feat: single bounded retry layer for agent LLM calls

[SkyeAv]

Replaces three stacked, literal-"429"-matching retry layers with exactly one bounded, jittered retry layer shared by the inner agent, reflexion, and judge — with credentials redacted at the retry boundary. The old stack burned ~16 min per failing call and retried permanent errors.

Retry layer

• agent.make_retrying_model(...) wraps any smolagents Model.generate through net.retry_transient with injectable sleep/rng/secrets: 4 attempts, sleeps summing to ≤ 45.0 s per call (LLM_MAX_TOTAL_BACKOFF).
• Provider retryers disabled: build_model returns bare providers with retry=False (OpenAI additionally client_kwargs={"max_retries": 0}); LiteLLM's backend-internal retry is not exposed by LiteLLMModel.__init__ and stays outside this control. The CLI composes retry INNER of distilling at all three model roles, passing secrets=(resolved_key,).
• Worst case: 29 logical calls per article (20 agent + 3 reflexion + 6 judge) × 45 s = 1,305 s ≈ 21.8 min, inside the 90-minute budget.

Coded failures and redaction

• LlmTransientError (llm-transient) is raised on exhaustion and persisted via the fix: signal coded supervisor skips #166 error_code field; errors.error_code_of's cause-walk preserves the code through smolagents AgentGenerationError wrappers.
• errors.redact_secrets(text, secrets=()) (named-group regex with quote backreference; exact configured secrets first) is applied at the retry log kwargs and both transient-error constructors — a quoted-JSON credential cannot leak into logs or state.
• Metadata preserved: RetryingModel explicitly copies model_id/flatten_messages_as_text/custom_role_conversions/client (the Model base defines them, so __getattr__ cannot delegate) and overrides supports_stop_parameter.

Testing

• uv run pytest tests/test_agent_assembly.py -q --no-cov -n 0 → 21 passed; tests/test_errors.py → 11 passed (redaction: exact secrets, quoted JSON, no over-redaction of benign text)
• CI-equivalent base env (no [agent] extra): smolagents-touching tests guard with pytest.importorskip
• Full gate on the stack tip: uv run pytest -q → 1565 passed, 15 skipped; ruff/format/pyright clean; Tier-2 passed after three rounds (metadata collisions, credential leakage, AgentGenerationError cause-walk all repaired)

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 1fe243c2-2d5e-4f48-abb7-dc69dd8badeb

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}