Built-in bounded auto-retry with schema degradation for structured-output subagents

Type: enhancement · Area: multi-agent workflows / structured output

Problem

The dominant cause of structured-output subagent failures is output shape, not the task: a schema that is too large/strict for a high-surface item, or output that runs out of room mid-JSON and fails validation repeatedly. Today the only recovery is manual — the author notices the failure, hand-edits the schema to be leaner, and relaunches. That is exactly the kind of mechanical recovery the harness should own. (Companion to the silent-drop bug filed separately.)

Proposed feature

An opt-in (ideally default) bounded auto-retry for agent({schema}) that, on retry-cap exhaustion, retries once with a degraded shape rather than the identical failing call:

• relax required to the minimal essential set,
• drop additionalProperties:false,
• apply/keep maxItems and string-length caps,
• optionally split the request.

Constraints:

• Bounded — at most one automatic degraded retry per item, so a genuinely broken task can't loop and burn tokens.
• Transparent — if the degraded retry succeeds, include it (optionally tagged "recovered"); if it still fails, surface it loudly. Auto-retry must reduce interruptions on the recoverable case without ever becoming a silent drop on the unrecoverable one.

Why it helps everyone

Structured fan-out over uneven work is a common workflow shape. A built-in degrade-and-retry turns the single most common failure mode from "silent partial result" into "self-healed," and removes a class of manual babysitting that every author currently reinvents.

Filed from real multi-agent workflow usage.