Releases: wp-media/maestro
v0.6.2 — Transplant overhaul
What's new
/maestro:transplant — major improvements
Auto-discovery — the component table is gone. The analyst now scans agents/ and skills/ at runtime. Any new skill or agent added to Maestro is automatically included in the next transplant — no manual table update needed.
Inline KEEP_AS_IS — trivial file copies are handled by the orchestrator directly via sed, not by spawning writer agents. Writers only receive ADAPT, REWRITE, or MERGE work. Faster, cheaper runs.
Standalone skills transplanted — groom, challenge, qa, review are now included in every transplant.
.aiassistant/ support — if the target project has a legacy .aiassistant/ directory, the analyst reads it for project-specific customizations before making disposition decisions. The directory is deleted after the transplant completes (prompted, default: yes).
AGENTS.md / CLAUDE.md awareness — the analyst reads project AI convention files before assigning any disposition. Their content is authoritative over codebase inference for tool names, test commands, and workflow rules.
.claude/bin/ — workflow scripts centralised — dev-start.sh, dev-seed.sh, dev-down.sh now live in .claude/bin/ instead of bin/. Everything workflow-related is in .claude/.
Orphan detection (upgrade mode) — components present in the last transplant manifest but no longer in Maestro are surfaced as REMOVE candidates for user confirmation.
Quality of life — chmod +x .claude/bin/*.sh runs automatically; mkdir -p is derived dynamically from dispatch_plans; next-steps list is one step shorter.
Migration note
Manifests written by v0.6.1 and earlier (version "2") are auto-migrated on the next upgrade run — bin/dev-*.sh output paths are rewritten to .claude/bin/dev-*.sh before any diffs are computed. Legacy bin/dev-*.sh files in the target project are removed automatically when re-transplanting.
v0.6.1 — Standalone pipeline skills
What's new in v0.6.1
Standalone pipeline skills
Four new skills let you invoke individual pipeline steps directly, without running the full orchestrator. Each skill gathers its own context, runs the agent, and prompts before posting anything to GitHub.
| Command | What it does |
|---|---|
/maestro:groom <issue> |
Groom a single issue — produces a spec, optionally posts the summary |
/maestro:challenge <issue> |
Adversarially review the grooming spec before implementation |
/maestro:review [PR] |
Lead code review on the current branch or a given PR |
/maestro:qa [PR] |
QA validation — boots the environment, tests every acceptance criterion |
All four underlying agents (grooming-agent, challenger, lead-reviewer, qa-engineer) are unchanged — the standalone/pipeline distinction is handled entirely through invocation context.
Breaking changes in v0.6.0 (included in this release)
commands/ has been replaced by skills/ throughout. Each commands/<name>.md is now skills/<name>/SKILL.md. orchestrator/workflows/ is now orchestrator/runners/.
Transplanted projects (those generated by /maestro:transplant) using the old commands/ layout will need a path update pass — new transplants generate the skills/ layout automatically.
v0.5.1 — Transplant: config-free workflow & QA pass
What changed
Transplant pipeline — breaking improvements
Config file eliminated. Transplanted workflows no longer generate or depend on maestro.json. The orchestrator now has a ## Project Config block with values baked in at transplant time (REPO, TEMP_ROOT, BASE_BRANCH, test/lint/boot commands). Agents receive everything via prompt injection — no file reads at runtime.
TEMP_ROOT is now project-specific. The project interview asks for the working temp directory name (default .ai, e.g. .TemporaryItems). The analyst also checks .gitignore for existing patterns.
Universal de-brand pass. Every transplanted file (regardless of disposition) now strips Maestro config reads, maestro.json path references, and Maestro branding from prose.
QA pass added (Claude Opus). After all writer agents complete, a transplant-qa agent runs 6 checks — no Maestro references, no runtime config reads, orchestrator has constants block, WP content stripped for non-WP projects, stack consistency, and FIXME scan. A red verdict triggers targeted re-work on flagged components (max 2 rounds).
Cleanup prompt added. After QA green light, users are asked whether to remove upgrade artifacts (transplant-refs/, manifest, context doc). Choosing yes leaves a fully self-contained workflow with no Maestro footprint.
Upgrading
Re-run /maestro:transplant <target> on any previously transplanted project to get the config-free layout. The upgrade path will detect and merge your team's customizations.
v0.5.0
Full Changelog: v0.4.3...v0.5.0
v0.4.3
v0.4.2
v0.4.1
What's changed
New
/maestrohub command — type/maestrowith no arguments to get an organised map of all available commands, grouped by purpose. Great entry point for new users.
Improvements
- Command renames for clarity:
sprint-planner→sprint,wordpress-compliance→compliance,po-changelog→changelog - All 13 command descriptions sharpened to clean single lines — reads well in the autocomplete picker
- README rewrite — orchestra voice restored ("One baton. Any orchestra."), universal positioning (works with any project, not just wp-media), Podium reduced to a one-liner
- Cross-references updated in
backend-agent,frontend-agent,lead-reviewer,docsto use new command names
Removed
bin/auto-update.shand itsSessionStarthook — redundant now that the Claude Code plugin marketplace handles updates natively
v0.3.1.1
v0.3.1
What's changed
Fix
- pr-agent: fall back to the Maestro-bundled PR template (
.claude/commands/issue-workflow/refs/pr-template.md) when no project-specific template is found at.github/refs/pr-template.md. Previously, Step 5 carried a hardcoded structure that diverged from the actual template.
Podium v0.3.0 — Zero-token agent observer dashboard
What's new in v0.3.0
🔭 Podium — Agent Observer Dashboard
Maestro now ships with Podium, a real-time observability dashboard for every Claude Code session. Built on a WP Media branded fork of Claude Code Agent Monitor (MIT).
Zero token cost. Hooks fire at the harness level — outside the LLM turn, no context consumed.
Quick start:
```bash
/podium setup # register hooks (once per project, restart CC after)
/podium start # http://localhost:4820
```
What it shows:
- Live agent activity with readable session names (from your first typed message)
- Real-time tool feed — every Bash, Read, Write, Agent call as it happens
- Analytics — token usage by model, tool frequency, cost tracking across all sessions
- Conversation view — reads JSONL transcripts, shows your full chat history
- Health metrics — runtime, storage, cache rate, subagent effectiveness
Design: WP Media gold #FED23A accent, warm linen light theme, navy dark theme, light/dark toggle persisted to localStorage.
🪝 Zero-token hook system
podium/hook.mjs— fires on all Claude Code events, POSTs to dashboard in <1spodium/install.mjs— registers hooks in.claude/settings.json- Works on any project — detects TEMP_ROOT from
maestro.jsonor falls back to.maestro/
📋 New skills
| Skill | What it does |
|---|---|
/podium setup |
Register Claude Code hooks (once per project) |
/podium start |
Start dashboard at http://localhost:4820 |
/podium stop/restart/status/logs |
Manage the server |
/maestro:onboard-project |
One-command project setup — generates maestro.json, installs hooks |
🚫 HTML log disabled by default
"html_log": false is now the default in maestro.json. The token-expensive workflow-log.html is skipped — Podium replaces it. Set "html_log": true to restore the legacy HTML log.
Breaking changes
None. All existing projects continue to work. Podium is opt-in via /podium setup.
Credit: Podium dashboard UI is a fork of Claude Code Agent Monitor by Son Nguyen (MIT). All WP Media branding, Maestro integration, and design improvements are original.