Release: v0.19.0

Added

• dev-workflows-fullstack plugin. A single plugin combining backend and frontend workflows (30 skills, 25 agents). Install this instead of both dev-workflows and dev-workflows-frontend to avoid duplicate skill descriptions competing for context budget.
• deprecations field in marketplace.json. Plugin entries can declare per-skill / per-agent deprecation notices that are mechanically prepended to the description frontmatter during pnpm sync. Idempotent: re-running sync does not duplicate the prefix.

Changed

• recipe-fullstack-build and recipe-fullstack-implement in dev-workflows now carry a (Deprecated) prefix in their description, pointing users to dev-workflows-fullstack. The recipes still work in this release; they will be removed from dev-workflows in the next release.
• README Quick Start table now points fullstack users to dev-workflows-fullstack (single plugin) instead of installing both. Migration notice and updated duplicate-skills warning added.

Migration

If you do fullstack work, migrate now. The fullstack recipes will be removed from dev-workflows in the next release, so switch from installing both plugins to installing dev-workflows-fullstack:

/plugin uninstall dev-workflows@claude-code-workflows
/plugin uninstall dev-workflows-frontend@claude-code-workflows
/plugin install dev-workflows-fullstack@claude-code-workflows
/reload-plugins

Pure backend or pure frontend users: no action required. Continue using dev-workflows or dev-workflows-frontend as before.

Notes

• The fullstack recipes (/recipe-fullstack-implement, /recipe-fullstack-build) continue to work from dev-workflows in this release for backward compatibility. They will be removed in the next release because keeping them in both dev-workflows and dev-workflows-fullstack defeats the purpose of consolidation.
• Claude Code limits skill descriptions to roughly 2% of the context window. Installing dev-workflows and dev-workflows-frontend together causes 12 shared skills to be registered twice (once per plugin namespace), which can exceed the limit and silently drop skills. dev-workflows-fullstack consolidates them into a single namespace.

Release: v0.18.0

Added

• external-resource-context skill (cross-cutting). Captures access methods for resources outside the repository (design source, design system, API schema, IaC source, secret store, visual verification environment) in a two-tier file. docs/project-context/external-resources.md holds environment-stable facts; per-feature External Resources Used sections in UI Spec / Design Doc carry feature-specific identifiers. Distributed in dev-workflows, dev-workflows-frontend, and dev-skills.
• ui-analyzer agent (frontend). Reads the project's external-resources file, fetches design / design-system / guideline sources using the access methods declared there (URL, file path, MCP, etc.), and analyzes existing UI code (component structure, props patterns, CSS layout, state matrices, display conditions, i18n format, accessibility, generated artifact readiness). Designed to run in parallel with codebase-analyzer. Uses disallowedTools so the parent session's full tool set carries through.
• /recipe-front-adjust recipe (frontend). Coordinates UI adjustment on already-implemented features: external-resource hearing → UI fact gathering → user-confirmed write set → scale judgment → optional work plan → edit / verify / refine loop in the parent session → quality-fixer-frontend → commit per adjustment unit.

Changed

• recipe-front-design runs the external-resource hearing protocol and invokes codebase-analyzer plus ui-analyzer in parallel before document creation. Both outputs feed ui-spec-designer and technical-designer-frontend.
• recipe-fullstack-implement and monorepo-flow.md add the hearing step and the parallel ui-analyzer invocation to both Large and Medium scale flows. technical-designer-frontend now accepts a UI Analysis input alongside Codebase Analysis, with code: and ui: prefixes on fact_id to disambiguate facts in the Fact Disposition Table.
• UI Spec template gains an External Resources Used section.
• Design Doc template gains an External Resources Used subsection under Background and Context.
• Implementation and quality agents (task-executor, task-executor-frontend, quality-fixer, quality-fixer-frontend) consult the project-tier external-resources file when their task references resources recorded there.

Release: v0.17.1

Refreshes the task-analyzer skills metadata index to match the current SKILL.md content and adds an automated consistency check.

Highlights

• Skills index refreshed — Three drifted entries fixed: coding-principles section list, plus integration-e2e-testing and test-implement tags / typical-use updated to reflect the two-lane E2E split (fixture-e2e and service-integration-e2e) introduced in 0.17.0.
• New deterministic consistency checker — scripts/check-skills-index.mjs walks every entry in skills-index.yaml, extracts H2 headings from each corresponding SKILL.md, and exits non-zero on any order or text mismatch. Reports yaml vs md diff per index for fast triage.
• Pre-commit wiring — Available as pnpm check:skills-index and registered in lefthook.yml; fires on any change to skills/*/SKILL.md or skills-index.yaml.

Notes

• Scope of the checker is intentionally narrow: sections are deterministically checkable; tags / typical-use are subjective and remain manual.
• Patch release — no behavior change to recipes, agents, or other skills.

Release: v0.17.0

Strengthens the workflow so design and verification context reaches the executor and the implementation is observable from the start.

Highlights

• New /recipe-prepare-implementation — Verifies the work plan is implementable and resolves gaps via Phase 0 tasks before build, or exits no-op when readiness criteria already pass. Distributed in dev-workflows.
• Implementation Readiness gate — work-planner emits an Implementation Readiness: header on every work plan; recipe-*-build and recipe-*-implement halt-and-confirm when status is pending or escalated.
• E2E split into two lanes — fixture-e2e (browser harness with mocked backend, MAX 3, runs alongside the UI feature) and service-integration-e2e (against a live local stack, MAX 1-2, final phase only). Independent budgets and ROI thresholds.
• UI Spec and Connection Map propagation — work-planner produces task-mapping tables and task-decomposer routes the relevant rows into each task's Investigation Targets, so design context reaches the executor.
• Design-side update path in review recipes — Per-finding routing (code-side fix / DD update / skip) replaces the single fix-all toggle. DD updates run before code-side fixes so satisfied findings can be dropped.
• Consumed Task Set in build recipes — A single restricted task pattern is computed up front and reused for the readiness gate, autonomous execution, and final cleanup, preventing stale or cross-recipe task files from being picked up.
• Final Cleanup pattern — Every task-creating recipe deletes its consumed task files at completion. README adds a FAQ recommending users gitignore docs/plans/.

Refinements

• Anchored examples for language-agnostic skills (RTL+MSW shown as React/TS example among possible stacks; Playwright remains the default browser harness).
• Atomic Design coverage targets become opt-in in frontend skills.
• Agents and skills no longer name specific recipes; cross-layer wording is abstract.
• Plugin manifests bumped to 0.17.0.

Compatibility

Semver minor — no breaking change to recipe invocations. The acceptance-test-generator JSON output schema replaces generatedFiles.e2e with generatedFiles.fixtureE2e and generatedFiles.serviceE2e; downstream consumers within these plugins are updated in the same release.

Release: v0.16.17

Optimize agent prompts for Claude Opus 4.7

Tightens subagent prompt contracts so each one — output protocol, gates, scope, file boundaries, intermediate state — is unambiguous and model-independent. The changes target Opus 4.7's stricter, more literal instruction-following: vague phrases, redundant output instructions, and silent contract gaps that earlier models smoothed over now leave execution open to interpretation, so they have been removed.

Highlights

• Output Protocol: every JSON-returning agent emits a single raw JSON object (no fence, no surrounding prose) as the final message.
• task-executor / task-executor-frontend: Phase Entry Gate is limited to true preconditions; Step 1 / Step 2 / Exit gates take over mid-flow checks. New File Scope Constraint and out_of_scope_file escalation prevent silent edits outside the task's allowed file set (Target Files, the task file, the work plan, and Provides: paths).
• technical-designer / -frontend: explicit Gate 0/1/2/3 ordering with inline annotations, and physical subsection order rearranged to match Gate order so document order equals execution order.
• security-reviewer: introduces suspected_risk. Uncertain findings are downgraded with confidence instead of being silently discarded, and requiredFixes is reserved for high-confidence items.
• subagents-orchestration-guide: Basic Flow split into a Planning flow and a separate task execution cycle. Placeholder substitution and context-isolation rules are now explicit. direct implementation removed in favor of routing every implementation through task-executor, including at Small scale.
• Recipes: replaced the misleading [SYSTEM CONSTRAINT] block with a focused Scope Boundary for Subagents section, applied only to the eight recipes that actually invoke implementation or fix loops (recipe-implement, -build, -add-integration-tests, -review, and their frontend / fullstack counterparts). Design, planning, and diagnose recipes drop the boilerplate entirely.
• Skill portability: agent names removed from skill bodies (templates, principles) and from cross-agent references inside agents; subagents no longer reference each other by name. Dependency direction is restored to recipe → agent → skill, with skills knowing nothing about specific agents.
• skills-index.yaml: refreshed to match the actual ## sections of every SKILL.md (resolves several pre-existing drifts).

Upgrade notes

• If auto-update is enabled for this marketplace, restart Claude Code and run /reload-plugins when prompted to apply the changes.
• Otherwise, run /plugin marketplace update claude-code-workflows to refresh the catalog, then reinstall each plugin you use (/plugin install dev-workflows@claude-code-workflows, and similarly for dev-workflows-frontend and dev-skills). Finish with /reload-plugins to activate the new content without restarting.
• No interface changes for end users; existing flows behave the same. The differences are in how subagents interpret their own instructions under Opus 4.7.

Release: v0.16.16

Fixed

• dev-skills no longer loads agents or recipe-* skills it does not declare. Previously the plugin loader would scan the whole repo when source: "./" was used, leaking every agent and recipe skill into dev-skills. The bundled plugins now live in their own subdirectories so the loader only sees what each plugin should expose.

Plugin contents (unchanged from intent, now actually enforced)

Plugin	Agents	Skills
dev-workflows	20	20
dev-workflows-frontend	20	18
dev-skills	0	9

Upgrading

Reinstall or update the plugins to pick up the new version. No configuration changes required.

Notes

This release works around two open Claude Code regressions (anthropics/claude-code#53948, anthropics/claude-code#13344). Once either is fixed upstream, the bundled plugins can return to a single shared source layout.

v0.16.15

Fix: fresh installs failing with "Path not found" errors

Restructures the marketplace so that fresh installs no longer end up with empty agents/ and skills/ directories. Previously, each plugin lived in a subdirectory (backend/, frontend/, skills-only/) with symlinks to the shared top-level agents/ and skills/. Claude Code's installer drops symlinks whose targets resolve outside the plugin root, which left every declared agent unresolvable on a clean install.

marketplace.json

• All three in-repo plugins (dev-workflows, dev-workflows-frontend, dev-skills) now use source: "./" with strict: false
• Per-plugin agents and skills arrays are declared inline, pointing at the shared top-level directories
• No symlinks involved — the repo's top-level agents/ and skills/ are the only physical copy

Repository layout

• Removed the backend/, frontend/, and skills-only/ proxy directories along with their per-plugin plugin.json
• README updated to reflect the flattened structure

Upgrade notes

• Existing users: run /plugin update (or claude plugin update) once. The version bump invalidates the previous cache and reinstalls a working layout.
• New users: claude plugin install <plugin>@claude-code-workflows now works on a clean macOS / Linux setup without manual workarounds.

Release: v0.16.14

Fact Disposition Framework

Existing-behavior facts surfaced by codebase analysis are now explicitly traced through to Design Docs, reviewed for completeness, and checked for cross-layer consistency.

codebase-analyzer

• focusAreas repurposed as disposition targets with fact_id, evidence, and factsToAddress fields
• New Step 4.5: disposition target enumeration (input fields, call sites, branching cases, data shapes, error paths, external dependencies, operational cases)
• Cardinality guidance: 5-15 entries per analysis, grouped by coherent fact unit

technical-designer / technical-designer-frontend

• Added "Fact Disposition" required section: one table row per focusArea with disposition (preserve / transform / remove / out-of-scope), rationale, and evidence
• Added "Prior-Layer Verification" input parameter for fullstack flows

document-reviewer

• Added codebase_analysis input parameter
• Gate 0: structural check for Fact Disposition Table presence and focusArea coverage
• Gate 1: disposition value validation, rationale completeness, evidence carry-through

design-sync

• Added "Disposition conflict" detection type: same fact_id across Design Docs with different disposition values → critical severity

Design Doc template

• Added "Fact Disposition Table" section (Fact ID, Focus Area, Disposition, Rationale, Evidence)

Fullstack Flow Improvements

monorepo-flow

• Backend Design Doc is now verified by code-verifier before frontend Design Doc authoring begins
• prior_layer_verification result explicitly passed to the frontend designer
• Layer Context templates updated; Task Cycle section deduplicated

Orchestration Improvements

subagents-orchestration-guide

• Handoff Contracts formalized as HC-01 through HC-06
• Basic Flow consolidated from numbered step lists into compact table + rules

Release: v0.16.13

Quality Assurance Mechanism Pipeline

codebase-analyzer

• Added qualityAssurance output section that discovers linters, validators, and domain-specific constraints covering the affected files

technical-designer

• Extended Standards Identification Gate with quality assurance mechanism adoption decisions (adopted vs noted with reason)

Design Doc / plan / task templates

• Added "Quality Assurance Mechanisms" section to Design Doc template with coverage scope
• Added QA Mechanisms table with Covered Files column to plan template
• Added QA Mechanisms section to task template

work-planner / task-decomposer

• work-planner extracts adopted QA mechanisms from Design Doc and propagates to plan header
• task-decomposer uses file coverage overlap to deterministically match mechanisms to tasks

quality-fixer / quality-fixer-frontend

• Added optional task_file input parameter for supplementary quality gate hints
• Added taskFileMechanisms field to approved and blocked responses for auditability
• All orchestration recipes now always pass task file path to quality-fixer

Reference Representativeness

coding-principles

• Added "Reference Representativeness" section: verify patterns and dependencies from nearby code are representative across the repository before adoption

ai-development-guide

• Extended Pattern 5 (Insufficient Existing Code Investigation) with reference representativeness check

task-executor

• Added Reference Representativeness check during implementation with repository-wide verification
• Added dependency_version_uncertain escalation type when repo state alone is insufficient to determine the appropriate version

Release: v0.16.12

Diagnose Flow: Failure Point Model

investigator / verifier / solver

• Replaced hypothesis-based reasoning with a failure point model in the diagnose workflow
• Investigator now identifies concrete failure points instead of generating abstract hypotheses
• Verifier validates failure points against evidence rather than ranking hypotheses
• Solver derives solutions from verified failure points with clearer causal chains

E2E Test Generation Fix

acceptance-test-generator

• Fixed ROI calculation that made E2E test selection mathematically impossible (max score ~3.16 against threshold of 50)
• Simplified ROI to raw Value Score (0-120) for same-type candidate ranking
• Replaced blanket push-down rule ("already integration-tested → no E2E") with journey-aware retention
• Added reserved E2E slot for user-facing multi-step journeys, emitted regardless of ROI threshold
• Added structured generation report with nullable generatedFiles.e2e and e2eAbsenceReason

integration-e2e-testing skill

• Added environment-agnostic multi-step user journey definition (Web, mobile, CLI, API)
• Added user-facing vs service-internal classification for reserved slot eligibility
• Added ROI calculation examples showing selection outcomes

work-planner

• Added E2E Gap Check in common planning path (fires for both Strategy A and B)
• Gap Check respects e2eAbsenceReason to skip warning on intentional E2E absence

Downstream contract updates

• Updated orchestration guide to treat nullable E2E as valid (not an error)
• Updated recipe-plan, recipe-front-plan, recipe-implement, recipe-fullstack-implement to propagate e2eAbsenceReason through the handoff chain