[LinuxDevil]

What

New src/providers/resilience.ts, exported from src/providers/index.ts and so from the package root:

• withRetry(provider, { maxRetries = 2, backoff: { initialMs, maxMs, factor, jitter }, retryOn, onRetry, signal, timeoutMs }) returns an LLMProvider whose generate() and stream() retry transient failures with exponential backoff.
  • Retryability comes from the existing compactProviderError() classification, or .compacted on a CompactedLLMProviderError, through the new exported isRetryableProviderError(). Rate limits, timeouts, network errors and 5xx responses are retried. Auth failures, invalid requests and context-length errors are not.
  • A provider's retryAfterMs replaces the backoff delay.
  • It never retries an abort. The wrapper's signal or the call's signal stops retrying, including during the backoff wait.
  • A stream() call is retried only when it rejects. An error inside a stream that was already returned is not retried.
• withFallback(providers, { fallbackOn, onFallback }) tries the providers in order and rethrows the last error when all fail.
  • name/defaultModel report the provider that served the latest call. supportsTools/supportsStreaming/getModels ask the first provider.
  • Each fallback runs on its own defaultModel. An explicit model goes only to the first provider.
  • It composes with retry: withFallback([withRetry(a), withRetry(b)]).
• resilientProvider(provider, config) applies the LLMProviderConfig.maxRetries and timeout fields, which were declared but never read. timeout is a per-attempt limit (timeoutMs).
• The mock provider's private abortableDelay moves to src/providers/abortableDelay.ts so the retry backoff can reuse it (fallow flagged the copy as a clone). The only change in behavior: it now rejects at once when the signal is already aborted.
• Docs: a "Provider retries and fallback" section in docs/configuration.md with a type-checked snippet, plus a row in the Providers table in docs/api-overview.md. llms-full.txt is regenerated.

src/execution/retry.ts is not reused, and it is left untouched:

• Its backoff sleep ignores the abort signal.
• Its delay hint reads only the legacy RateLimitError, or a fixed 60s for any 429.
• Its default isRetryableError retries any LLMProviderError without a status code, which includes compacted auth failures.

Reusing it would have meant changing its behavior for existing callers. Its removal is a separate ticket.

Why

LOU-V7 (Epic V, run loop) closes a parity gap with eve's AI Gateway fallbacks and open-harness's withRetry middleware. This is the provider-layer half. The agent-level createAgent({ retry, fallbackModels }) options and the typed retry stream event are LOU-V7.2.

Acceptance criteria

• withRetry retries generate and stream on retryable errors using the SDK's provider-error classification
• retryAfterMs honored (fake-timer test)
• Non-retryable errors are not retried; aborts are never retried; abort during backoff stops retrying
• withFallback switches providers, reports onFallback, and rethrows the last error when exhausted
• name/defaultModel reflect the active provider; capabilities delegate to the first
• Wrappers compose (withFallback([withRetry(a), withRetry(b)]), stream path)
• LLMProviderConfig.maxRetries/timeout honored via resilientProvider()
• Exported from src/providers/index.ts (and the root via export * from './providers')
• Tests in src/providers/resilience.test.ts (17 tests)
• Docs section with type-checked snippet; llms files regenerated
• No edits to createAgent.ts, AgentExecutor.ts, agentEvents.ts, resume.ts, execution/retry.ts

How verified

• npx tsc --noEmit: no errors
• npm run lint: ✖ 433 problems (0 errors, 433 warnings), the same count as main
• npm run test:types: Tests 19 passed (19), Type Errors no errors
• npx vitest run: Test Files 133 passed (133), Tests 1969 passed | 3 skipped (1972)
• npm run build && npm run build --workspace=packages/create-loushy-agent: both exit 0
• npm run docs:verify-snippets -- --skip-build: all 85 snippet(s) type-check against src/; 6 also run cleanly
• npm run docs:llms:check: passes
• npm run test:coverage: passes (resilience.ts 99.54% lines). npm run fallow: ✓ No issues found, ✓ 0 above threshold. The remaining 30 duplicated lines are the existing OpenAI/Anthropic provider clone.

Note: the built-in providers still run the ai SDK's own internal retries (2 by default) inside each withRetry attempt. This is documented in the new section.

🤖 Generated with Claude Code

https://claude.ai/code/session_016bY29ibQgv3MDdjAbUT4YR

---

Generated by Claude Code