Troubleshooting

Model ID Invalid or “Agent Failed Before Reply”?

A clean recovery flow for broken model settings, invalid fallback IDs, and “all models failed” errors.

If your agent suddenly stops responding after model changes, you’re usually dealing with config validation — not a broken account.

Common trigger: changing the primary model and fallback models at the same time, then one ID is typed wrong or no longer available.

What this error usually means

5-minute recovery workflow

1) Revert to one known-good model (no fallbacks)

Temporarily remove every fallback model. Keep only one model that previously worked in your setup.

2) Run a tiny canary test

Send a minimal prompt like “Reply with OK.” If this fails, the issue is still base config/provider, not prompt quality.

3) Validate fallback IDs one-by-one

Add one fallback, test. Add the next, test. Do not add 3+ at once or you won’t know which one broke it.

4) Check provider/key limits

If you see 403 key limit errors, the model ID may be fine — the key quota is not. Fix quota/key first.

5) Keep a rollback block

Save your last known-good model config in a note so you can restore fast after experiments.

Good vs risky model-change pattern

Risky (hard to debug)

Change primary model + add 3 new fallbacks + switch provider key in one edit.

Safe (debuggable)

Step 1: Change primary model only → test Step 2: Add fallback #1 → test Step 3: Add fallback #2 → test

Important: if a model worked last week but fails now, the model may have been renamed, deprecated, or your key scope changed.

High-friction edge cases (that look random but aren’t)

Loop-break sequence

1) Set ONE known-good model (no fallbacks) 2) Start a fresh session (or restart runtime if required) 3) Run canary: "Reply with OK" 4) Re-add exactly one fallback 5) Re-test before adding another

Copy/paste prompt to self-heal config

Diagnose my model config without guessing. 1) Tell me the exact primary model currently set. 2) List all fallback model IDs exactly as configured. 3) Flag any invalid or suspicious IDs. 4) Propose the smallest safe rollback to one known-good model. 5) After rollback, run a tiny canary prompt: "Reply with OK" and report result.

When to escalate to support

Support packet (fast)

Bottom line: model failures are usually config drift, not agent failure. Roll back to one known-good model, test small, then reintroduce fallbacks one at a time.