Feature request: Built-in Fallback Model for Subagents

[marcinkardas]

Problem

When a subagent's primary model fails (rate limit, overloaded, API error), the entire subagent crashes and propagates the error to the main session. This requires manual intervention and disrupts the workflow.

Common scenario:

• GPT-5.5 rate limit reached → reviewer fails → main session must handle manually

Proposed Solution

Add fallback_model field to agent .md frontmatter:

---
name: my-reviewer
model: openai-codex/gpt-5.5
fallback_model: anthropic/claude-sonnet-4-6
thinking: high
---

When the primary model fails with a retryable error (429, 529, overloaded_error), pi-subagents-lite automatically retries with the fallback model.

Benefits

• Resilient workflows: Multi-agent pipelines continue running despite API issues
• Cost optimization: Use cheaper primary model, fallback to reliable secondary
• Developer experience: No manual intervention needed for transient failures

Implementation Notes

Retry logic should:

• Only retry on retryable errors (rate limit, overloaded, timeout)
• Preserve all other parameters (thinking level, max turns, tools, etc)
• Log the fallback attempt for debugging
• NOT retry on non-retryable errors (invalid API key, model not found, etc)

[AlexParamonov]

Hi @marcinkardas . Thanks for the suggestion! Feel free to open a PR if you'd like to implement this yourself, I would be happy to review it.

Personally I use a different approach: I have a dedicated proxy that switches the active model based on my own logic, rate limits, so I haven't needed to build this into subagentst.
That said, if it'd be useful to you and others, I'd welcome a PR adding it. Thanks again!