[tang-vu]

Why this matters

When the switchboard is in observe/disabled mode, the LiteLLM success callback owns Token Spy telemetry. LiteLLM 1.90.7 reports Anthropic cache creation under usage.prompt_tokens_details.cache_creation_tokens; ODS currently reads only usage.cache_write_tokens, so a real 4,096-token cache creation is sent to /api/ingest/routed as zero. The cache-write history loses that activity.

Read the normalized detail field through the existing bounded counter conversion. Keep the previous top-level key for callers without the detail, and preserve an explicit measured zero. Input/output totals, cache reads, metadata-only payloads, queue behavior and the switchboard's exactly-once suppression remain unchanged. This corrects telemetry attribution; it does not change pricing or reinterpret total input tokens.

Overlap check

Searched all open/closed PR titles, bodies and changes to ods_token_spy_callback.py, plus LiteLLM cache creation/write telemetry. #2035 introduced routed telemetry; #2086 corrects endpoint labels; #2354 and related variants cover worker cleanup; #3755 concerns non-finite durations. #2519 changes cached-input billing in Token Spy's OpenAI provider, not this callback mapping. No existing PR preserves the normalized cache-creation field here.

Contract and validation

Provider completion → LiteLLM's normalized Usage/ModelResponse → async_log_success_event → bounded queue → authenticated metadata-only POST → existing Token Spy ingest counter/storage contract. The pinned source Usage mapping and Anthropic transformation were checked.

• New callback tests exercise the actual success hook, worker and outgoing HTTP request: creation count, explicit zero versus legacy value, legacy compatibility, invalid/bounded counters, unchanged input/output/read counts and absence of prompt/answer text. Three cases failed before the change.
• pytest -q ods/extensions/services/litellm/tests — 11 passed, 1 opt-in skipped in the ordinary Python 3.11 environment.
• ODS_TEST_LITELLM_PIN=1 LITELLM_LOCAL_MODEL_COST_MAP=True pytest -q ods/extensions/services/litellm/tests with litellm==1.90.7 — 12 passed, including a real ModelResponse(usage=Usage(cache_creation_input_tokens=4096, ...)) reaching the HTTP ingest payload.
• pytest -q ods/extensions/services/token-spy/tests/test_routed_telemetry.py — 9 passed.
• Scoped pre-commit and git diff --check passed.

This tests the exact LiteLLM Python release and the outgoing HTTP boundary using MockTransport; it does not call a paid provider, launch the full proxy image, or claim an end-to-end cloud billing test. Shared Python code applies across hosts; executed evidence is Linux/WSL. No stored events are rewritten. Revert restores the prior mapping without migration. Base is public-beta c03e0809, which now includes eight earlier batch fixes merged upstream. Remaining shared baseline gate limitations are documented for the batch.

Follow-up validation

Merged the current beta base in commit 31d60c2. The scoped diff still contains only the LiteLLM callback and its regression test. All 12 pinned-LiteLLM tests passed on this exact candidate. GitHub checks on this head: 32 success, 4 skipped. The earlier unrelated RemoteProvider frontend failure is no longer present; no duplicate UI fix was added (canonical follow-up: #4570).

Batch integration evidence

This exact candidate is included in the 30-PR integration receipt, on code integration commit a38a38a2303724b2b2a3bfba6324fc40a67154ba over public-beta d49ad4bbc760be319de7f4c9b9a93b1f18fe3922. The receipt lists all candidate SHAs, suggested merge order and additive catalog/README reconciliation.

The integration passed 2,858 API tests (3 skipped), 1,069 UI tests, build/lint, schema and four backend Compose plans. The combined live suite passed 52 cases and hit one Uptime Kuma Docker-start timeout; the unchanged Kuma cases passed in a separate rerun on the same commit. The negative cold-start observation remains recorded. Root/WSL fixture failures and the existing macOS fixture PR #3160 are explicitly separated from passing smoke/simulation checks. These receipts do not qualify every hardware platform or a simultaneous service load.

The report distinguishes completed CI from the queued Memos follow-up. Gitea #4909 remains draft for independent migration review.