ADR β ACS Failure & Expansion Policy (v0) β
Decision
- On provider timeout/errors, do not auto fallback to L1.
- First, retry within per-provider time slice, possibly with reduced top_k.
- If project data is missing/insufficient, emit warning and propose explicit L1 expansion.
- Return structured error to the agent with options; agent decides next step.
Context
- Priorities: Quality > Strict token budget > Latency (comfort).
- Streaming disabled in v0; batch replies.
Consequences
- Predictable token spend; avoids silent scope creep via hidden L1.
- Slightly higher perceived latency in error cases, but transparent control.
Record
- Errors contain: code, message, retriable, attempt/max_attempts, suggested options.
- Provider API adds
warnings[]
; L1 expansion requiresexpansion:true
.