[ajianaz]

What

Part 1 of 3 of #209: the failover mechanism in tole-core. Part 2 adds the real-provider integration test, part 3 the env/config wiring and docs.

Why

An unattended mission parks resumable when its provider has an outage, but makes no progress until a human resumes. A reserve provider turns most outages into a non-event, and durable sessions make the switch auditable.

Changes

• failover.rs (new): FailoverProvider wraps an ordered chain (Link { label, provider }, primary first). On an "unavailable" error (connection failure, timeout, dropped stream, HTTP 5xx or 429) the same step is retried on the next provider; request errors (401/403/404/400, malformed responses) fail loud and never fail over. The answering provider stays active (no flapping); the primary is retried after retry_primary_after (default 5 min). When every provider fails, the last error leads (the turn loop's transient-retry classification reads it) followed by the earlier ones.
• provider.rs: new default method last_route() (None for a single provider: no behaviour change).
• turn.rs: the route is stored under the namespaced tole_route key of the step's usage row, next to tole_wire.
• usage_report.rs: a row holding only tole's own tole_* keys is not a provider step (an empty object still is, the pre-token efficiency: lazy tool registration + tool-result trimming policy #211 contract), so route rows never change step accounting.
• CHANGELOG.

Testing

• 6 unit tests: error classification, failover and route report, request errors never reach the reserve, sticky reserve and primary retry after the cooldown, all-fail message order, single-provider transparency; plus a usage-report test for route-only rows.
• fmt, clippy -D warnings, cargo test --workspace, cargo check -p tole-core --no-default-features pass; cora review --staged clean.
• Real OpenAiProvider error text and a full turn are covered by the integration test in part 2 (passes locally).

Refs #209

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-code · BYOK · MIT