Add Codex multi-limit support

[lkliu7]

Pro subscriptions have separate limits for GPT-5.3-Codex-Spark; updated to ping and monitor this limit.

Changes

• Add Codex models config support so multiple Codex models can be pinged on each reset
• Read model-specific Codex usage from additional_rate_limits when available
• Show configured Codex models in status and bg status
• Update English and Chinese README/config examples for multi-model Codex setup

[wavever]

Thanks for this, @lkliu7 — really solid work. It builds cleanly, go vet is quiet, all tests pass (including the two new ones), both READMEs are updated, and the Trigger refactor into triggerOne/buildCodexArgs is nice. The model → models config is a good, backward-compatible addition.

I'd like one change before merging, plus a couple of optional nits.

Please fix: empty model no longer falls back to Codex's default

On main, an empty model means "let Codex pick its own default model" — Trigger only appends -m when c.cfg.Model != "". That behavior is still documented (README and the default config comment: "Empty = use the Codex default model").

With this PR, CodexModels() hardcodes a fallback:

if cfg.Model != "" { return []string{cfg.Model} }
return []string{"gpt-5.4-mini"}

so model = "" now always injects -m gpt-5.4-mini. That silently breaks the documented escape hatch and removes the only way for users whose plan doesn't include gpt-5.4-mini to fall back to Codex's default model. (The PR also drops the "Empty = ..." note from the embedded defaultTOML comment while the README still implies it.)

Could you either:

• preserve a way to express "no -m" (e.g. let an empty model entry skip the flag in buildCodexArgs), or
• if forcing gpt-5.4-mini is intentional, update all the docs (README EN/zh + config comment) to drop the "Empty = Codex default" promise consistently?

Optional nits (not blocking)

1. Model matching in pickAdditionalWindows — the bidirectional strings.Contains(n, m) || strings.Contains(m, n) plus the hardcoded spark special-case can mis-match (e.g. gpt-5.4 matching a gpt-5.4-codex limit). It works for today's data since additional_rate_limits is small, but matching on the exact limit_name/id would be more robust and let you drop the spark special-case.
2. status.go re-loads config inside the loop (config.Load() per Codex provider, error ignored), and the per-model expansion logic is duplicated with bg.go. The Codex provider already knows its models via effectiveModels(); exposing that would avoid the re-read and the duplication.

For context (no action needed): the scheduler still tracks only the primary 5h window, so every configured model is pinged when that window resets — a model with an independent reset time isn't scheduled separately. That matches the PR's stated goal, just noting it so the limitation is documented.

Thanks again — happy to merge once the empty-model case is sorted out.

[lkliu7]

I rebuilt the feature by treating GPT-5.3-Codex-Spark as if it were its own provider, while reusing Codex auth, CLI transport, hooks, and usage endpoint plumbing. Still disabled by default, but possible quality-of-life features for later: on the first ping, if a pro subscription is detected, ask the user if they want spark enabled; auto-archive ping conversations.

[wavever]

Thanks for the thorough rework, @lkliu7 — this is a much cleaner design. All my earlier points are addressed:

• Empty-model escape hatch restored — NewCodex keeps the original behavior (no -m when model is empty), so "Empty = use the Codex default model" still holds. Spark defaults its own model to gpt-5.3-codex-spark.
• Robust limit matching — normalizeCodexLimitName does an exact normalized match against additional_rate_limits[].limit_name, dropping the fragile bidirectional Contains/spark special-case.
• No status/bg duplication — Spark is a first-class provider, so it flows through the existing status/bg status rendering and gets independently scheduled on its own 5h window.
• Safe by default — disabled out of the box (with a test asserting it), so existing users don't get an extra quota-consuming ping on upgrade.

Build, go vet, and the full test suite all pass locally, and the new tests cover the spark window selection, the missing-limit error path, dry-run model, and provider selection. Merging now.