dh-cli 0.17.0__tar.gz → 0.17.2__tar.gz

{dh_cli-0.17.0 → dh_cli-0.17.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: dh-cli
-Version: 0.17.0
+Version: 0.17.2
 Summary: Dayhoff Labs developer CLI
 Author-email: Dayhoff Labs <dev@dayhofflabs.com>
 License: # PolyForm Noncommercial License 1.0.0

{dh_cli-0.17.0 → dh_cli-0.17.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "dh-cli"
-version = "0.17.0"
+version = "0.17.2"
 description = "Dayhoff Labs developer CLI"
 requires-python = ">=3.11"
 readme = "README.md"

{dh_cli-0.17.0 → dh_cli-0.17.2}/src/dh_cli/bedrock/commands.py

@@ -389,8 +389,9 @@ def bedrock_cost(
       dh bedrock cost --format csv | csvlook         # further analysis
 
     Exit codes:
-      0   success (within 10% of Cost Explorer, or --no-reconcile)
-      1   unknown model ID, bad pricing file, or S3 read error
+      0   success (within 10% of Cost Explorer, or --no-reconcile); unpriced
+          models are warned about on stderr and counted at $0
+      1   bad pricing file or S3 read error
       2   reconciliation drift > 10% AND absolute gap >= $1
     """
     # Defaults use UTC dates to match the S3 day-partition layout
@@ -445,13 +446,17 @@ def bedrock_cost(
                 if cr.classify_arn(rec.get("identity", {}).get("arn", "")).principal_name == my_handle
             )
         report = cr.build_report(records, pricing, group_by=group_by)
-    except cr.UnknownModel as exc:
-        click.echo(f"error: {exc}", err=True)
-        sys.exit(1)
     except Exception as exc:
         click.echo(f"error: failed to process logs: {exc}", err=True)
         sys.exit(1)
 
+    for model_id, calls in sorted(report.unpriced.items(), key=lambda kv: -kv[1]):
+        click.echo(
+            f"warning: no pricing for {model_id} ({calls} call{'s' if calls != 1 else ''}); "
+            "counted at $0 under an UNPRICED: row. Add it to pricing.yaml.",
+            err=True,
+        )
+
     # Header is only useful for humans — skip in csv mode so the output
     # is a clean csv stream that can be piped.
     if output_format != "csv":

{dh_cli-0.17.0 → dh_cli-0.17.2}/src/dh_cli/bedrock/cost_report.py

@@ -61,6 +61,8 @@ class ReportRow:
 @dataclass
 class Report:
     rows: list[ReportRow] = field(default_factory=list)
+    # modelId -> call count for records whose model has no pricing entry.
+    unpriced: dict[str, int] = field(default_factory=dict)
 
 
 @dataclass(frozen=True)
@@ -115,6 +117,12 @@ _BASE_MODELS = (
     "haiku-4-5",
     "fable-5",
     "fable-5-1",
+    "gpt-6-astra",
+    "gpt-6-sol",
+    "gpt-6-luna",
+    "gpt-5.6-sol",
+    "gpt-5.6-terra",
+    "gpt-5.6-luna",
 )
 
 
@@ -185,6 +193,7 @@ def build_report(
     if group_by not in _VALID_GROUP_BY:
         raise ValueError(f"group_by={group_by!r} is not one of {sorted(_VALID_GROUP_BY)}")
     agg: dict[tuple, dict[str, Any]] = {}
+    unpriced: dict[str, int] = {}
     for rec in records:
         # Bedrock emits records for failed validations / throttles with
         # `errorCode` set and no `input` / `output` fields. No tokens
@@ -193,8 +202,15 @@ def build_report(
             continue
         arn = rec.get("identity", {}).get("arn", "")
         principal = classify_arn(arn)
-        base = resolve_base_model(rec["modelId"])
-        rec_cost = cost_of(rec, pricing)
+        # An unpriced model gets its own $0 `UNPRICED:` row and is counted in
+        # `Report.unpriced`, rather than aborting the whole report.
+        try:
+            base = resolve_base_model(rec["modelId"])
+            rec_cost = cost_of(rec, pricing)
+        except UnknownModel:
+            base = f"UNPRICED:{rec['modelId']}"
+            rec_cost = 0.0
+            unpriced[rec["modelId"]] = unpriced.get(rec["modelId"], 0) + 1
         if group_by == "user+model":
             key = (principal.principal_type, principal.principal_name, base)
         elif group_by == "user":
@@ -236,7 +252,7 @@ def build_report(
         for (ptype, pname, model), s in agg.items()
     ]
     rows.sort(key=lambda r: r.estimated_cost_usd, reverse=True)
-    return Report(rows=rows)
+    return Report(rows=rows, unpriced=unpriced)
 
 
 _COLUMNS = (

{dh_cli-0.17.0 → dh_cli-0.17.2}/src/dh_cli/bedrock/pricing.yaml

@@ -57,3 +57,38 @@ models:
     output_per_1m: 4.00
     cache_write_per_1m: 1.00
     cache_read_per_1m: 0.08
+  # OpenAI models reached via `bedrock-runtime` (these write invocation logs;
+  # mantle traffic does not). Global-profile rates as billed in Cost Explorer,
+  # Sep 2026; `us.`/in-region calls bill ~10% higher, same as Anthropic.
+  gpt-6-astra:
+    input_per_1m: 10.00
+    output_per_1m: 50.00
+    cache_write_per_1m: 12.50
+    cache_read_per_1m: 1.00
+  # GPT-6 Sol/Luna: Marketplace rate card `*_global_standard`, 2026-09-25;
+  # cache write is the 30-minute tier.
+  gpt-6-sol:
+    input_per_1m: 2.00
+    output_per_1m: 10.00
+    cache_write_per_1m: 2.50
+    cache_read_per_1m: 0.20
+  gpt-6-luna:
+    input_per_1m: 0.10
+    output_per_1m: 0.50
+    cache_write_per_1m: 0.125
+    cache_read_per_1m: 0.01
+  gpt-5.6-sol:
+    input_per_1m: 4.00
+    output_per_1m: 20.00
+    cache_write_per_1m: 5.00
+    cache_read_per_1m: 0.40
+  gpt-5.6-terra:
+    input_per_1m: 2.00
+    output_per_1m: 12.00
+    cache_write_per_1m: 2.50
+    cache_read_per_1m: 0.20
+  gpt-5.6-luna:
+    input_per_1m: 0.20
+    output_per_1m: 1.20
+    cache_write_per_1m: 0.25
+    cache_read_per_1m: 0.02

{dh_cli-0.17.0 → dh_cli-0.17.2}/tests/bedrock/test_build_report.py

@@ -100,21 +100,29 @@ def test_aggregation_per_principal_and_model(pricing):
     assert haiku_row.input_tokens == 1000
 
 
-def test_missing_pricing_raises(pricing):
-    """An unrecognized modelId surfaces as UnknownModel, not silent 0."""
-    from dh_cli.bedrock.cost_report import UnknownModel, build_report
+def test_unknown_model_is_reported_as_unpriced_row(load_fixture, pricing):
+    """An unrecognized modelId gets its own visibly-unpriced row.
+
+    Contract change (2026-09-25): this used to raise UnknownModel, which let
+    one unpriced model abort the whole report — four times, most recently a
+    handful of `us.openai.gpt-5.6-luna` calls blanking out ~$40k/month of
+    attribution. The unknown model's tokens stay visible under an
+    `UNPRICED:` label and the rest of the report is still priced; it is
+    never folded into a real model's row.
+    """
+    from dh_cli.bedrock.cost_report import build_report
 
-    records = [
-        {
-            "modelId": "global.anthropic.claude-someshinynew-99-v42",
-            "input": {"inputTokenCount": 10, "cacheReadInputTokenCount": 0, "cacheWriteInputTokenCount": 0},
-            "output": {"outputTokenCount": 10},
-            "identity": {"arn": "arn:aws:iam::074735440724:user/dev-cursor-dma"},
-        }
-    ]
+    records = [load_fixture("G_unknown_model.json"), load_fixture("A_cache_write.json")]
 
-    with pytest.raises(UnknownModel):
-        build_report(records, pricing)
+    report = build_report(records, pricing, group_by="model")
+    rows = {r.model: r for r in report.rows}
+
+    unpriced = rows["UNPRICED:global.anthropic.claude-someshinynew-99-v42"]
+    assert unpriced.estimated_cost_usd == 0.0
+    assert unpriced.invocations == 1
+    assert unpriced.input_tokens == 10
+    assert unpriced.output_tokens == 10
+    assert rows["opus-4-7"].estimated_cost_usd == pytest.approx(0.0255)
 
 
 def test_error_records_are_skipped(load_fixture, pricing):

{dh_cli-0.17.0 → dh_cli-0.17.2}/tests/bedrock/test_cli_exit_codes.py

@@ -48,24 +48,20 @@ def test_cost_exit_0_on_success(load_fixture):
     assert result.exit_code == 0, result.output
 
 
-def test_cost_exit_1_on_unknown_model(load_fixture):
-    with (
-        patch("dh_cli.bedrock.cost_report.walk_logs") as walk,
-        patch(
-            "dh_cli.bedrock.cost_report.fetch_cost_explorer_total",
-            return_value=0.0,
-        ),
-    ):
-        walk.return_value = iter([load_fixture("G_unknown_model.json")])
-        result = _invoke(
-            [
-                "--start",
-                "2026-04-17",
-                "--end",
-                "2026-04-17",
-            ]
-        )
-    assert result.exit_code == 1, result.output
+def test_cost_unknown_model_warns_and_still_reports(load_fixture):
+    """An unpriced model must not abort the report (it used to exit 1).
+
+    The known-model rows still print, and stderr names the unpriced model
+    and its call count so the gap can't pass unnoticed. Any dollar gap it
+    causes is still caught by the Cost Explorer reconcile (exit 2).
+    """
+    with patch("dh_cli.bedrock.cost_report.walk_logs") as walk:
+        walk.return_value = iter([load_fixture("G_unknown_model.json"), load_fixture("A_cache_write.json")])
+        result = _invoke(["--start", "2026-04-17", "--end", "2026-04-17", "--no-reconcile", "--group-by", "model"])
+    assert result.exit_code == 0, result.output
+    assert "opus-4-7" in result.stdout
+    assert "global.anthropic.claude-someshinynew-99-v42" in result.stderr
+    assert "1 call" in result.stderr
 
 
 def test_cost_exit_2_on_drift(load_fixture):

{dh_cli-0.17.0 → dh_cli-0.17.2}/tests/bedrock/test_resolve_base_model.py

@@ -52,6 +52,29 @@ import pytest
         ("anthropic.claude-opus-5-5", "opus-5-5"),
         ("us.anthropic.claude-opus-5-5", "opus-5-5"),
         ("global.anthropic.claude-opus-5-5", "opus-5-5"),
+        # OpenAI models on `bedrock-runtime` do write invocation logs (only
+        # mantle traffic doesn't). Seen in the dev bucket 2026-09, in both the
+        # bare and inference-profile-ARN shapes.
+        ("us.openai.gpt-5.6-luna", "gpt-5.6-luna"),
+        (
+            "arn:aws:bedrock:us-east-1:074735440724:inference-profile/global.openai.gpt-5.6-luna",
+            "gpt-5.6-luna",
+        ),
+        ("openai.gpt-5.6-sol", "gpt-5.6-sol"),
+        ("global.openai.gpt-5.6-terra", "gpt-5.6-terra"),
+        ("global.openai.gpt-6-astra", "gpt-6-astra"),
+        (
+            "arn:aws:bedrock:us-east-1:074735440724:inference-profile/global.openai.gpt-6-astra",
+            "gpt-6-astra",
+        ),
+        ("global.openai.gpt-6-sol", "gpt-6-sol"),
+        ("us.openai.gpt-6-sol", "gpt-6-sol"),
+        ("openai.gpt-6-sol", "gpt-6-sol"),
+        ("global.openai.gpt-6-luna", "gpt-6-luna"),
+        (
+            "arn:aws:bedrock:us-east-1:074735440724:inference-profile/global.openai.gpt-6-luna",
+            "gpt-6-luna",
+        ),
     ],
 )
 def test_resolve_base_model_known(model_id: str, expected_base: str):
@@ -117,6 +140,30 @@ def test_opus_55_is_priced_below_opus_5():
         assert pricing["opus-5-5"][rate] < pricing["opus-5"][rate]
 
 
+@pytest.mark.parametrize(
+    "base, rates",
+    [
+        # Global-profile rates as billed in Cost Explorer, Sep 2026
+        # (`USE1-MP:USE1_*_global_standard` unit cost). Sol's regional rate
+        # fell from $5.50/$33 in Aug to $4.40/$22 in Sep; Sep is current.
+        ("gpt-6-astra", (10.00, 50.00, 12.50, 1.00)),
+        # GPT-6 Sol/Luna: Marketplace rate card, `*_global_standard`,
+        # 2026-09-25 (cache write is the 30-minute tier).
+        ("gpt-6-sol", (2.00, 10.00, 2.50, 0.20)),
+        ("gpt-6-luna", (0.10, 0.50, 0.125, 0.01)),
+        ("gpt-5.6-sol", (4.00, 20.00, 5.00, 0.40)),
+        ("gpt-5.6-terra", (2.00, 12.00, 2.50, 0.20)),
+        ("gpt-5.6-luna", (0.20, 1.20, 0.25, 0.02)),
+    ],
+)
+def test_openai_runtime_models_are_priced(base, rates):
+    from dh_cli.bedrock.cost_report import default_pricing_path, load_pricing
+
+    pricing = load_pricing(default_pricing_path())
+    keys = ("input_per_1m", "output_per_1m", "cache_write_per_1m", "cache_read_per_1m")
+    assert tuple(pricing[base][k] for k in keys) == rates
+
+
 def test_resolve_base_model_unknown_raises():
     from dh_cli.bedrock.cost_report import UnknownModel, resolve_base_model