fix(agent): walk cause chain when extracting wrapped API error bodies (salvage of #14287 by @LeonSGP43)

[teknium1]

Summary

Wrapped API errors now keep the nested body message, so transient 402 usage limits classify as rate_limit (retryable) instead of being misread as terminal billing failures.

Root cause: classify_api_error() extracted the status code via _extract_status_code() (which walks __cause__/__context__) but the body via _extract_error_body() (which only inspected the top-level exception). A wrapped 402 kept the nested status code but lost the nested "try again in 5 minutes" body, so _classify_402() saw no transient signal and defaulted to billing/non-retryable.

Changes

• agent/error_classifier.py: _extract_error_body() now walks the __cause__/__context__ chain (max depth 5), mirroring _extract_status_code().
• tests/agent/test_error_classifier.py: covers nested-body extraction and the full wrapped-402 → rate_limit classification path.

Validation

Case	Before	After
Wrapped 402, nested body "Usage limit, try again in 5 min"	billing, retryable=False	rate_limit, retryable=True
Wrapped 402, nested body "credit balance too low"	billing	billing (no regression)

Targeted suite: 164/164 pass. E2E verified with real imports against the live classification path.

Salvage of #14287 by @LeonSGP43 (cherry-picked onto current main, authorship preserved). Closes #14195. Duplicates: #14219 (@liuhao1024, first submitter, already closed), #14349 (@sgaofen), #14787 (@Tranquil-Flow).

Infographic

[github-actions]

🔎 Lint report: hermes/hermes-7ce5d388 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11482 on HEAD, 11482 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 6032 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.