Summary

When context overflow occurs and auto-compaction cannot recover (reason: "nothing to compact"), runEmbeddedPiAgent currently hard-fails even if fallback models are configured.

This prevents the fallback chain from doing its job in a common failure mode (request too large for current model).

Current behavior

• Model returns context/window overflow (e.g. 413 / request_too_large)
• Auto-compaction runs and fails to reduce enough context
• Runner throws a normal error path and exits
• No fallback model attempt is made

Expected behavior

If fallback models are configured, the runner should escalate this failure as a FailoverError so the outer loop can try the next fallback model/profile.

Why this matters

Fallback chains are specifically meant to recover from model-specific constraints (including context window limits). Hard-failing here defeats reliability for long-running sessions.

Proposed fix

In the overflow/compaction-failure branch in src/agents/pi-embedded-runner/run.ts:

• Build a user-facing overflow message
• If fallbackConfigured, throw FailoverError with:
  • status: 413
  • code: "context_overflow" | "compaction_failure"
  • provider, model, profileId
• Keep existing non-fallback behavior unchanged

Also add coverage in src/agents/pi-embedded-runner/run.overflow-compaction.test.ts:

• Assert FailoverError is thrown when compaction fails and fallback is configured.

Version

Observed on OpenClaw v2026.2.15 with local reliability patches.