Refresh web console controls and Codex quota display

[MapleStoryIdle]

Summary

• Refresh the web session header and composer controls with a layout inspired by Codex Desktop.
• Move status, model, reasoning, permission, and auxiliary composer actions into compact controls and popovers.
• Add Codex subscription quota retrieval through the CLI/hub RPC path and show 5h and weekly quota indicators in the header.
• Cache Codex quota queries per session and model so model changes do not reuse stale quota buckets.

Testing

• bun typecheck
• bun run test
• bun run typecheck:web
• cd web && bun run test src/components/AssistantChat/ComposerButtons.test.tsx src/components/AssistantChat/HappyComposer.modelEffort.test.tsx src/components/AssistantChat/StatusBar.test.ts src/api/client.test.ts

AI assistance

• Implemented with Codex.

[github-actions]

Findings

• [Major] Refetch Codex limits when the selected model changes — getCodexSubscriptionLimits now selects a model-specific limit bucket, but the web query is keyed only by sessionId and is infinitely stale. After switching between a model with its own bucket and a regular Codex model, the header badge can keep showing the old model's quota until a later thinking transition happens. Evidence: web/src/hooks/queries/useCodexSubscriptionLimits.ts:22, related selector context cli/src/modules/common/codexSubscriptionLimits.ts:76.
  Suggested fix:

  // SessionHeader.tsx
  const codexLimitsState = useCodexSubscriptionLimits({
      api,
      sessionId: session.id,
      model: session.model ?? null,
      enabled: session.active && session.metadata?.flavor === 'codex',
      thinking: props.status?.thinking ?? session.thinking
  })
  
  // useCodexSubscriptionLimits.ts
  queryKey: [
      ...queryKeys.sessionCodexSubscriptionLimits(sessionId ?? 'unknown'),
      args.model ?? null
  ],

Summary

• Review mode: initial
• One issue found: Codex quota display can become stale after model changes because the client cache key does not include the model. Residual risk: web UI paths are not covered by automated tests in this run.

Testing

• Not run (automation: bun is not installed in this runner)

HAPI Bot

[github-actions]

[MAJOR] This query needs to vary with the selected model. The CLI selector now chooses a model-specific bucket when model matches limitId/limitName, but this cache key only contains sessionId and staleTime is infinite. If the user switches between Spark and a regular Codex model, the header badge can keep rendering the previous model's quota until a later thinking transition refetches it.

Suggested fix:

// include model in the hook args from SessionHeader
model: session.model ?? null,

// and key by it here
queryKey: [
    ...queryKeys.sessionCodexSubscriptionLimits(sessionId ?? 'unknown'),
    args.model ?? null
],

[github-actions]

Findings

• [Minor] Avoid fetching hidden Telegram-only header data — SessionHeader returns null in Telegram, but the new Codex limits hook runs before that return and is enabled for every active Codex session. Opening a Codex session in the Mini App now triggers /codex-subscription-limits and an RPC that can spawn codex app-server, even though the badge is never rendered. Evidence: web/src/components/SessionHeader.tsx:211, web/src/components/SessionHeader.tsx:246.
  Suggested fix:

  const telegramHeaderHidden = isTelegramApp()
  const codexLimitsState = useCodexSubscriptionLimits({
      api,
      sessionId: session.id,
      model: session.model ?? null,
      enabled: !telegramHeaderHidden && session.active && session.metadata?.flavor === 'codex',
      thinking: props.status?.thinking ?? session.thinking
  })
  
  if (telegramHeaderHidden) {
      return null
  }

Summary

• Review mode: follow-up after new commits
• Prior stale-model-cache issue no longer reproduces because the query key now includes model. One minor issue remains: Telegram opens can still trigger the new quota RPC for a header that is not displayed. Residual risk: this review did not run the full browser UI manually.

Testing

• Not run (automation)

HAPI Bot

[github-actions]

[MINOR] This hook still runs before the Telegram early-return below, so active Codex sessions opened in the Mini App will fetch /codex-subscription-limits and route an RPC that can spawn codex app-server, even though SessionHeader returns null and never renders the quota badge in Telegram.

Suggested fix:

const telegramHeaderHidden = isTelegramApp()
const codexLimitsState = useCodexSubscriptionLimits({
    api,
    sessionId: session.id,
    model: session.model ?? null,
    enabled: !telegramHeaderHidden && session.active && session.metadata?.flavor === 'codex',
    thinking: props.status?.thinking ?? session.thinking
})

if (telegramHeaderHidden) {
    return null
}