Provider rate-limit pacing

Specsmith's provider adapters can pace requests before a 429, using rolling request-per-minute and token-per-minute windows. This is supporting runtime behavior for Grace and benchmarks, not a separate credit-management product.

The public CLI does not expose rate-limit administration. Prefer the coding agent or provider dashboard for account limits. Grace reads configured endpoint and provider settings automatically and reports actionable retry guidance.

Project-specific overrides may be stored in .specsmith/model-rate-limits.json; runtime counters stay in .specsmith/model-rate-limit-state.json and should remain untracked.

Python API

from pathlib import Path

from specsmith.rate_limits import BUILTIN_PROFILES, load_rate_limit_scheduler

scheduler = load_rate_limit_scheduler(Path("."), BUILTIN_PROFILES)
reservation = scheduler.acquire(
    "openai",
    "configured-model",
    estimated_input_tokens=5_000,
    max_output_tokens=2_000,
)

# After the provider call:
scheduler.record_success(
    reservation,
    actual_input_tokens=4_800,
    actual_output_tokens=1_900,
)

On a provider error, call record_rate_limit; it parses a provider wait hint, reduces concurrency, and returns a bounded retry delay. Limits remain estimates: the provider account is authoritative.