WebDev Agent Kit v0.4.0

WebDev Agent Kit v0.4.0

Added

• Added stable goal, acceptance-criterion, and execution-slice identifiers for durable standard and deep planning contracts.
• Added explicit ENABLER slices that must name the approved downstream work they unlock.
• Added a compact clarification gate for standard and deep goals that asks at most three material questions, one at a time, and integrates accepted answers into the goal contract.
• Added durable coverage maps linking acceptance criteria to execution slices, verification evidence, and canonical planning states.
• Added a deterministic read-only planning analysis mode with stable findings,
  severity-based handoff gates, and checks for coverage, scope, ownership,
  conflict, approval, and verification gaps.
• Added post-implementation convergence with evidence-backed gap types, stable
  findings, and append-only durable-plan passes.
• Added shared review, loop, checkpoint, decision, and resume fields that keep
  goal, criterion, slice, convergence, and verification state aligned.
• Added a 13-case planning-integrity eval suite plus Codex and Claude Code parity
  coverage for portable, read-only plan analysis.
• Added deterministic planning-integrity validation, in-memory negative fixtures,
  and blocking PR and release CI gates for planning drift.
• Added generated-target planning parity checks for shared rules, templates,
  skill bodies, alias artifacts, and context budgets.
• Added 48 cross-client planning smoke checks covering eight portable workflow
  contracts across all six generated runtime targets.

Changed

• Release target inspection now distinguishes legitimate templates/project/**
  source templates from forbidden root project/** runtime state.
• Tagged release validation now requires non-empty Unreleased notes while ordinary branch validation continues to allow a clean post-release section.
• Goal and execution plan templates now preserve criterion and slice identity across implementation, retry, and resume workflows without adding planning ceremony to lightweight work.
• Goal planning now uses verified context before asking and skips repeated, low-impact, safely inferable, or purely implementation-level questions.
• Checkpoints, decisions, compaction, and loop contracts now preserve the same goal, criterion, and slice identifiers instead of creating parallel completion models.
• Execution planning now selects an explicit create, resume, or analyze
  mode and keeps analysis findings outside durable plan and source files.
• Execution planning now supports converge mode without changing application
  files, rewriting prior plan history, or adding empty success passes.
• Independent review now judges active criteria and coverage evidence without
  creating slices; bounded loops and checkpoints reuse the same plan identifiers.
• Planning evals validate static fixture contracts and cross-client expectations;
  they do not claim to prove live model behavior.
• Aligned manifests, eval fixtures, installation guides, and release validation
  with the 0.4.0 release candidate.

Installation

Verify the selected archive against SHA256SUMS. Extract Codex and Cursor archives from the project root; they create .agents/ automatically. Extract Claude archives as the native webdev-agent-kit/ plugin root. See docs/install/README.md for the client-specific entrypoint.

WebDev Agent Kit v0.3.0

WebDev Agent Kit v0.3.0

Added

• Added machine-readable runtime contracts for canonical Codex, Claude Code, and Cursor targets.
• Added scripts/validate_target_contracts.py to enforce canonical target, alias, entrypoint, skill-root, output, and validator declarations.
• Added docs/architecture/runtime-target-contracts.md as the human-readable target contract.
• Added a native Claude Code plugin manifest and structural artifact validator.
• Added a compact portable runtime core, evidence-gated React/TypeScript profile, and thin Claude Code, Codex, and Cursor adapters.
• Added runtime-layer validation for client neutrality, context budgets, manifest alignment, and generated adapter selection.
• Added a centralized eight-level policy precedence contract and validator for forbidden absolute authority claims.
• Added eight cross-client policy conflict evals covering every adjacent precedence boundary.
• Added a targeted README read/edit policy, technical evidence hierarchy, and six cross-model README behavior evals.
• Added a test change policy and eight evals separating existing-test maintenance from new tests and infrastructure.
• Added full 19-skill capability coverage, strict capability metadata validation, and eleven cross-client capability behavior evals.
• Added deterministic budgets for the always-on entrypoint, runtime layers, skill-discovery descriptions, and Claude skill prelude.
• Added fact-based output evals and deterministic final-answer budgets, including a 180-word cap for lightweight workflows.
• Added a 12-case GPT/Claude parity suite covering routing, context, README, tests, precedence, capabilities, verification, output, onboarding, and scope control.
• Added release-tag, installation-guide, and generated release-notes validation for the 0.3.0 release path.
• Added independent Codex project, Codex plugin, Claude plugin, and Cursor target validators plus source/generated version consistency checks.
• Added deterministic release archive construction and extracted archive fixtures for native paths, exclusions, links, checksums, and root-instruction safety.
• Added common/windows-shell-sandbox-rules.md for Windows PowerShell package-manager fallbacks and sandbox-blocked build, dev-server, and browser verification.
• Added trigger and output eval coverage for Windows shell and sandbox verification blockers.
• Added trigger and output eval coverage for unauthorized component, helper, hook, or function test authoring during frontend implementation.

Changed

• Replaced independent client target entries with three canonical targets and explicit compatibility aliases.
• Defined Claude Code as a native plugin target and VS Code client packages as aliases of their canonical runtimes.
• Claude Code now builds as a self-contained native plugin with portable skill frontmatter, stripped source graph metadata, and no Codex-only plugin or UI metadata.
• Legacy claude and vs-code-claude outputs now reuse the canonical claude-code artifact exactly.
• Claude project-policy onboarding now uses the exact @.agents/AGENTS.md import and requires approval before creating or merging host instructions.
• Runtime routing now loads core, project profile, client adapter, local conventions, and task skill as separate progressive-disclosure layers.
• Generated targets now ship the portable core, default profile, and exactly one canonical client adapter; Claude skills load those layers explicitly.
• Runtime skills, profiles, adapters, and project facts now resolve instruction conflicts through one shared precedence contract.
• README may now be read for relevant intent, setup, onboarding, audit, and drift questions, but it is never sufficient technical proof and cannot be edited without an explicit current user request.
• Relevant existing tests may now be run and maintained for confirmed contract changes; new tests remain explicit-request-only, infrastructure remains approval-gated, and routine skipped-test output is removed.
• README now documents Windows shell and sandbox verification behavior as a user-facing boundary.
• Lint, bounded retry, smart verification, rendered visual QA, and linter skill rules now stop repeated verification attempts after Windows shell or sandbox blockers and require honest blocked-check reporting.
• Tool capability metadata now represents approved out-of-sandbox fallback and repeated sandbox-blocked dev-server limits for rendered visual evidence.
• Codex agents/openai.yaml files now contain only UI and invocation metadata; optional MCP and native-provider candidates live exclusively in the client-neutral capability manifest.
• Current OpenAI API, ChatGPT Apps SDK, and Codex lookups now use conditional openai_platform_docs routing: prefer the callable official OpenAI Developer Docs MCP, fall back to official web docs, and skip unrelated frontend work.
• Capability availability now requires current-registry, validated-profile, or user-supplied-reference evidence; package, config, and provider names are explicitly non-proof.
• Replaced the long always-on policy catalog with a 19-skill compact index and progressive-disclosure routes while preserving detailed rules in owning common docs and skills.
• Compressed all skill discovery descriptions and stripped source graph frontmatter from every generated target.
• Final responses now default to Changed, Verified, and Blocked facts while omitting request echo, skill narration, empty headings, and raw logs.
• Aligned bundle-manifest.json, .codex-plugin/plugin.json, and tool-capabilities-manifest.json with the 0.3.0 next-release metadata.
• Expanded bundle manifest target metadata and schema validation to include codex, claude, claude-code, cursor, vs-code-codex, and vs-code-claude generated targets.
• Release workflow and full validation orchestrator now run schema validation with --strict-graph.
• New skill templates now quote the placeholder description value so generated drafts do not start with invalid YAML.
• Test authoring rules now explicitly block component, hook, function or unit, integration, E2E, snapshot, fixture, mock, and visual regression tests unless the user explicitly asks and approves the exact scope.
• Codex archives extract runtime under .agents/; Cursor archives place shared runtime under .agents/ and native rules under root .cursor/; Claude archives retain their native plugin root.
• Installation guides now follow the extracted archive roots, prevent .agents/.agents nesting, and document persistent Claude skills-directory plugins.
• Tagged releases now fail when the tag differs from source version metadata and use the validated Unreleased changelog section as release notes.
• Release runs now validate notes without writing generated Markdown into the source tree, then create RELEASE_NOTES.md only during artifact packaging.
• Archive validation now requires all stable/versioned pairs, byte equality, exact checksum inventory, and duplicate-free checksum entries.
• Root README installation, archive-root, and evidence-policy guidance now matches the validated 0.3.0 runtime contracts.
• Runtime targets no longer include human-facing root files: README.md, CHANGELOG.md, CONTRIBUTING.md, or SECURITY.md.
• Runtime targets no longer include examples/.
• Codex and VS Code Codex targets keep .codex-plugin/; Claude Code, VS Code Claude, Cursor, and legacy Claude targets do not include it.
• Repository-side installation guides remain human-facing and are excluded from generated runtime targets.

Installation

Verify the selected archive against SHA256SUMS. Extract Codex and Cursor archives from the project root; they create .agents/ automatically. Extract Claude archives as the native webdev-agent-kit/ plugin root. See docs/install/README.md for the client-specific entrypoint.

WebDev Agent Kit v0.2.3

Restore VS Code Claude install guide

WebDev Agent Kit v0.2.0

This release turns WebDev Agent Kit from a Codex/Claude portable skill bundle into a more complete multi-client agent operating kit with capability-first MCP handling, smarter verification, stricter frontend anti-patterns, native client installation paths, and release-ready download artifacts.

Highlights

• Added client-specific packages for Codex, Claude Code, Cursor, VS Code Codex, and VS Code Claude.
• Added a capability-first MCP/tool model instead of assuming specific server names.
• Added smart verification budgeting to prevent over-testing small changes.
• Added strict rules preventing agents from reading or editing host-project README/docs during normal work.
• Added stronger frontend anti-patterns, including useCallback misuse, parallel status booleans, implicit return types, function declarations, and test authoring by default.
• Added Russian installation guides for each supported client.
• Added README download links for stable latest release artifacts.

Release Artifacts

Download the package that matches your client:

• webdev-agent-kit-codex.tar.gz
• webdev-agent-kit-claude-code.tar.gz
• webdev-agent-kit-cursor.tar.gz
• webdev-agent-kit-vs-code-codex.tar.gz
• webdev-agent-kit-vs-code-claude.tar.gz

Versioned copies are also published for this release, together with:

• SHA256SUMS

Added

Client-specific package targets

The build and release flow now produces dedicated packages for:

• Codex
• Claude Code
• Cursor
• VS Code Codex
• VS Code Claude

The release workflow also keeps legacy generated targets for current validators.

Capability-first MCP/tool model

Added:

• tool-capabilities-manifest.json
• common/tool-capability-model.md
• common/mcp-installation-policy.md

Skills now describe required capabilities such as:

• project_files
• current_library_docs
• web_platform_docs
• rendered_visual_evidence
• repo_metadata

Instead of assuming that a named MCP server exists, agents must detect whether the current client actually provides the required capability. If not, they must use an allowed fallback and clearly report confidence impact or blocked checks.

Native client adaptation

Added:

• common/client-adaptation-policy.md
• templates/project/client-profile.md
• templates/project/mcp-profile.md
• root pointer templates for Codex, VS Code Codex, Cursor, and Claude Code

Onboarding now records client/tool facts in local-only project overlays instead of mixing host-client state into reusable bundle rules.

Smart verification budget

Added:

• common/smart-verification-budget.md
• common/diff-impact-verification-rules.md
• common/mcp-availability-detection-rules.md

Small CSS-only changes, such as background color or decorative mask updates, should no longer trigger full-repo lint, full typecheck, dev server startup, route discovery, or browser QA by default.

Repeated visual failures can still escalate to targeted rendered visual QA when the task requires it.

Host README/docs boundary

Added:

• common/host-project-documentation-boundary.md
• common/anti-patterns/no-host-readme-or-docs-edits-without-approval.md
• scripts/check_readme_boundary.py

Agents must not read, use as context, route from, derive facts from, or edit a host project’s README/docs during normal implementation, bugfix, refactor, onboarding, visual QA, or review work.

README/documentation edits require an explicit documentation task and approved scope.

New frontend anti-patterns

Added or tightened anti-pattern rules for:

• no implicit return types
• no function declarations by default
• prefer object destructuring
• no parallel lifecycle/status booleans
• no test authoring by default
• no console file reads before filesystem/project-file capability
• no host README/docs edits without approval
• strict useCallback boundary

useCallback is now explicitly forbidden by default unless there is a proven identity-sensitive consumer, a real observed issue, and a safe dependency explanation.

Pattern library manager

Added:

• skills/pattern-library-manager

This skill maintains reusable approved patterns and anti-patterns with bad/good examples, source-sourcing rules, graph links, and eval updates.

Link checking

Added:

• scripts/check_links.py
• .github/workflows/link-check.yml

Local Markdown links are checked on PR/push. External link checks run on a scheduled workflow.

Russian installation guides

Added draft Russian guides under docs/install/**:

• Codex
• Claude Code
• Cursor
• VS Code Codex
• VS Code Claude

These guides explain archive choice, native pointers, MCP capability behavior, fallback behavior, onboarding via адаптируйся, and validation expectations.

Changed

• mcp-toolchain-manager now reports capabilities, providers, missing tools, fallbacks, approval needs, and confidence impact.
• project-onboarding-adapter now detects installed client targets and creates local-only client/MCP profiles.
• frontend verification rules now use a cost-aware verification ladder.
• README has been rewritten as a human-facing release/download guide and explicitly excluded from runtime policy.
• Release workflow now publishes stable and versioned client archives.
• validate_skill_pack.py now includes README-boundary validation.
• CHANGELOG.md and AUDIT_AND_OPTIMIZATION_PLAN.md now reflect the release-prep state.

Validation

Before merging the release-prep PR, the following checks passed:

• skill-pack-ci
• quality-ci
• link-check

The release workflow validates:

• schemas
• source bundle
• skill eval fixtures
• generated targets
• Codex target
• Claude target
• excluded paths
• release artifacts and checksums

Upgrade Notes

For existing projects:

1. Download the package that matches your agent client.

2. Replace reusable .agents/** bundle files.

3. Preserve local-only project overlays such as:

   • .agents/project/**
   • project/client-profile.md
   • project/mcp-profile.md

4. Keep root pointers small:

   • Codex / VS Code Codex / Cursor: AGENTS.md
   • Claude Code / VS Code Claude: CLAUDE.md

5. Run:

адаптируйся

6. Review any proposed local overlay changes before accepting durable project facts.

Do not copy host-project README/docs into reusable bundle rules, and do not let agents edit project documentation unless that documentation task is explicit and approved.

Known Notes

• GitHub Wiki pages still need to be published manually from the repo-side drafts in docs/install/**.
• A dedicated installed-bundle update skill was planned but not included in this release because the current implementation needs a cleaner follow-up pass.
• The VS Code Claude target uses the Claude-style root pointer shape.

Можно еще сделать короткую версию для GitHub Release textarea, если хочешь меньше текста.

WebDev Agent Kit v0.1.0

WebDev Agent Kit portable release artifacts.

Artifacts:

• Codex target: webdev-agent-kit-codex-${GITHUB_REF_NAME}.tar.gz
• Claude target: webdev-agent-kit-claude-${GITHUB_REF_NAME}.tar.gz
• Checksums: SHA256SUMS

Validation performed:

• schema validation
• source bundle validation
• skill eval validation
• portable target build
• Codex target validation
• Claude target validation
• excluded-path inspection