aiti-model-check

❝ Is the model home? Knock and ask. ❞

Model availability management extension: sends one minimal probe message ("ping" → expects "ok") to every model in models.json via a suggest-like lightweight direct connection, and reads the result: 200 = available / 429 = rate-limited (parse Retry-After, X-RateLimit-Reset, error text → predict recovery time) / anything else = error. Each probe is recorded into history.jsonl + state.json, and a background sentinel periodically re-tests non-available models → the actual recovery time is observed, identifying availability windows.

• Model list merges two sources: each time it opens, "explicit models.json config ∪ pi ModelRuntime snapshot" is merged — the snapshot is built at extension startup, so providers/models newly added via the wizard afterwards are supplemented from models.json even if not in the snapshot (source: models.json), without relying on a rebuild; entries the runtime has still defer to the runtime (store completion + baseUrl/api normalization). The check list is never based solely on its own historical state
• Probing: probe all models in one click (concurrency 2 · 15s timeout · ~8 tokens each · multi-key providers check per key by default probe.perKey), supporting openai-completions / anthropic-messages / openai-responses interfaces
• 429 recognition: Retry-After (seconds/HTTP date), X-RateLimit-Reset, error body text ("Weekly usage limit reached. Resets in 4 days.", "GLM 套餐已到期"…) → expected recovery time
• Records: history.jsonl full history + state.json per-model latest status / 20-entry timeline / consecutive-429 count / recoveredAt recovery time
• Background sentinel (off by default; enable via watch.enabled): every 60s re-tests only 429/error models; far-horizon rate limits auto-downsample by watch.farFactor → automatically identifies "when it recovers"
• Runtime observation + per key (out-of-band, zero probing): per-key status for multi-key providers; tail of pi-models-combo proxy logs + 429/402/401/5xx ingested from direct-connection session events (key attribution, 1h/24h windows, notify-only, no marking); POST /blocked-execution reports of executions that "hit massive 429s and were genuinely blocked" → mark the model unavailable (runtime.minBlockCount default 3); OAuth subscription types get a provider-level status check (e.g. qoder quota/login state)
• Rate-limit expiry reminders (recover, 2026-09-19): models rate-limited long-term (lasting ≥ threshold, default 72h = 3 days) get a reminder when the expected recovery time recorded with the 429 comes due — pure local time check, zero probing: browser notification + panel "⏰ To verify" list + per-model/provider 🔔 switches (model level > provider level > recover.notifyDefault, POST /recover-notify); due checks default to one round per 30min with idempotent catch-up on startup and panel load (POST /recover-check).
• Agent tools (8, see below; the whole group disabled by default): AI calls them directly, e.g. "check model availability"
• ai-helper help points (skills, invocable from the 🤖 dialog): model-check (check availability) / model-config (adjust config), see the aiti-helper integration guide (internal doc)
• 📡 Web panel: status overview (✅/🔴/⚠️/⚪ + expected recovery) / one-click probing / history timeline
• mc:avail event-driven: 429/recovery broadcast mc:avail frames via aiti-event-bus.publish (persistent rate limits throttled per key to 2min, eliminating event storms); model-wizard banner consumes the data directly with no reverse lookup, no polling in the background
• Servicification: server features consolidated into ctx.features, client cross-extension services on dual tracks (ctx.services first + legacy global fallback), see ext-services

Agent tools

aiti_model_status          # Status overview (reads latest state, fast; includes per-key status/runtime errors/due reminders list)
aiti_model_keys            # Multi-key check (provider.apiKeys listed masked, per key available/429/latency)
aiti_model_availability    # Any available model (all-unavailable verdict + recent blocked execution records)
aiti_model_probe           # Probe now (all by default, filterable by provider/model; combo groups not probed directly) + record
aiti_model_report_blocked  # Report executions blocked by rate limiting (marked only when ≥ runtime.minBlockCount and genuinely blocked)
aiti_model_history         # Probe history (identifies availability windows)
aiti_model_balance         # Balance/plan usage (public balance APIs, cached ≤10min)
aiti_model_config          # Read/write config

The 8 tools are disabled by default (features tool.aiti_model_* / agentTools.enabled default false); enable them from the extension manager or the ai-helper 🤖 help point.

Config config.json (aiti-model-check dataDir)

{
  "modelsPath": "~/.pi/agent/models.json",
  "probe": { "timeoutMs": 15000, "concurrency": 2, "maxTokens": 8, "perKey": true },
  "watch": { "enabled": false, "intervalSec": 60, "onlyNotOk": true, "farThresholdMin": 30, "farFactor": 30 },
  "history": { "maxEntriesPerModel": 500 },
  "state": { "windowSize": 20 },
  "runtime": { "minBlockCount": 3 },
  "recover": { "enabled": true, "minLimitedHours": 72, "checkIntervalSec": 1800, "notifyDefault": true }
}

Full design: internal plan document