Summary

ProxyException.__init__ stores the human-readable message on self.message but never calls super().__init__(self.message), so Exception.args stays empty and str(exc) returns "".

Logging integrations such as StandardLoggingPayloadSetup.get_error_information (consumed by proxy_track_cost_callback, Datadog, OpenTelemetry, S3, etc.) call str(original_exception) to populate StandardLoggingPayloadErrorInformation.error_message. The empty string then ends up in logs/telemetry for every ProxyException-based failure — most visibly for 401 auth errors raised before a provider is selected (llm_provider blank is expected in that case).

Fix

In ProxyException.__init__, immediately after self.message = str(message), call super().__init__(self.message). Five extra lines (one functional + four comment lines). The existing code remap logic ("No healthy deployment available" → 429, tag-routing → 401) is preserved because it runs later on self.code, not on args.

Evidence

Reproduced on clean main (commit 06f6cfc) and re-run on this branch.

Before — clean main, real ProxyException flowed through the real logging chain:

=== BEFORE FIX: ProxyException on clean main (commit 06f6cfc) ===
exc.message  = "key not allowed to access model. This key can only access models=['gpt-3.5-turbo']. Tried to access gpt-4"
str(exc)     = ''
exc.args     = ()

StandardLoggingPayloadSetup.get_error_information() returned:
  error_message = ''
  error_class   = 'ProxyException'
  error_code    = '401'
  llm_provider  = ''

>>> BUG: error_message is empty even though exc.message is populated <<<

After — patched branch, same call:

=== AFTER FIX: ProxyException on patched branch (feat/lit-3094-proxyexception-args) ===
exc.message  = "key not allowed to access model. This key can only access models=['gpt-3.5-turbo']. Tried to access gpt-4"
str(exc)     = "key not allowed to access model. This key can only access models=['gpt-3.5-turbo']. Tried to access gpt-4"
exc.args     = ("key not allowed to access model. This key can only access models=['gpt-3.5-turbo']. Tried to access gpt-4",)

StandardLoggingPayloadSetup.get_error_information() returned:
  error_message = "key not allowed to access model. This key can only access models=['gpt-3.5-turbo']. Tried to access gpt-4"
  error_class   = 'ProxyException'
  error_code    = '401'
  llm_provider  = ''

After — four real ProxyException shapes through the full proxy logging chain, all now propagate to payload.error_message:

--- 401 auth - restricted key ---
  payload.error_message     = "key not allowed to access model. This key can only access models=['gpt-3.5-turbo']. Tried to access gpt-4"
  payload.error_code        = '401'

--- 401 auth - invalid token ---
  payload.error_message     = 'Authentication Error, Invalid proxy server token passed.'
  payload.error_code        = '401'

--- 429 no healthy deployment (routing remap) ---
  payload.error_message     = 'No healthy deployment available, passed model=foo'
  payload.error_code        = '429'   # routing remap preserved

--- 400 invalid request ---
  payload.error_message     = 'missing required field: model'
  payload.error_code        = '400'

Regression tests added in tests/test_litellm/proxy/test_proxy_types.py (all pass on the patched branch):

tests/test_litellm/proxy/test_proxy_types.py::test_audit_log_masking PASSED
tests/test_litellm/proxy/test_proxy_types.py::test_internal_jobs_user_has_proxy_admin_role PASSED
tests/test_litellm/proxy/test_proxy_types.py::test_proxy_exception_non_string_message_coerced PASSED
tests/test_litellm/proxy/test_proxy_types.py::test_proxy_exception_populates_standard_logging_error_message PASSED
tests/test_litellm/proxy/test_proxy_types.py::test_proxy_exception_routing_code_override_still_works PASSED
tests/test_litellm/proxy/test_proxy_types.py::test_proxy_exception_str_returns_message PASSED
tests/test_litellm/proxy/test_proxy_types.py::test_proxy_exception_to_dict_unchanged PASSED
======================== 7 passed, 2 warnings in 9 s =========================

Five new regression tests cover:

1. str(ProxyException) == message (the immediate bug)
2. Full StandardLoggingPayloadSetup.get_error_information(...) chain returns the populated message
3. to_dict() shape unchanged
4. "No healthy deployment available" → 429 code remap still works
5. Non-string message is still coerced

Notes for reviewers

• Existing # NOTE: DO NOT MODIFY THIS comment in ProxyException is about the OpenAI-shape mapping (code, type, param, openai_code, to_dict() keys). This change does not touch any of that — it only populates Exception.args so str(self) works.
• Pushed via the GitHub Contents API (current bot token lacks repo/workflow scopes for git push), one commit per file plus one cleanup commit. Net diff is the two-file change above; intermediate commit messages are noisier than the final state.
• Tooling friction (sandbox volatility) made the cleanup commit (959288c) necessary; the cumulative effect against the base branch is +5 / 0 in _types.py and +74 / 0 in the test file.

Fixes LIT-3094.

Verification (ship-pr)

• change-type: litellm/proxy/_types.py (exception class, backend logging surface) + tests/test_litellm/proxy/test_proxy_types.py (tests). Required evidence: chained call output through the actual logging code path. PASS — see Evidence section.
• evidence present: PASS (3 fenced blocks: clean-main repro, patched-branch verification, four real ProxyException shapes through StandardLoggingPayloadSetup.get_error_information chain).
• image sanity: N/A (no images — terminal/Python captures inlined).
• image content matches caption: N/A.
• backend evidence from running system: PASS — outputs came from real Python imports of both the unpatched-main branch and the patched branch, exercising the same code path StandardLoggingPayloadSetup.get_error_information that proxy_track_cost_callback, Datadog, OpenTelemetry, and S3 callbacks call inside the running proxy.