This commit is contained in:
2026-07-07 16:09:28 +02:00
parent 32c8bbe0a9
commit 5083efb150
42 changed files with 2556 additions and 2806 deletions
+3 -3
View File
@@ -18,7 +18,7 @@ All application code is under `devplacepy/`: `devplacepy/routers/`, `devplacepy/
Default to **PLAN** mode. Investigate the area, then return a layer-by-layer implementation plan and STOP - do not write code until the invocation approves the plan or explicitly asks you to implement directly ("implement", "just do it", "no plan needed"). Once approved (or when invoked in implement mode), build the whole feature, then validate. Never run the test suite; never perform any git write operation.
## Operating protocol
1. **Understand before writing.** Read the router, template, matching tests, the relevant `CLAUDE.md`/`AGENTS.md` domain section, and trace the existing data flow (input model -> router -> data helper -> HTML and JSON response) before proposing anything. Reuse beats re-implementation: find the canonical helper/partial/component and use it.
1. **Understand before writing.** Read the router, template, matching tests, the relevant nested `CLAUDE.md` (each subsystem directory has its own, e.g. `devplacepy/services/devii/CLAUDE.md`) and the root `CLAUDE.md` for any cross-cutting rule, and trace the existing data flow (input model -> router -> data helper -> HTML and JSON response) before proposing anything. Reuse beats re-implementation: find the canonical helper/partial/component and use it.
2. Use Grep/Glob for discovery; read the relevant range, not whole large files. Never repeat a grep or re-read a file you already read.
3. Match the surrounding code: its naming, structure, comment density (none), and idioms. A new feature must be indistinguishable in style from the area it lives in.
4. Build the fan-out coherently in one pass - changing one layer and forgetting a connected one is the cardinal failure here.
@@ -26,7 +26,7 @@ Default to **PLAN** mode. Investigate the area, then return a layer-by-layer imp
## Research the task before designing (codebase first, web when external)
Investigation is two passes, in order:
1. **Codebase pass (always).** Read the router, template, matching tests, and the relevant `CLAUDE.md`/`AGENTS.md` section; trace the existing data flow (input model -> router -> data helper -> HTML and JSON response); find the canonical helper, partial, or component to reuse. Never design from assumption when the answer is in the repo.
1. **Codebase pass (always).** Read the router, template, matching tests, and the relevant nested `CLAUDE.md` (plus the root `CLAUDE.md` for cross-cutting rules); trace the existing data flow (input model -> router -> data helper -> HTML and JSON response); find the canonical helper, partial, or component to reuse. Never design from assumption when the answer is in the repo.
2. **Web pass (whenever the feature touches anything outside this repo).** If the work integrates a third-party API or protocol, a library's correct usage, a new dependency, a file format, standard, or spec, external provider or model behavior, or a security consideration, run a focused WebSearch/WebFetch pass BEFORE designing. Pull the authoritative, current contract - exact endpoints, parameters, request and response shapes, auth, limits, version differences, and known bugs or quirks - and cite the sources in your plan. Prefer official docs and corroborate version-specific details. Do not design an external integration from memory: one wrong assumption about the external contract (a field name, an auth header, a documented bug such as a query-param that must be avoided) silently breaks the feature. Skip this pass only for purely internal features with no external surface.
When the external contract and the internal system must meet (for example an external API mirrored onto an internal store), resolve every mismatch in the plan - identity and ownership mapping, allowed-value or type differences, failure and partial-failure handling - before writing code.
@@ -41,7 +41,7 @@ A DevPlace feature is one data source fanning out into several consumers, all fr
5. **View** - templates extend `base.html` (page CSS in `extra_head`, page JS in `extra_js`); import the shared `templates` from `devplacepy.templating`, never instantiate `Jinja2Templates`. Wrap every static asset URL in `static_url(...)`/`assetUrl(...)`. Reuse partials (`_avatar_link.html`, `_user_link.html`, `_sidebar_search.html`) and the shared frontend utilities (`Http`, `Poller`, `JobPoller`, `OptimisticAction`, `FloatingWindow`, the `dp-*` components) - never hand-roll fetch/polling. JS is ES6 modules, one class per file, on `app`. Dates are DD/MM/YYYY via `format_date`.
6. **Agent + docs (the most-forgotten layers)** - if a user could ask Devii to do it, add an `Action` in `services/devii/actions/catalog.py` with auth flags matched to the route guard (and a declared `confirm` boolean for any irreversible action added to `CONFIRM_REQUIRED`). Add a `docs_api.py` `endpoint()` entry (params + `sample_response`) for every public/auth endpoint; add a prose page to `routers/docs/pages.py` `DOCS_PAGES` when warranted. State-changing actions need an audit event (`events.md` key, `category_for`, recorder call at the mutation point).
7. **SEO** - public pages build `base_seo_context` and the right JSON-LD; add to `routers/seo.py` sitemap when indexable.
8. **Docs of record** - update `README.md` (product-facing) and `AGENTS.md` (deep companion) for any new route/config/dependency/mechanic; update `CLAUDE.md` only when a NEW architectural rule or convention is introduced.
8. **Docs of record** - update `README.md` (product-facing) and the relevant nested `CLAUDE.md` (deep companion for the subsystem you touched - create one if the directory doesn't have one yet) for any new route/config/dependency/mechanic; update the root `CLAUDE.md` only when a NEW cross-cutting architectural rule or convention is introduced, and add a row to its "Subsystem map" table if you created a new nested `CLAUDE.md`.
9. **Tests (a hard project requirement, never optional)** - the DevPlace suite is one test file per endpoint, ~932 tests, split into three tiers with the directory tree mirroring the URL/source path. Every feature gets a test in EVERY tier it exercises: `tests/unit/` for a new data/query helper (pure in-process, `local_db` or no fixture, path mirrors the SOURCE module - `devplacepy/utils.py` -> `tests/unit/utils.py`); `tests/api/` for a new JSON or HTML route (HTTP integration against the live uvicorn subprocess via `app_server`/`seeded_db`, path mirrors the endpoint - `POST /auth/login` -> `tests/api/auth/login.py`) - but when a route depends on an in-process injected fake or a module-level singleton the separate uvicorn subprocess cannot see (the Gitea client via `runtime.set_client(fake)`, or any other `set_client`/monkeypatched backend), test it IN-PROCESS instead with `from starlette.testclient import TestClient; TestClient(m.app)`, the fake set in the test process, and auth via a `create_session(uid)` `session` cookie, asserting JSON with `Accept: application/json` (the `tests/api/issues/` files are the canonical example); `tests/e2e/` for a new interactive UI flow (Playwright `page`/`alice`/`bob`, path mirrors the endpoint - `GET /admin/ai-usage` -> `tests/e2e/admin/aiusage.py`). A route or feature with no test in any tier is incomplete. Follow the required patterns (`wait_until="domcontentloaded"` on every `goto`/`wait_for_url`, scoped selectors, `try/finally` restore of any flipped global setting, the shared fixtures, `test_`-prefixed functions in non-prefixed files, born-live `deleted_at`/`deleted_by` on raw soft-delete inserts) and create any missing package directories (`__init__.py`). WRITE them; validate each by a clean import only; NEVER run them.
When a layer is intentionally absent (an internal route with no public docs, a route Devii must never call), say so explicitly in the plan with the rationale rather than fabricating the layer.