feat(cli+tui): provider quota and rate-limit in status bar

[rafaumeu]

Summary

Closes #53306

Adds provider quota/cost tracking to both CLI and TUI status bars with a dual-source approach:

Changes

agent/account_usage.py (+87 lines)

• Add _fetch_zai_account_usage() — fetches Z.AI (Zhipu) token quota via the monitoring endpoint /api/monitor/usage/quota/limit
• Base URL heuristic in fetch_account_usage(): when the provider name is not directly recognised (e.g. xai-oauth hitting api.z.ai), resolves to zai by inspecting the host. Falls back to custom:zai for API key resolution.

cli.py (+46 lines)

• Quota/cost segment in _get_status_bar_snapshot() using dual-source pattern:
  1. fetch_account_usage() as primary (queries provider quota APIs)
  2. rate_limit_tracker.format_rate_limit_compact() as fallback (parses rate-limit headers from last response)
• Timezone-safe reset timer: uses datetime.now(timezone.utc) and forces timezone.utc on naive reset_at values, so the countdown is correct regardless of system timezone (tested across UTC-3, UTC, UTC+9).

tui_gateway/server.py (+41 lines)

• Same quota block in _get_usage() for the TUI WebSocket payload (quota_pct, quota_reset, quota_rl_text fields).
• Same timezone-safe UTC comparison.

Design decisions

• All quota fetches are wrapped in try/except — failure is silent, status bar simply omits the quota segment.
• The rate_limit_tracker fallback ensures providers without a quota API still show useful info.
• Provider normalisation uses a whitelist of known provider names; unrecognized names with a matching base_url host get resolved automatically.

Tested

• Z.AI quota endpoint returns TOKENS_LIMIT and TIME_LIMIT with percentage and nextResetTime (epoch ms).
• Timezone simulation confirms identical countdown across São Paulo (UTC-3), UTC, and Tokyo (UTC+9).
• Syntax check passes (ast.parse) on all 3 files.
• Lint: no new pyright errors (pre-existing type issues only).

[alt-glitch]

This was generated by AI during triage.

Implements feature request #53306 (all-provider quota/rate-limit status-bar segment). Related to the Codex-only status-bar cluster (#18958, #34978); broader all-provider scope, not a duplicate. Maintainers may want to consolidate.

[tonydwb]

Code Review Summary

Verdict: Approved

Good addition of provider quota and rate-limit display in the status bar. The new _fetch_zai_account_usage function follows the established pattern for other providers (OpenRouter, etc.) with proper error handling and timeout.

Looks Good

• Follows established pattern for provider usage fetching
• Proper timeout and error handling
• No security concerns
• Single-concern feature addition

---

Reviewed by Hermes Agent

[rafaumeu]

Multi-provider rate-limit support

Research findings

Tested all 8 providers in our config with real API calls. Results:

Provider	Dedicated quota endpoint	Rate-limit headers
Z.AI	/api/paas/v4/charges/token	No
OpenRouter	/api/v1/credits	No
Anthropic	OAuth usage endpoint	No
Codex	Internal	No
Groq	None	x-ratelimit-{limit,remaining,reset}-{requests,tokens}
Mistral	None	x-ratelimit-{limit,remaining}-{req,tokens}-minute
Cerebras	None	x-ratelimit-{limit,remaining}-{requests,tokens}-{minute,hour,day}
SambaNova	None	x-ratelimit-{limit,remaining,reset}-requests-day
NVIDIA	None	None
Google	None	None
Cloudflare	None	None

What was done (3rd commit in this PR)

Since Groq, Mistral, Cerebras and SambaNova don't have dedicated quota/usage endpoints but do send rate-limit headers on chat completions, the right approach is to make the existing rate_limit_tracker parser multi-provider aware:

1. Extended parse_rate_limit_headers() — added _best_bucket() helper that tries canonical format first, then provider-specific header naming variants (req-minute, requests-minute, tokens-minute, requests-day, etc.)
2. SambaNova daily buckets promoted to hour-level when no hourly data exists
3. Fixed format_rate_limit_compact() call without argument — both cli.py and tui_gateway/server.py fallback paths now properly fetch agent.get_rate_limit_state() (or _rate_limit_state attribute) before formatting

Tested with real headers

All 5 provider formats parse correctly:

• Groq: RPM: 14399/14400 | TPM: 6.0K/6.0K
• Mistral: RPM: 187/188 | TPM: 625.0K/625.0K
• Cerebras: RPM: 4/5 | RPH: 149/150 | TPM: 30.0K/30.0K | TPH: 1000.0K/1.0M
• SambaNova: RPH: 19/20

NVIDIA, Google, and Cloudflare don't expose rate-limit data at all — they will show nothing in the status bar (no API to call).

[teknium1]

Thanks for extending the existing account-usage and rate-limit work. The gap is real: current main only surfaces these details through /usage (cli.py:9647-9710), not the status bars.

Problems

• The TUI portion is payload-only: tui_gateway/server.py:3171 adds fields, but the PR changes no ui-tui/ files. Current Usage has no quota fields (ui-tui/src/types.ts:170-184) and StatusRule does not render them (ui-tui/src/components/appChrome.tsx:426-440).
• The CLI renderer synchronously calls the fetcher through cli.py:4702; the new fetchers use 10-second httpx.Client calls (for example agent/account_usage.py:687-692). This conflicts with #53306's requirement that render paths avoid network I/O.
• The cache is process-global rather than provider/session keyed (cli.py:52, tui_gateway/server.py:3063), so one session can receive another session's quota. The TUI also retries empty results every usage update because {} is falsy at tui_gateway/server.py:3063.
• The new helpers infer provider from agent.model (cli.py:57, tui_gateway/server.py:3068) instead of the canonical agent.provider used by current /usage (cli.py:9691).

Suggested changes

• Refresh keyed snapshots asynchronously and keep rendering read-only.
• Use agent.provider, add the Ink contract/rendering, and cover parser, cache, and UI behavior with tests.

Automated hermes-sweeper review.

[teknium1]

This cache is process-global and has no provider, credential, or session key. A second agent rendered within 60 seconds will receive the first agent's quota snapshot. Cache per active provider/account identity instead.

[teknium1]

Use agent.provider as the primary input here. The established /usage path does that at current cli.py:9691; deriving it from a model string drops provider identity whenever the model lacks a provider/model prefix.

[teknium1]

An unavailable provider produces {} at line 3107, but {} is falsy, so this guard refetches on every _get_usage invocation instead of observing the TTL. Test cache initialization separately from cache contents.

[teknium1]

This only adds fields to the backend payload. The PR changes no ui-tui/ files; current ui-tui/src/types.ts has no quota fields and ui-tui/src/components/appChrome.tsx does not render them, so the advertised TUI status-bar segment remains invisible.