[fylorn]

Site and docs for the joint release of ThinkWatch Lite 2026.10.29, ThinkWatch Core 0.83.1 and ThinkWatch Enterprise 3.10.0. To be merged once the three releases are published.

Lite docs (English and Chinese)

• Features: routing rules set Reasoning (Off, or low to max) in place of extended thinking, with the safe-mode note for rules saved before 2026.10.29; Off is written the way each upstream's format writes it, also after a conversion, and a refused level is resent once with one the upstream accepts (the lowest listed for a refused none) and remembered; temporary failures of every upstream are retried after 2, 4, 8, then 16 seconds, a single upstream included; Wait at most covers queueing and retrying (0 turns both off); the attempt chain shows Retry n, Waited … s and effort changes, and Traffic shows Retrying; request details and conversations render Markdown.
• Failover and load balancing: a new note on retrying after a temporary failure (what counts, waits, what is not retried, how it shows); the waiting note and the 400 about the reasoning effort; written for 2026.10.29.
• Install and update, FAQ: updates hand the running gateway over to the new version, which finishes the requests in flight; Homebrew upgrades and Restart Gateway work the same way; the window comes back as it was; updating to 2026.10.29 itself still uses the previous way.
• Use Claude Code with GLM, DeepSeek or Kimi: setting Reasoning in a rule, for example to max.
• Plugins: a retry after a temporary failure runs the request hooks again.
• Overview: the routing and failover line.

Product pages and Core docs

• Lite page: the routing card and the update line; Core page and Core overview: retries in failover before the first byte; Enterprise features: per-model reasoning and retries; llms.txt.
• Core documents snapshot (src/data/core-docs) refreshed from the v0.83.1 tag: the retry and reasoning sections of the configuration reference, set.reasoning replacing set.thinking.

Enterprise docs

• Configuration reference: a Gateway table with gateway.retry_wait_secs.
• API reference: new sections on temporary upstream failures and on the model's reasoning effort (models.reasoning_effort), including how off is written for each upstream format, the resend of a refused level for every model, the retries and effort log fields and gateway_provider_retry_total.
• Architecture: the retry module description.

The site builds and its internal links resolve.

🤖 Generated with Claude Code