Repository navigation
Releases
Orchestra 5.5.0
A skipped check is no longer a pass.
- A route miss has confidence 0. Doctor and verify print NOT_RUN when the check did not run.
- Memory rows store a sha256 of the task text.
orchestra memory scrubrewrites an older raw task field. - Codex is a default host. The Antigravity CLI skill path is opt-in.
- Browser regression loads
ow-qa. Browserize and author stay commands. - The Go suite on this commit's tree was 795 passed in 26 packages. The paired eval was not run. The claim stays none.
The 5.0.0 tag stays on 24ad53c.
Orchestra 5.0.0
Site recon, app recon, scrape, UI assembly, waves, and MCP audit are routes. The eval did not run a live model trial, so none of those skills is a measured win. orchestra mcp-serve is experimental. The claim stays none.
Orchestra 4.7.0
A prepared handoff is not a passing build. A skipped visual check is not a pass. orchestra verify exits non-zero on a hard failure.
The route table is docs/ROUTING.md. Adapters stay short. Hooks warn. Sync refuses a symlink and writes Claude Code MCP servers to ~/.claude.json.
No comparison claim is enabled.
Orchestra 4.6.0
The pin is 4.6.0. Spec review is egm-draft, egm-review, and egm-resolve. unslop, build-awwwards-quality-sites, floorplan, anatomy, webapp-testing, and mcp-builder keep their upstream names. Sync moves a retired skill name into a backup. Plugin manifests live at the repo root. Mermaid fences in the README and docs are parsed in CI. The logo is gone.
The landing record was rerun on composer-2.5-fast. The Orchestra arm had AGENTS.md. Desktop Lighthouse performance was 100, 99, and 99 on the raw pages and 99, 98, and 99 on the Orchestra pages. Accessibility was 100, 96, and 100 raw, and 100, 100, and 100 on Orchestra. Axe found landmark-one-main and region on the second raw page and none on the Orchestra pages. A blind judge called the first trial a tie and preferred the Orchestra page on the other two. Both arms could see the global skills. compare.go stays disabled.
CI: https://github.com/mahik504/orchestra-workflow/actions/runs/42466472744
Hygiene: https://github.com/mahik504/orchestra-workflow/actions/runs/42466472903
Orchestra 4.5.0
The pin is 4.5.0. Orchestra-owned skills are ow-*. Upstream names stayed. Sync hashes a skill and leaves a local edit in place. orchestra compile prints the route as JSON.
Three bugfix trials on composer-2.5-fast tied: both arms changed Add from a + b + 1 to a + b, and go test passed. Three landing trials both built. Layer A captured desktop and mobile, axe reported no violations, and Lighthouse accessibility was 100. Performance was 97, 98, and 99 on the raw arm, and 93, 100, and 99 on the Orchestra arm. A blind judge preferred the Orchestra page twice and called the third a tie. No comparison claim is enabled.
CI: https://github.com/mahik504/orchestra-workflow/actions/runs/42411579261
Hygiene: https://github.com/mahik504/orchestra-workflow/actions/runs/42411579260
Install is in INSTALL.md. The default mark is assets/brand/ow-mark.svg. GitHub cannot set the social preview from the API. Upload assets/brand/social.svg yourself after exporting a PNG.
Orchestra Workflow 4.1.0
4.1.0 keeps the eight-stage conductor and names only the skills, MCPs, plugins, and libraries a task loads. Design domains and technology rules sit on the existing routes. Migration and rollback are in kit/ROLLBACK.md.
Orchestra 3.3.5
Changelog
[3.3.5] - One contract, open skills, kind-based design bar
The pin in VERSION is 3.3.5. This is not a claim that output beats a raw host. 3H has not scored a website.
- One contract in
AGENTS.md. Prompt-forge is step 0. Projectkindsets the design bar. - Skills are the folders in
skills/.skills-local/stays off git.superpowers-planningandunlazyare removed from the source set. - Engineering skills: systematic debugging, test-driven development, git worktrees, verification before completion, writing skills, Ponytail.
- Hosts sync from
kit/sync-ides.ps1. The doctor checks version, skill presence, and the private-vault heartbeat schedule without printing secrets. - Graphify on one local code graph scored 7 of 10 file-finding questions. It is not the code map. fallow's Windows download did not produce a runnable binary.
- Rollback tag for the pre-change tree is the local tag
pre-3.3.5. Public rollback remainsv3.3.1until this tag is pushed.
[3.3.2-3g] - Mechanisms through the measurement pilot
Prerelease. The contract pin stays 3.3.1. This is not the stable 3.3.2 release.
- Mission, worker capsule, derived budget, failure classes, and recovery around the existing eight-stage engine.
- Layer A Playwright on a running app. Layer B only on those captures. Flagship work gets a few lightweight directions, then one contract.
- Capability selection is the minimum set the stages consume. Brain notes do not become the run source of truth.
- T1 and T3 host pairs.
claim: none. The clean T3 pair made the same 183-bytesum.gochange and Orchestra was slower. That fixture does not score a website. - Phase 3H (product-class comparison) and phase 3I (
VERSIONbump and stable tag) are not done. - Rollback to the governance release remains
v3.3.1.
[3.3.1] - Governance: visual evidence gate, named-reference mode, four routes
Same OS. One conductor. Not a bigger toolbox. Version is 3.3.1.
- Design Lab has two locks:
CONTRACT_APPROVED(lockedDESIGN.md) thenAPPROVED(golden stills). Product UI stays refused until stills. ADESIGN.mdis not visual evidence. - Named visual references take exactly 3 evidence-backed translations. Open-ended PREMIUM still uses 23 cards. Pasted contracts may skip the survey; named sites may not.
- Four capability routes + four job-loaded skills:
visual-forensics,visual-golden-still,fresh-context-review,final-showcase. Pack:templates/packs/premium-portfolio.md. - Resource status enum extended (
PROJECT_ONLY,CATALOG,EXPERIMENTAL,PRIVILEGED,AUTH_REQUIRED,OBSOLETE). Named candidates reclassified only. No GSD/ECC/MCP bulk install. - Docs:
protocols/VISUAL_GOVERNANCE.md,docs/3.3.1/. Keep / Rework / Kill / Park.
[3.3.0] - Richer graph, four hosts, two-stage Design Lab, free-resource honesty
Evolve 3.2. Same OS. Curated 40 skills. Contract stays open. Version stays 3.3.0.
- Four-host sync: Cursor, Antigravity, Claude Code, jcode (
~/.jcode/skills). Allowlist fromregistries/host-stack.json. - Design Lab: 23 short cards → one full
DESIGN.md. Named override outranks the graph. Custom paste wins. - Catalog: GetLayers
PURCHASE_PENDING. Refero MCP AUTH_REQUIRED / PRO; Refero public pages FREE. 21st OPTIONAL_FREE_MCP (2 installs/day). Magic UI FREE MCP + OSS. Tailkit MCP PAID; public Tailkit RESOURCE. - 13th capability
interaction-components(one route across kits, not a skill per library). Tint/adapt onreverse-engineering. - Free research: Godly, Land-book, Lapa, SaaSFrame. OSS 3D: three.js, R3F, drei, postprocessing, scroll-rig. Prompt catalogs as references. devops-skills quarantined (no 88 dump).
- Docs: RESOURCE_INVENTORY A–Q, RESOURCE_HEALTH, DESIGN_RESOURCE_ROUTING, manual-setup 9-field, eight architecture notes plus three committed README SVGs (GitHub Mermaid treats
graph/clickas keywords). - Global skills added (wrappers, not dumps): refero-design, hallmark, unlazy, react-best-practices, gsap-core, supabase, postgres-best-practices, diagram-generator, pretty-mermaid, skill-security-check.
- Job packs in
templates/packs/. README states architectural facts — no fake A/B, no invented /10 scores. - Rollback:
v3.2.0(9d6900d) and the 3.1 zip. Uncommitted 3.2 hygiene (YAML stubs, Go module rename) lands here.
[3.2.0] - Intelligence planes on the same OS
Evolve 3.1. Do not rebuild. Keep the 30 skills. Contract stays open.
- Research: Scrapling is the primary public crawler; Firecrawl on-demand; Crawl4AI catalog fallback. Playwright stays on our app.
- Design: Designer SOS (S.A.V.E., six-pass, Keep/Kill/Park, Vault/Live/Anatomy); named source tiers; GetLayers on premium/3D/shader only.
- Quality: Unlazy-style verification ledger; Taste → DESIGN.md → Impeccable → Playwright → SOS → one consultant packet.
- Catalog (not global install):
playwright-cli,crawl4ai,serena,jcode. Optional telemetry fields on rows. Cheap trigger matching inAGENTS.md. - Ingest: graph miss → research → one named capability → wait for update.
templates/custom-skill.md. - Executors: jcode as a fifth host adapter under
AGENTS.md. Does not replace Cursor / Antigravity / Claude Code. - Taste: on-disk Stitch wrapper vs upstream v2 experimental — do not
npx skills addthe pack. - README rewritten for the evolve story. Overlay + Antigravity adapters identify as 3.2.0.
- Post-tag hygiene (working tree after
v3.2.0/9d6900d): drop duplicate YAML catalogs; rename Go module togithub.com/mahik504/orchestra-workflow/runtime; genericize personal from-strings. - Catalog: Soup is CURATED_OPTIONAL for ML fine-tune jobs (not a 31st skill). Anti-Slop AJ + blader/humanizer + Graphify catalogued; none dumped into the 30.
- jcode on this operator PC: OmniRoute default;
auto/best-coding/auto/best-reasoning/auto/best-fast/auto/best-freealiases. Barejcodeis enough after PATH.
[3.1.0] - Control plane
One contract. Design Lab is a gate. Evidence-first completion.
- Overlay, Cursor, Antigravity, and Claude Code share the same 3.1.0 rule.
- Public registry stripped of personal catalog rows. MCP health is explicit.
orchestra doctorwarns if Antigravity science or data-agent-kit plugins are Global.
Routing. The classifier was a keyword stub returning task-stub-001; it now scores every
capability row against the brief and produces a structured re-brief (type, archetype, quality bar,
platform, research depth, verify depth, hard constraints). Every row in
design-resource-graph.json gained trigger_conditions, skip_conditions, quality_bar, and
risk_rank — a route that can never decline itself is a default in disguise. Declined routes are
reported with the skip condition that fired. Two close routes produce exactly one question; silence
takes the lower risk_rank and logs assumed <capability>, no response.
Design Lab is a lock, not a warning. verify.DesignLab blocks writes to anything a browser
renders while the gate is PENDING, including on dry runs, which previously slipped past the
synthesize halt. Directions must number 2 or 3 and name a source for typography, colour, and motion.
Rejections persist to .orchestra/design-lab/rejected-directions.json, fingerprinted by stack, so a
renamed rejected direction cannot be re-offered. Bypass is allowed but requires a note.
See protocols/DESIGN_LAB_PROTOCOL.md.
No second conductor. Deleted the unreferenced internal/kernel and internal/planner packages,
which duplicated the pipeline's approval and execution logic.
Isolation. A fresh clone keeps resource memory at .orchestra/memory/resource-memory.json.
The public method repo does not ship a live memory/resource-memory.json. Set ORCHESTRA_HOME
to a private Brain so overlay and outcomes resolve there instead of the clone.
Phase 7 evidence. Twelve capability rows. research-paper loads orchestra-docs and does not arm the Design Lab. Restaurant / school SaaS / 3D portfolio resolve to three different resource graphs. A premium brief that names no libraries still surfaces research, design, and motion. pixi.js is reported as unknown technology rather than forced into a route. References acquire on_demand via web_fetch; GSAP is project_scoped_install; global npm stays blocked. Live fixture screenshots at 1440×900, 768×1024, 390×844; click wrote a hold; heading contrast 14.43:1; no horizontal overflow.
User-added resources. orchestra add --intent "..." inspects a URL, infers kind
(skill / dependency / MCP / plugin / subagent / reference / adapter), and writes a Brain overlay
at memory/added-resources.json. It does not edit registries/resources.json or the capability
graph. Matching tasks activate the overlay row; non-matching tasks skip it. Recorded outcomes in
memory/resource-memory.json update overlay routing: last failure suppresses auto-activation.
That loop is experience → evaluation → memory → routing update → future selection. It is not
reinforcement learning. orchestra lifecycle prints the 15-step proof for one URL.
README. Rewritten against the code. The previous version cited capability and domain names that
do not exist in the graph, and described research as querying live design indexes when
ResearchCoordinator defaults to offline fixtures. CI and Hygiene badges sit on the README.
The eight-stage architecture diagram from ARCHITECTURE.md is inlined there.
Doctor. Playwright is checked as npx playwright --version (split argv). Missing Playwright
is [OPTIONAL] — visual jobs use a project install or the agent browser. Global npm stays blocked.
Engine install. kit/install-local-engine.ps1 / .sh persist ORCHESTRA_HOME as a user env
and go install orchestra onto ~/go/bin. Chat still uses the markdown contract without the
binary. docs/getting-started.md treats classify / plan as optional.
Catalog. build-your-own-x rationale is a public bookmark se...
Orchestra 3.3.2-3g
Engineering prerelease through the measurement pilot. VERSION is still 3.3.1. claim is none. The clean T3 pair made the same 183-byte sum.go change and Orchestra was slower. 3H and the stable v3.3.2 tag are not done. See CHANGELOG.md.
Orchestra 3.3.1
Governance release. Visual evidence gate, named-reference mode, 17 routes, 44 skills. Rollback target for the 3.3.2 prerelease.
Orchestra 3.3.0
Orchestra 3.3.0 keeps one control plane and four host adapters.
- Richer capability graph (13 routes, trigger + skip). interaction-components searches kits; it is not a skill per library.
- Free resource integration: Magic UI OSS + free MCP, 21st catalog (2 installs/day after login), Refero public pages, Tailkit public pages, Aceternity public categories, OSS 3D stack, prompt catalogs as references.
- Four-host sync (Cursor, Antigravity, Claude Code, jcode).
- Design Lab unchanged: 23 cards → one DESIGN.md. Tint keeps structure.
- Research, repo intelligence (Serena optional), quality/verification, diagrams, resource provenance.
- GetLayers remains PURCHASE_PENDING. Refero MCP and Tailkit MCP are AUTH_REQUIRED. Do not treat configured as HEALTHY without host confirmation.