gemba

               

⭐ Star us on GitHub — it helps people find this project and tells us we're solving a real problem.

       

In lean manufacturing, gemba (/ˈɡem.bə/ 現場) is "the actual place" — the factory floor, where real work happens. A gemba walk is when leadership observes the work directly, not through reports, and leaves actionable feedback as they go. Gemba is built around that metaphor.

Gemba features Kanban like planning, execution, and management of complex vibe coding projects in a single pane of glass. Organize, review, and dispatch large, parallel batches of work and monitor real progress towards milestones, all with built-in coaching and drift detection. Start a new project with a guided conversation or import your existing work — see Getting Started for a quick start.

Table of Contents

• 🚀 About
• 🧱 Architecture
• 🎯 Two-Axis Work Planning and Dispatch
• 🚦 Status
• ✨ What's New
• 📸 Screenshots
• 📦 Project Layout
• 🏁 Getting Started
• 🛠️ Development
• 📚 Documentation
• 🤝 Feedback and Contributions
• 📜 License

🚀 About

Gemba was born out of frustration with the current state of vibe coding. Massively parallel, "headless" agentic software development was technically possible but it was cumbersome. Before Gemba, a single app to manage the process using concepts and terms familiar to developers did not exist - concepts like milestones, epics, and Kanban planning to tie them together.
Systems like Beads made it possible for agentic pipelines to discover, claim and report work progress but these systems were not functionally integrated with execution — ordering, dispatching, and monitoring the work.

This left developers to face several obstacles in projects of any appreciable scale:

• Planning is disconnected from execution — issues in Jira, Beads, or Todo.txt, agents in a terminal, output in git.
• Priority is invisible. The dep graph says what can run; nothing says what should run next.
• Expertise is siloed — a side-of-the-desk project needs product judgment, architecture, UX review, QA gating, release mechanics, security, and the operator doesn't have time to personally don each hat.
• State is fragile — git, Beads, LLM sessions, and sidecar artifacts each move on their own timeline; "undo back to yesterday" doesn't exist.
• Orchestration tooling is tied to one runtime — playing nicely with one tracker or one agent framework means a rewrite when either moves.

Gemba addresses all five: a browser-based UI for walking the floor of an agentic project, seeing the work at the right grain, directing it, with a roster of configurable LLM specialists a click away for the expertise the operator doesn't have time to personally deliver for every decision.

In effect, Gemba creates a "side of the desk" experience for running vibe coding projects, meaning complex multi-milestone projects spanning weeks can be tackled while minimizing the cognitive load and freeing you up to keep the creative flow going. It keeps you in the driver's seat for what remains critical - high level planning, review, and course correction - while allowing fully automated agentic software development to take care of the execution.

The only hard requirement is a data plane. Beads fulfills that out of the box, so the minimum working deployment is gemba serve and a browser pointed at it — no orchestrator, no scheduling infrastructure required. For teams that only need Beads viewing and management, gemba serve --beads-only starts a deliberately smaller mode: no project, GitHub, or agent runtime required. It opens the Beads board in a Flat list by default, keeps Cascade and Graph views available, and records create / edit / delete / state-change actions in a Beads history ledger. Add --beads-read-only when the same views should be inspection-only: Gemba shows a Beads-read-only status pill and rejects every mutation before it reaches the Beads adaptor. For the lowest-friction trial path, the self-contained Docker quickstart image builds Gemba plus bd, seeds the sample project, and starts Beads-only mode with one command: make quickstart-run. For a real unseeded container, make docker-run builds the standard Docker image with Gemba, the sentinel CLIs, bd, git/ssh, and mounted /data + /work volumes. Native terminal orchestration is bundled so operators who want agent sessions don't need to install anything extra: gemba serve --orchestration=native drives tmux / iTerm2 / Terminal.app sessions directly, surfaces permission prompts and HITL requests in the SPA, and correlates bd mutations back to the session that made them. Gemba also supports existing orchestration solutions like Gas Town, which can picked at launch using --orchestration=<name> when you need its specific scheduling or isolation semantics.

🧱 Architecture

Plane	Required	Default	What it gives you
WorkPlane	✅	Beads (bd)	The data — work items, sprints, evidence, DoD
OrchestrationPlane	❌	Native (tmux / iTerm2 / Terminal.app)	Agent sessions, HITL, dispatch, escalations

• WorkPlane is required. Gemba won't boot without one — the SPA has nothing to render.
• OrchestrationPlane is optional. Running without it is a supported mode: Gemba becomes a human-driven Kanban over the WorkPlane. Starting sessions, surfacing escalations, and agent dispatch light up when an adaptor is wired. Use --beads-read-only when the WorkPlane itself should reject writes.
• Native is the default OrchestrationPlane. gemba serve --orchestration=native is the happy path and needs no external daemon; it drives terminal sessions directly.

Gemba's ranking layer has two distinct systems for distinct jobs. Selection (internal/planner/selection/) is a pure-Go dispatch-time scorer that decides which ready bead a session should take next; it runs every dispatch loop and powers the /coach affinity grid. epic_order (the PM persona's skill) is an LLM-driven planning consult that ranks candidate epics for sprint composition; it lives at /sprints and produces narrative recommendations with confidence scores. See Dispatch vs Planning.

⬆ back to top

🚦 Status

Milestone 3 — Native orchestration shipped (April 2026). Gemba runs end-to-end against a Beads rig with native terminal orchestration out of the box; work items, escalations, and session state all round-trip through the SPA. The Gas Town adaptor is optional.

Active work lives in the project's Beads rig (bd list). Top-level open epics include token spending management (gm-root.14), the UI/SPA build-out (gm-e12), and cross-cutting features (escalation surfacing, evidence v2, DoD v2 — gm-e11). Design docs capture the durable architectural decisions; adaptor authoring references live in Adaptors.

🏁 Getting Started

New project (starting from scratch)

The primary path for operators starting a new project. Install Gemba, run the server, and open a browser — Gemba redirects to /new on first run, which lands on /board and opens the unified Create-project modal:

make build                                               # or: brew install GembaCore/tap/gemba-core (once taps ship)
./bin/gemba serve
# -> http://127.0.0.1:7666  (redirects to /new on first run)

The Create-project modal scaffolds a new project (directory + git init + .gemba/workspace.toml + beads database + initial commit) in one step — no LLM required. After ratify, the new project becomes the active workspace and you land on the empty board, ready to add work items by hand or via the optional Plan with the Onboarder → CTA (only shown when an LLM client is configured; see Configuration). The Onboarder is a conversational planner at /onboard that emits a fully-populated Milestone → Epic → Bead tree in one ratification.

You can start a new project at any time — even from an existing workspace — using the + affordance next to the project picker in the top bar.

Existing beads databases. If Gemba finds a beads database that isn't yet bound to a Gemba workspace (no .gemba/workspace.toml, or no git repo), the project picker labels it with a needs setup badge. Click it and choose either Create a new git repo here or Move it into an existing git repo.

Advanced: importing existing work

If you have an existing Jira project, Beads workspace, or source-code repo you want to bring into Gemba, use the legacy advanced import wizard at /bootstrap while the Settings/import consolidation lands. For existing Beads databases, the project picker's needs setup bind flow is the current preferred path.

For the common case of pointing Gemba at an existing Beads rig and rendering it immediately:

make build
./bin/gemba serve --beads-dir <path-to-your-beads-rig>
# -> http://127.0.0.1:7666

Gemba reads the rig, renders the Kanban, tails state changes, and wires the native terminal orchestrator by default. If you want a read-only / human-driven board with no session controls, pass --orchestration=none.

End-to-end setup (both --beads-dir CLI mode and writable --dolt-url direct-SQL mode), explicit --beads-read-only, expected banner output, and troubleshooting: Running Gemba against your work items.

Native terminal orchestration

Native is the default orchestration mode. Spell it explicitly when you want the command to document that choice:

./bin/gemba serve --beads-dir <rig> --orchestration=native

On boot, the native adaptor auto-detects an available terminal backend (tmux in priority order, then iTerm2, then Terminal.app) and exposes /sessions in the SPA. Clicking "New session" for a bead or dragging a runnable bead into In Progress:

1. Provisions a git worktree (git worktree add -b bead/<id>).
2. Runs gemba install-bridge in the worktree (idempotent merge into .claude/settings.local.json; installs the hook stanza + the gemba-mcp MCP server).
3. Spawns a terminal pane with the chosen agent (Claude Code by default; shell-only for operators driving the bead by hand).
4. Injects the project + epic + bead preamble through core/prompt.Envelope, including the interaction-profile section that governs question / blocker behaviour.

Permission prompts and HITL approvals that Claude Code raises surface live in the SPA; operator answers route back to the terminal as input. Coach and Manager skill output (## Questions / ## Blockers via gemba-ask) surfaces on the same escalations surface with kind/channel distinguished.

Native, with parallelism

Per-agent parallelism lets a single session of a capable agent type carry multiple concurrent beads (gm-root.16). The default .gemba/agents.toml (auto-seeded by ratify since gm-root.24) already declares this for claude:

[[agent]]
name           = "claude"
binary         = "claude"
preamble       = "claude_md"
hooks          = "claude_code"
intra_parallel = true
max_parallel   = 4

Tune the cap or add a second agent type by editing the file. The schema reference is in Agent setup.

The dispatcher tries to reuse an existing capable session before spawning a new pane; the SPA renders an n/max pill per pane plus a global in-flight counter. Full detail: the Parallelism guide.

Autonomous dispatch (sticky session pools)

Per-agent parallelism is the capacity axis. Pools are the continuity axis: long-lived sessions that survive across beads, carry warm context, and pick up ready work on their own — no human drag, no fresh spawn per bead.

Phase 0 is the default: no pool.toml, no daemons, identical behavior to the basic flows above. Pools are opt-in.

Native quickstart (one persona, two pool members, agents.toml already has claude with max_parallel = 4; add codex when you want OpenAI Codex CLI sessions):

# pool.toml
[pool]
default_persona = "engineer-claude"
default_floor = 0.5
reserved_for_manual = 1

[pool.routing]
epic = "engineer-claude"
task = "engineer-claude"

[pool.local.engineer-claude]
size = 2
agent_type = "claude"

./bin/gemba serve --beads-dir <rig> --orchestration=native \
  --pool-config ./pool.toml

/api/pools shows live pool state; /settings/pools is the adaptor-aware SPA editor (path picker, dropdowns, live TOML preview, clamp warnings).

Gas Town: same shape but the scope axis is the rig name and the SPA editor imports rigs/personas from gt agents and shells to gt rig create / gt polecat create for adaptor-owned changes (gm-s47n.17 wires the SPA buttons; CLI works today).

[pool.gemba.engineer]
size = 3
agent_type = "claude"

[pool.lume.engineer]
size = 1
agent_type = "claude"

./bin/gemba serve --beads-dir <rig> --orchestration=gastown \
  --pool-config ./pool.toml

Full quickstart with verification + troubleshooting: Autonomous dispatch quickstart. Operational deep dive: Pool sizing and MaxParallel. Design: docs/design/session-pool.md

• docs/design/pool-editor.md.

TLS

# Self-signed: ephemeral cert + fingerprint banner on boot
./bin/gemba serve --beads-dir <rig> --tls-self-signed

# Operator-supplied chain
./bin/gemba serve --beads-dir <rig> --tls-cert ./cert.pem --tls-key ./key.pem

Optional: Gas Town, LangGraph, other orchestrators

When you want the scheduling or isolation semantics those stacks provide, swap the orchestration flag:

./bin/gemba serve --beads-dir <rig> --orchestration=gastown

Gas Town's session lifecycle is implemented end-to-end (gt sling for dispatch, gt unsling/gt release for end, gt convoy list for session enumeration, gt mail for escalations) — see gm-e7.9. Pause/Resume + ClaimNextReady remain KindUnsupported where gt's CLI lacks the primitive (tracked as gm-e7.10 / gm-e7.11).

Native, Gas Town, Gas City, LangGraph, CrewAI, OpenHands, Devin, and Factory adaptors are mutually exclusive — exactly one or zero OrchestrationPlane adaptors per gemba serve process.

Shader interop smoke (gm-root.5)

scripts/shader-interop.sh exercises the full encode → bd-store → decode round-trip against a live gemba serve + bd backend:

make build
bash scripts/shader-interop.sh

# Optional — also fire `gt sling` and watch for the polecat to hook.
SHADER_INTEROP_SLING=1 bash scripts/shader-interop.sh

⬆ back to top

✨ What's New

Recent (May 2026)

• Scope status pills on live session cards — operational context cards now show top-level health of the session scope: Git clean/dirty, upstream sync (synced, ahead, behind, diverged), and GitNexus analysis freshness against the checked-out HEAD. Dirty worktrees mark analysis stale because the graph cannot include uncommitted source.
• Beads-only mode (gm-etq2 / gm-4u4l) — gemba serve --beads-only runs Gemba as a Beads viewer and manager without a project or orchestration plane. The board starts in Flat view with milestones, epics, decision beads, and work beads in one ordered list; Cascade shows the wrapper hierarchy; the shared Order control sorts by modified, created, edited, or ID; Graph remains available. Status shows Beads health and remote setup actions, while the RHP adds a Beads history tab backed by a JSONL session manifest. Create, edit, hard-delete, milestone wrapper, decision tag, and milestone tag authoring all stay available. --beads-read-only is the inspection-only variant: it implies Beads-only, switches the Status pill to Beads-read-only, hides write affordances, and hard-blocks create/edit/delete/state mutations.
• Self-contained Docker quickstart — Dockerfile.quickstart, docker-compose.quickstart.yml, and make quickstart-run package Gemba with the bd CLI and a seeded sample project. A new user can run the container, open http://localhost:7666, and explore Flat, Cascade, details, Status, Beads history, and Graph without installing the local developer toolchain.
• Standard Docker server image — Dockerfile, docker-compose.yml, and make docker-run package unseeded Gemba with bd, git/ssh, and the sentinel CLIs for mounted worktrees or Dolt URL deployments.
• Codex native agent driver — .gemba/agents.toml can declare a codex agent type using binary = "gemba-codex-driver" and preamble = "codex_exec". The driver runs codex exec --json, exposes session-scoped gemba-mcp tools for cooperative status, escalation, and skill-output telemetry, emits fallback gemba-state lifecycle frames, and closes the bead on success. It is intentionally one-shot: auto-dispatch starts a fresh Codex session when no active Codex work exists rather than reusing an idle pane.
• Wrapper cascade dispatch — dragging an epic or milestone into In Progress marks the wrapper active, stages backlog descendants, and dispatches currently unblocked runnable leaf beads up to the configured pool limit. Wrapper state reflects descendant state rather than running as agent work itself.
• Simplified Board lanes + card status pills — the default kanban now presents Ready → In Progress → Done. Ready combines canonical unstarted and staged work; cards expose the precise state with Next up / Staged pills plus actionable signals such as Triage, Needs input, Review, and Ready. Backlog is still available through /refine and the board toggle; canceled work stays in list/filter views.
• Recent is now a Board named view (gm-uipx.18) — the standalone /recent route remains as a deep link, but the sidebar no longer treats Recent as a top-level concern. Use the Board's View → Recent chip for the last-24h created-items list, or Done · 7d for recent completions.
• Autonomous dispatch (sticky session pools) (gm-s47n.10/.11/ .12/.16) — long-lived agent sessions per (scope, persona) with SessionReady idle state, in-place recycle (no respawn), an idle-pane reaper, and a per-pool dispatch daemon that runs Layer-5 selection over the ready set. Adaptor-aware editor at /settings/pools (path picker, dropdowns, live TOML preview, clamp warnings). Phase 0 zero-delta: opt-in by writing pool.toml. Quickstart: Autonomous dispatch.
• Native orchestration is the default happy path (gm-native.x, matured across the cycle) — --orchestration=native auto-detects tmux / iTerm2 / Terminal.app, provisions a git worktree per session, idempotently merges the bridge stanza into Claude Code's settings.local.json, and injects a project + epic + bead preamble through core/prompt.Envelope. Permission prompts and HITL approvals from Claude Code surface live in the SPA; operator answers route back to the terminal as input. No external daemon required.
• Gas Town adaptor — session lifecycle, escalations, cost synthesis (gm-e7.x) — gt sling / gt unsling / gt convoy list / gt peek / gt mail / gt escalate close all wired through the Gas Town orchestration adaptor (gm-e7.9), plus cost-meter synthesis from transcript tokens (gm-e7.4) and live escalation listing (gm-e7.5). Pause/Resume + ClaimNextReady remain KindUnsupported where the gt CLI lacks the primitive (tracked as gm-e7.10 / gm-e7.11).
• Triage workspace /refine (gm-3ofd + descendants) — dedicated triage + refinement surface for state_category=backlog beads, distinct from the execution kanban. Density-rich tabular layout with age, suggested-epic, blockers, and dispatch_status columns; defer + dismiss actions with notes (gm-mw5n); single
  • bulk drop-into-epic (gm-ju5o); persona-driven milestone creation during refinement (gm-yjst). Replaces the ?layout=list&view=backlog crutch.
• Escalation visibility (gm-e11.3 and friends) — Triage pills on bead/epic cards when open escalations target that work, a Board-level banner with scope-aware count and link to the dedicated escalations page, and a hand-off dispatcher skill (gm-e11.8.7) so PM-class personas can route blockers to the right reviewer. Escalations are advisory by design — they surface, they don't halt.
• Right-Hand Panel (RHP) detail-tab system (gm-root.22.x) — unified URL-driven tab system replaces the legacy drawer infrastructure across Epic / WorkItem / RecommendOrder views. URL codec with kind-replace/stack semantics for deep-link navigation; per-route Help tab pinned in the panel; shared <TabBar> component for in-pane nav (gm-e12.19.7). Epic and WorkItem details include a Milestone → Epic → Work item breadcrumb when ancestry is known.
• Insights with /api/v1/metrics/series time-series proxy (gm-e12.17.1, gm-e9m0) — /insights ships a three-tile MVP on top of recharts (sprint burn-down, dispatch rate, escalation rate) backed by a Prometheus query proxy with capability-manifest gating. /api/v1/metrics/series is the public surface; the SPA hides tiles when no Prometheus URL is configured.
• Milestones across the kanban + RHP (gm-98sq / gm-935r / gm-o2k9 / gm-mqiz / gm-4se1 / gm-yyo8) — milestone child-epic panel inside the RHP detail view, drag-an-epic-onto-a-milestone to re-parent, milestone dropdown on epic detail, auto-close + notify when the last child closes, color stripe + M-pill on EpicCard, and selective milestone context threaded into the preamble for personas working under one.
• Workflows API surface (gm-e12.22.2) — /api/workflows lists active workflow runs and templates from the WorkPlane; underlying beads adaptor exposes the cooked + active distinction so the SPA can render the Workflow Library and Active runs without template/wisp polluting the Plan / Backlog surfaces.
• Per-agent parallelism (gm-root.16) — agent types declare intra_parallel + max_parallel in .gemba/agents.toml; a try-reuse-before-spawn dispatcher policy co-locates beads on capable panes; SPA shows per-pane pills + a global in-flight counter. See the Parallelism guide.
• TLS support (gm-e5.3) — --tls-self-signed for instant HTTPS with a fingerprint banner, or --tls-cert / --tls-key for operator-supplied chains.
• Provider-aware agent detail view (gm-e12.15) — /agents/:id switches its panel based on Workspace.kind (worktree, container, k8s_pod, vm, exec, subprocess) so the affordances match the runtime.
• AgentGroup board (gm-e12.8) — /agent-groups renders one card per group with mode-dispatched visuals (static / pool / graph).
• Sprint roster (gm-e11.5) — /sprints list + /sprints/:id detail. Token-budget rollups deferred to gm-root.14.1.
• Evidence synthesis library (gm-e11.6) — git log + GitHub PR API + CI status collectors with parallel execution and partial-failure tolerance. has_evidence capability gates + synthesized-marker on beads (gm-t4af).

See the GitHub commit log for the full history.

Coming soon

• Insights expansion — the current /insights MVP renders three Prometheus-backed tiles. Follow-up work adds more series (stuck sessions, token spend, escalation backlog) and returns Insights to first-order navigation once the proxy chain is mature enough for everyday use.
• Workflow — beads molecules surfaced as a first-class Gemba concept (Library / Active runs / Authoring + epic-gate enforcement). See docs/design/workflows.md for the ratified UX.

📦 Project Layout

.
├── cmd/
│   ├── gemba/                    # Cobra root (serve, doctor, version, adaptor test, ...)
│   ├── gemba-bridge/             # hook-shim subprocess Claude Code (+ friends) invoke per lifecycle event
│   ├── gemba-state/              # session-status sentinel CLI (ready / working / prompting / stalled)
│   ├── gemba-ask/                # Coach/Manager question/blocker sentinel CLI
│   ├── gemba-codex-driver/       # one-shot Codex exec driver for native sessions
│   └── gemba-mcp/                # MCP-tool server variant of gemba-ask + gemba-state
├── core/                         # adaptor-agnostic types: WorkItem, AgentRef, Relationship, ...
├── internal/
│   ├── server/                   # chi router, handlers, OpenAPI spec
│   ├── events/                   # SSE hub, GembaEvent schema, OTEL propagation
│   ├── auth/                     # bind policy, token, TLS, OIDC interface
│   ├── transport/                # api / jsonl / mcp adaptor hosts
│   ├── evidence/                 # shared evidence-synthesis library (git log, gh PR, CI)
│   └── adapter/
│       ├── noop/                 # in-memory reference (both planes)
│       ├── bd/                   # WorkPlane: Beads (CLI + direct Dolt SQL modes)
│       ├── native/               # OrchestrationPlane: native tmux / iTerm2 / Terminal.app (default)
│       ├── mock/                 # OrchestrationPlane: deterministic in-process acceptance runner
│       ├── gt/                   # OrchestrationPlane: Gas Town (optional)
│       ├── jira/                 # WorkPlane: Jira (forcing-function)
│       ├── langgraph/            # OrchestrationPlane: LangGraph (forcing-function)
│       └── gascity/              # OrchestrationPlane: Gas City (stub)
├── web/                          # Vite + React + TypeScript + Tailwind + shadcn/ui SPA
│   └── src/
│       ├── api/                  # codegenned client + types
│       ├── capabilities/         # capability-manifest readers + JSX gates
│       ├── pages/                # Board, Refine, Sessions, Sprints, AgentGroups, AgentDetail, ...
│       ├── components/           # Board, RHP detail tabs, sessions, project picker, ...
│       └── extensions/           # adaptor-namespaced widgets (gated by capability manifest)
├── testing/                      # public conformance harness (importable as github.com/GembaCore/gemba-core/testing)
│   ├── fixtures/                 # JSON fixtures used by the harness + scripts/
│   └── e2e/                      # Playwright specs + page objects
├── docs/
│   ├── adaptors/                 # per-adaptor authoring docs + conformance reports
│   ├── design/                   # durable conventions (parallelism-boundary, milestone-convention, ...)
│   ├── getting-started/          # operator-facing guides
│   ├── api/                      # OpenAPI 3.1 spec served at /api/openapi.json
│   └── img/                      # diagrams referenced from docs + README
├── docs-site/                    # Astro Starlight build of docs/, deployed to GitHub Pages
├── branding/                     # banner + social-preview PNGs + Pillow generator
├── scripts/                      # dev/CI shell helpers (run.sh, shader-interop.sh, ...)
├── .github/workflows/            # CI, e2e, docs, release
├── embed.go                      # //go:embed of web/dist (must live above web/, can't be moved deeper)
├── Makefile
└── go.mod

⬆ back to top

🛠️ Development

The repo ships a single Makefile covering the full dev/build/release loop. From a fresh clone:

make help         # list every target with one-line docs
make dev          # Vite + Go server, both with hot reload (needs: air, pnpm)
make build        # build the SPA, embed it, produce ./bin/gemba
make test         # go test -race ./... + pnpm test --run
make lint         # golangci-lint + pnpm lint
make release      # local goreleaser snapshot (multi-OS/arch tarballs in ./dist)
make clean        # remove bin/, dist/, web/dist/*, web/node_modules

Tooling expected on PATH:

• Go (see go.mod for the minimum version)
• pnpm — frontend package manager
• air — Go hot reload for make dev (go install github.com/air-verse/air@latest)
• golangci-lint — make lint
• goreleaser — make release

[!TIP] make dev runs Vite (port 5173) and the Go server (port 7666) with hot reload on both sides. The Vite dev server proxies /api and /events to the Go process so you can edit and see changes end-to-end without manual rebuilds.

⬆ back to top

📚 Documentation

The full docsite is published at https://gembacore.github.io/gemba-core/.

Where	What
Getting Started	Operator-facing guides — running against your work items, parallelism configuration
Adaptors	Per-adaptor authoring docs + conformance reports — how to write a new WorkPlane / OrchestrationPlane
Design	Durable architectural decisions — parallelism boundary, milestone convention, Gemba walk (review of work in progress), persona PPPP
Agents	Per-role agent operating docs
UI spec	The SPA spec — every surface, every affordance, every test-id

[!NOTE] Gemba captures decisions as decision-type beads with a D#: numbered prefix (matching the milestone convention M1 / M2). Every decision is paired with a design doc that captures the ratified contract; the linkage is CI-enforced via make lint-decisions and a --docs-only check on every PR. See Decision process for the lifecycle (draft → in_review → ratified | rejected) and worked examples.

⬆ back to top

🤝 Feedback and Contributions

We've tried to make every architectural seam — adaptor boundaries, capability gating, transport plurality — explicit and inspectable. Where we got it wrong, your feedback is the fastest path to fixing it.

[!IMPORTANT] Bug reports, feature suggestions, and architecture critiques are all welcome. The high-leverage things to share: what you tried, what you expected, and what actually happened, in roughly that order.

• 🐛 File an issue — bugs, feature requests, design feedback
• 💬 Open a discussion — questions, design conversations, "how would you …" threads
• 🔧 Pull requests — start with a draft if the change is non-trivial; we'd rather review the design before the code

If you're writing a new adaptor, the conformance harness is the shortest path to "does it work": see the WorkPlane authoring guide or the OrchestrationPlane authoring guide.

Git hooks (optional)

The repo ships a versioned pre-commit hook at scripts/git-hooks/pre-commit that runs prettier on staged Markdown files (--prose-wrap always --print-width 72) and re-stages anything it reformats. Install once per clone:

make install-hooks       # → git config core.hooksPath scripts/git-hooks
make uninstall-hooks     # restore default

Tolerant: the hook skips silently if web/node_modules/.bin/prettier isn't present (e.g. you haven't run pnpm install in web/ yet) and never blocks a commit.

⬆ back to top

📜 License

MIT © Gemba contributors.

This means: use it, fork it, ship it, no attribution required beyond the LICENSE file in your distribution. Pull requests welcome under the same terms.

⬆ back to top