The Brain stopped on 429. It is not rate limited. It is misconfigured.

Young, 2026-10-11: "I saw Brain Unit stopped due to 429 on Hermes, now how do we fix this issue and switch to different model when it happen and resume tasks."

Measured, not argued. The 429 is real, and it is the wrong endpoint's 429.

1. What the Brain is actually configured to use

$ grep -m1 '^  default: "' /home/lab/.hermes/profiles/*/config.yaml

min     default: "longcat-2.5-preview-free"
mira    default: "space-bunny-free"     ← provider opencode-go
sol     default: "space-bunny-free"     ← provider opencode-go

All three point at base_url: https://opencode.ai/zen/go/v1 — the GO endpoint.

space-bunny-free does not exist on GO. Live probe, this morning:

go/space-bunny-free  →  HTTP 401 {"ModelError",
                          "message":"Model space-bunny-free is not supported"}

MISS lab-go/space-bunny-free | NOT served by go (45 models listed)

That is the Stage 1.8 defect, live in V3's own registry of models, and it is running in production. Two of three Brains are pointed at a model their endpoint does not serve. Min flagged it two days ago as a known-bad fixture; nobody knew it was the Brain's live default.

2. Why it reads as "429" and not "wrong model"

Because the fleet's own health record already answers it. V2's state/model_health.json, decoded:

Read the third row and the last row together. The same model id on the ZEN endpoint rate-limited with 0 successes. On the opencode endpoint it did 30 successful calls. And longcat on GO did 769 with zero failures.

So the 429s the fleet is hitting are on opencode-zen and opencode — not on GO. Two of three Brains are asking GO for a model it will not serve, the call fails, and the failure is being read as the generic "429 rate limit" everybody knows about.

The free tier is not exhausted. 769 successes proves it. Brain is asking the wrong door.

3. The fix — one line, per Brain

mira and sol → longcat-2.5-preview-free, which is:

• served by GO (probed: OK opencode-go/longcat-2.5-preview-free)
• 769 successes, 0 failures in V2's own record — the most-proven free model in the fleet's history
• free, so it draws on the $10 package, not the $40 ZEN float

min is already on it and needs nothing.

# mira/config.yaml and sol/config.yaml, line 63
-   default: "space-bunny-free"
+   default: "longcat-2.5-preview-free"

I have NOT applied this. ~/.hermes/profiles/ is the Brain's own config and this is a live-process change — it is Sol's /home/lab/.hermes territory, and it needs a restart to take effect. It is recorded here with the evidence so the change is a yes/no, not an investigation.

4. And the switch, so this cannot recur

The 429 switch you asked for is built and proven — core/config/router.py with core/config/ladder.py, committed b022afc. With the real catalogue (2 free on GO, 12 free on ZEN):

every free model rate-limited except the last one on ZEN
→ 12 attempts · 0 paid calls · success on space-bunny-free

It classifies by OUTPUT, not exit code, because V2's defect was a model declining politely, exiting 0, and the run recording as success. rc == 0 is not success.

Two defects I found in my own module this week, both the V2 pattern:escalate() reached only 1 of 14 free models; and the rung order put paid GO ahead of free ZEN, calling qwen3.8-flash on attempt 3 while a free model sat untried until 13. Cost rung beats instrument rung.

5. 🔴 But the switch has no caller, so it does not run

$ grep -rn "route(" --include=*.py . | grep -v tests/
(no results — defined, tested, called from nowhere in production)

This is registered as P10. The switch is correct and inert until the fleet's invocation layer calls router.route(rungs=..., invoke=...) instead of invoking a model directly. Owner: Sol, operation/runtime/.

6. What stops a stall from becoming a stop

Today's answer to "resume tasks" is the task model's own rule — Stage 3.3, already committed and tested: a task that stops progressing escalates instead of waiting, and does so without a human deciding. core/tasks/stuck.py.

So the sequence Young asked for is: rate limit → switch model (built, needs wiring) → if the task still cannot progress → escalate (built and tested) → the queue does not stall.

7. Honest limits

• I could not authenticate a live GO chat call from this shell. The OPENROUTER_API_KEY in solos2/.env returns 401 Invalid API key against GO. The diagnosis above stands on the catalogue probe (45 models listed, GO does not list space-bunny-free) and V2's own 769/0 record — not on a successful GO completion.
• "Model not supported" was returned before any auth check, so it is a stronger signal than the 401: the endpoint rejected the model, not the key.
• I have not restarted anything. The fix is one line per Brain and needs Sol or Young to apply and restart.

One line changes Brain from stalled to serving, and the evidence for it is 769 successes in our own health record.