Docs for Lite 2026.10.11, core 0.67.1 and ThinkWatch 3.3.0

[fylorn]

Site docs for the joint release: ThinkWatch Lite 2026.10.11 (core 0.67.1, ThinkWatch-Core#303–#306, #308; ThinkWatch-Lite#299, #300 and the no-response timeout and abort change) and ThinkWatch 3.3.0.

Lite

• features.md (en, zh-CN):
  • Traffic and sessions: a request in progress can be stopped with Abort request in its menu, and every request in progress in a session with Abort session in the session's menu or at the top of its panel, after a confirmation. The upstream call is dropped, the client receives an error, the request is marked Aborted in grey, and its upstream is not paused for it; turns on a Responses API WebSocket connection cannot be aborted. The attempts show No response · 300 s («无响应超时 · 300 秒») for an upstream given up at the no-response timeout, and Aborted («手动中止») for the attempt under way when a request was aborted, in place of Start timed out. Long upstream, model and key names are cut short and shown in full on hover (ThinkWatch-Lite#300).
  • Routing and failover: an upstream that sends no content for the No-response timeout (300 seconds by default) is given up on, keep-alive pings not counting; before anything has reached the client the request moves to the next upstream, with none left the client receives a timeout error, and once the answer has started it ends with an error. This replaces the Move to the next upstream when the start times out switch.
  • Settings › Failover: the no-response timeout (300 seconds by default, 30 to 3,600) replaces the start wait and the switch.
  • Upstreams: the Models section of the New upstream and Edit upstream dialogs ends with the same input as Add models…, for an upstream with or without a list; models added there are saved with the upstream and, with Selected models, enabled at once (ThinkWatch-Lite#299). The sentence about the dialog replacing the manual model list is gone, since the free-text list no longer exists. Clients that mark no prompt-cache breakpoints, such as Codex, get them from the gateway on requests converted for an Anthropic Messages upstream, or for a Claude model on Bedrock that AWS lists for prompt caching (Claude 3.5 Sonnet v2, 3.7 Sonnet, 4.5 and later); an upstream that refuses them receives the request again without them (ThinkWatch-Core#306, #308).
• failover-and-load-balancing.md (en, zh-CN): now follows 2026.10.11.
  • When the next upstream is tried lists sending no content within the no-response timeout, in place of the 15-second Wait for the answer to start.
  • Slow starts becomes No response: what counts as content, when the clock starts and restarts, failover before content and the error after it, the same wait for whole answers, the 15-second hold after which streamed answers get their headers and keep-alive comments (Gemini clients the headers only), the latency sample that ranks a silent upstream lower, the attempt label, and that long-thinking models may need a longer timeout.
  • Pauses: a request aborted on the Traffic page does not count against its upstream.
  • Same model name: missing models can also be added in the Models section of the member's edit dialog.
• codex-other-models.md (en, zh-CN):
  • Step 1: a relay's missing models are typed into the box at the end of Models; the Manual list it named no longer exists.
  • New note Tools, history and compaction: from 2026.10.11, every tool Codex declares reaches the upstream, including those in its input, and calls come back under Codex's names; local shell calls, tool search and reasoning-effort changes are converted; compaction works through the upstream's summary, while OpenAI-encrypted compactions are refused with an error; mid-conversation instructions stay in place so the prompt cache keeps hitting (ThinkWatch-Core#304, #305).
  • Cost: the gateway marks prompt-cache breakpoints for Claude (tools, system prompt, last two user turns), priced at the price sheet's cache rates. On Bedrock only the Claude models AWS lists for prompt caching get them (Claude 3.5 Sonnet v2, 3.7 Sonnet, 4.5 and later; older ones such as Sonnet 4 get none), and an upstream that refuses them receives the request again without them.
• Not changed: remote-core.md and the other pages state no core version, so the protocol 43 requirement needs no edit, and no page describes per-release upgrade effects, so the cleared request history and the removed failover keys (which put an old configuration into safe mode with a repair offer) are left to the release notes. compare-cc-switch.md, overview.md and the Lite page say a request moves on when an upstream fails before the answer starts, which stays true.

Core docs

• src/data/core-docs refreshed with CORE_DOCS_REF=v0.67.1 pnpm core-docs. Since v0.66.0 the configuration reference replaces failover.stream_start_wait_secs and failover.next_on_slow_start with failover.idle_timeout_secs (the section on quiet upstreams, the example and the table row), counts an idle timeout among the failures and a manual abort among those that are not, and adds a paragraph on the prompt-cache marks for requests converted for Claude, limited on Bedrock to the Claude models AWS lists, with the resend when an upstream refuses them (v0.67.1). Until the v0.67.1 GitHub release is published, a plain build reads v0.67.0 as the latest release and warns that the copy differs; with CORE_DOCS_REF=v0.67.1 the build reports "Core documents from v0.67.1".

ThinkWatch

• deployment-guide.md (en, zh-CN): the pinning example in 3.3 and the rolling update in 9.2 set 3.3.0. The other pages describe format conversion and cost tracking only in general terms, so nothing about caching or billing needs an edit.

pnpm install --frozen-lockfile and pnpm build pass.

🤖 Generated with Claude Code