What this does

Closes #6211 — when a conversation reaches the token / context-window limit, the dashboard now guides the user to start a new session instead of leaving a bare error bubble.

When the most recent agent-chat turn fails with a token / context-window or length / quota limit, the chat view renders an inline guidance banner with a one-click "Start a new session" action.
The action reuses the existing useCreateAgentSession mutation + handleNewSession flow (create session → navigate to it), so there is no new endpoint and no duplicated session-creation logic.

Investigation findings

• The agent chat lives in crates/librefang-api/dashboard/src/pages/ChatPage.tsx.
  A failed send (HTTP ApiError or a WebSocket error event) is surfaced by attaching the error string to the assistant ChatMessage.error and rendering it as a red bubble — that error string is the only limit signal available on the chat surface in main.
• There is no structured context-exhaustion signal in main: session types carry context_window_tokens, and models carry context_window, but there is no exceeded / limit_reached / usage-percentage field on the chat path. So a reliable detector has to read the error text rather than a backend flag.
• useCreateAgentSession({ agentId, label? }) (in src/lib/mutations/agents.ts) already exists and already invalidates agentKeys.sessions / agentKeys.detail / sessionKeys.lists in its onSuccess; handleNewSession in ChatPage.tsx already wires it to a navigate. The banner reuses both as-is.
• Channels have no conversation/chat surface in the dashboard (ChannelsPage.tsx + lib/queries/channels.ts are config / QR-login / topology only). The issue mentions channel conversations, but there is nothing to attach a banner to there, so this change is honestly scoped to the agent session chat view and that limitation is called out below.

The change (minimal + honest about limits)

• isContextLimitError(text) — a small, case-insensitive heuristic over the error string, matching the phrases real providers and the kernel use for this class of failure (context window / max tokens / "too long" / "string too long" / quota / rate limit / 429, …).
  It is a frontend heuristic, not a fabricated backend field — a false positive only offers a new session, it never blocks the user.
• LimitReachedBanner — guidance banner (styled like the existing CompactionSummaryBanner) with the one-click action, gated on the last message so it clears automatically as soon as the user sends again.
• en / zh / uk locale strings (chat.limit_reached_*). The issue is in Chinese, so the zh copy mirrors the issue's wording.
• ChatPage.limit.test.ts pins the classifier against real-world provider error strings.

No backend contract was invented. When #6215 (in-flight: context-usage indicator + quota-error classification) lands a structured signal, isContextLimitError can be swapped for it without touching the banner.

Relationship to #6215

#6215 (fix/quota-classify-and-context-usage) is related and possibly overlapping — it adds a context-window usage indicator and quota-error classification.
This PR is based on origin/main and does not depend on or duplicate it; it builds the smallest sensible guidance UI on what exists in main today.
If #6215 lands first, the maintainer can reconcile by feeding its structured signal into isContextLimitError (or replacing the heuristic outright).

Verification

All run from crates/librefang-api/dashboard with pnpm (the project's package manager):

• pnpm install --frozen-lockfile — PASS (lockfile unchanged).
• pnpm typecheck (tsc --noEmit) — PASS.
• pnpm lint (eslint .) — PASS (0 errors; only pre-existing baseline warnings, none in the touched files).
• pnpm test (vitest) — PASS, 83 files / 825 tests, including the new ChatPage.limit.test.ts and the locale-parity test.
• pnpm build (vite) — PASS (built in ~0.7s; only the pre-existing chunk-size informational note).

Note: pnpm test:i18n-parity reports 6 pre-existing extra keys in uk.json (*_count_few / *_count_many plural forms for mcp / network / prompts); confirmed present on origin/main and unrelated to this change. The newly added chat.limit_reached_* keys are at full parity across en / zh / uk.

Scope / honest limitation

• Scoped to the agent session chat view — channels are config-only in the dashboard and have no conversation surface to attach guidance to.
• Detection is a frontend error-string heuristic, by necessity, since main exposes no structured context-exhaustion signal on the chat path (see feat(api): context-window usage indicator + honest quota-error classification #6215).

Closes #6211