loomrail 0.1.0-alpha.1 → 0.1.0-alpha.3

package/README.md

@@ -1,244 +1,127 @@
 <div align="center">
   <img src="docs/assets/brand/loomrail-wordmark.svg" alt="Loomrail" width="360" />
-  <p><strong>The local control plane for accountable AI software teams.</strong></p>
+  <p><strong>AI agents work. You decide.</strong></p>
+  <p>
+    <a href="https://loomrail.github.io/loomrail/">Website</a> ·
+    <a href="docs/guides/GETTING-STARTED.md">Quick start</a> ·
+    <a href="docs/guides/GETTING-STARTED.ru.md">Быстрый старт</a> ·
+    <a href="docs/README.md">Documentation</a>
+  </p>
   <p>
     <a href="https://github.com/loomrail/loomrail/actions/workflows/ci.yml"><img src="https://github.com/loomrail/loomrail/actions/workflows/ci.yml/badge.svg" alt="CI" /></a>
-    <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-5e6ad2" alt="Apache 2.0 license" /></a>
+    <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-6173ff" alt="Apache 2.0 license" /></a>
     <img src="https://img.shields.io/badge/status-pre--alpha-c58b20" alt="Pre-alpha status" />
     <img src="https://img.shields.io/badge/Node.js-24.19-43853d" alt="Node.js 24.19" />
   </p>
 </div>
 
-Loomrail is a local-first workspace for planning, running, and supervising complete software-delivery workflows
-across coding agents such as Codex and Claude Code. It is designed around tasks, evidence, budgets, review, and human
-decisions instead of disconnected chat sessions.
-
-> [!IMPORTANT]
-> Loomrail is an early pre-alpha. The local kernel, authenticated browser session, SQLite state, audit log, and
-> cross-platform CI are real. The Workbench now runs a restart-safe synthetic Discovery → Plan → Implement → Review
-> → QA → Acceptance workflow with durable budgets, evidence, Human Requests, and owner Decisions in English and
-> Russian. A live Codex session now runs inside a Git worktree cut for the task it works on, so all six stages can
-> reach a real repository; the Claude Code adapter still serves DISCOVERY, PLAN and REVIEW only.
+Loomrail is a local control plane for AI-assisted software work. It keeps the task brief, workflow state, Human
+Requests, budgets, evidence, and final owner decision durable across agent sessions instead of treating chat history as
+the source of truth.
 
 <picture>
   <source media="(prefers-color-scheme: dark)" srcset="docs/assets/screenshots/workbench-dark.png" />
   <source media="(prefers-color-scheme: light)" srcset="docs/assets/screenshots/workbench-light.png" />
-  <img src="docs/assets/screenshots/workbench-light.png" alt="Loomrail Workbench with a Kanban delivery board and task inspector" width="100%" />
+  <img src="docs/assets/screenshots/workbench-light.png" alt="Loomrail Workbench showing delivery state, a task contract, and owner activity" width="100%" />
 </picture>
 
-## Why Loomrail
-
-- **Task-centric delivery.** Keep discovery, planning, implementation, review, QA, and acceptance on one accountable
-  route.
-- **Local by default.** The daemon binds to loopback, state lives in local SQLite, and the browser uses a one-time
-  authenticated bootstrap session.
-- **Human control.** Questions, approvals, budgets, recovery decisions, and acceptance stay visible and explicit.
-- **Auditable work.** Commands are idempotent and state changes are recorded as append-only events.
-- **Cross-platform baseline.** macOS and Windows run the same blocking verification and browser smoke tests.
-
-## Current checkpoint
-
-| Area          | Today                                                                                                                    | Next                                           |
-| ------------- | ------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------- |
-| Local runtime | Loopback daemon, CLI launcher, one-time browser session, installable tarball                                             | Published package and desktop installer        |
-| State         | Tasks, runs, budgets, recovery, typed evidence, acceptance packages, Decisions, append-only Events                       | Retention and restore hardening                |
-| Workbench     | Persisted board, workflow cockpit, command summary, evidence matrix, owner acceptance, EN/RU, light/dark                 | Full Attention Inbox and richer workflow views |
-| Agents        | Capability-checked provider contract, live Codex/Claude CLI adapters, per-task Git worktrees, and on-demand change diffs | The Claude Code adapter on the write path      |
-| Projects      | Bundled demo repositories, plus any local Git repository registered by absolute path                                     | Per-project guardrails and permissions         |
-| Platforms     | macOS and Windows CI are green                                                                                           | Clean-machine acceptance and hardening         |
-
-## How it is intended to work
-
-```mermaid
-flowchart LR
-  Brief[Task brief] --> Plan[Delivery plan]
-  Plan --> Build[Implementation agents]
-  Build --> Review[Independent review]
-  Review --> QA[Browser QA and evidence]
-  QA --> Accept{Human acceptance}
-  Accept -->|Approved| Done[Done]
-  Accept -->|Changes requested| Plan
-  Guardrails[Rules · permissions · budgets] -. constrain .-> Plan
-  Guardrails -. constrain .-> Build
-  Guardrails -. constrain .-> Review
-```
-
-Loomrail is the control plane around this route. It does not replace the coding agents; it gives their work a shared
-model, clear permissions, recoverable state, and an inspectable history.
+> [!IMPORTANT]
+> Loomrail is public pre-alpha software. New projects use **Auto**: an installed, authenticated Codex or Claude Code
+> CLI can be selected for new sessions and spend provider quota. To guarantee a zero-quota first run, choose **Mock**
+> in **Settings → AI provider** before starting the workflow. Loomrail never installs or signs into a provider,
+> enables permission-bypass flags, commits, pushes, merges, or deploys for you. A task worktree is not an
+> operating-system sandbox.
 
-## Install
+## Install and run safely
 
-For the complete route from installation through the first Human Request, budget decision, acceptance, live provider,
-change review, restart, and state backup, use the [English user guide](docs/guides/USER-GUIDE.md) or the
-[руководство на русском](docs/guides/USER-GUIDE.ru.md).
+Requirements: Node.js `>=24.19 <25`, macOS or Windows, and a browser on the same machine. Linux is best effort.
 
-Loomrail ships as a single package: a bundled launcher, the prebuilt Workbench, the SQLite migrations and the bundled
-fixture projects. Pre-alpha releases use the explicit `next` channel so they are never installed as a stable release:
+Start in a new empty directory, not inside a repository you care about:
 
 ```bash
+mkdir loomrail-evaluation
+cd loomrail-evaluation
 npm install loomrail@next
-npx loomrail --port 4176
-```
-
-To verify a source revision before it reaches the registry, build and install the exact release tarball instead:
-
-```bash
-pnpm pack:release
-npm install ./dist-release/loomrail-0.1.0-alpha.1.tgz
-npx loomrail --port 4176
+npx loomrail
 ```
 
-Install it globally with `npm install -g` instead if you want `loomrail` on your `PATH`. Either way the launcher
-starts on loopback and opens a one-time authenticated URL; add `--no-open` and it prints that URL instead, so a
-same-machine browser can still sign in without being opened automatically.
-
-`pnpm test:release` performs exactly this install into an empty project using only the public registry, and runs on
-macOS and Windows in CI. See the [release guide](docs/RELEASE.md) for the full procedure.
+The launcher binds to `127.0.0.1`, opens a one-time authenticated URL, and stores state in local SQLite. Keep the
+terminal open and stop Loomrail with `Ctrl+C`.
 
-## Run from source
-
-There is no desktop installer yet. To develop Loomrail, or to try the current checkpoint without building a package,
-run the repository directly.
-
-### Requirements
-
-- Node.js as pinned in [`.nvmrc`](.nvmrc)
-- Corepack
-- macOS or Windows
+If the browser must not open automatically:
 
 ```bash
-git clone https://github.com/loomrail/loomrail.git
-cd loomrail
-nvm use          # or: fnm use
-corepack enable  # installs the pnpm version pinned by packageManager
-pnpm install --frozen-lockfile
-pnpm dev
+npx loomrail --no-open --port 4176
 ```
 
-`pnpm dev` builds the workspace, starts Loomrail on an available loopback port, and opens a one-time authenticated URL
-in the default browser. Stop it with `Ctrl+C`.
-
-For a fixed port after the first build:
-
-```bash
-pnpm build
-pnpm start --port 4176
-```
-
-Use `pnpm start --no-open --port 4176` when the browser should not open automatically; the launcher then prints the
-one-time sign-in URL for a browser on the same machine. That URL signs in a single browser, expires after 60 seconds,
-and is replaced on every restart. `LOOMRAIL_DATA_DIR` can point a development run at an isolated data directory.
-
-| Platform | Default local state                                   |
-| -------- | ----------------------------------------------------- |
-| macOS    | `~/Library/Application Support/Loomrail/state.sqlite` |
-| Windows  | `%LOCALAPPDATA%\Loomrail\state.sqlite`                |
-
-## Repository
-
-```text
-apps/
-  cli/       # local launcher and authenticated browser bootstrap
-  daemon/    # loopback API, commands, events, and SQLite lifecycle
-  web/       # React Workbench
-packages/
-  contracts/             # shared schemas and transport contracts
-  domain/                # deterministic WorkItem and workflow decisions
-  persistence-sqlite/    # SQLite repositories, queue, and migrations
-  context-assembly/      # what a provider session is told, and in what order
-  workspace/             # Git process boundary, repository inspection, worktrees
-  provider-core/         # provider lifecycle and capability boundary
-  provider-mock/         # deterministic synthetic provider scenarios
-  provider-codex/        # the real `codex` CLI as a child process
-  provider-claude-code/  # the real `claude` CLI as a child process
-  workflow-engine/       # versioned workflow template validation
-  ui/                    # shared product primitives and patterns
-docs/        # product, architecture, security, design, plans, and evidence
-```
+Open the printed URL on the same machine within 60 seconds. `--no-open` does not enable remote access.
 
-The daemon owns state and capability boundaries. The web app never receives raw provider credentials and never talks
-to an agent, a shell or Git directly: every one of those crossings goes through the daemon.
+For a global launcher, use `npm install -g loomrail@next` and then `loomrail`. The project-local route above is
+recommended for evaluation because it keeps the selected pre-alpha channel visible.
 
-## Roadmap
+## First run
 
-- [x] **M0 — Foundation:** monorepo, contracts, CI, public-readiness rules
-- [x] **M1 — Walking skeleton:** CLI → daemon → authenticated browser UI
-- [x] **M2 — Local kernel:** SQLite state, idempotent commands, append-only events, macOS/Windows gate
-- [x] **M3 — Real task cockpit:** authenticated API client, persisted projects/work items, editing, EN/RU, activity
-      replay and secure reconnect guidance
-- [x] **M4 — Mock delivery workflow:** restart-safe dispatch queue, Human Request, Decision, and resumable task
-      pipeline
-- [x] **M5 — Budgets and recovery:** explicit limits, pause/resume, crash recovery
-- [x] **M6 — Acceptance:** typed Review/QA evidence, criterion matrix, owner-only final approval, audit surface
-- [ ] **M7 — Public checkpoint:** packaged launcher and clean-install gate are in place; remaining work is hardening
-      and the first published release
+1. Choose **Initialize demo workspace**.
+2. Open **Settings → AI provider** and choose **Mock** for the zero-quota walkthrough.
+3. Create a task with a concrete outcome and observable acceptance criteria.
+4. Move it to **Ready** and start the workflow.
+5. Answer the blocking Human Request and approve the explicit mock budget increase.
+6. Inspect Review and QA evidence, then accept the delivery or return it to work as the owner.
 
-Real Codex and Claude Code execution has landed, and milestone E1 has since given it somewhere to work. Before
-dispatching a work item's first agent stage, Loomrail cuts a Git worktree for that work item on a branch of its own and
-runs the CLI there — for every stage but your own acceptance decision, because a review reads the change it judges and a
-plan is worth more when it can read the code it plans against — so **the Codex adapter now serves all six stages** under
-`codex exec -s workspace-write`. The Claude Code
-adapter still declares DISCOVERY, PLAN and REVIEW only: its write path has never been exercised against the real CLI
-here, and one adapter's evidence is not taken as proof about the other. A stage an adapter does not declare is refused
-to you as a blocking question rather than dispatched, and Loomrail never enables a permission-bypass flag on any code
-path.
+The task, request, budget, evidence, and decision survive page reloads and Loomrail restarts. The
+[quick start](docs/guides/GETTING-STARTED.md) walks through the route in detail.
 
-Your own checkout is never the working directory. Worktrees are cut under the Loomrail data directory
-(`<data>/workspaces/<project>/<work item>`), the branch is deleted only while it still is the one Loomrail cut, and a
-workspace whose directory has disappeared is reconciled at startup instead of being quietly reused.
+## Your repository and live providers
 
-Everything downstream of the edit itself remains outside the current checkpoint: Loomrail commits nothing, pushes
-nothing and merges nothing, so a stage's work stays on its worktree's branch for you to inspect and dispose of. Plugins,
-remote mode and desktop packaging are outside it too. What comes next — project guardrails and extensibility — is
-decomposed in the [post-Phase-0 plan](docs/plans/06-post-phase-0-decomposition.ru.md).
+After the mock route works, the owner guide explains repository registration, Project Constitution review, task
+worktrees, change inspection, backup, and recovery:
 
-### Pointing Loomrail at a repository
+- [Owner guide](docs/guides/USER-GUIDE.md)
+- [Руководство владельца](docs/guides/USER-GUIDE.ru.md)
+- [Reproducible full-route example](docs/examples/full-route/README.md)
+- [Security and trust boundaries](docs/security/THREAT-MODEL.md)
 
-A Project no longer has to be one of the two bundled demos. Open **Settings → Projects**, give it the absolute path
-of a Git repository on this machine, and that repository becomes a Project you can create tasks against; the
-directory's own name becomes the project name. The path has to be absolute — a relative one would resolve against
-whatever directory the daemon happened to start in — and it has to be a repository's top level, because registering a
-subdirectory would branch the repository enclosing it without you having chosen that.
+Install and authenticate the provider CLI yourself, then start Loomrail normally. In **Settings → AI provider**, keep
+**Auto** to use an available signed-in CLI or choose Codex, Claude Code, or Mock explicitly for that project. Use
+**Check again** after installing or signing in; no extra launch command is required. `LOOMRAIL_PROVIDER` remains an
+optional process-wide override for automation and troubleshooting. Read the owner guide and threat model before
+exposing a repository to either live CLI.
 
-Nothing stops you pointing it at this checkout, and that is deliberate. What keeps it safe is the shape of the work
-rather than a refusal: the agent writes only inside a worktree cut outside the repository, your working copy, index
-and checked-out branch are untouched, and nothing is ever pushed. Loomrail does add the worktree's bookkeeping and
-its own `loomrail/…` ref to your `.git`, and creates one commit — the carry-in snapshot that branch starts from —
-but it never moves or deletes a ref you made. Be aware of what travels with it, though — everything you have not committed is carried
-into the worktree, including untracked files the repository does not ignore, and the agent has network access in that
-same tree. The [threat model](docs/security/THREAT-MODEL.md) records this as an accepted risk rather than a solved one.
+Context7 is different from an AI provider: its exact-pinned MCP server ships with Loomrail. In **Settings → MCP
+connections**, choose **Review bundled Context7**; no global install or `npx` command is needed. Loomrail still requires
+you to approve the exact local process and grant its two read-only tools. Context7 documentation queries leave your
+machine, so never include secrets, personal data, or proprietary code.
 
-Once a stage has cut one, the task card shows the workspace itself: which repository it came from, the branch, the
-base commit and the worktree path, so you can open the tree in your editor or run `git diff` against it yourself.
+## Current boundary
 
-### Choosing a provider
+- Local browser UI, loopback daemon, and local SQLite state.
+- Auto-discovered or explicitly selected providers, plus a deterministic Mock route with no provider quota.
+- Project-scoped local MCP connections and a bundled, owner-approved Context7 preset.
+- A typed read-only tool SDK at `loomrail/plugin-sdk`; local registration still uses explicit C1 consent and grant.
+- Existing-repository registration, readiness checks, per-task worktrees, change inspection, and owner-approved
+  Project Constitution.
+- Explicit new-project creation from a fixed local recipe, with exact file review, durable recovery, and local Git
+  initialization. Loomrail does not install dependencies or create a commit.
+- No desktop installer, remote access, cloud sync, team accounts, automatic Git publishing, or complete OS sandbox.
 
-One adapter serves every stage a daemon dispatches, for the life of the process. It is chosen by a single environment
-variable, read once at startup:
+The versioned product scope lives in [Product decisions](docs/product/PRODUCT-DECISIONS.ru.md) and the
+[Master plan](docs/product/MASTER-PLAN.ru.md). Historical implementation plans remain under `docs/plans/`; they are
+engineering records, not a public roadmap.
 
-| `LOOMRAIL_PROVIDER` | What runs                                                                             |
-| ------------------- | ------------------------------------------------------------------------------------- |
-| unset, or `MOCK`    | The deterministic mock adapter. No real agent runs; every stage completes on its own. |
-| `CODEX`             | The real `codex` CLI, as a child process.                                             |
-| `CLAUDE_CODE`       | The real `claude` CLI, as a child process.                                            |
-
-The values are case-sensitive. A value Loomrail cannot read falls back to `MOCK` — a typo must not stop the daemon from
-starting — but it is never silent: the launcher and the daemon log both name the value and the accepted spellings,
-because the mock completes stages successfully and you would otherwise watch a whole delivery run believing a live agent
-did it. The launcher also says when a selected adapter's CLI is not installed on this machine.
-
-The CLIs authenticate themselves: Loomrail adds nothing to the child's environment and never handles your provider
-credentials.
-
-The task card shows the files in that worktree that differ from its starting snapshot and reads one unified diff only
-when you expand that file. This is inspection, not Git authority: Loomrail still does not commit, push, or merge those
-changes.
+## Develop from source
 
 ```bash
-LOOMRAIL_PROVIDER=CODEX loomrail
+git clone https://github.com/loomrail/loomrail.git
+cd loomrail
+nvm use
+corepack enable
+pnpm install --frozen-lockfile
+pnpm dev
 ```
 
-## Development
+`pnpm dev` builds the workspace, starts Loomrail on loopback, and opens a one-time authenticated browser session.
+
+Before contributing:
 
 ```bash
 pnpm verify
@@ -246,20 +129,9 @@ pnpm exec playwright install chromium
 pnpm test:e2e
 ```
 
-`pnpm verify` runs formatting, public-tree safety checks, linting, type checks, and the test suite. Changes to the CLI,
-daemon session flow, or web shell should also pass the browser smoke test on both blocking platforms.
-
-Start with [CONTRIBUTING.md](CONTRIBUTING.md). Product and engineering sources of truth:
-
-- [Release guide](docs/RELEASE.md)
-- [Master plan](docs/product/MASTER-PLAN.ru.md)
-- [Product decisions](docs/product/PRODUCT-DECISIONS.ru.md)
-- [Architecture overview](docs/architecture/OVERVIEW.md)
-- [Phase 0 implementation plan](docs/plans/00-phase-0-implementation-plan.ru.md)
-- [Threat model](docs/security/THREAT-MODEL.md)
-- [Component system](docs/design/COMPONENT-SYSTEM.md)
-- [Brand guide](docs/design/BRAND.md)
-- [Localization contract](docs/design/LOCALIZATION.md)
+See [CONTRIBUTING.md](CONTRIBUTING.md), the [architecture overview](docs/architecture/OVERVIEW.md), and the
+[release guide](docs/RELEASE.md). Plugin authors should start with the
+[Plugin SDK guide](docs/guides/PLUGIN-SDK.md) or its [Russian version](docs/guides/PLUGIN-SDK.ru.md).
 
 ## License