jerry/buzz
1
0
Fork 0
mirror of https://github.com/block/buzz.git synced 2026-10-31 14:31:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: preserve DeepSeek max effort on Databricks

Browse Source
This commit is contained in:
Atish Patel
2026-08-18 13:31:24 -05:00
parent ebb0730929
commit a075619f47
6 changed files with 47 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
scripts/model-capabilities.json
View File

@@ -392,12 +392,32 @@
"normalization_policy": "none"
}
],
"_comment_databricks_v2_known_models": "Curated fallback catalog for Databricks v2: the model IDs offered when an authenticated api/ai-gateway/v2/endpoints call succeeds but returns an empty list (see catalog.rs authenticated_empty_v2_catalog). Both IDs also have full exact_records, so once offered they classify precisely. Uniqueness enforced by validate_manifest().",
"_comment_databricks_v2_known_models": "Curated fallback catalog for Databricks v2: the model IDs offered when an authenticated api/ai-gateway/v2/endpoints call succeeds but returns an empty list (see catalog.rs authenticated_empty_v2_catalog). Every ID also has a full exact_record, so once offered it classifies precisely. Uniqueness enforced by validate_manifest().",
"databricks_v2_known_models": [
"databricks-gpt-5-5",
"databricks-claude-opus-4-7"
"databricks-claude-opus-4-7",
"databricks-deepseek-v4-flash-0731"
],
"exact_records": [
{
"provider": "databricks_v2",
"raw_model_id": "databricks-deepseek-v4-flash-0731",
"registry_label": "DeepSeek V4 Flash 0731",
"thinking_mode": "none",
"supported_efforts": [
"low",
"high",
"max"
],
"default_effort": "high",
"databricks_v2_wire_route": "mlflow-chat",
"normalization_policy": "openai-standard",
"_provenance": "live Databricks staging endpoint validation plus provider-advertised effort set",
"source": "Direct staging invocation with reasoning_effort=max returned HTTP 200 and reasoning content on 2026-08-18; model metadata advertises low|high|max",
"_reconciliation": "adopt",
"_reconciliation_note": "The concrete-unknown DatabricksV2 fallback clamps max to xhigh. This endpoint accepts max, so an exact record is required to preserve the requested leaderboard condition.",
"_reconciliation_doc": "Live POST /serving-endpoints/databricks-deepseek-v4-flash-0731/invocations on 2026-08-18 with reasoning_effort=max"
},
{
"provider": "databricks_v2",
"raw_model_id": "databricks-gpt-5-4-mini",

18
scripts/normative-corpus.json
View File

@@ -1312,6 +1312,24 @@
"registry_label": "GPT-5.6 Luna"
}
},
{
"id": "dbv2-deepseek-v4-flash-max-probe",
"provider": "databricks_v2",
"raw_model_id": "databricks-deepseek-v4-flash-0731",
"_note": "Probes that the validated DeepSeek endpoint preserves max instead of inheriting the concrete-unknown max-to-xhigh clamp.",
"expect": {
"thinking_mode": "none",
"supported_efforts": [
"low",
"high",
"max"
],
"default_effort": "high",
"databricks_v2_wire_route": "mlflow-chat",
"normalization_policy": "openai-standard",
"registry_label": "DeepSeek V4 Flash 0731"
}
},
{
"id": "dbv2-gpt-5-6-terra-exact-record-probe",
"provider": "databricks_v2",

2
scripts/run-tests.sh
View File

@@ -115,7 +115,7 @@ run_unit_tests() {
# buzz-agent model-capabilities corpus: the Rust half of the cross-language
# drift guard. model_capabilities.rs embeds scripts/model-capabilities.json +
# scripts/normative-corpus.json via include_str! and replays all 103 vectors
# scripts/normative-corpus.json via include_str! and replays all 104 vectors
# as pure in-process tests (no infra). Mirrors the nextest path in
# `just test-unit` — the two lists must stay in step.
run_test_step "buzz-agent unit tests" \