Release Notes / v1.0 → now

What's in.
What's next.

Skills, MCP integrations, and the roadmap for a design-first agentic stack. A living page — updated on every major evolution.

01 Skills by agent

EVE

Optional

Discovery · pre-Plan

  • user-researchLivePreps interviews & usability tests before the brief.
  • research-synthesisLiveTurns raw notes & transcripts into evidence, or a stakeholder restitution.
  • design-critiqueLiveChecks how competitors solve this — evidence, not opinion.

RAY

Live

Architect & Strategist · Plan

  • product-idea-captureLiveFrames raw ideas into a spec seed before a cycle starts.
  • architectureLiveWrites ADRs for T3 decisions that lock scope.

BOB

Live

Builder & Quality Director · Build

  • figma-useLiveReads & writes the Figma canvas via the plugin API.
  • figma-generate-designLiveTurns an approved direction brief into a Figma screen.
  • figma-generate-libraryLiveStands up design tokens & component sets.
  • figma-code-connectOptionalMaps Figma components to code components.

ANALYZER

Live

Product QA & CX · Review

  • design-critiqueLiveScores UX & design-system fidelity against the brief.
  • code-reviewLiveGrades technical quality before the verdict.
  • accessibility-reviewRoadmapA WCAG pass on shipped UI as part of CX scoring.

SHIP

Optional

Delivery · post-Review

  • deploy-checklistOptionalPre-deploy verification gate before going live.
  • scheduleRoadmapSchedules the release announcement.
02 MCP integrations

Available now

Figma

Bidirectional design context — read design into code and write to the canvas.

endpoint · figma-console-mcp / official Figma plugin

Penpot

A second design-tool path in design-workflow, parallel to Figma, not a degraded fallback — its own named execute_code recipes, every gap against Figma's tools named rather than faked.

endpoint · official Penpot MCP

Mobbin

600k+ screens from shipped, real-world products — searchable reference for ANALYZER & BOB.

claude mcp add mobbin --scope user --transport http https://api.mobbin.com/mcp

requires a Mobbin paid plan

Miro

Collaborative whiteboard — pull boards, frames & sticky notes as live references for EVE's discovery and RAY's planning.

endpoint · official Miro MCP

Roadmap

Refero

142k+ real product screens & 12k+ user flows with structured UX/UI metadata — concrete pattern research for BOB's aesthetic gate.

claude mcp add --transport http refero https://api.refero.design/mcp

requires a Refero paid plan ($17/mo) · documented as optional in BOB_aesthetic_gate.md, not yet connected

Savee

Curated visual moodboard library — search, boards & image analysis for palette and composition, ahead of BOB's aesthetic brief.

claude mcp add --transport http --scope user savee https://mcp.savee.com/mcp

requires a Savee paid plan ($9/mo) · documented as optional in BOB_aesthetic_gate.md, not yet connected

Custom inspiration MCP

A locally-curated index with semantic search over your own reference library.

status · planned

Any HTTP- or stdio-compatible MCP server works natively with Claude Code — the stack inherits it automatically.

03 Changelog
2026-06v1.0 — Initial release. RAY / BOB / ANALYZER / EVE / SHIP.
2026-06Figma MCP — bidirectional design-to-code bridge.
2026-06Mobbin MCP — shipped-screen references now connectable.
2026-08Miro MCP — boards & frames as live references for discovery.
2026-08Documentation site — a searchable docs hub at /docs: Get Started, the five agents & core concepts.
2026-08/pds — the adaptive conductor. One entry point that runs the whole cycle and adapts to your level (user_level): guided & pedagogical for newcomers, terse for experts. Bootstraps the context files, proposes pre-argued aesthetic directions, and glosses the vocabulary inline — so anyone, junior or expert, can use the stack.
2026-08Naming unification. RAY / ANALYZER are the canonical commands (/jo, /do kept as aliases); commands are now thin loaders over single-source system-prompts — docs and runtime finally tell one story.
2026-08v3.1 — What's in the repo is now what's in the box. npx pds-stack install was shipping a stack without its conductor, without BOB's Quality Brief gate, and without a single slash-command file — /ray simply did not exist on a fresh install. The installer now ships the conductor flow, the aesthetic gate, the design resources, and real entry points for Claude Code, Cursor, Gemini CLI, VS Code / Copilot and Codex CLI. A parity check runs on every publish and fails the release if any file a loader references is missing from the package.
2026-08The gates became mechanical. Two git hooks ship with the install: commit-msg refuses product code that carries no Ref: feature_<id> or whose spec is not VALIDATED; pre-commit enforces ADR-004, the line cap and tsc --noEmit. Thresholds come from your STACK.md, existing hooks are never overwritten, --no-verify still works. "No code without a validated spec" was a sentence nothing checked — now something checks it.
2026-08v3.4 — two gates that each worked, disagreeing. Pulse run 5 drove a full cycle in junior mode from three empty context files to a verdict. The interview holds: five beginner answers produced a usable vision and roadmap. Six findings came out of it, one unlike anything the earlier runs surfaced — an acceptance criterion, frozen at validation, forbade a colour literal outside the token map; the Quality Brief approved four steps later authorised one; the builder obeyed the brief and breached the spec, with both gates behaving exactly as specified. Nothing compared them. BOB now checks the brief against the spec's visual criteria and routes a conflict back to RAY, because changing a frozen criterion is a new cycle. The bootstrap's two other blockers are closed too, and a fifth release check verifies that a cited section actually exists in the file that carries it. Score ~92 → ~89: the drop is what the run was for.
2026-08v3.3 — one language, and a verdict band that stops lying. The conductor flow, the aesthetic gate and all 24 command loaders were French-only regardless of language_agents, so an English project got a French form to fill in. They are English now, like every other agent prompt — the dial controls what an agent says, never which file it reads. Fixing that surfaced a second problem: the band publicly named "SHIPPED WITH NOTES" describes a feature ANALYZER explicitly does not commit. Thresholds untouched; every surface now states the commit rule beside the band. Pulse run 4 confirmed the v3.2 fixes on a fresh install — the score is back to ~92/100 and all four release checks are green for the first time.
2026-08v3.2 — the self-test found two rules that had never been run. Pulse run 3 drove a full feature cycle through everything that did not exist at the previous run. Three axes went up: the spec gate refused a commit against an unvalidated spec on its first real use, and ANALYZER caught a defect the builder had self-passed by running the logic rather than re-reading it. The score fell from ~91 to ~86 anyway. /pds hunted for placeholder markers the shipped English templates did not contain, so two of three empty context files passed as complete — the exact failure the conductor exists to prevent. And the proof rule told BOB to run assertions "with whatever the project can already execute", which in a bundler-resolution project is nothing. Both fixed; the run, the findings and the lost points are published in full at /docs/self-test.
2026-08A fourth release check: detection rules must match their target. Existence checks could never have caught the conductor bug — both files shipped, both were packaged, they simply did not agree with each other. The parity check now verifies that the markers a rule looks for actually appear in the files it looks at, alongside the three passes it already ran: nothing referenced is missing, no template has drifted from its source, and the self-test benchmark still describes the current gates.
2026-08v3.1.1 — guards stop judging code you did not write. A fresh project without a .gitignore stages its own node_modules on the first commit, and the hooks were flagging ADR-004 violations inside dependency code. node_modules, .next, dist, build and friends are now excluded unconditionally. A guard that blocks on someone else's code gets switched off within a day, which costs more than the guard was worth.
2026-08The stack now tests itself, in public. _stack-test-pulse/ — two full feature cycles driven end to end through every gate, every agent signal transcribed, scored across 8 axes — is committed rather than living untracked on one machine. The parity check warns when a gate-defining file has moved since the last recorded run, so a stale benchmark is visible at publish time instead of being discovered later.
2026-08Proof over inspection. BOB now writes exactly one assertion per code-decidable acceptance criterion and runs it before handing over — no test framework installed, no coverage target. Criteria are reported proven, unproven or failed; "believed OK" is banned, because that is precisely the phrasing under which a real defect once reached QA. ANALYZER re-runs the assertions instead of trusting the output.
2026-08One front door. /pds is the documented way in — install ends on it, and the context files are filled by interview instead of by homework. The legacy aliases /jo, /do and /analyze are removed: three names for two agents cost every newcomer more than they saved. /eve and /ship ship as real commands with their modules.
2026-08Savee & Refero MCP — documented as optional visual/pattern research sources in BOB's aesthetic gate (Savee for mood & palette, Refero for real UX patterns & flows). Both paid, both inert until connected.
2026-09EVE re-wired to real discovery skills. EVE no longer routes intake through product-idea-capture — it hands off to user-research, research-synthesis and design-critique depending on the ask, and a stray V3 residue that sent EVE's handoff back to RAY instead of DIRECTION is fixed. RECHERCHE now sits upstream of DIRECTION and a HANDOFF step between PRODUIRE and JUGER, both optional, neither a gate.
2026-09Penpot MCP — a second design-tool path in design-workflow, live end to end (backend detection, connection precondition, named execute_code recipes run against a real file, review branch). Deliberately parallel to Figma rather than a fallback: its own vocabulary, and every capability gap named instead of faked — Penpot exposes one general-purpose execute_code tool where Figma's bridge ships typed ones, so the recipes are built and documented, not assumed.
2026-09v4.0 · V5 — the conversation is disposable. We measured nine real sessions of our own stack: 75 % of the spend was the main conversation re-reading its history (up to 669k of context), 20 % one agent looping with full-size screenshots. The prompts were never the problem. So the memory now carries the state too: each gate writes state_<feature>.md and /pds reprendre starts a fresh session from it. The direction runs on Opus at high effort; the build on Sonnet at medium, with a hard turn cap. Verification moves to text (Playwright CLI), the anti-slop floor becomes mechanical (impeccable), chat output is short by default, and npm run tokens measures it all. ADR-014 · ADR-015.
2026-09v4.1 — the fixes, proven one by one. Pulse run 6 was the first measured V5 cycle: ~87/100, main conversations down to 10 % of the spend (≤ 62k of context), but the build still at 78 %. The image guard lived in an agent's frontmatter, where it never fired for a subagent; it now sits in .claude/settings.json, scoped to bob-build, and a live run showed it blocking a Read inside the subagent. The installer now ships memory/, the scripts and their npm commands, without overwriting anything you already have. RAY cuts a spec into lots of ~50 turns, with a fresh builder per lot. A hand edit in components/ui/ is now a blocker. Waived rules stay waived, and the conductor routes changes instead of making them. Each fix was checked with its own targeted test rather than a full pulse re-run.
2026-09v4.1.1 — a guard is proven by running it. check-parity no longer looks for the image guard's text: it executes the hook on three cases and fails the release if bob-build can read a PNG. The installer drops a question that only two of four brief types could answer. Then a validation cycle on a fresh install, Sketch lane, measured with npm run tokens: 2.1M tokens read in total, bob-build in 19 turns and 43k of context at most, the main conversation at 98k — no ADR-014 target exceeded.
2026-09v4.3 — a context thermostat. The relay threshold stops being a rule someone has to remember. A UserPromptSubmit hook reads the context size on every prompt: at 120k it asks Claude to finish the step and propose the relay, at 150k the relay comes first. You see one line in the chat; your prompt is never blocked. check-parity runs it against three fake transcripts. (ADR-018)
2026-09v4.2 — session 0, then one session per agent. /pds cadrer gathers a multi-feature project once — target, constraints, references — without crossing a gate, and writes a single project_<name>.md of about 60 lines plus a state file per feature. It ends on a dispatch plan: a title and a prompt to paste for each session, which you open yourself. A conversation that runs long is relayed to a fresh session through the state file, never forked — a fork carries the whole history. (ADR-017)