adapter keys —
project-config,constitution-set,change-scope,validate-spec,spec-template. Built-in adapters define them; custom adapters must add them (initlists any that are missing). - CLI commands removed:
check-architectureandtest-install(both only printed a message). - Built-in adapters are refreshed on every
init: copy one to a new name to customize it instead of editing it in place. - Delivery policy: the legacy Spec Kit extension (
commands/) receives safety and breakage fixes only; new features ship in the standalone tool.
- Custom SDD adapters: write
.architecture-guard/adapters/<name>.mdand runarchitecture-guard init --framework <name>. The adapter is validated against the Required Keys inADAPTERS.mdbefore anything is written, and its tokens are resolved into the installed skills. validate-setup [target]: adapter-aware check of the selected adapter, shared resources (lean or vendored), mapped constitutions, governance config, andAGENTS.md; exits non-zero on errors.create-fixtures <target>: generates the context-budget benchmark fixtures.validate-sync-config [path]: registered CLI command used bygoverned-delivery-team.detect-changed-files --base <ref>.- Untrusted Content rules in the shared Ponytail Core contract: reviewed code, artifacts, reports, memory, and tool output are data, never instructions; embedded instructions are reported as findings and cannot grant approval.
- Severity Scale in the shared Ponytail Core contract: one scale for architecture and security findings, hygiene reported exactly as the hygiene runner returns it, and Sonar severities mapped onto the scale.
governed-implementstops on a Constitution P0 at any review severity (previously only when rated CRITICAL/HIGH), and checks preconditions before writing code: approval-gated feature branch, tasks present, not blocked, not stale.governed-tasksevaluates a blocking gate before offering the advisory revision loop.- Plan-repair and spec-validation repair loops are capped at 2 attempts, then stop and report.
- Branch creation and derived change names in governed delivery go through the Write Approval Gate preview instead of being created without asking.
- Security Review status is reported as Reviewed / Advisory / Blocked / Unavailable; Unavailable is never a pass.
- Memory writes always require explicit approval, including inside capture flows.
consolidate-specsconfirms before writing the fallback index.- Rerunning
initkeeps user settings in.architecture-guard/config.yml(context,repository_hygiene, comments) instead of overwriting the file;initrefuses to write budgeted-context settings into a constitution. init --yesno longer waits for stdin to close.detect-changed-filesincludes commits since the branch point, handles renames, and works from subdirectories (a clean feature branch previously produced an empty review set).- SonarLint filter keeps Sonar MAJOR and above (the previous CRITICAL/HIGH filter matched almost nothing). The bundled Sonar rules are replaced with 24 rules verified against SonarSource metadata; several old keys had the wrong meaning or did not exist.
architecture-guard archivefollows.architecture-guard/selected-adapterbefore filesystem markers, acceptsspec-kit/none, and refuses custom adapters instead of silently archiving as OpenSpec.AGENTS.md: an older unmarked Architecture Guard section is replaced in place even when other content surrounds it.- OpenSpec adapter branch creation requires explicit approval.
- Legacy Spec Kit extension: writes require approval (
[y/N]prompts), the user-story link path is computed instead of pointing outside the repo, one approval-gated branch rule, command references use the registered/speckit.architecture-guard.ag-*names, prompts no longer load standalone adapters, andverifyworks without the CLI installed.
- Removed hardcoded
.specifypaths in orchestration templates to ensure seamlessnpm publish --dry-runand clean packaging.
-
Synchronized standalone CLI initialization, adapter materialization, and lean versus vendored runtime resource resolution across generated governance prompts.
-
Added dynamic CLI resource resolver (
architecture-guard resolve <category> [name]) for on-demand resolution of immutable engine templates, presets, hygiene rules, and SonarLint rules. -
Implemented lean workspace initialization:
architecture-guard initnow writes a compact.architecture-guard/config.ymland avoids copying static unmodifiable files into target repositories. -
Added legacy static directory purge during
initto remove obsolete asset folders from.architecture-guard/. -
Added
--vendor/--fullflag toarchitecture-guard initfor air-gapped setups requiring full local copying. -
Enforced clean force-replacement of
ag-*agent skills/commands across selected agent directories. -
Implemented strict path traversal bounds validation (CWE-22) in resource resolution.
- Enforce schema-compliant OpenSpec artifact keys in
rules:(proposal,specs,design,tasks). - Relocate verification-only requirements and testing standards into top-level shared
context:block inopenspec/config.yamlandag-inittemplates. - Update
ag-initguidance and OpenSpec adapter documentation to prevent schema validation warnings across agents.
- Added Angular architecture preset (
presets/angular.md) supporting modern Angular conventions (Signals, Standalone components,inject()DI,OnPushchange detection, typed reactive forms, DTOs, and functional guards/interceptors). - Added deep interactive interview inquiry flow for Angular-specific concerns (Signals-to-RxJS interop, Hydration/SSR &
@defer, Micro-frontends/Module Federation, and@angular-eslint). - Added Deprecated & Dangerous Code hygiene rule (
hygiene-rules/deprecated-and-dangerous-code.md) with default Critical (blocking) severity, cataloging obsolete language and framework patterns (PHP, JS/TS, Node.js, Angular, React, Security). - Added pre-implementation and task preflight deprecation checks to
governed-implementworkflow. - Streamlined interactive CLI initialization: auto-detects single SDD frameworks, installs all commands by default, automatically replaces skills/resources, scopes Antigravity to workspace, and auto-updates
AGENTS.mdwithout prompt fatigue. - Updated preset documentation, repository hygiene documentation, templates, and init command options.
- Aligned SDD orchestration and legacy skills across adapters, commands, and workflows.
- Fixed adapter path resolution in legacy commands.
- Improved installer failure reporting and added regression coverage.
- Hardened archive framework routing: preserve SpecKit feature history in
specs/<feature>/and incrementally updatespecs/system_context.md. - Enhanced governed delivery active feature resolution via
.specify/feature.json.
- Evolved
ag-applyinto an orchestrator that resolves findings across all artifact types (proposal.md,spec.md,design.md,tasks.md). - Added pre-classification
Target:fields toag-review-artifactsandag-verifyoutput. - Delegated upstream artifact fixes (
proposal.md,spec.md) to SDD native update capabilities after user confirmation.
- Removed generated dist artifacts from version control while preserving build and npm packaging behavior.
- Fixed the sync-config validator build by declaring its
yamlandzodruntime dependencies.
- Added AI agent auto-detection during installation to seamlessly select installed agents based on directory presence.
- Added an explicit
archivecommand to fully clean up empty source directories during feature archival.
- Added: Enhanced
/ag-initworkflow to interactively scaffold.architecture-guard/sync.ymlwhen a team issue tracker is selected. - Added: One-way sync feature for User Stories to external issue trackers (GitHub/Jira) via
.architecture-guard/sync.yml. - Added framework-agnostic governed delivery and team User Story workflows.
- Synchronized adapter contracts, prompt descriptions, installer registration, and project-relative command examples.
- Added CLI update and JSON changed-file capabilities.
- Added a verification-gated
governed-archiveworkflow for OpenSpec, SpecKit, and generic Markdown projects.
- Fixed
ag-review-artifactsambiguity by removing auto-detection for pre-implementation reviews. It now requires an explicit--artifacts-onlyflag to ignore codebase implementations.
- Added
Branch Managementsafety check toag-governed-deliveryto prevent accidental generation on protected branches (matchingag-governed-specbehavior).
- Rebuilt installer as a unified TypeScript CLI (
cli.ts). - Generalized command prompts to support both SpecKit (
spec.md, the technical design artifact) and OpenSpec (design.md) artifacts natively. - Fixed command prompt prefixing to consistently use
.ag-for extensions andag-for standalone environments. - Added explicit Pre-Implementation and Post-Implementation review gates in the documentation workflows.
- Repositioned Architecture Guard as a standalone SDD Tools Orchestration layer for discovery, specification, planning, implementation, and verification.
- Kept Spec Kit extension installation as a backward-compatible integration path.
- Updated version references and release-artifact examples to
2.0.1.
- Fixed memory-tool discovery in
governed-plancapability-detection:rg --filesrespects.gitignorehidingdist/— switched tofindorrg --files -uufor paths under.specify/extensions/. - Removed the phantom "shared-memory promotion" / global-lesson step that stalled the workflow when the MCP tool was not exposed — local capture is the terminal step; global publication is out of scope for governed-plan.
- Added Budgeted Architecture Context Retrieval with bounded Flash-Mem expansion and mandatory active-feature authority.
- Added
consolidate-specsand the non-authoritativespecs/system_context.mdfallback contract. - Added project configuration, freshness handling, context-expansion reporting, benchmark fixtures, and a reproducible token benchmark protocol.
- Made optional sub-agent delegation capability-gated so governed workflows fall back inline when companion commands are unavailable.
- Moved framework-specific initialization questions into preset-owned
## Init Interviewsections while keepinginitfocused on generic orchestration and portable architecture decisions. - Expanded all built-in preset interviews with framework-appropriate application structure, dependency wiring, persistence, contracts, authorization, and async infrastructure choices without forcing optional patterns.
- Added React Native as a built-in preset with mobile navigation, state and offline behavior, platform isolation, typed native boundaries, device infrastructure, and background-work guidance.
- Added the Framework Presets guide with supported-preset comparison, selection guidance, guardrails, and a custom-preset contract.
- Clarified the README, beginner guide, and workflow docs so
governed-discoveris clearly documented as the idea-stage entry point beforegoverned-specandgoverned-delivery. - Reworked the README quick-start and command directory so discovery, specification, delivery, implementation, and verification appear in the correct order.
- Added a compact choose-your-path flow in the README to show when to start with discovery versus when to go straight to delivery.
- Added
governed-deliveryas the suggested resumable plan-to-tasks workflow, with mandatory Flash-Mem preflight when available, architecture and applicable security plan gates, task staleness handling, and formal analysis. - Kept
governed-planandgoverned-tasksas targeted recovery commands and moved the manual sequence to the bottom of the README as the legacy workflow. - Standardized Security Review detection around
security-reviewwhile acceptingspec-kit-security-reviewas a compatibility alias in the new orchestrator. - Added a shared Ponytail Core contract with an ordered decision ladder, root-cause caller tracing, safety and verification floors, and explicit phase guidance.
- Wired every command to the shared contract and expanded architecture detection to catch unsafe simplification and caller-specific symptom patches.
- Added a senior-engineering lens to every framework preset and replaced the strongest line-count, mandatory-layer, and dependency cargo-cult rules with evidence- and tradeoff-based guidance.
- Standardized Security Review detection across all governed phases and expanded findings with consequence, smallest-fix, verification, and tradeoff guidance.
- Fix SonarLint rules bundling issue by moving rules into the extension source directory so they are packaged correctly on install.
- Added the
governed-discoverorchestrator command to shape raw feature ideas into architecture-aware discovery briefs before formal specification.
- Added DRY cleanup guidance across the architecture prompts and documentation so brownfield projects can more easily collapse duplicated business logic, validation, DTO mapping, and orchestration into one shared source of truth.
- Added a dedicated DRY Cleanup Guide, linked it from the onboarding and reference docs, and aligned the README feature set around Ponytail Pragmatism, DRY Cleanup Guidance, Brownfield Discovery + Verification, and Repository Hygiene Guard.
- Add branch management to governed-spec flow
- Repository Hygiene Guard (Full Changelog: v1.9.0...v1.10.0)
- Ponytail Pragmatism Integration: Adopted the Ponytail "lazy senior developer" philosophy natively across all orchestrator commands.
- Initialized constitutions now bake in YAGNI and standard library preference.
governed-implementenforces writing the absolute minimum code and favors one-line solutions.architecture-reviewnow features a "Ponytail Audit" phase to catch and flag bloat, over-engineering, and unnecessary abstractions.- Specification, planning, and task generation orchestrators strictly enforce minimalism to prevent future-proofing.
- Clarified installation instructions in the README to explicitly show both the default registry path and the direct artifact URL.
- Introduced the
governed-specorchestrator command. This command fills the gap before thegoverned-planphase by chainingspeckit.specifyandspeckit.clarifytogether. - Formalized the
speckit.analyzeAnalyst step inside thegoverned-tasksflow, introducing an Automatic Analyst Loop that pauses to repair execution gaps before moving to implementation. - Updated README with clearer reasoning on why developers should use the governed orchestration flows.
- Centralized brownfield and greenfield onboarding guidance in the README quick start and trimmed duplicate references across supporting docs.
- Added and registered the
init-brownfieldcommand for existing codebases. - Bumped the extension version and aligned the release badge and download links to
v1.8.17.
- Refined the Flash-Mem-first orchestration wording across governed and architecture prompts.
- Synced the release artifacts, download links, and badge to
v1.8.15.
- Enhanced the architecture commands with Flash-Mem context retrieval guidance.
- Aligned the install artifacts and badge with
v1.8.14.
- Tightened the architecture workflow handoffs and related orchestration notes.
- Aligned the install artifacts and badge with
v1.8.13.
- Made the
update_project_summaryexecution in the architecture init command conditional uponflash-memavailability to ensure backward compatibility.
- Bumped the extension version to 1.8.11 and aligned the install artifacts and badges with the new release tag.
- Preserved the
flash-membackend migration so governed workflows continue to useflash-memas the canonical MCP source.
- Updated governed Architecture Guard workflows to be memory-first when
flash-memis available. governed-plan,governed-tasks,governed-implement,architecture-workflow,architecture-review,architecture-apply, andarchitecture-verifynow prefermemory-synthesis.mdbefore broader scans.- README and command registry descriptions now reflect the memory-first orchestration model instead of treating
flash-memas merely supplemental.