manifest v6.26.1

🐛 Patch Changes

• ce10922: Stop adding five-minute cache breakpoints to Anthropic Messages requests that already use one-hour breakpoints (Claude Code), which Anthropic rejected with a 400, and explain in the Claude Code setup card how to get a 1M context window through the gateway.
• 9b0a662: Fix a 500 on Chat Completions requests when Autofix heals an OpenAI subscription (Codex) call by switching the model.
• 6173ff2: Bound routing momentum memory by session count and session-key length.
• d228930: Bump CODEX_CLI_VERSION from 0.154.0 to 0.156.1 to track the current Codex CLI release, so OpenAI subscription model discovery can list the newest Codex models (e.g. gpt-6-sol, gpt-6-luna). Older client_version values receive an older model subset.
• aa6734c: Make the gateway's per-tenant (M201) and per-IP (M202) rate caps configurable with MANIFEST_RATE_MAX_REQUESTS and MANIFEST_IP_RATE_MAX_REQUESTS, alongside the in-flight (M203) cap already set by MANIFEST_CONCURRENCY_MAX.
• bf9702c: Editing a custom provider: "Fetch models" now reuses the saved API key when you don't retype it. The saved key is only sent to the provider's own base URL.
• 64918e7: Make the per-harness enabled-providers endpoints resolve the active harness when a deleted harness had the same name, so reading, enabling, and disabling providers no longer act on the deleted one.
• 813e058: Show a loading state while a routing fallback reorder is saving, and block further drags on that tier until it lands
• e458112: Stop the Limits page from flashing the cloud email card on self-hosted installs, and keep the remove-provider confirmation honest when the refresh that follows a successful delete fails.
• fca2948: Keep Pro while Stripe retries a failed renewal, so the dashboard offers Manage billing instead of a second checkout.
• 8b4d9d4: Stop showing the "You've used all 10,000 requests" notice on the Upgrade page to Pro users who arrive from an old request-limit link.