Skip to content

CHANGELOG.md

Latest commit

 

History

History
148 lines (125 loc) · 8.63 KB

CHANGELOG.md

File metadata and controls

148 lines (125 loc) · 8.63 KB
Raw

Changelog

1.5.2

  • The configured Provider Switch now owns the shared agent URL. A restored Provider Tile with the localhost default cannot redirect the whole monitor away from the real agent. Standalone tiles retain their own URL support.
  • Settings replies prefer the SDK settings envelope over an empty param, allowing cold-start hydration without opening or focusing the inspector.
  • Added fresh-process startup tests against the real Node app, bundled SDK, WebSocket host and HTTP agent; no inspector/input event is used.

1.5.1

  • Working provider editor. Provider rows now use the Ulanzi PI's required input + label checkbox structure, so selected/unselected state is visible and clickable instead of appearing as a read-only text list.
  • Selected providers have explicit ↑/↓ controls. Their saved providerIds order is now the actual deck cycle order instead of being discarded in favour of the agent response order.
  • Providers omitted from an older allow-list remain visible under Available providers. This makes OpenAI recoverable by ticking it; All / reset order restores all returned providers in agent order.
  • Added pure editor-model and ordered-cycle regression tests plus a real Chromium Property Inspector interaction check.

1.5.0

  • Ollama Cloud usage limits. The agent now reads the authenticated GET /api/usage session and weekly fractions and sends them as live gauges. The API supplies no reset timestamps, so the deck deliberately shows no made-up countdown. Profile/renewal metadata remains a non-blocking POST /api/me read.
  • OpenCode Go provider. Reads OPENCODE_GO_API_KEY or OpenCode's local auth JSON and renders the official rolling 5-hour, weekly and monthly usage windows. The new optional Monthly Provider Tile slot shows the third window.
  • Per-switch provider selection. Provider Switch settings now list the providers returned by the agent as checkboxes; each switch cycles only its selected allow-list. Existing saved switches retain their all-provider behavior.
  • Treat Hermes's structured logged_in:false Nous response as the supported configured-fallback state, not as a malformed no json fetch failure. Retry login discovery every 15 minutes and log the fallback notice once per agent process, while keeping real helper failures visible with the exit status.

1.4.0

  • OpenAI/Codex usage provider. The agent now reads the live ChatGPT account usage payload used by the official Codex CLI and exposes its real rate-limit windows as an OpenAI provider. It reads Codex auth but never refreshes, rewrites, or logs it.
  • Window labels are duration-aware (5H, DAY, WEEK, etc.). If an account has only the current weekly window, the first tile shows its plan and the second shows the weekly ring gauge; dual-window accounts populate both gauges.
  • A bounded tail-read of recent Codex rollout JSONL supplies a stale fallback on cold start when the live usage request is temporarily unavailable. The existing per-provider last-good cache handles later transient failures.
  • Added Python probe tests and a synthetic D200H-safe OpenAI render test.
  • The live preview now treats an upstream provider error as a valid, asserted error-card state while still requiring the new OpenAI provider to be live.
  • Fixed pack.sh returning a false failure after successful zip creation under pipefail (tail | head SIGPIPE); archive listing now reads cleanly to EOF.

1.3.0

Studio-restart resilience (plugin-only; agent unchanged).

Root cause (observed live 2026-07-02): after an Ulanzi Studio restart the Property Inspector still shows the saved Agent address, but Studio does not re-deliver the stored key settings to the plugin's Node backend — so the sampler silently fell back to the 127.0.0.1:9890 default and every key showed "agent off" until the user re-saved any setting by hand.

  • Persist the agent address. Every applied address change is written to plugin/monitor/.agent-url.json (fail-open); on startup the sampler begins from the persisted address instead of the localhost default. A Studio restart that drops the settings no longer takes the deck down.
  • Actively pull missing settings. When a key is (re-)added without its stored params, the plugin now calls getSettings and consumes the didReceiveSettings reply — at most once per key, so never-configured keys can't loop.

1.2.1

Nous freshness + observability (agent-only; plugin unchanged from 1.1.0).

  • Faster live refresh: INF_NOUS_LIVE_TTL default 300s → 60s, so the tile tracks the account API within ~1 min instead of up to 5 min. (Residual gap to the portal dashboard is Nous-side eventual consistency — /api/oauth/account trails the portal UI slightly and can't be eliminated programmatically.)
  • Freeze is now visible, not silent: if the live fetch fails, the agent still serves the last-good value but now (a) logs nous live … outcomes to the journal (journalctl -u inf-agent), and (b) flags stale + live_age_s once the last successful fetch ages past NOUS_STALE_AFTER (3×TTL) → the switch tile shows the amber dot. Previously a failing helper froze the value with no indication.

1.2.0

  • Live Nous balance & plan (agent). The agent now reads the real portal subscription + purchased credits (e.g. Plus / $18.93 / $22-mo) by delegating to hermes's own get_nous_portal_account_info() run in the hermes venv — the sanctioned path that refreshes + persists the single-use token with hermes's locking (the agent still NEVER calls the refresh endpoint itself). Cached INF_NOUS_LIVE_TTL (default 300s). Precedence: live > INF_NOUS_* env fallback > rate-limits. New env: INF_NOUS_HELPER_PY, INF_NOUS_HELPER_CWD, INF_NOUS_LIVE_TTL. (Plugin unchanged from 1.1.0 — this is an agent-only release; the bundled plugin zip is bumped for version parity.)

1.1.0

Bug-fix release from on-device feedback.

  • Ring gauge double segment (fix): the progress arc used stroke-dasharray with pathLength="100", but the D200H's SVG renderer ignores pathLength — so the normalised dash was read in real user units and tiled a spurious SECOND arc. Now drawn as an explicit <path … A …> arc (one definite start/end), renderer-agnostic.
  • Transient "usage HTTP 4xx" on tiles (fix): the agent now keeps a per-provider last-good cache. When a probe briefly fails — e.g. Claude's OAuth token 401s while Claude Code rotates it — the tile shows the last good value marked stale (amber dot on the switch key) instead of an error card. Only a never-succeeded provider shows an error.
  • Nous showed "Free" (fix): the hermes nous-portal.json token is the hermes-agent free-tier identity (and was logged out/expired), so the old probe mislabeled the real account as Free. The probe no longer infers "Free" from the stale JWT. Plan & balance now come from INF_NOUS_PLAN / INF_NOUS_BALANCE overrides (authoritative) or the live GET /api/oauth/account only when a valid token exists. The agent NEVER calls the Nous refresh endpoint (single-use; only hermes may rotate it — reuse revokes the whole session). The balance tile now shows the plan as its caption.

1.0.0

Initial release — sibling of ulanzi-system-monitor for inference providers.

  • inf-agent (stdlib Python): one GET /providers snapshot for Claude, OpenRouter, Nous and Ollama Cloud, refreshed every 60s.
    • Claude: session (5h) + week (7d) utilisation % and resets, via the Anthropic OAuth usage/profile endpoints (reads the Claude Code token, never refreshes it). Plan headline (e.g. "Max 20x").
    • OpenRouter: balance (credits) + spend today/week/month (key).
    • Nous: tier, spend and rate limits decoded locally from the portal JWT (free tier).
    • Ollama Cloud: plan + billing-period renewal (no usage API exists → no session/week %).
    • Token guard (INF_AGENT_TOKEN), bind address, poll interval, provider list and credential paths all configurable via env.
  • Plugin (reuses the System Monitor interface model):
    • Provider Switch cycler key — each press selects the next provider.
    • Provider Tile — Primary/Secondary slot renders the active provider: session/week ring gauges for limit providers, balance/spend value cards for balance providers, with graceful fallbacks (plan/renewal, rate limits).
    • Single ProviderSampler with the no-wedge guaranteed-settle fetch from RemoteSampler; per-key independent rendering keyed by stable actionid.
    • Dark/Light themes; configurable refresh (5/10/30s).
  • Added robot + swap-horizontal to the MDI icon subset for provider glyphs.