Claude Code Tooling

Reference of the Claude Code configuration versioned in this repository, and of the MCP servers it expects you to provide yourself.

Versioned files

Path	Role
CLAUDE.md	Project instructions loaded into every session
.claude/settings.json	Project hooks, permissions and MCP server enablement
.claude/local-qa.md	Recipes for running QA and tests on a dev machine, loaded on demand by /check and /sprint
.claude/skills/<name>/SKILL.md	The four project skills below
.claudeignore	Paths kept out of Claude's context (dependencies, lock files, binaries, secrets)
.github/workflows/claude.yml, claude-code-review.yml	Claude on GitHub (see below)

There is no project-level .mcp.json: MCP servers are configured per user (see MCP servers).

Settings (.claude/settings.json)

Key	Value
includeCoAuthoredBy	false: no AI attribution in commits
enableAllProjectMcpServers, enabledMcpjsonServers	enable an openapi-spec MCP server when one is defined
permissions.allow	read-only shell commands, git, gh, make, composer, docker compose, curl to localhost, web search
permissions.deny	rm, and reading .env, api/.env.local*, api/config/secrets/*, pwa/.env*
permissions.defaultMode	acceptEdits

Hooks

Event	Matcher	Effect
PreToolUse	Write\|Edit	Blocks any edit to a path containing .env, schema.d.ts, compose.override, vendor/, node_modules/, .claude/settings.json or .claude/settings.local.json
PostToolUse	Write\|Edit	For a .php file, runs make php-cs-fixer then make rector; for a .ts/.tsx file, runs make prettier. Skipped for files under .claude/worktrees/

The formatting targets do not read a file argument: make php-cs-fixer and make rector process the whole api/ and provisioner/ trees, and make prettier only checks formatting.

Skills

Command	What it does
/pick <issue-number> [base-branch]	Implements a GitHub issue end to end: branch, code, tests, pull request, then up to 3 surveillance cycles (CI, review comments) until the PR is READY or NEEDS ATTENTION. Never merges.
/sprint <sprint-number>	Implements every issue of a sprint in parallel with one worktree agent per issue, picks the model per issue (Opus by default, Sonnet for mechanical changes), runs the same bounded loop on each PR and records the result in TRACKING.md through a dedicated branch
/check [qa\|test\|all]	Runs the QA and/or test suites, fixes failures and retries, 3 runs at most, and reports the real output. Default all
/close [sprint-number]	Closes a sprint: removes local worktrees and branches after confirmation, fast-forwards main, runs a retrospective that proposes configuration changes without applying them

Claude on GitHub

Trigger	Workflow	Effect
@claude pick [base-branch] in an issue comment	claude.yml, job pick	Same flow as /pick, in CI and without Docker: creates feature/<issue>, implements, opens the PR, follows CI
@claude <instruction> in an issue or PR comment, a review, or an issue body/title	claude.yml, job claude	Follows the free-form instruction
Pull request opened, synchronized or marked ready for review	claude-code-review.yml	Automated review in Conventional Comments format; the bot resolves its own threads once a fix is pushed

MCP servers

None is versioned. These are the servers the project workflow assumes, configured in your user settings or as Claude Code plugins:

Server	Used for	Source
GitHub	Issues, pull requests, reviews and CI results (mcp__github__*)	https://github.com/github/github-mcp-server
Context7	Up-to-date library documentation (Symfony, API Platform, Next.js, Expo...)	https://github.com/upstash/context7
Playwright	Driving a browser against the local stack	https://github.com/microsoft/playwright-mcp
openapi-spec	Loading the backend's OpenAPI document as context	for example https://docs.apidog.com/apidog-mcp-server, pointed at https://localhost/docs.jsonopenapi

openapi-spec needs the local stack to be running. The name matters: it is the one .claude/settings.json enables.

This backend also exposes its own MCP server at /mcp, for end users' AI agents rather than for development. See Connect an AI agent and MCP tools.