sw-init

name	sw-init
description	Initializes Specwright in a project. Detects stack, asks about practices, creates constitution and charter, configures quality gates and hooks.
argument-hint
allowed-tools	["Bash","Read","Write","Glob","Grep","AskUserQuestion"]

Specwright Init

Goal

Set up Specwright in this project by understanding how the user works, what they're building, and what quality standards they expect. Produce configuration and anchor documents that will guide all future work. Tracked project artifacts should be shared across developers and agent sessions via Git; runtime session state stays local to each clone or worktree. New interactive installs should prefer project-visible runtime roots under .specwright-local/, while git-admin roots under .git/specwright/ remain compatibility-only.

Inputs

• The codebase (scan for language, framework, dependencies, test runner)
• The user (ask about practices, vision, quality expectations)

Outputs

When complete, ALL of the following exist:

• {projectArtifactsRoot}/config.json -- detected + configured project settings
• {projectArtifactsRoot}/CONSTITUTION.md -- development practices the AI must follow
• {projectArtifactsRoot}/CHARTER.md -- technology vision and project identity
• {repoStateRoot}/work/ -- shared work root, initialized and empty
• {worktreeStateRoot}/session.json -- initialized detached top-level session for the current worktree

Optional (created if the user opts in):

• {projectArtifactsRoot}/TESTING.md -- testing strategy: boundaries, infrastructure, mock allowances
• Hooks set up if the user wants them
• Operator follow-up: tell the user to run /sw-status to confirm the active runtime roots and detached session state.

Quality gates are configured in config (all six default to enabled; user may disable).

Constraints

Worktree context (LOW freedom):

• Check worktreeContext per protocols/context.md.
• If linked, explain that tracked project artifacts under {projectArtifactsRoot} are shared across developers and agent sessions via Git, while runtime session state stays local to each clone or worktree under {repoStateRoot} and {worktreeStateRoot}.
• If the shared layout already exists, reuse the tracked project artifacts and shared runtime roots, then create or repair only this worktree's {worktreeStateRoot}/session.json.
• If the shared layout does not exist yet, say that /sw-init will create the tracked project-artifact root plus repo-level runtime state under the Git common dir and a session root for the current worktree. This is a warning, not a block.

Detection (MEDIUM freedom):

• Scan codebase: language(s), framework(s), package manager, test runner, linting/formatting, git workflow, CI/CD. Read dependency manifests. Don't guess what you can detect.

Survey (MEDIUM freedom):

• After detection, survey the codebase using Glob/Grep/Read: directory structure, entry points, module dependencies, conventions, integration points, gotchas.
• Produce {projectArtifactsRoot}/LANDSCAPE.md per protocols/landscape.md format. User approves before saving.
• Optional — if user declines, skip. LANDSCAPE.md is never required.

User conversation (HIGH freedom):

• Ask the user about things you CANNOT detect from the codebase:
  • What is this project? Who uses it? (→ CHARTER.md)
  • Testing philosophy and coverage expectations? (→ CONSTITUTION.md)
  • Security requirements? (→ gate config)
  • Code review standards? (→ CONSTITUTION.md)
  • Any practices or patterns they insist on? (→ CONSTITUTION.md)
• Use AskUserQuestion with concrete options based on what you detected.
• Batch related questions. Maximum 3-4 questions per interaction.
• Don't ask about things the codebase already answers.

Constitution creation (HIGH freedom):

• The constitution captures the user's development practices as clear rules.
• Rules should be specific and actionable, not vague aspirations.
• Bad: "Write clean code." Good: "All public functions must have error handling."
• The user must approve the constitution before it's saved.

Charter creation (HIGH freedom):

• The charter captures the project's identity and vision.
• What is this project? What problem does it solve? Who are the consumers?
• What are the architectural invariants (things that won't change)?
• What technologies are foundational (not up for debate)?
• Keep it concise -- one page, not a business plan.
• The user must approve the charter before it's saved.

Testing strategy creation (HIGH freedom):

• After constitution and charter are approved, generate {projectArtifactsRoot}/TESTING.md.
• Follow protocols/testing-strategy.md for document structure and boundary classifications.
• Ask the user about testing boundaries using AskUserQuestion:
  • "What external services does this project call?" (payment APIs, email, auth providers, etc.)
  • "What's your test database strategy?" (in-memory, testcontainers, shared test DB, none)
  • "Are there any rate-limited or cost-attached APIs?" (metered APIs, slow services)
  • "Any other expensive or unreliable dependencies?"
• Batch these into 1-2 questions based on what stack detection reveals. Don't ask about services the codebase doesn't use.
• Generate TESTING.md with required sections:
  • Boundaries: Classify each detected dependency as internal (test with real components), external (mock with contracts), or expensive (mock with rationale)
  • Test Infrastructure: What test database, fixtures, containers, or test servers are available based on detected stack and user answers
  • Mock Allowances: Which dependencies may be mocked, with explicit rationale
  • Test Commands (added after gate configuration): If tiered commands were captured during gate config, append this section to TESTING.md with the actual commands. See protocols/testing-strategy.md Test Commands section. Omit if no tiers configured.
• The user must approve TESTING.md before it's saved.
• If the user declines or skips: do not create TESTING.md. Constitution testing rules remain the sole authority. TESTING.md is recommended but not required.

Git workflow configuration (MEDIUM freedom):

• Detect workflow by scanning branch names, remotes, CI files. Present detected strategy with confidence.
• Confirm via AskUserQuestion: strategy (trunk-based/github-flow/gitflow/custom), target-role defaults and freshness checkpoints, branch prefix, merge strategy, PR required, commit format.
• Treat runtime mode as an explicit Git policy choice. Ask whether clone-local runtime state should stay under git-admin roots or use project-visible roots, and recommend project-visible for Claude-oriented installs unless the user has a strong reason to keep runtime files hidden under .git/.
• Ask whether optional auditable work artifacts stay clone-local or are published under a tracked work-artifact root. Recommend a tracked work-artifact root under .specwright/works for new interactive installs unless the user explicitly prefers clone-local-only artifacts. Keep runtime mode separately from tracked work-artifact publication; project-level anchor docs remain project artifacts and runtime session state stays local-only.
• Store git.targets and git.freshness in config.json, seeding branch-role defaults and freshness checkpoints from the detected workflow strategy while preserving baseBranch as a compatibility alias and without requiring users to define a custom branch DSL.
• Store the resulting settings in the config.json git section per protocols/git.md. Record git.runtime.mode / git.runtime.projectVisibleRoot separately from tracked work-artifact publication, and record any optional work-artifact publication choice in config instead of inferring it from .git/ paths or symlinks.
• If old git schema detected: offer migration with sensible defaults.

Configuration (LOW freedom):

• Write {projectArtifactsRoot}/config.json with detected and configured values.
• Create {projectArtifactsRoot}/research/ and {projectArtifactsRoot}/learnings/. If survey produced LANDSCAPE.md, write it under {projectArtifactsRoot}.
• Create {repoStateRoot}/work/ for shared runtime work records and {worktreeStateRoot}/ for the current worktree session.
• Create {worktreeStateRoot}/session.json as a detached top-level session for the current worktree. Initialize attachedWorkId: null, current branch when available, mode: "top-level", and fresh lastSeenAt.
• If a prior runtime-only install exists, treat it as migration input only: import the tracked docs/config the user keeps back into {projectArtifactsRoot}, then write repaired runtime state to {repoStateRoot} and {worktreeStateRoot}.
• Follow protocols/state.md for state file format.

Gate configuration (MEDIUM freedom):

• All six gates default to enabled: build, tests, security, wiring, semantic, spec. Semantic is WARN-only with graceful degradation (no tools required). The user may disable any gate. Write gates as object format in config.json: { "gates": { "build": { "enabled": true }, ... } }. Configure thresholds per user expectations.
• If a test runner is detected, batch these additional questions:
  • "Do you have integration tests against real infrastructure? What command runs them?" → Populate commands.test:integration in config.json. Skip if user answers none.
  • "Do you have smoke or E2E tests? What command?" → Populate commands.test:smoke. Skip if user answers none. Empty or declined answers are not written to config (tiers are optional).

Backlog configuration (MEDIUM freedom):

• Batch with gate configuration question (both are quality infrastructure).
• Ask: "Where should Specwright track tech debt, deferred work, debug findings, and audit items?"
  • markdown — writes to {projectArtifactsRoot}/BACKLOG.md (default, always available)
  • github-issues — creates GitHub Issues via gh CLI (requires gh auth login)
• If github-issues selected: ask for label name (default: specwright-backlog).
• Store as backlog.type and backlog.label in config.json.
• If user skips or is unsure: default to markdown. Backlog config is always optional.

Experimental features (LOW freedom):

• If user indicates interest in experimental features: add experimental section to config.json.
• Schema: { "experimental": { "agentTeams": { "enabled": false, "maxTeammates": 3, "requirePlanApproval": true } } }
• All experimental features default to disabled. Don't prompt unless the user asks.

Protocol References

• protocols/state.md -- workflow.json initialization
• protocols/context.md -- config.json format
• protocols/git.md -- git config field reference and strategy definitions
• protocols/landscape.md -- codebase reference document format

Failure Modes

Condition	Action
Existing Specwright tracked/runtime roots detected	Ask user: repair or reconfigure the existing install, or abort
No dependency manifest found	Ask user about language and framework directly
User unsure about practices	Suggest sensible defaults based on detected stack, let them adjust
Old config.json git schema detected	Show diff of old vs new fields. Offer migration with AskUserQuestion. Preserve existing values, add new fields with defaults.
Compaction during init	Check which files exist, resume from next missing artifact