This project uses [Task](https://taskfile.dev/) as a unified command runner. All build, dev, test, lint, and docker commands can be run from the repo root via `task <command>`. Run `task --list` to see all available commands.
Task `desc:` fields should describe **what** the task does, not **how** it does it. Keep them generic and stable: don't reference implementation details like aliases, internal helpers, mode flags, or which other task delegates to which. The description is for users picking a command from `task --list`, not a changelog of refactors.
A comment must carry information the code cannot. If a reader could derive it from the code in front of them, delete it.
Comment the current state. Not what the code used to do, not what changed, not why it changed: git holds that. Where history explains the shape, state the reason instead, so "this used to reimplement the modal internals" becomes "thin wrapper over the shared Modal: duplicating its portal and focus trap is how dialogs drift apart". Future state goes in a TODO with an issue.
Write a comment when it does one of these four jobs:
- **Contract.** What a caller must know that the signature cannot say: preconditions, invariants, units, ownership and lifetime, thread-safety, error semantics, side effects. Document the contract of everything a caller outside the file can reach, and nothing else. Goes on the type/method/module as Javadoc, JSDoc, or a docstring.
- **Why.** The constraint the code satisfies, the bug it avoids, the alternative rejected and the reason.
- **Hazard.** "Must stay in sync with X", "order matters because Y", "do not remove, it prevents Z".
- **Map.** A short orientation at the top of a genuinely complex file: what it owns, and what it deliberately does not.
Never write:
- A comment that restates the next line. `// Handle drag start` above `handleDragStart` is noise.
- Section banners or position markers: `// --- Types ---`, `// Helpers`, `// =====`.
- Step narration in a function body (`// Step 1:`, `// Then we`). If the steps need labels they need names: extract functions. Numbering a genuinely numbered thing, like a wizard step, is fine.
- Commented-out code. Delete it.
- Doc tags that restate the signature. `@param blob - The blob` says nothing; omit the tag rather than pad it.
- Docs on self-explanatory members with no constraint to state.
Two tests before keeping a comment:
- **Delete it.** Is any information lost? If not, it stays deleted.
- **Could a name carry it instead?** A better identifier, an extracted function, or a named constant beats a comment. Prefer the code change.
A comment at the end of a line usually decodes that line, and that is worth keeping: `{0x25, 0x50} // "%PDF"`, `50L * 1024 * 1024 // 50 MB`. The rules that compare a comment against the code below it do not apply there, but a trailing TODO or a trailing bit of history is judged like any other.
A reference is supplementary, never load-bearing: the comment must survive deleting it. `// See #1234` is a dead end; `// saving first loses every annotation (#6865)` is not. Prefer a spec (`RFC 3161`) or CVE where one applies.
A TODO needs an issue, not an owner: `// TODO(#1234): re-enable the gate once account syncing lands`. If it is not worth an issue, it is not worth a TODO. A question is not a TODO.
A comment block over ~12 lines outside a file or type header usually means the code needs restructuring, or that the prose is product documentation and belongs in the docs repo.
`task comment-lint` checks the mechanical part of this on the lines you add, and runs inside `task pre-commit`. Reasoning, worked examples and the linter's own rules: @devGuide/CODE_COMMENTS.md
After modifying any files in the project, you must run the relevant `task check` command that covers that area of the code. For example, when editing frontend files run `task frontend:check`; for Python engine files run `task engine:check`; for Java backend files run `task backend:check`.
- **Example compose files**: Located in `exampleYmlFiles/` directory
### Security Mode Development
Set `DOCKER_ENABLE_SECURITY=true` environment variable to enable security features during development. This is required for testing the full version locally.
The engine is a Python reasoning service for Stirling: it plans and interprets work, but it does not own durable state, and it does not execute Stirling PDF operations directly. Keep the service narrow: typed contracts in, typed contracts out, with AI only where it adds reasoning value. The frontend calls the Python engine via Java as a proxy.
#### Python Commands
All engine commands run from the repo root using Task:
-`task engine:check` — run all checks (typecheck + lint + format-check + test)
-`task engine:fix` — auto-fix lint + formatting
-`task engine:install` — install Python dependencies via uv
-`task engine:dev` — start FastAPI with hot reload (localhost:5001)
-`task engine:test` — run pytest
-`task engine:lint` — run ruff linting
-`task engine:typecheck` — run pyright
-`task engine:format` — format code with ruff
-`task engine:tool-models` — generate `tool_models.py` from the Java OpenAPI spec
The project structure is defined in `engine/pyproject.toml`. Any new dependencies should be listed there, followed by running `task engine:install`.
#### Python Code Style
- Keep `task engine:check` passing.
- Use modern Python when it improves clarity.
- Prefer explicit names to cleverness.
- Avoid nested functions and nested classes unless the language construct requires them.
- Prefer composition to inheritance when combining concepts.
- Avoid speculative abstractions. Add a layer only when it removes real duplication or clarifies lifecycle.
- Deserialize into Pydantic models as early as possible.
- Serialize from Pydantic models as late as possible.
- Do not pass raw `dict[str, Any]` or `dict[str, object]` across important boundaries when a typed model can exist instead.
- Avoid `Any` wherever possible.
- Avoid `cast()` wherever possible (reconsider the structure first).
- All shared models should subclass `stirling.models.ApiModel` so the service behaves consistently.
- Do not use string literals for any type annotations, including `cast()`.
#### Python Configuration
- Keep application-owned configuration in `stirling.config`.
- Only add `STIRLING_*` environment variables that the engine itself truly owns.
- Do not mirror third-party provider environment variables unless the engine is actually interpreting them.
- Let `pydantic-ai` own provider authentication configuration when possible.
#### Python Architecture
**Package roles:**
-`stirling.contracts`: request/response models and shared typed workflow contracts. If a shape crosses a module or service boundary, it probably belongs here.
-`stirling.models`: shared model primitives and generated tool models.
-`stirling.agents`: reasoning modules for individual capabilities.
-`stirling.api`: HTTP layer, dependency access, and app startup wiring.
-`stirling.services`: shared runtime and non-AI infrastructure.
-`stirling.config`: application-owned settings.
**Source of truth:**
-`stirling.models.tool_models` is the source of truth for operation IDs and parameter models.
- Do not duplicate operation lists if they can be derived from `tool_models.OPERATIONS`.
- Do not hand-maintain parallel parameter schemas when the generated tool models already define them.
- If a tool ID must match a parameter model, validate that relationship explicitly in code.
**Boundaries:**
- Keep the API layer thin. Route modules should bind requests, resolve dependencies, and call agents or services. They should not contain business logic.
- Keep agents focused on one reasoning domain. They should not own FastAPI routing, persistence, or execution of Stirling operations.
- Build long-lived runtime objects centrally at startup when possible rather than reconstructing heavy AI objects per request.
- If an agent delegates to another agent, the delegated agent should remain the source of truth for its own domain output.
#### Python AI Usage
- The system must work with any AI, including self-hosted models. We require that the models support structured outputs, but should minimise model-specific code beyond that.
- Use AI for reasoning-heavy outputs, not deterministic glue.
- Do not ask the model to invent data that Python can derive safely.
- Do not fabricate fallback user-facing copy in code to hide incomplete model output.
- AI output schemas should be impossible to instantiate incorrectly.
- Do not require the model to keep separate structures in sync. For example, instead of generating two lists which must be the same length, generate one list of a model containing the same data.
- Prefer Python to derive deterministic follow-up structure from a valid AI result.
- Use `NativeOutput(...)` for structured model outputs.
- Use `ToolOutput(...)` when the model should select and call delegate functions.
#### Python Testing
- Test contracts directly.
- Test agents directly where behaviour matters.
- Test API routes as thin integration points.
- Prefer dependency overrides or startup-state seams to monkeypatching random globals.
-`frontend/editor/.env` — core and shared vars (base, loaded in every mode)
-`frontend/editor/.env.proprietary` — proprietary-only vars, e.g. the admin portal's SaaS/account-link keys (layered on top of `.env` in proprietary mode)
- These files are committed to Git and must not contain private keys
- Local overrides (API keys, machine-specific settings) go in uncommitted sibling `.env.local` / `.env.saas.local` / `.env.desktop.local` files — Vite automatically layers them on top
- Never use `|| 'hardcoded-fallback'` inline — put defaults in the committed env files
-`task frontend:prepare` creates empty `.local` override files on first run; pass `MODE=saas` or `MODE=desktop` to also create the mode-specific `.local` file
Before touching colours or theming (tokens, dark mode, accent colours), read @frontend/editor/src/core/theme/README.md — it explains the palette/`--c-*` token system and the rule that literal colours live only in `primitives.css`.
- Building layer-specific override that wraps a lower layer's component
- Example: `import { AppProviders as CoreAppProviders } from "@core/components/AppProviders"` when creating proprietary/AppProviders.tsx that extends the core version
The `@app/*` alias automatically resolves to the correct layer based on build target (core/proprietary/saas/desktop/cloud) and handles the fallback cascade — see "Frontend `cloud/` Layer" below for the full per-flavor order.
#### Frontend `cloud/` Layer
`@app/*` resolves through a per-flavor cascade — first existing file wins (shadow/override):
`cloud/` MUST NOT import `@supabase/*`, `@tauri-apps/*`, raw `fetch`, `window.location`, `localStorage`, `sessionStorage`, or `import.meta.env.VITE_*` (all enforced by the linter). It reaches platform-specific things only via `@app/*` seams: `services/apiClient`, `auth/session.getAccessToken`, `auth/supabase`, `platform/openExternal`, `services/billing`, `hooks/useSaaSMode` — each provided per-platform in `saas/` and `desktop/`.
**Cloud feature flags on desktop.** The local `AppConfigContext` reads `/api/v1/config/app-config` from the LOCAL bundled backend, so cloud-only flags (`aiEngineEnabled`, `premiumEnabled`, …) are never seen on desktop. To read the cloud's view, use `useSaasAppConfig()` (`desktop/hooks/useSaasAppConfig.ts`, backed by the general `saasAppConfigService` — SaaS-mode-only, public endpoint, native HTTP, 5-min cache). It returns `null` outside SaaS mode, so cloud features stay off in local/self-hosted and the server keeps the on/off switch (no desktop release needed to flip a flag). Gate a feature behind a per-platform seam — e.g. `useAiEngineEnabled()` (core reads `useAppConfig()`, desktop reads `useSaasAppConfig()`) — rather than hardcoding the flag on.
