Summary

When a session is pinned to a model that only exists in agents.defaults.model.fallbacks (not as primary), the fallback chain is silently reduced to length 1 via deduplication. The result is fallbackConfigured:false in the decision log — the session has no safety net despite a fallback chain being configured globally.

Steps to reproduce

1. Configure: agents.defaults.model.primary: venice/claude-sonnet-4-6, fallbacks: [openai/gpt-5.5]
2. Pin the main session to openai/gpt-5.5 (e.g. via model switch)
3. Trigger a timeout on openai/gpt-5.5

Expected: fallback chain is consulted; at minimum the user gets an error from exhausted fallbacks with options
Actual: fallbackConfigured:false, chain_exhausted after 1 attempt, silent failure

Root cause (traced in source)

model-fallback-Bpl6-0ev.js line 662:

const hasFallbackCandidates = candidates.length > 1;

When effectivePrimary is openai/gpt-5.5 and the configured fallbacks list is [openai/gpt-5.5], the createModelCandidateCollector dedup set (seen) eliminates gpt-5.5 as a fallback candidate because it was already added as the primary. candidates.length === 1 → hasFallbackCandidates = false.

The fallback inclusion logic (lines 530–545) correctly detects that the effective primary differs from the configured primary and checks whether the effective primary exists in the fallback list — but this check serves to include fallbacks, not to expand them. The dedup step then collapses the list back to 1.

User-facing surprise

Pinning reduces resilience, which is the opposite of user expectation. A user who pins to a fallback model for testing assumes the global fallback chain remains operative.

Proposed fix

When a session pin's effective primary matches a model in the configured fallback list, expand the candidate list with the next-tier fallback (e.g. if the pin is openai/gpt-5.5 and the global fallbacks are [openai/gpt-5.5], include the global primary venice/claude-sonnet-4-6 as a fallback candidate, or at minimum warn that the fallback chain is collapsed).

A simpler mitigation: when candidates.length === 1 and isPrimary && effectivePrimary !== configuredPrimary, add the configured primary as a fallback candidate.

Evidence from logs

model_fallback_decision: decision=candidate_failed
requestedProvider=openai, requestedModel=gpt-5.5
attempt=1, total=1
fallbackConfigured=false
fallbackStepFinalOutcome=chain_exhausted

Version

OpenClaw 2026.5.3-1 (2eae30e), macOS 26.4.1 arm64