English | 中文

Positioning and First Scenarios (P0 Direction)

Direction-setting record, 2026-10-07: establishes company packaging tool + AI plan quota as Beacon's two first-class scenarios, divides the Provider system into three categories — Status / Usage / Machine — with the Generic HTTP Provider as the general-purpose integration capability. This document fixes the direction and priorities, without touching the already-scheduled M5 acceptance (merge into mvp-issues after B-803/B-804 close out).

1. Positioning

Beacon = the individual developer's local "status & quota" radar — a Status & Action Center, not a DevOps Dashboard.

It does not replace GitHub, Jenkins, the company packaging system, Claude, or Codex; it only takes "what I need to know right now / what I need to do right now" from these systems and brings it uniformly onto the Windows desktop:

• Signal: the unified abstraction of status, quota, and lifecycle (already carried by the existing WidgetState / Severity / LifecycleState)
• Action: the next step after a failure (view logs / rerun / cancel) delivered straight to the desktop

2. First Batch of Three Real Scenarios

Beacon
│
├── 🏗 Build / CI          公司打包工具（首位）、GitHub Actions、Jenkins、GitLab CI
├── 🤖 AI Usage            Claude、OpenAI/Codex、Gemini、其他 AI 服务
└── 🖥 Developer Environment 服务器、Docker、Agent、开发机（第二阶段）

• The company packaging tool ranks ahead of GitHub: the daily flow of real commits, queueing, packaging, and uploading. Beacon merely abstracts it into a BuildProvider (Build #18231 · 🟡 Building · Game-A · develop · a82f91c · 18m), with no need to understand the whole internal system.
• Failure is an Action scenario: 🔴 Build Failed → [查看日志] [重新打包].

3. Provider Taxonomy (Three Categories + General-Purpose Integration)

Provider
├── Status Provider    GitHub Actions、公司打包、Jenkins、GitLab、Docker …
├── Usage Provider     Claude、OpenAI/Codex、Gemini …
└── Machine Provider   服务器、PC、Agent

Principles:

1. No hard-coded interfaces for internal company systems — the company packaging tool implements the Status Provider to adapt to its own API, converts into the unified model and enters Core; if the system is replaced, Core does not move.
2. Generic HTTP Provider (the killer capability): instead of writing a C# plugin for every internal system, integrate any HTTP API via configuration. Landed form: component types http.status (status light: status word list → level mapping) and http.quota (quota number card: total/used/remaining/reset field mapping, see §8) — JSON config in widgets.json (connection endpoint + credentials + point paths):

{
  "type": "http.status",
  "connectionId": "ci-local",
  "config": {
    "label": "company-build",
    "status_path": "$.status",       // $.a.b[0] point-path extraction
    "summary_path": "$.message",
    "url_path": "$.html_url",
    "success_values": "success,ok,passed,healthy",
    "warning_values": "running,building,pending,queued",
    "error_values": "failed,failure,error,critical"
  }
}

Landed (batch 3, supplement 42, 2026-10-12): the YAML provider DSL is implemented — a connection configured in Settings with a dsl_path pointing at a declaration file enables it, no config means zero change (the existing widget.Config path is byte-for-byte equivalent). Semantics: the declaration layer only supplies default values for data extraction (status/summary/url paths, status word lists, quota paths and units); instance values in widget.Config still win (the declaration layer does not seize instance override authority); label/icon/thresholds belong to the component instance and do not go into the DSL; actions go through the existing Action executor chain (open.url/http/webhook), and retry (max_attempts 1–10, backoff_seconds 0–300) lands in the Runner retry parameters (default 1 attempt = no retry; retries share the same total timeout budget; user cancel/timeout does not retry). Zero third-party dependencies: a built-in minimal YAML subset parser (indentation mappings / - lists / quotes / inline comments); syntax outside the supported subset (flow, block scalars, Tab, multiple documents) fails loud with a line number, a bad file lands on a Degraded card face, and action loading failures are logged as warnings and skipped. Example:

# 连接 Settings: dsl_path: C:\Beacon\packer.yaml
provider: company-packager
mapping:
  status_path: $.job.state
  success_values: green,passed
actions:
  - id: restart
    type: http
    name: 重启打包任务
    url: https://packer.local/api/jobs/42/restart
    method: POST
    confirm: true          # http/webhook 默认 true；open.url 默认 false
    retry:
      max_attempts: 3      # 1–10，默认 1（不重试）
      backoff_seconds: 5   # 0–300，默认 5
  - id: dashboard
    type: open.url
    url: https://packer.local/dashboard

Whether the internals are homegrown, legacy, or "garbage" systems — anything with an HTTP API can be integrated into Beacon. Credentials are always credentialRef + DPAPI (plaintext tokens strictly forbidden; the same security constraint as connection secrets).

1. The GitHub Provider is the reference implementation: the GitHub Actions REST API (workflow runs status/rerun/cancel/logs) lands directly as GitHubProvider; the company packaging system lands as HttpProvider — together they validate whether the abstraction is truly general.

4. AI Usage as a First-Class Concept: UsageMetric

Status: conceptual design, not landed as a standalone type — the current usage components are carried directly by WidgetState (payload + threshold severity), with no UsageMetric structure introduced; the model below is a direction record for a later abstraction.

The essential difference between CI and AI quota:

CI：   失败 → 处理
AI：   额度 → 持续消耗 → 接近上限 → 预测耗尽 → 提醒 → 切换模型 / 等待恢复

That is why AI Usage stands alone as a Usage Provider, and no unified percentage field is defined — different services have completely different limit semantics (the Claude API has organization-level limits like RPM and input/output tokens per minute; subscription-based consumption windows are yet another scheme):

UsageMetric
├── name / current / limit / unit
├── window        （1h / 5h / weekly / monthly）
├── resetAt?
├── percentage?
└── status

The same model can express:

Source	Expression
Claude	Input tokens 82M / 100M, window = 1h
Codex	Usage 72%, window = weekly
OpenAI API	Cost $18.42 / $50, window = monthly

Advanced direction (predictive Signal, where Beacon gets genuinely interesting):

Claude   82% used   预计 47 min 后达到限制
Codex    71% used   weekly 8h 后重置

Active alerts: ⚠ Claude 5h quota 82% used / ⚠ Codex weekly 91% / ✓ quota recovered.

5. P0 (First-Batch List After Direction Confirmation)

1. Company Build Provider (Status Provider adapting the company packaging API)
2. GitHub Actions Provider
3. Generic HTTP Provider (YAML mapping + actions)
4. Usage Provider abstraction (UsageMetric)
5. Claude Usage
6. OpenAI/Codex Usage
7. Notification (reuse the existing notification chain)
8. L0 / L1 / L2 (reuse the existing chains)

When done: Beacon becomes a tool you genuinely use yourself every day.

6. Mapping to the Existing Architecture

Direction	Current state (RFC-001 / code)
Status Provider	IConnectionProvider + IWidgetProvider already in place (GitHub connection = reference implementation)
Unified Signal	WidgetState (id / title / severity / lifecycle / updatedAt / …) + StatusAggregator
Secrets	credentialRef + ISecretStore (DPAPI, plaintext tokens strictly forbidden)
Refresh	RefreshScheduler layered scheduling (the window granularity of status and quota is reusable)
Notification	NotificationEngine + Toast (B-5xx already wired)
Actions	IActionExecutor + ActionRunner (rerun/cancel-type actions already have a landing point)
Usage Provider, Generic HTTP	Landed: http.status / http.quota / bigmodel.usage / claude.usage / kimi.coding / deepseek.balance / ark.usage / mimo.usage / codex.usage / copilot.usage / opencode.usage / qwen.usage / moonshot.balance / sensenova.tokenplan (§8, detailed progress)
Jenkins / company packaging slot	Landed (2026-10-11): jenkins.build — the open-source standard JSON API (batch 3, supplement 31; the suspended item of batch 3, supplement 27 closed)
Machine Provider	Partially landed (2026-10-11 batch 3, supplements 32/33/36): machine.net local network (IP/↑↓ rate/cumulative traffic) + machine.git local Git repos (↑↓·changed·stash, one card per repo) + machine.local local resources (CPU/memory/disk + thresholds), all zero-credential zero-network; remote machine/server management (one of the three categories in §5) is not done, per the frozen record

7. Scheduling Agreement

• Close out M5 first: B-803 real-machine walkthrough, B-804 clean Win11 VM — acceptance must not be interrupted;
• After close-out, merge the §5 P0 into mvp-issues.md as an additional group, and re-rank subsequent phase priorities by P0;
• The existing 43 issues keep their content unchanged — only priority adjustment and additions.

8. Landed (P0 Progress)

• Generic HTTP Provider (P0 #3): component type http.status — point-path extraction $.a.b[0].c, configurable status-word → level mapping;
• Generic quota number card (acceptance feedback): component type http.quota — map arbitrary quota JSON fields onto the board: total_path required, used_path/remaining_path pick one of the two (the latter auto-derives used), reset_path auto-detects epoch seconds/milliseconds/ISO, unit/warn_percent (default 60)/error_percent (default 90) configurable; WidgetState.Progress renders as the progress bar at the bottom of the tile (quota fraction filled in the same color), payload stores raw numbers (% belongs to the display layer); the bigmodel module on the settings page attaches 「一键添加 GLM 额度组件（示例）」, and after the connection test passes it prompts onboarding via the component wizard — binding flow: connection test → choose component type → choose connection → configure fields → pin the display (with source annotation).
• Zhipu GLM plan quota (P0 #5, ahead of schedule): component type bigmodel.usage — GET https://open.bigmodel.cn/api/monitor/usage/quota/limit, raw Key auth (Authorization header without Bearer), data.limits[] normalized into the 5h/weekly/monthly three windows, threshold alerts; summary and payload carry the nearest reset time (the monitor API's native nextResetTime epoch ms, 5h window preferred, local timezone MM-dd HH:mm); for the Z.AI international site, just switch the connection Endpoint to https://api.z.ai/api/monitor/usage/quota/limit and it works;
• Claude Code local usage (P0 #5): component type claude.usage — reads local ~/.claude/projects/**\/*.jsonl (the same-source data as ccusage, ccusage.com semantics), aggregates tokens and costUSD by window; zero credentials zero network, daily_cost_limit optional threshold alert (no limit stays Success forever);
• Kimi For Coding plan remaining (P0 #5): component type kimi.coding — default GET https://api.kimi.com/coding/v1/usages Bearer auth (sk-kimi-* Key, a separate account system from the Moonshot open platform); the console's real endpoint (captured live by the user, 2026-10-11) is POST www.kimi.com/apiv2/kimi.gateway.billing.v1.BillingService/GetUsages (Cookie auth; add http_method=POST/auth_header=Cookie/http_body in Settings to switch). Response usages[] (scope=FEATURE_CODING) entries detail=period window, limits[]=5h window, totalQuota into payload, values are strings (parsing is tolerant); the old shape (top-level usage+limits[]+membership.level) is a compatibility fallback. Windows normalized to 5h/weekly, threshold alerts by remaining percentage (warn 30 / error 10); card face/progress bar follow the used semantics (user directive, 2026-10-13: used=100−remaining back-computed, consistent with bigmodel/ark's Progress semantics). Credential model: the Cookie is copy-pasted by the user from the browser's F12 into the settings credential box (DPAPI storage, replayed on request); Beacon does not build QR-scan/login flows; after session expiry it is re-entered once (same approach as the MiMo console Cookie);
• DeepSeek open platform balance (P0 #5): component type deepseek.balance — GET https://api.deepseek.com/user/balance, Bearer auth, balance_infos[] string amounts, is_available=false goes straight to Error, alert when balance breaches the lower bound (warn 20 / error 5, by balance currency) (verified against real data, 2026-10-09: CNY balance, is_available=true, shape matched field by field);
• Volcano Ark Coding Plan Pro quota (P0 #6): component type ark.usage — the Volcano control-plane OpenAPI GetCodingPlanUsage (open.volcengineapi.com, the Volcano variant of V4 signing; credentials are the console AK/SK pasted once as AccessKey:SecretKey and only ever into DPAPI — the inference ARK_API_KEY was tested and rejected 400 by the control plane, so the two cannot stand in for each other), Result.QuotaUsage[] normalized into 5h/weekly/monthly three windows of used percentage (the API only gives percentages; absolute numbers do not come through this endpoint); empty array = unsubscribed/reclaimed, rendered as a 「无套餐」 Info card, not an error;
• Xiaomi MiMo usage card (one of the nine AI Usage providers): component type mimo.usage — three-tier degradation: ① console plan usage (endpoint evidenced in the front-end bundle 2026-10-10, response captured live by the user 2026-10-11: GET platform.xiaomimimo.com/api/v1/tokenPlan/usage, Cookie api-platform_ph auth + plan-manage referer; response data.monthUsage/usage blocks + items[], the percent field is a 0-1 ratio, not 0-100; used/limit computed by us into exact percentages; the earlier 「/usage 等 7 端点全 404」 came from live testing of the inference-domain API — the console domain is the exception) ② local counting (Settings usage_endpoint, defensive parsing of total_calls/window_calls/window_minutes → 「本机累计 N 次 · 近M分 K 次」) ③ neither present → quota slot shows 「无额度口」 + plan-manage guidance, model count demoted to a Summary sub-line (user directive: model count is a catalog, not a quota, and must not occupy the quota slot). Cookie credential model same as Kimi: browser copy-paste + DPAPI replay, no QR-scan login, re-enter on expiry;
• OpenAI Codex local stats (one of the nine AI Usage providers): component type codex.usage — reads token_count events from local ~/.codex/sessions/**/rollout-*.jsonl (total_token_usage is a within-session cumulative; take the last record per file and sum), zero credentials zero network, card face honestly labeled 「本地统计」 (the ChatGPT OAuth official usage endpoint is not isomorphic with the "connection + credentials" model, not integrating it for now);
• GitHub Copilot plan quota (eighth of AI Usage): component type copilot.usage — the same quota endpoint as the official editor extension GET api.github.com/copilot_internal/user (PAT needs the copilot scope; Editor-Version / Copilot-Integration-Id both headers proven required by live testing, and the api.githubcopilot.com domain likewise only works with both), quota_snapshots three slots — premium requests (1,500/month on Pro per live testing, included) threshold-checked by used percentage (default 80/95, configurable), chat/completion unlimited show ∞; Pro monthly labeled with the official price $10/month, quota_reset_date is the quota reset day (the subscription renewal day is not returned by this endpoint, and is not fabricated); 401/403/404→Degraded, 5xx→Offline;
• OpenCode Go usage (ninth of AI Usage): component type opencode.usage — direct connection to the official Console Budgets API (the documented official interface, v2 docs/console/api/budgets): GET opencode.ai/console/api/v2/budgets/members (oc_sk Key Bearer, reading Budgets requires the All permission; v1 has migrated to v2, official announcement says v1 is retired after the transition period). Amounts are micro-cents decimal strings (1 dollar = 100 million, the official precision-safe convention), limit_micro_cents null = no limit; multi-member aggregation (spend summed / any null means no limit / any exceeded means over); the v2 member object adds additional_budget_rules (an array of shared/model-specific budget rules) — the provider takes only the main budget's four fields (limit/spent/exceeded/resets_at); unknown fields are naturally tolerated during parsing (JsonDocument+TryGetProperty, not strict); official exceeded is a single vote of Error; resets_at follows the UTC calendar month; the card-face 「Go $10/月」 is display-layer plan labeling (the user's subscription semantics), payload stores only the API's real numbers; 401/403/404→Degraded, 5xx→Offline; card face/progress bar (user directive, 2026-10-13): when there is a limit the payload carries percent (used % converted from the limit) and Progress = used fraction 0-1 (same semantics as ark/bigmodel); exceeded fills the progress bar (aligned with ark's 「已尽=满格」); no limit keeps $spent · 无上限, no fabricated percentage;
• Alibaba Qwen Bailian Token Plan (tenth of AI Usage, 2026-10-09): component type qwen.usage — Token Plan (Coding Plan dedicated Key with sk-sp- prefix, Base URL token-plan.cn-beijing.maas.aliyuncs.com, OpenAI-compatible /compatible-mode/v1). The official quota-query REST endpoint is not open (the FAQ only has the console 「我的订阅」 page; community evidence cc-switch#7484 — the dedicated Key only authorizes model calls; /models returns 200 in live testing while console query endpoints are not authorized), after four-tier probing it degrades honestly: by default show real /models model-catalog data + 「额度口径官方未开放」; fill a custom endpoint in connection Settings/form 「用量端点」 (GET + Bearer), and the response is parsed with defensive key names (percent/used_percent, remaining_percent back-computed, remaining/total_credits, reset_at/reset/period_end ISO or second-level epoch) to produce a quota card (used % progress bar + threshold alerts
• reset time). Window semantics: personal edition a fixed 7-day window, team edition monthly (official FAQ), reset follows the endpoint data;
• Moonshot open platform balance (eleventh of AI Usage, 2026-10-10): component type moonshot.balance — open-platform balance monitoring (a separate account system from the Kimi For Coding plan), GET api.moonshot.cn/v1/users/me/balance Bearer auth, response {code:0, data:{available_balance}}; balance ≤ warn_below (default 20) → Warning, ≤ error_below (default 5) → Error, a response without balance data is treated as Degraded (connection test uses the same rule; a proxy returning an error body does not count as reachable);

• SenseTime SenseNova Token Plan credit pool (twelfth of AI Usage, user directive 2026-10-13): component type sensenova.tokenplan — console pool-level quota GET platform.sensenova.cn/lite/console/v1/tokenplan/pool-usage, Bearer console JWT (a session token, not an API Key; carries exp and expires); response plan + pools[], each pool window_5h/window_7d (limit/used/remaining/reset_at) plus grant (grant_balance/nearest_grant_expiry); worst pool's worst window wins (highest used; ties broken by pool order → 5h → 7d) driving the main value/ progress bar/level; summary shows all pools together; tile value slot shows 「池名 窗名 已用%」; grant-expiry warning (within 7 days or remaining fraction ≤10%) never below Warning; JWT expiry 401 → Degraded with explicit re-acquisition guidance, not silent empty data; multiple accounts = multiple connections;

• Jenkins build status (open-source CI, the master-plan §13 company packaging tool slot): component type jenkins.build — the open-source Jenkins standard JSON API (works for any instance, not vendor-specific): connection Endpoint = instance base URL, credentials optional (anonymous read) or user:apiToken Basic auth; the component's job_path escapes each job/<name> segment (Folder nesting supported). One request /job/.../api/json?tree=… is graded by color (blue→Success, yellow→Warning unstable, red/dark_red→Error, aborted→Cancelled, gray/notbuilt →Info not built, disabled→disabled; _anime suffix = building → Running) — job-level lastBuild does not include result (proven with a real fixture from weekly.ci.jenkins.io), no second hit is chased. The connection test hits the root view /api/json; only 2xx and JSON count as reachable (a reverse proxy answering 200 with an HTML shell counts as Degraded; a proxy false-positive does not count as reachable);

• Local network (2026-10-11, batch 3 supplement 32): component type machine.net — reads only local NetworkInterface (zero credentials zero network): IP + real-time ↓↑ rate (adjacent sample counter delta ÷ time delta, machine tier 15s; the first refresh has no baseline so only IP and cumulative values appear) + cumulative traffic (since boot, into payload for the L3 detail); auto card pick = Up + non-loopback + has IPv4, gateway-having preferred; a named NIC that does not exist counts as Degraded; does not touch the frozen remote Machines management;
• Local Git repos (2026-10-11, batch 3 supplement 33, user decision 「git 都搞」): component type machine.git — one card per repo beacon ↑2↓5 · 改 7 · stash 1: git CLI triple read-only (status --porcelain -b / rev-list @{u}...HEAD / stash list, zero network zero credentials; no upstream means no second hit); behind>0 or diverged → Warning (yellow only when action is needed; ahead-only waiting to push is Info); the top 5 changed-file previews (renames take the right side of ->) go into payload for the L3 detail; if 32 DIP cannot fit them, still no pagination;
• Local resources (batch 3 supplement 32 decision receipt 「本机的cpu 内存 磁盘空间」, landed in batch 3 supplement 36): component type machine.local — CPU/memory/disk as three percentages in one line (zero credentials zero network): CPU = delta of adjacent cumulative ticks (first refresh has no baseline so no CPU; on fallback the baseline is re-recorded); Linux reads /proc (stat first line + meminfo MemAvailable, old-kernel MemFree fallback), Windows goes through GetSystemTimes + GlobalMemoryStatusEx, disk uniformly DriveInfo (empty = system disk/root partition; named-but-missing or not-ready → Degraded, fail-loud); the three thresholds default to 90% (over the line lights yellow; Signal is not graded), compared against the rounded display value (what you see is what alerts);
• Herald integration (2026-10-11, batch 3 supplements 35/37; user decision: incidents+feed first, deliveries to follow in later batches): connection type herald — Endpoint = base URL, credentials optional (anonymous works if auth is off) or an API Key via the X-API-Key header; the test hits the always-public endpoint /api/v1/status (only 2xx+JSON counts as reachable); herald.incidents event ledger (open unresolved >0 lights yellow, contains critical lights red; lookback view stays Info) + herald.feed subscription feeds (public-category RSS or private token feed — the token is the credential, goes to DPAPI, never into config; unread = new guid since the last refresh, first refresh only records the baseline) + herald.deliveries delivery failures (/api/v1/logs?status=failed, failed>0 lights yellow = mail sent but not delivered, needs human attention; the log status vocabulary pending/success/failed/shadow is a hardcoded four states; retry-exhausted dead is also written failed at the log layer with no distinction, and shadow is naturally excluded by the status filter). Herald is the user's own project (cuihairu/herald, a unified subscription and delivery hub);
• Global check frequency (Settings → General): AppConfig.PollIntervalSeconds (0 = follow each component's refresh-tier policy table, the default) — the component-level 「检测间隔」 takes priority; components without one follow the global; five tiers (by tier 15s/30s/1m/5m); changes rebuild the schedule registration through change detection (no change, no action — avoiding a casual settings-page tweak triggering a full re-pull); takes effect immediately and is logged;
• Component type short names (tile label/L2 chip unified vocabulary): WidgetTypeNames (Core, pure function, unit-testable) — previously the PinTile label fallback took the type's tail segment (ark.usage → tile name showed 「usage」, a field name impersonating a display name; user live test, 2026-10-09); now the type prefix maps to the connection-type short name (Volcano Ark/Zhipu/Xiaomi MiMo/Alibaba Qwen/…), and the QuickPanel group chip's hardcoded GLM/ARK is merged into the same table; unknown prefixes still fall back to the type tail (no throw);
• Ark 5h window primary rule (bug batch 3, revised by user directive 2026-10-13): the ark.usage tile main value/level/progress bar originally followed the Coding Plan 5-hour window semantics with the session tier as primary; now unified to worst-window-wins — level/progress bar/main value take the highest-used window (5h/weekly/monthly) with the window name labeled; any window ≥100% goes to the exhausted branch, constant Error; Summary carries 「最差窗 x% · 其余窗 · 重置」, weekly/monthly demoted to payload detail for L3; Zhipu/Qwen and other per-window rules are unaffected;
• PowerToys-style configuration center: the settings page's left-side module catalog (each module with its own icon + on/off switch); only enabled modules show their config page; enable/disable lands in connections.json's enabled; the host skips refresh, and residual panel state is cleaned up immediately.
• L0 standalone floating window form (B-701 extension): settings offer a 「悬浮形态」 pick-one — host panel (default, single window with multiple tiles) / standalone floating window (one window per pinned component, freely draggable on the desktop; position recorded per component as PinLayout.FloatingX/Y physical pixels, never-dragged ones cascade in order); both forms share the always-on-top sentinel, click drills into L2, disabled modules disappear instantly; switching takes effect after restart. Product decision (numeric-type floating: configurable; default on, user directive 2026-10-11): standalone floating windows for information-dense components (trend charts/status-light groups, descriptor FloatingSupported) always take effect; numeric/quota types (descriptor FloatingOptIn) are controlled by the settings' 「数量悬浮窗」 master switch — default on (numericFloatingEnabled defaults to true; 2026-10-08 had decided default off, overturned by user directive), turning it off immediately retracts and persists back to host-panel carrying. The current component library has no FloatingSupported types yet (status types are always panel-carried); the switch only gates the numeric standalone floating window path.

Research Conclusions (Integration Deferred)

• Alibaba Cloud Bailian (DashScope/Qwen): ~~No independent balance/usage query interface… high integration cost, low information gain, deferred~~ Overturned by the 2026-10-09 Token Plan (Coding Plan) — the pay-as-you-go conclusion no longer applies: the Token Plan dedicated Key only authorizes model calls, and the official quota-query REST endpoint is still not open; qwen.usage has landed as 「目录真数据 + 自定义用量端点」 (see §8, tenth of AI Usage); pay-as-you-go users' account balance still goes through the Alibaba Cloud BSS OpenAPI QueryAccountBalance (deferral unchanged).
• Xiaomi MiMo: the API open platform (platform.xiaomimimo.com) is open, model weights are MIT open-source, but no public balance/usage query interface has been seen (2026-10-10 recheck: /usage /quota /balance and 7 endpoints all 404, responses lack rate-limit headers) — mimo.usage has landed as 「无额度口 + 本机计数」 (see §8); once the official one opens, it upgrades isomorphically like bigmodel/kimi. Side note: the Xiaomi key has no local caller (litellm/relay); litellm's mimo-v2.6-flash-free goes through the DeepSeek serverless relay, not a direct Xiaomi connection.
• Moonshot open platform balance: ~~Endpoint confirmed; integrate later isomorphic to deepseek~~ moonshot.balance landed on 2026-10-10 (see §8, eleventh of AI Usage, isomorphic to deepseek).
• OpenAI Codex usage: the community goes through ChatGPT OAuth (chatgpt.com/backend-api/wham/usage, 5h/weekly/monthly limits + Credits), which needs local OAuth credential wrapping and is not isomorphic with the current "connection + credentials" model — the third tier, local session JSONL stats, has landed (codex.usage, see §8); the OAuth endpoint is scheduled separately.

Configuration Examples (config directory %APPDATA%/Beacon/)

Connection connections.json (credentials store only credentialRef; the Key body lives in DPAPI, entered on the settings page):

[
  {
    "id": "ci-local",
    "type": "http",
    "endpoint": "http://192.168.1.10:8080/api/status",
    "credentialRef": "conn:ci-local",
    "enabled": true,
    "settings": { "auth_header": "X-Token", "auth_prefix": "" }
  },
  {
    "id": "zhipu",
    "type": "bigmodel",
    "endpoint": null,
    "credentialRef": "conn:zhipu",
    "enabled": true
  },
  {
    "id": "claude-local",
    "type": "claude",
    "endpoint": null,
    "enabled": true
  },
  {
    "id": "kimi",
    "type": "kimi",
    "endpoint": null,
    "credentialRef": "conn:kimi",
    "enabled": true
  },
  {
    "id": "deepseek",
    "type": "deepseek",
    "endpoint": null,
    "credentialRef": "conn:deepseek",
    "enabled": true
  },
  {
    "id": "ark",
    "type": "ark",
    "endpoint": null,
    "credentialRef": "conn:ark",
    "enabled": true
  },
  {
    "id": "mimo",
    "type": "mimo",
    "endpoint": null,
    "credentialRef": "conn:mimo",
    "enabled": true,
    "settings": { "usage_endpoint": "" }
  },
  {
    "id": "codex-local",
    "type": "codex",
    "endpoint": null,
    "enabled": true
  },
  {
    "id": "copilot",
    "type": "copilot",
    "endpoint": null,
    "credentialRef": "conn:copilot",
    "enabled": true
  },
  {
    "id": "opencode",
    "type": "opencode",
    "endpoint": null,
    "credentialRef": "conn:opencode",
    "enabled": true
  }
]

http connection settings: auth_header (default Authorization), auth_prefix (default "Bearer ", can be emptied); bigmodel connection endpoint can be empty (the official monitor API is used automatically), auth defaults to raw Key (no prefix); claude connection is zero-credential zero-network (endpoint can point at a custom session directory); kimi/deepseek endpoints can be empty to use the official APIs, Bearer auth; ark credentials are the console AK/SK (AccessKey:SecretKey pasted once, into DPAPI only); mimo endpoint can be empty to use the official inference domain, usage_endpoint points at a custom usage endpoint (counting/percentage shapes produce numbers; unrecognized shapes show explicit Degraded — user directive 2026-10-13: no passing remote raw text onto the card face); codex is zero-credential zero-network (endpoint can point at a custom session directory); copilot credentials are a GitHub PAT (needs the copilot scope), endpoint can be empty to use the official quota endpoint (both headers of the official extension carried automatically); opencode credentials are an OpenCode Key (oc_sk; reading Budgets needs the All permission), endpoint can be empty to use the official Console Budgets API.

Component widgets.json:

[
  {
    "id": "http.status:pack-a",
    "type": "http.status",
    "connectionId": "ci-local",
    "refreshTier": "ci",
    "pinned": true,
    "config": {
      "label": "打包机A",
      "status_path": "$.status",
      "summary_path": "$.message",
      "url_path": "$.html_url",
      "success_values": "success,ok,done,passed,healthy,finished",
      "warning_values": "running,building,pending,queued,in_progress,deploying,starting",
      "error_values": "failed,failure,error,critical,broken"
    }
  },
  {
    "id": "http.quota:custom",
    "type": "http.quota",
    "connectionId": "ci-local",
    "refreshTier": "ci",
    "config": {
      "label": "自定义额度",
      "total_path": "$.data.total",
      "used_path": "$.data.used",
      "reset_path": "$.data.reset_at",
      "unit": "tokens",
      "warn_percent": "60",
      "error_percent": "90"
    }
  },
  {
    "id": "bigmodel.usage:GLM",
    "type": "bigmodel.usage",
    "connectionId": "zhipu",
    "refreshTier": "ci",
    "config": { "label": "GLM", "warn_percent": "60", "error_percent": "90" }
  },
  {
    "id": "claude.usage:claude",
    "type": "claude.usage",
    "connectionId": "claude-local",
    "refreshTier": "static",
    "config": { "label": "Claude", "days": "1", "daily_cost_limit": "35" }
  },
  {
    "id": "kimi.coding:kimi",
    "type": "kimi.coding",
    "connectionId": "kimi",
    "refreshTier": "ci",
    "config": { "label": "Kimi", "warn_percent": "30", "error_percent": "10" }
  },
  {
    "id": "deepseek.balance:deepseek",
    "type": "deepseek.balance",
    "connectionId": "deepseek",
    "refreshTier": "default",
    "config": { "label": "DeepSeek", "warn_below": "20", "error_below": "5" }
  },
  {
    "id": "ark.usage:ark",
    "type": "ark.usage",
    "connectionId": "ark",
    "refreshTier": "ci",
    "config": { "label": "方舟", "warn_percent": "60", "error_percent": "90" }
  },
  {
    "id": "mimo.usage:mimo",
    "type": "mimo.usage",
    "connectionId": "mimo",
    "refreshTier": "ci",
    "config": { "label": "MiMo" }
  },
  {
    "id": "codex.usage:codex",
    "type": "codex.usage",
    "connectionId": "codex-local",
    "refreshTier": "static",
    "config": { "label": "Codex", "days": "1" }
  },
  {
    "id": "copilot.usage:copilot",
    "type": "copilot.usage",
    "connectionId": "copilot",
    "refreshTier": "default",
    "config": { "label": "Copilot", "warn_percent": "80", "error_percent": "95" }
  },
  {
    "id": "opencode.usage:opencode",
    "type": "opencode.usage",
    "connectionId": "opencode",
    "refreshTier": "default",
    "config": { "label": "OpenCode Go $10/月", "warn_percent": "80", "error_percent": "95" }
  }
]

The http.status status word lists can be omitted entirely (the default values above are used); unknown status words fall back to Warning (better to over-report than to miss). The bigmodel.usage summary looks like GLM Coding Pro · 5h 窗口已用 42.5% · 周 8% · 月 3%; worst-window-wins (user directive 2026-10-13): level/progress bar/main value always take the highest-used window with the window name labeled, and the warn/error thresholds still apply to the worst window's used % (semantics unchanged). The claude.usage summary looks like Claude · 今日 · $1.00 · 2.5K tok · 12 条 (when costUSD is missing only tokens are reported; no self-pricing). The kimi.coding summary looks like Pro · 周 已用95% · 5h 已用5% (the community endpoint carries the plan tier; the console endpoint has no tier, and the head falls back to the label or 「Kimi」); card face/multi-window/progress bar follow the used semantics (user directive 2026-10-13, unified with bigmodel/ark: used = remaining back-computed), worst-window-wins: level/progress bar/main value take the lowest-remaining (= highest-used) window with the window name labeled; thresholds warn 30 / error 10 are still judged by remaining percentage (lower is worse; semantics unchanged). The deepseek.balance summary looks like DeepSeek · ¥110.00（含赠 ¥10.00）, currency CNY/USD adaptive. The ark.usage summary looks like 方舟 · 月 92% · 5h 42.5% · 周 8% · 待重置 (the API only gives percentages; worst-window-wins, user directive 2026-10-13: level/progress bar/main value take the highest-used window with the window name labeled, other windows follow, reset follows the worst window; any window ≥100% → exhausted branch, constant Error + full bar); the mimo.usage summary looks like MiMo · 本机累计 128 次 · 近60分 6 次 · 官方无额度接口（本机计数） (with a counting source) or MiMo · 官方无额度接口 · 8 模型可用 (model count as sub-line, quota slot 「无额度口」); the codex.usage summary looks like Codex · 今日 · 8 会话 · 12.3K tok · 本地统计 (days defaults to 7; days=1 shows 「今日」); the copilot.usage summary looks like Copilot · Pro $10/月 · 高级 1/1500 · 聊天 ∞ · 补全 ∞ · 11-01 重置 (level takes the premium requests' used percentage against the threshold); the opencode.usage summary looks like OpenCode Go $10/月 · $0.00 · 无上限 · 11-01 重置 (when there is a limit, 已用 / 上限 is shown in the same segment).