v0.7.0

Added

• Add a provider-usage cache and detached-refresh framework, with the first
  adapter for DeepSeek's official API balance. The active model resolves to
  its provider table, cache files are isolated by a one-way API-key
  fingerprint, the render hot path never performs network I/O, and custom
  DeepSeek-compatible proxies are refused so credentials only reach the
  fixed official balance endpoint. The footer renders currency as compact
  balance text (DeepSeek Balance ¥N.NN) instead of inventing a quota
  percentage or relying on unexplained abbreviations. DeepSeek also uses the
  local all-agent cost path, so ready facts compose as
  DeepSeek Balance ¥N.NN · Session Cost ≈¥N.NN for CNY accounts (and the
  matching USD form for USD accounts). Cost selects the official price table
  from the balance response's currency and remains visible by itself when the
  balance is unavailable but that account currency is known; an unknown
  currency fails closed instead of guessing the symbol.
• Add local all-agent session-cost estimates for supported models on the
  official direct DeepSeek, OpenAI, and Anthropic APIs. A dedicated content-free
  usage.record ledger reconstructs main and subagent history without
  disturbing live metrics, waits for every wire to catch up, and fails closed
  for unknown models or compatible proxies. The footer renders the full brand
  and scope (DeepSeek Session Cost ≈¥N.NN / OpenAI Session Cost ≈$N.NN /
  Anthropic Session Cost ≈$N.NN),
  keeping estimates distinct from balances, admin billing, and subscriptions.

Changed

• Format CNY session costs with two decimal places once they reach one fen,
  while retaining additional precision below that boundary so small nonzero
  costs do not collapse to ¥0.00.

Fixed

• Follow in-session effort and model switches from the per-request ground
  truth: hosts stamp every llm.request wire row with the thinkingEffort
  and modelAlias the request actually ran with, so a switch that emits no
  new config.update/profile.bind row — the host's own footer (e.g. the
  line-2 context figures) can lag behind here — now updates the HUD on the
  next request. The one-time backfill scan version is bumped so sessions
  tracked by earlier HUD builds re-project effort and model from their
  request journal too.