Codex Radar Sync

# CodexRadar Sync

Use this skill when the user asks to keep Codex Radar Sentinel aligned with CodexRadar, mentions CodexRadar data format changes, asks why a CodexRadar-backed field is missing, or asks to prepare a release that depends on CodexRadar live data.

## Goal

Keep the macOS menu bar app mapped to the latest public CodexRadar site and endpoint behavior:

- `https://codexradar.com/`
- `https://codexradar.com/current.json` (legacy; may redirect to homepage)
- `https://codexradar.com/feed.xml` (legacy; may redirect to homepage)
- `https://codexradar.com/api/model-ratings` (community model ratings)
- `https://codexradar.com/data/intelligence-efficiency.json` (distributed IQ, cost, and runtime points)
- `https://api.codexradar.com/api/v1/radar-insights` (scenario recommendations and degradation alerts)

## Workflow

1. Fetch the homepage, `current.json`, `feed.xml`, `api/model-ratings`, `data/intelligence-efficiency.json`, and `api.codexradar.com/api/v1/radar-insights`. Note the retrieval date, root keys or redirect target, changed field types, new visible site sections, and any new public links or APIs.
2. Compare the live payloads with `Sources/CodexRadarCore/RadarModels.swift`, `Sources/CodexRadarCore/NotificationPolicy.swift`, `Sources/CodexRadarSentinel/DashboardMenuView.swift`, and `Sources/CodexRadarSentinel/StatusMetric.swift`.
3. Fix decoding before changing UI. JSON fields that may evolve from integer to decimal should use compatible numeric types and a display formatter.
4. Map only useful new CodexRadar capabilities into the macOS app. Prefer clear menu-bar value, compact menu detail, or low-noise notification behavior over exposing raw endpoint complexity.
5. Add or update tests. For live endpoint compatibility, update `Tests/CodexRadarCoreTests/LiveCodexRadarContractTests.swift`.
6. Update README screenshots and docs only after the app renders correctly in Chinese and English.
7. Maintain `PROMPTS.md`: append the triggering user prompt and map it to clickable commit links. Commit messages for prompt-driven work should include `Prompt-Id: N`.

## Release Gate

Before creating or pushing a release, run:

```bash
./scripts/check_release_readiness.sh VERSION
```

This checks live CodexRadar endpoints, runs Swift tests with live contract checks enabled, rebuilds the app, refreshes real status/menu screenshots, packages the release, and verifies checksum plus DMG integrity.

Also inspect the generated screenshots in:

- `docs/assets/zh/status-normal.png`
- `docs/assets/zh/menu-full.png`
- `docs/assets/en/status-normal.png`
- `docs/assets/en/menu-full.png`

If any menu-bar segment shows `--` while CodexRadar has a visible value on the website, treat it as a release blocker.

## Current Known Contract Notes

- As of 2026-06-15, CodexRadar says reset prediction, speed-window reminders, and historical windows are retired. `current.json` may return JSON again with official entitlement events; `feed.xml` may still redirect to homepage HTML.
- When JSON endpoints are unavailable, `CodexRadarClient` falls back to parsing the homepage Model IQ SVG `<title>` values and synthesizes a compatible `RadarCurrent` with `window_open = false`.
- Do not treat every `window_open = true` as a speed window. Current JSON can use `window_open` for official entitlement/reset-card events; only explicit speed/速蹬 wording should trigger speed-window UI and notifications.
- As of app v0.1.28, live UI should treat CodexRadar as a model-quality source first. The menu-bar `signal` metric is still the persisted key, but its user-facing label/value are Quality/质量 from Model IQ unless a legacy speed window or local limit is active.
- As of app v0.1.29, the dropdown should expose the public model-quality details CodexRadar shows on the homepage: runtime, cost, cache hit rate, and community rating. Keep these in the menu, not the default status title.
- As of app v0.1.30, CodexRadar HTTP requests must use `AppConstants.requestTimeoutSeconds`; otherwise one stuck endpoint can block future polling cycles and leave the status bar stale during an active window.
- Legacy CodexRadar schema v2 embedded Prediction and model IQ in `current.json`; keep those decoders because older fixtures and possible future JSON restoration still depend on them.
- `model_iq.latest.score` / homepage IQ values can be decimal, for example `62.5`; do not decode IQ as an integer.
- As of 2026-07-14, CodexRadar describes the 5h limit as temporarily inactive and renders only the active 7d Quota Radar column. `current.json` may still carry derived `five_h` row values while `basis_window_label` is `7d`; follow the basis label for UI visibility instead of exposing values the site intentionally hides.
- The local Codex app-server may return only a 10,080-minute weekly window while 5h is paused. Never infer 5h from the shortest available window; show local 5h UI only for an explicitly returned window near 300 minutes so it can disappear and return dynamically.
- As of 2026-07-17, Model IQ uses distributed community runs across roughly 80-110 tasks per model configuration. `cost_usd` and `wall_seconds` are totals for all selected tasks; user-facing cost and time must prefer `average_cost_usd`, `average_task_seconds`, and `average_task_time_human`.
- The distributed homepage chart publishes fallback values in IQ-view circle `aria-label` attributes rather than the old SVG `<title>` format. Keep both parsers, restrict the new parser to `data-model-iq-chart-view="iq"`, and ignore duplicate value/cost/time chart circles.
- `model_iq.data_source.url` points to the public distributed radar. Surface it as an optional menu link, while keeping the default status title compact.
- As of 2026-07-22, the homepage renders its 19-point Intelligence Efficiency matrix from `data/intelligence-efficiency.json`; raw homepage HTML no longer contains the generated model cards. Merge this lightweight same-origin payload into `current.json`, preserve richer matching fields such as cache hit rate, and use it as the Model IQ fallback when the static homepage chart is absent.
- As of 2026-07-26, the homepage loads scenario recommendations and degradation alerts from `https://api.codexradar.com/api/v1/radar-insights`. Treat every declared `schema` as a strict `1` compatibility gate, accept empty recommendation/alert lists, ignore unknown optional fields, and keep the last valid result on endpoint, decoding, or timestamp-regression failure. Fetch independently at most once per 10-minute monotonic cache window, never attach local Codex credentials, and keep these insights out of the default status title and notification stream.