Working on this repo

Read current context first

Before planning work, read docs/plan.md for the decisions and the reasoning behind them, and docs/implementation.md for the dated record of what actually went wrong and how it was resolved — that one is the faster read when debugging, because most entries describe something that looked like one bug and turned out to be another. Then docs/NEXT_STEPS.md for current recommendations and known deferred fixes. Mobile interaction work also has its own detailed plan in docs/mobile_enhance.md. The CDN-to-Vite move is docs/vite_build.md. Light/dark theme constraints are docs/plan_light_dark_themes.md. Daily MA alert toggles are docs/plan_dma_alerts.md.

Tests earn their place by catching a real bug

When a bug is found, ask whether a unit test could reasonably have caught it. If yes, write that test with the fix. If no — a rendering artefact, a browser quirk, a data-source oddity — say so and don't add one.

The bar is "would this have failed before the fix, and would it fail again if someone reintroduced it". Tests that restate the implementation, assert constructor defaults, or exercise paths nothing depends on are noise; they make the suite slow to run and expensive to change, which is how a suite stops being trusted.

What has actually paid off here: bar aggregation and bucket boundaries, the store's replace-vs-append rules, level and alert arithmetic, parsing real market-data payloads (fixtures are trimmed real responses, not invented), and the invariants that would otherwise be silent — a comment must never become a level, a tick must never overwrite a settled bar, volume must be counted once.

Name the test after the failure, not the function: test_a_tick_cannot_overwrite _a_settled_bar beats test_put.

Where things run

The agent works on a remote machine over SSH. The user's browser runs on a different machine. Consequences, all learned the hard way:

You cannot see the user's screen, console, or cursor. Screenshots and pasted console output are the only window into it. Browser extensions that drive "your" Chrome do not help — they attach to the machine the browser is on.
Headless Chromium here renders on server hardware: different screen, window size and device pixel ratio from the user's. "Works in my headless run" is not evidence that it works for them. When a UI bug will not reproduce, match their viewport and deviceScaleFactor explicitly before concluding anything.
The dev stack is served to them over the network (e.g. hera.local:8010), which is the same app the headless browser reaches as http://api:8000.

When a visual bug resists reproduction, prefer putting the numbers on screen in the app over asking for another console paste — one screenshot then carries the whole diagnosis.

Verify UI in a real browser

Chart bugs are invisible from the outside — the API, the socket and the frontend source can each be correct while the screen is wrong. Drive the Playwright container against the dev stack:

docker exec -i chart-playwright-1 node - <<'EOF'
const { chromium } = require('/usr/lib/node_modules/playwright');
// launch with args:['--lang=en-US'] — see below
EOF

Always launch Chromium with args: ['--lang=en-US']. The container has no usable locale, so Chromium reports en-US@posix, Intl throws, and the chart renders as a blank canvas that looks exactly like a broken app.

window.__chart is a deliberate debug handle. Querying it separates "the data is missing" from "the data is off-screen" — which is how a viewport bug that three passing API checks had missed was finally found.

Diagnostic mode

Chart geometry bugs live in the browser, which is usually on a different machine from whoever is debugging them. Rather than asking for console pastes:

open the chart with ?diag=1          # remembered until ?diag=0
docker compose logs api | grep SNAPDBG

With it on, every snap the trendline tool computes is posted to /api/debug/snap and logged server-side — the cursor's time, price and x, the snapped time and price, how many bars were held, the first and last bar, and the chart's width. Throttled to about one a second. It reads the client's own numbers, which is exactly what "works in my headless run" cannot tell you.

Extend it when the next geometry puzzle appears; the endpoint takes whatever fields SnapReport declares.

?diag=1 also exposes Capture diagnostic. The uploaded PNG URL at /api/debug/captures/{id} is deliberately public: its 72-bit id is the handoff from a browser to an agent on a different machine. Capture upload and metadata remain authenticated. Inspect only a URL the user explicitly shares, then immediately DELETE /api/debug/captures/{id}. The server also expires captures after 24 hours and caps the directory at 50 files.

After the browser's required share picker closes, capture waits five seconds so the user can restore a hover tooltip. Alt+Shift+C starts the same delayed flow without clicking the status-bar button.

Diagnostic mode also shows a compact projection readout for visible manual trendlines: historical/future canonical price changes, their screen slopes, and whether the future canvas point exists. Include it in a capture when a line looks kinked at the live edge; it separates bad geometry from a bad renderer.

Keep the two documents current

This is a running system under continual change, not a build being executed, so both live documents decay unless updating them is part of finishing the work — not a tidy-up afterwards.

docs/plan.md — when a decision changes, change it here. A plan that contradicts the code is worse than no plan, because someone believes it. If you find a section describing behaviour that no longer exists, that is a bug in the document; fix it in the same commit that revealed it.
docs/implementation.md — append when a fix was not obvious. The bar is "would this have saved me an hour": wrong theories that were measured and killed, the evidence that settled it, the thing that looked like one bug and was another. Not every fix. A log of trivia stops being read, and then the useful entries go unread too.

Rule of thumb: if you needed a measurement to be sure, write down what it was. Git records what changed; these record why it was hard.

Direction of travel

Two live planning documents, both written to be refactored toward rather than implemented in one go:

docs/async_refactor.md — nothing blocks the event loop. P0 and P1 are done; /api/status reports loop_lag_ms, and a rise there is the signal.
docs/multi_user.md — separate people with their own drawings and alerts, authenticated by OIDC. Read it before adding state to Runtime: new state is either genuinely shared (market data) or belongs to a user, and knowing which now is much cheaper than untangling it later.
docs/vite_build.md — pin and hash the frontend, stay on Coolify, do not split components on the way. A production Dockerfile first, then Vite; never a root package.json while nixpacks is still the builder.

Do not build local user accounts. The destination is OIDC, so password storage would be written and then deleted.

Running tests

docker exec chart-api-1 sh -c "cd /app && python -m pytest -q"

pytest + pytest-asyncio, declared in requirements-dev.txt. Tests live in tests/, import from app.*, and use tmp_path for anything that persists. Async paths are driven with asyncio.run(...) directly rather than async test markers.

Things that will cost you an hour

Never write scratch .py files into the repo root. It is bind-mounted, so --reload restarts the app, and startup takes ~82 seconds. Pipe throwaway scripts over stdin instead: docker exec -i chart-api-1 python - <<'EOF'. Screenshots into artifacts/ are safe; only .py triggers the reloader.
Dev and production keep separate drawing stores. Dev writes data/manual_lines.json; production has its own Coolify volume. A fix that "didn't land" is often the other store.
Rebuild the image after touching requirements.txt. The bind mount makes source edits look live while an added dependency is simply absent.
A deploy resets alert cooldowns, so production may re-alert on whatever price is sitting on. There is no durable state yet.
Times are epoch seconds, UTC, everywhere. Only the display is localised — never shift the stored values.
Do not hardcode how many bars a client gets. A leftover 1000 on the snapshot made 1m look empty past ~1am while the store held 5,000. The store cap is max_bars_per_tf. The next step is a visible-window fetch, not another silent number.

Stay cheap

Performance is a product feature, not a later cleanup. The app already pays for a live stream, 5k bars, and a canvas. New work must not add cost on the hot path unless the screen or an alert has to change.

Nothing extra on the event loop each closed bar or tick. Watch loop_lag_ms. CPU stays in the threadpool or off the loop — see docs/async_refactor.md.
Do not grow a payload because it is easier than asking what the client needs. Caps are named settings, not leftover literals.
Crosshair move is for the cursor (OHLC, drawing tooltip). Do not rebuild overlays that only depend on the viewport.
Overlay redraws go through scheduleOverlays. Force when data or size changed; let the range-key skip no-ops.
window pointermove/up attach only while a tool is armed or a drag is live. Do not leave them on for the life of the page.
Prefer one rAF over N DOM rebuilds. Read offsetWidth only if the layout actually changed.
A feature that needs a per-frame or per-tick loop needs a reason, and an off switch.

9.8 KiB Raw Blame History