v1.5.1 — Case-Insensitive Figma Steering Discovery

Bugfix

Fixes case-sensitive Figma steering file discovery that missed mixed-case filenames (e.g., feature-CSI-FS-001A-figma.md).

Changes

• criteria/figma-design-match.md — Applicability check uses grep -i figma instead of case-sensitive glob
• agents/pr-reviewer/agent.md — Same case-insensitive fix for steering file discovery
• agents/pr-orchestrator/agent.md — New Step 2b pre-reads Figma steering files and injects content into the reviewer prompt, eliminating file discovery failures

Upgrade

npm install -g @yepengfan/agent-registry@1.5.1

v1.5.0 — Automated Figma-DOM Design Verification

What's New

Replaces screenshot-based Figma design verification with automated element-level extraction and diffing. Every meaningful visible element is extracted from both Figma and the rendered DOM, then compared property-by-property with tolerance thresholds — catching mismatches that screenshots miss.

New Scripts

• scripts/figma-extract.js — Figma Plugin API extraction (browser context, placeholder-based)
• scripts/dom-extract.js — DOM getComputedStyle extraction for Playwright browser_evaluate
• scripts/design-diff.js — Node.js CLI for deterministic element mapping + tolerance-based diffing

New Skills

• figma-inspect — Extract a complete element inventory from a Figma screen
• design-verify — Compare a Figma inventory against a rendered page

Updated

• figma-design-match criterion — 4-phase algorithm: Figma inventory → DOM inventory → Map+Diff → Structured report
• pr-reviewer — Uses new skills, outputs structured figma_value/dom_value/fix_hint per design issue
• pr-fixer — Consumes structured mismatch data with actionable fix hints
• pr-orchestrator — Updated workflow example format

CLI

• New agent-registry root subcommand for agents to locate the registry installation directory
• scripts/ added to npm package files array

Install / Upgrade

npm install -g @yepengfan/agent-registry@1.5.0

v1.4.2

Fixes

• Increase max_rounds multiplier from 3x to 10x required_clean (default 9 → 30)
• Add postinstall hook to auto-install agents on npm install

v1.4.1 — README on npm + Criteria column

Fixes

• README.md now included in npm package — was missing from v1.4.0 due to omission from files list
• Agent table updated — added Criteria column showing which quality gates each agent evaluates
• Updated descriptions — pr-reviewer (two-pass Figma verification) and pr-orchestrator (review-fix loop, auto-detect PR, --rounds N)

v1.4.0 — Rendered Screenshot Verification & Review-Fix Loop

What's New

Two-Pass Rendered Screenshot Verification (#14)

• Pass 1 (PRIMARY): Uses Playwright MCP to capture actual rendered pages from localhost, then compares side-by-side with Figma screenshots — catches i18n value mismatches, data overflow, DS component state gaps, and CSS specificity overrides that code reading cannot detect
• Pass 2 (SUPPLEMENTARY): i18n value verification (actual string values, not just keys), runtime data awareness (API response types vs column widths), DS component state verification
• Graceful fallback: When Playwright can't access the page, falls back to code-only with pass: false and recommends manual visual QA
• Severity escalation: Suggestions flagged 3+ rounds without fix get escalated to must-fix
• 5 new mismatch patterns: i18n text mismatch, data overflow, DS component state gap, CSS specificity override, persistent suggestion decay

Review-Fix Loop (#14)

• pr-orchestrator loops until N consecutive clean runs (default 3, configurable via --rounds N)
• Auto-detect PR: When no PR number is given, auto-detects from current branch
• Environmental blocker handling: Fixer-unfixable issues (MCP unavailable, auth blocked) trigger early exit with manual QA recommendation instead of wasting rounds
• Prior Round Context: Accumulated suggestions passed to reviewer across rounds for severity escalation
• Max rounds safety cap: required_clean * 3 (default 9) prevents infinite loops
• Inline PR comments mandatory: Reviewer always posts inline comments as authoritative issue record

Documentation

• README updated with criteria tables, profile tables, and updated agent descriptions

Stats

• 20 review rounds on PR #14, 24 issues found and fixed
• 72 tests passing
• 10 criteria, 2 profiles

v1.3.0 — Task-Aware Criteria & Figma Design Verification

What's New

Task-Aware Criteria with Repo Profiles (#12)

• Repo profiles (profiles/frontend.md, profiles/backend.md) define detection rules and map task types to criteria
• Task type detection — automatically detects feature/bugfix/refactor from PR branch prefix and title conventions
• 7 new criteria: has-regression-test, has-test-coverage, no-behavior-change, no-accessibility-regression, figma-design-match, no-breaking-api-change, has-migration-safety
• lib/profiles.js module with detectTaskType, loadProfiles, detectProfile, resolveCriteria
• Criteria override support: --criteria +name, --criteria -name, --criteria name1,name2
• Zero breaking changes — repos without profiles fall back to reviewer's default criteria

Figma Design Verification for PR Reviewer (#13)

• Figma verification workflow step — conditional on .sdd/steering/feature-*-figma.md files
• Figma MCP + Playwright MCP for side-by-side visual comparison (Figma screenshot vs rendered localhost)
• Design Fidelity checklist with 7 common mismatch patterns (container widths, button grouping, missing elements, section ordering, conditional elements, component styles, spacing tokens)
• Design-specific severity — design must-fix issues count toward the zero-must-fix-issues gate
• Frontend Design Conventions — DS component usage, Figma compliance, layout rules

Criteria System (#11)

• Criteria framework for measuring agent work quality
• Gate vs advisory distinction — gates block merge, advisories are reported only
• criteria/ directory with markdown-based criterion definitions
• Criteria injection into agent prompts at install time

Security

• Path traversal protection for criteria names in profile files and --criteria overrides

Stats

• 10 criteria (9 gate, 1 advisory)
• 2 repo profiles (frontend, backend)
• 72 tests passing

v1.2.0 — Composable Behavior System

What's New

Composable Behavior System

Agents can now equip discipline rules via frontmatter. Behaviors are defined once in behaviors/ and injected into agent prompts at install time — no runtime dependencies.

6 Built-in Behaviors

• verification-gate — Run verification after changes, only commit on pass
• evidence-based-claims — Never claim success without fresh evidence
• no-blind-trust — Verify findings before acting on them
• independent-output-verification — Verify sub-agent outputs independently
• safe-revert-on-failure — Revert changes that fail verification
• structured-pushback — Push back on incorrect findings with reasoning

New CLI Command

• agent-registry update — Reinstall all agents to pick up behavior/prompt changes

Agent Behavior Assignments

• pr-fixer: 5 behaviors (maximum discipline for code changes)
• pr-orchestrator: 2 behaviors (evidence-based claims + output verification)
• pr-reviewer: 1 behavior (evidence-based claims)

Full Changelog

• Frontmatter parser validates behaviors field
• Installer injects behavior content between <!-- behaviors:start/end --> markers
• Discovery shows behaviors in list and status commands
• README updated with full behaviors documentation
• 100 tests passing (55 frontmatter + 45 integration)

v1.1.1

What's Changed

• Fix npm publish: restore registry auth with NPM_TOKEN by @yepengfan in #7

Full Changelog: v1.1.0...v1.1.1

v1.1.0

What's Changed

• Add repository URL for npm provenance by @yepengfan in #8

Full Changelog: v1.1.1...v1.1.0