BasicallyMythos — Claude Code reroute observability + fallback choice

See every time Claude Code reroutes your request to a different model — and choose what the fallback model is.

When Fable 5 trips a safety safeguard, Claude Code re-runs that request on Opus. This setup makes it re-run on Kimi K3 instead — automatically, mid-session, while everything else stays on your Anthropic subscription. And whether or not you change the fallback at all, the proxy logs every reroute, so the downgrade is finally visible.

No third-party routers, no packages. One ~100-line dependency-free Node proxy + two settings keys.

[2026-07-21T09:01:57Z] anthropic model=claude-fable-5 POST /v1/messages -> 429
[2026-07-21T09:02:10Z] KIMI      model=kimi-k3        POST /v1/messages -> 200

Every KIMI line is a silent downgrade that got caught — and answered by the model you chose.

Why

• Safeguard reroutes are silent. No banner, no warning — the response just comes back from a different model than the one you picked.
• The fallback model should be your choice. The reroute target is a client-side config value; this proxy lets you point it at an open model.
• Zero trust surface. Zero dependencies, loopback-only, your Anthropic auth passes through untouched, the Kimi key never leaves the keychain.

Quick start

cp kimi-reroute-proxy.cjs ~/.claude/
security add-generic-password -s moonshot-kimi -a "$USER" -w "sk-kimi-YOUR_KEY"
cp com.kimi-reroute-proxy.plist ~/Library/LaunchAgents/   # edit YOUR_USERNAME + node path first
launchctl load ~/Library/LaunchAgents/com.kimi-reroute-proxy.plist
# then merge settings-snippet.json into ~/.claude/settings.json and restart Claude Code

Verify with /status (Base URL should read http://127.0.0.1:8787) or the curl test in Manual setup.

Roadmap

Version	What
v0.2	stats command — shareable reroute summary, formatted for screenshots
v0.3	Linux systemd unit + Windows instructions
v0.4	Multi-provider fallback presets (GLM / DeepSeek / OpenRouter)
v1.0	Tests, semver, docs freeze

Watch releases to follow along.

FAQ

Is this against Anthropic's ToS? It runs locally against your own session, on your own machine. It logs rather than hides, forwards flagged requests to a different provider, and impersonates nothing. No API keys are shared.

Does it touch my Anthropic auth or quota? No. Anthropic traffic passes through byte-identical with your normal auth. Only requests already flagged for rerouting take the Kimi branch.

Windows/Linux? macOS today (launchd + keychain). Linux systemd and Windows instructions are on the roadmap — issues and PRs welcome.

Why Kimi K3? Open weights, 1M context, #4 on Agent Arena (level with Opus 4.8). Any Anthropic-compatible endpoint works — see the env overrides below.

How it works

Claude Code's safeguard fallback is client-side: when a request is flagged, the CLI re-issues it using the model id in ANTHROPIC_DEFAULT_OPUS_MODEL, sent to the same ANTHROPIC_BASE_URL as everything else (it's process-wide — you can't natively point one model at a different provider).

So:

1. ANTHROPIC_DEFAULT_OPUS_MODEL=kimi-k3[1m] — the reroute now asks for Kimi.
2. ANTHROPIC_BASE_URL=http://127.0.0.1:8787 — all traffic exits via a tiny local proxy.
3. The proxy splits by model id:
   • kimi* → Kimi's Anthropic-compatible gateway (api.kimi.com/coding), with your Kimi key injected from the macOS keychain
   • everything else → transparent pass-through to api.anthropic.com (your normal auth/subscription untouched)

Result: normal Fable session on Anthropic; the moment a safeguard fires, that one request lands on Kimi K3 (1M context) instead of Opus. The transcript shows the model-switch notice as usual.

Gotchas this handles (learned the hard way)

• Kimi subscription keys (sk-kimi-...) are NOT Moonshot platform keys. They 401 on api.moonshot.ai / api.moonshot.cn. The correct endpoint for Kimi Code subscription keys is https://api.kimi.com/coding/v1/messages (full Anthropic Messages format, streaming included).
• The gateway rejects Claude Code's [1m] suffix ("Please set model id as k3"). Keep kimi-k3[1m] in settings so Claude Code budgets a 1M context — the proxy strips the bracket suffix before forwarding.
• switchModelsOnFlag must be true in settings.json, or the safeguard reroute never fires at all (flagged requests just error).
• Keep your session model on Fable, not Opus. The env var remaps the whole Opus alias — if your main model is Opus, your main session rides to Kimi too.

Setup

Prereqs: macOS, Node (any recent version), Claude Code, a Kimi Code subscription API key (kimi.com console → API Keys).

1. Proxy — copy it into place:

   cp kimi-reroute-proxy.cjs ~/.claude/kimi-reroute-proxy.cjs

2. Key — store it in the keychain (never on disk / in settings):

   security add-generic-password -s moonshot-kimi -a "$USER" -w "sk-kimi-YOUR_KEY"

3. Keep it running — edit com.kimi-reroute-proxy.plist: replace YOUR_USERNAME, adjust the node path if which node differs. Then:

   cp com.kimi-reroute-proxy.plist ~/Library/LaunchAgents/
   launchctl load ~/Library/LaunchAgents/com.kimi-reroute-proxy.plist

4. Claude Code settings — merge settings-snippet.json into ~/.claude/settings.json (the two env keys + switchModelsOnFlag: true).

5. Verify — new claude session, run /status: Base URL should read http://127.0.0.1:8787. Direct proxy test:

   curl -s -X POST http://127.0.0.1:8787/v1/messages \
     -H "content-type: application/json" -H "anthropic-version: 2023-06-01" \
     -H "x-api-key: x" \
     -d '{"model":"kimi-k3[1m]","max_tokens":16,"messages":[{"role":"user","content":"say ok"}]}'

   A Kimi completion back = the reroute path works end to end.

Rollback

Remove the two env keys from ~/.claude/settings.json, then:

launchctl unload ~/Library/LaunchAgents/com.kimi-reroute-proxy.plist

Files

File	Purpose
kimi-reroute-proxy.cjs	The proxy. Zero dependencies, binds 127.0.0.1 only.
com.kimi-reroute-proxy.plist	launchd service (auto-start + keep-alive).
settings-snippet.json	The Claude Code settings to merge.

Overrides via env (in the plist if needed): KIMI_PROXY_PORT (default 8787), KIMI_MOONSHOT_HOST / KIMI_MOONSHOT_PREFIX (default api.kimi.com + /coding; Moonshot platform keys would use api.moonshot.ai + /anthropic).

Security notes: the proxy binds loopback only; your Anthropic auth passes through untouched; the Kimi key is read from the keychain at request time and only ever attached on the Kimi branch.

---

Built by Mythos — build updates and Claude Code internals at @TEA_Resistance on X. Issues and PRs welcome. If it catches a reroute for you, a star helps others find it.