AGENTS.md · git:20261006.08c70dd · 2026-10-06 · sha256 4b51ccb6bc8abbe0
AGENTS.md git:20261006.08c70ddA
Immutable. This exact content never changes and is served at /api/v1/blob/4b51ccb6bc8abbe0.
# AGENTS.md — PostHog Wizard Instructions for all agents (and humans) working in this repo. This is the single source of truth; [`CLAUDE.md`](CLAUDE.md) just points here. User-facing docs: https://posthog.com/docs/ai-engineering/ai-wizard The PostHog wizard (`npx @posthog/wizard`) is a CLI that adds PostHog to a user's project using an AI agent. It authenticates the user, detects their framework, runs an agent that integrates the SDK and instruments events, and walks the user through their first dashboard. All from the terminal. ## Design discipline This codebase follows a specific design discipline: **product knowledge never enters infrastructure code.** The runner pipeline, the TUI store, the detection loop, and the prompt assembler are machinery. They don't know what PostHog is. They don't know what a framework is. They execute a pipeline driven by typed configuration surfaces. Each domain has a dedicated boundary: - **Frameworks** → `FrameworkConfig` in `src/programs/frameworks/<name>/` - **Integration knowledge** → markdown skills in the [context-mill](https://github.com/PostHog/context-mill) repo - **Security policy** → YARA-X rules in the [warlock](https://github.com/PostHog/warlock) sibling repo. The wizard wires the scanner through SDK hooks and Pi tool events; see [security boundaries](.claude/skills/wizard-development/references/ARCHITECTURE.md#security-boundaries). Gateway admission and required safety prompts live in [ai-gateway](https://github.com/PostHog/ai-gateway). To disable scanning in the field without a release, see the kill-switch runbook: `docs/runbooks/warlock-kill-switch.md`. ONLY USE THIS IF ABSOLUTELY NECESSARY. - **Agent** → `src/agent/`. Layer source outside it imports it only through `@agent` (values) and `@agent/types` (types); see [src/agent/README.md](src/agent/README.md) - **Shared** → `src/shared/`, library code with no upward imports. It also holds process-global state: the log file, the analytics client, the exit cleanup list and the process's one OAuth login session (`@shared/oauth-session`); see [src/shared/README.md](src/shared/README.md) - **Host** → `src/host/` (`@host/*`), how the hosts end a run: `startHostExit`, `wizardAbort` and `registerShutdown`. Every `wizardAbort` call passes the presenter that shows its outro. A host resolves an exit code and only the CLI calls `process.exit`. It imports env and shared only. Headless, the TUI, the CLI and the e2e harness may import it. The agent and the programs may not; see [src/host/README.md](src/host/README.md) - **Programs** → one folder per program in `src/programs/<id>/`, which other layer source enters only through its `index.ts`, plus `runProgram`. A program is an agent run; see [src/programs/README.md](src/programs/README.md) and the [developer interfaces](docs/developer-interfaces.md) - **Tools** → `src/tools/` (`@tools`, one entry), the commands that run no agent: `mcp add`, `mcp remove`, `mcp tutorial`, `slack`, `doctor`, `provision`, `cli add` and `skill list`. A tool may import env, shared and the `@agent` entry (the MCP tutorial's prompt stream only), never the programs, host, TUI, headless or CLI, and no program imports it. Its screens live in `src/tui/tools/<id>/`, run by `runTuiTool`; every tool runner resolves an exit code. See [src/tools/README.md](src/tools/README.md) - **TUI** → the interactive host (`runTui`), screens, primitives and the screen store in `src/tui/`, which other layers enter only through its one entry, `@tui` (`src/tui/index.ts`); see [src/tui/README.md](src/tui/README.md). Each program's flow, deck and screens live in `src/tui/programs/<id>/`, and each tool's in `src/tui/tools/<id>/`, entered through its `index` module; the TUI core reaches them only through the TUI program and tool registries. A program's flow uses a step a tool also shows, such as the MCP install, by its core screen id - **Headless** → the host with no screens (`runHeadless`) and `LoggingUI` in `src/headless/`, which other layers enter only through its one entry, `@headless` (`src/headless/index.ts`); see [src/headless/README.md](src/headless/README.md). The two hosts share only the kind of session store they build, `SessionStore` in `@programs`, and the host layer; neither loads the other's code, and each calls `runProgram`; see [what calls what](README.md#what-calls-what) - **CLI** → commands, runners and the choice of host in `src/cli/`, the only layer that imports both the TUI and headless, each through its entry. `bin.ts` checks the Node version and loads `main.ts`, which imports the CLI through its entry, `@cli` (`src/cli/index.ts`); see [src/cli/README.md](src/cli/README.md) - **Layers** → one tsconfig project per layer folder, on TypeScript project references, and `pnpm typecheck` is `tsc -b`. A project's `references` name the layers it may import, so importing a file of any other fails with TS6307. Outside the TUI, `ink`, `react`, `@inkjs/ui` and `ink-testing-library` resolve to a fence that fails every import form. ESLint rules in `pnpm lint` close the paths the compiler can't see: a relative import that leaves its layer's folder and a deep alias outside its own layer; a `.tsbuild/` or `node_modules/` path in an import, re-export, `import()` or `typeof import()`; any import of `module` or `node:module`; a bare `require`; and triple-slash `path` references; see [layer boundaries](.claude/skills/wizard-development/references/ARCHITECTURE.md#layer-boundaries) and [import aliases](README.md#import-aliases). Code outside a layer imports it only through its public entries. Each layer's tests sit in its project, so a test deep-imports only its own layer and mocks another through its entry. Adding a new concern means finding the narrowest existing surface, not adding logic to the runner. Keep changes local to the boundary that owns them. ## Before making structural changes Read [wizard-development](.claude/skills/wizard-development/SKILL.md) first. It covers the design discipline, a decision framework for new extensions, and warning signs that a change is drifting off-pattern. Its references extend it: - [Architecture](.claude/skills/wizard-development/references/ARCHITECTURE.md) — runner, data flow, security boundaries, screen resolution - [Anti-patterns](.claude/skills/wizard-development/references/ANTI-PATTERNS.md) — failure modes and alternatives - [Maintaining skills](.claude/skills/wizard-development/references/MAINTAINING-SKILLS.md) — accuracy, references, and verification ## Skills available Five skills live under `.claude/skills/`. Read `wizard-development` first for any structural change; then load the relevant procedural skill: | Skill | When to use | | ---------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- | | [wizard-development](.claude/skills/wizard-development/SKILL.md) | Before any structural change. Design principles + decision framework. | | [adding-framework-support](.claude/skills/adding-framework-support/SKILL.md) | Adding or extending a framework integration. | | [adding-skill-program](.claude/skills/adding-skill-program/SKILL.md) | Adding a new skill-based program (e.g. a new product feature setup). | | [ink-tui](.claude/skills/ink-tui/SKILL.md) | Building or modifying TUI screens, layouts, and primitives. | | [exploring-the-wizard](.claude/skills/exploring-the-wizard/SKILL.md) | Running/driving/exploring the wizard headlessly (read_state/perform_action, TUI snapshots). | ## Agent execution policy Default new work to **Pi**, and prefer the **orchestrator** sequence. Linear execution remains useful for very simple tasks and legacy support. The Anthropic Agent SDK is a supported legacy fallback, deprecated as the default; retain it for major Pi vulnerabilities or gaps in support for new Anthropic models. This is the contribution policy; existing programs' bindings vary, and `DEFAULT_BINDING` is Pi + linear. Set new bindings explicitly and check sequence-specific hooks before migrating existing flows. See [execution policy and model admission](.claude/skills/wizard-development/SKILL.md#execution-policy-and-model-admission) for the gateway allowlists, required system prompt, and composition constraints. ## CLI command surface The CLI was overhauled to a smaller, extensible command surface. **Use the new command names.** Old names mostly no longer exist — only some are kept as aliases. | Old command | New command | Status | | -------------------------- | --------------------------- | ---------------------------------------------------------- | | `wizard integrate` | `wizard` (default flow) | command removed | | `wizard events-audit` | `wizard audit events` | moved into `audit` family | | `wizard audit` (single) | `wizard audit <subcommand>` | now a family — see [Audit subcommands](#audit-subcommands) | | `wizard audit-3000` | _removed_ | retired | | `wizard revenue` | `wizard revenue-analytics` | renamed (old `revenue` removed) | | `wizard upload-sourcemaps` | `wizard upload-source-maps` | renamed; `upload-sourcemaps` kept as alias | ### Audit subcommands `audit` is the only family with skill-backed subcommands today: | Subcommand | What it audits | | ----------------------------- | ---------------------------------------------------- | | `wizard audit events` | event capture quality + cost | | `wizard audit all` | comprehensive audit across every area (**default**) | | `wizard audit autocapture` | autocapture setup + cost | | `wizard audit feature-flags` | feature flag usage + cost | | `wizard audit identify` | `$identify` implementation | | `wizard audit session-replay` | session replay setup | | `wizard audit web-analytics` | web analytics setup (**wizard-native**, not a skill) | ### Commands vs. skills (the `audit [skill]` gotcha) A skill and a command are the **same machinery** — a context-mill skill becomes a command when its `cli:` block sets `role: command`. So `wizard audit events` _is_ the `audit-events` skill, just promoted. `wizard skill <skill-name>` ([`skill.ts`](src/cli/commands/skill.ts)) runs a skill that **wasn't** promoted. Two surfaces, one mechanism. So `wizard audit <subcommand>` is choosing an audit area — it is **not** asking for a skill name, despite `wizard audit --help` labelling the positional `[skill]` (a wizard-internal name we left as-is). Don't confuse it with the top-level `wizard skill` command. ### Where the surface is defined (source of truth) - **Registration:** `runCli` in [`src/cli/index.ts`](src/cli/index.ts) — the `.use()` chain wires each command, and [`main.ts`](main.ts) calls it once the Node check in [`bin.ts`](bin.ts) passes. Each command has its own file or folder in `src/cli/commands/`. - **Command shape:** [`src/cli/commands/command.ts`](src/cli/commands/command.ts) — the `Command` interface every command implements. - **Flat native commands** (e.g. `revenue-analytics`, `upload-source-maps`) are built with `nativeCommandFactory` ([`src/cli/commands/factories/native-command-factory.ts`](src/cli/commands/factories/native-command-factory.ts)) in a one-line command file, such as [`revenue.ts`](src/cli/commands/revenue.ts). - **Family commands** (e.g. `audit`) resolve subcommands at runtime against the `cliEntries` in `skill-menu.json`. Logic lives in [`src/cli/commands/dispatch-family.ts`](src/cli/commands/dispatch-family.ts). Adding a skill-backed subcommand is a **context-mill** release, not a wizard change. ### Commands vs. programs (don't confuse these) - A **command** is the word a user types (`audit`, `revenue-analytics`). - A **program** is the internal business logic (`posthog-integration`, `revenue-analytics-setup`) that a command invokes, and that other programs depend on via `requires: [...]`. - `posthog-integration` is a **program id, not a command**. It powers the default flow and is a dependency of most other programs. Do not treat it as a CLI command or reference it in CI as one. ### Adding a command alias (keep an old name working) Give the `Command.name` an array of `[newName, ...legacyNames]`. yargs treats the extra entries as aliases. See [`src/cli/commands/upload-sourcemaps.ts`](src/cli/commands/upload-sourcemaps.ts). Reserve aliases for names that external callers (users' scripts) may still use — when the only caller is one we control, update the caller instead. ## Commands ```bash pnpm install # Install dependencies pnpm try --install-dir=<path> # Run the wizard locally against a test project pnpm build # Compile TypeScript pnpm test # Unit tests (builds first) pnpm test:watch # Unit tests in watch mode pnpm test:e2e # End-to-end tests pnpm lint # Prettier + ESLint checks pnpm typecheck # tsc -b, one project per layer pnpm fix # Auto-fix lint issues pnpm dev # Build, link globally, watch for changes ``` Choose verification for the change: check links and formatting for docs; run `pnpm typecheck` and focused existing tests for code. Build when bundling or runtime behavior changes. `pnpm test` already builds; avoid building twice. Use nonmutating lint checks, and scope formatting fixes to edited files. Do not add tests for prose, compiler-enforced shapes, or duplicated implementation. Keep new code comments to one line; put longer explanations in linked docs. Local `--ci`, smoke-test, and full headless runs require two separate secrets: a PostHog personal API key and an already-issued gateway token supplied through `WIZARD_CI_GATEWAY_TOKEN_FILE`, plus the target project ID. Follow the [credential setup](docs/local-dev.md#credentials-for-local-ci-and-headless-runs). ### Local dev targets Four things can independently point at a local server — the wizard binary, context-mill (`:8765`), the MCP server (`:8787`), and PostHog (`:8010`). One flag per service (`--local-context-mill`, `--local-mcp`, `--local-posthog`), plus `--local-dev` for all three. Only non-production builds accept them, such as `pnpm try` and `pnpm build:ci`; production builds, including the published package and the `pnpm dev` link, reject them. Note `wizard mcp add --local` is **not** one of these — it writes a `posthog-local` entry into your editor's MCP config, and is unrelated to where a wizard run points. Full catalog: [`docs/local-dev.md`](docs/local-dev.md). ## Repository conventions - Write documentation as the current design. Describe behavior and usage directly; omit change history, migration narration, and implementation rationale. Apply the same style to PR descriptions. - TypeScript everywhere. Use `type` (not `interface`) for framework context types so they satisfy `Record<string, unknown>`. - Programs never use the UI. A program's `run` and `ciPreRun` get a runner context, and `onReady` gets a ready context. TUI code reports through the store it is handed, headless through its `LoggingUI`, and the CLI through `consoleLog`; there is no UI interface shared between hosts, and nothing looks a UI up. Never import the store directly from business logic. - Shared and host helpers take a sink or return data; there are no process-global sinks. An agent run's debug lines are info log progress, shown by the host's progress handler. An abort takes its presenter as its first argument: `abortOnScreens(store)` in the TUI, `printAbortOutro` in headless and before the TUI mounts. - Outside `src/agent`, import the agent through `@agent` or `@agent/types`, and every other layer through its entries. An entry re-exports only from layers its consumers also map, so no declaration turns to `any` in a consumer. Add to an entry rather than deep-importing; `pnpm typecheck` and `pnpm lint` reject a deeper path in layer source, tests, the e2e harness, scripts and docs examples. A test of another layer's internals belongs in that layer's tests, and a test helper beside the tests that use it. - Session mutations go through explicit store setters, which notify the store's subscribers. Never mutate `session` directly — nanostore holds a shallow copy. - State only the TUI reads (a screen's answer, what an overlay shows) lives in the TUI store (`TuiState`, read as `store.X`), never in the shared session. - The router resolves the active screen from the session and the TUI state. No imperative navigation (`goTo`, `navigate`, `push`) anywhere. - Never write secrets to source code or hardcode API keys. Use the `wizard-tools` MCP server (`check_env_keys` / `set_env_values`) for `.env` file operations. - Feedback / issues: wizard@posthog.com or [GitHub Issues](https://github.com/posthog/wizard/issues). ## Companion projects - **[context-mill](https://github.com/PostHog/context-mill)** — builds and publishes the markdown skills the wizard agent uses for framework-specific integration knowledge. Skills are decoupled from the wizard release cycle so docs and integration patterns can update independently. - **[wizard-workbench](https://github.com/PostHog/wizard-workbench)** — the development and testing environment. Houses framework test apps (Next.js, React Router, Django, Flask, Laravel, SvelteKit, Swift, TanStack, FastAPI) with no PostHog installed, plus an `mprocs`-driven local dev stack that runs context-mill + MCP + the wizard together with hot reload. Use this to develop and test wizard changes against real projects. - **[warlock](https://github.com/PostHog/warlock)** — the security scanner engine for PostHog's agentic flows. Bundles YARA-X rules for prompt injection, exfiltration, destructive operations, supply chain attacks, hardcoded secrets, and PII. Engine-only: it returns matches with category/severity/action metadata; the wizard decides how to respond. New security rules belong in warlock, not in the wizard.