cmpnd 0.68.0__tar.gz → 0.70.0__tar.gz

{cmpnd-0.68.0 → cmpnd-0.70.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: cmpnd
-Version: 0.68.0
+Version: 0.70.0
 Summary: DSPy observability and deployment SDK for cmpnd
 Project-URL: Homepage, https://cmpnd.ai
 Author-email: cmpnd <hello@cmpnd.ai>

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/_authored_import.py

@@ -263,7 +263,6 @@ def _program_facts(
     lm = packaging._extract_lm_object(program)
     lm_kwargs = getattr(lm, "kwargs", {}) if lm is not None else {}
     facts = ProgramFacts(
-        program_signature_hash=identity.compute_program_hash(program),
         entrypoint_kind=kind,
         module_class=type(program).__name__,
         litellm_model=getattr(lm, "model", None),

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/authored.py

@@ -43,7 +43,6 @@ class SourceCheckError(ValueError):
 class ProgramFacts:
     """Facts derived by importing the staged program without calling forward."""
 
-    program_signature_hash: int | None = None
     entrypoint_kind: str = "instance"
     module_class: str | None = None
     signature_name: str | None = None
@@ -72,11 +71,7 @@ class ProgramFacts:
 
     def deploy_body(self) -> dict[str, Any]:
         """Return fields the deployment create route accepts."""
-        from . import identity
-
         body: dict[str, Any] = {}
-        if self.program_signature_hash is not None:
-            body["program_signature_hash"] = identity.format_signature_hash(self.program_signature_hash)
         for name in (
             "module_class",
             "signature_name",

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/cli/commands/evals.py

@@ -9,7 +9,18 @@ from .. import api_client, deprecation, pagination
 
 
 def _add_list_args(p: ArgumentParser) -> None:
-    p.add_argument("--program", dest="program_name", help="Filter by program name")
+    p.add_argument(
+        "--program",
+        dest="program_name",
+        metavar="NAME",
+        help="Only the runs of the program with this name, current or retired",
+    )
+    p.add_argument(
+        "--module",
+        dest="module_class",
+        metavar="CLASS",
+        help="Only the runs that evaluated this module class, matched exactly (e.g. ReAct)",
+    )
     p.add_argument("--metric", dest="metric_name", help="Filter by metric name")
     p.add_argument("--status", help="Filter by status (running, completed, failed, cancelled)")
     pagination.add_args(p, with_page_aliases=True, default_limit=50)
@@ -63,6 +74,7 @@ def handle_list(client: api_client.ApiClient, args: Any) -> Any:
     body: dict[str, Any] = pagination.wire(args, "page_size")
     for attr, key in [
         ("program_name", "program_name"),
+        ("module_class", "module_class"),
         ("metric_name", "metric"),
         ("status", "status"),
     ]:

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/dataset_sync.py

@@ -15,7 +15,7 @@ from . import _http, configuration, identity, models
 logger = logging.getLogger(__name__)
 
 
-def check_dataset_exists(dataset_hash: str) -> bool:
+def check_dataset_exists(dataset_hash: str, config: configuration.Config | None = None) -> bool:
     """Check if a dataset with this hash already exists on the backend.
 
     Makes a GET request to /api/v1/datasets/check/{dataset_hash}.
@@ -23,14 +23,17 @@ def check_dataset_exists(dataset_hash: str) -> bool:
 
     Args:
         dataset_hash: The SHA256[:16] hash of the dataset.
+        config: The account to ask, which must be the account the run publishes
+            to. None asks the process-global account.
 
     Returns:
         True if the dataset exists, False otherwise (including on errors).
     """
-    try:
-        config = configuration.require_config()
-    except RuntimeError:
-        return False
+    if config is None:
+        try:
+            config = configuration.require_config()
+        except RuntimeError:
+            return False
 
     url = f"{config.endpoint}/api/v1/datasets/check/{dataset_hash}"
 

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/deployment.py

@@ -527,13 +527,10 @@ def deploy_module_under(
     # Package the module
     zip_bytes = packaging.generate_source_zip(module, metric, mcp_tools=pinned_mcp_tools)
 
-    # Compute program identity via the IR reflector on the live module (class
-    # tag included); the signature is kept around only for display metadata.
+    # The deployment names its program by id alone; the signature only describes it.
     sig = packaging._get_module_signature(module)
-    program_signature_hash = identity.compute_program_hash(module) or 0
 
     deploy_body: dict[str, Any] = {
-        "program_signature_hash": identity.format_signature_hash(program_signature_hash),
         # Declare the ZIP's exact length so the server can reject an oversized
         # deploy before the upload.
         "declared_content_length": len(zip_bytes),

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/eval_handler.py

@@ -10,7 +10,7 @@ import logging
 from datetime import datetime, timezone
 from typing import Any
 
-from . import context, exporter, helpers, identity, models
+from . import context, helpers, identity, models
 
 logger = logging.getLogger(__name__)
 
@@ -85,7 +85,7 @@ class EvalHandler:
         logger.debug(f"Extracted program: {program}, type: {type(program) if program else None}")
 
         if program:
-            eval_run.program_name = program.__class__.__name__
+            eval_run.module_class = program.__class__.__name__
 
             # Hash via the IR reflector on the live module so the class tag
             # discriminates correctly; signature_summary is still derived from
@@ -165,16 +165,23 @@ class EvalHandler:
             if not input_fields and devset and hasattr(devset[0], "_input_keys"):
                 input_fields = set(devset[0]._input_keys)
             eval_run._input_fields = input_fields
-            # Fire pre-flight check (synchronous, short timeout)
+            # Fire pre-flight check (synchronous, short timeout). It asks the
+            # account the run publishes to: a client's own, else the global one.
             from . import dataset_sync
 
             try:
-                eval_run._dataset_exists = dataset_sync.check_dataset_exists(eval_run.dataset_hash)
+                eval_run._dataset_exists = dataset_sync.check_dataset_exists(
+                    eval_run.dataset_hash, self._callback._own_config
+                )
             except Exception:
                 eval_run._dataset_exists = None  # Will include examples as fallback
 
+        exp = self._callback._current_exporter()
+        eval_run.program_id = self._program_id(exp)
+
         logger.info(
-            f"EvalRun created: program_name={eval_run.program_name}, "
+            f"EvalRun created: program_id={eval_run.program_id}, "
+            f"module_class={eval_run.module_class}, "
             f"metric_name={eval_run.metric_name}, "
             f"dataset_hash={eval_run.dataset_hash}, "
             f"dataset_size={eval_run.dataset_size}"
@@ -184,7 +191,6 @@ class EvalHandler:
         # the exporter (serve/observe, or deploy in-Lambda) assigns it. Obtain
         # it synchronously now, before any eval trace ships, since every
         # in-flight trace and the dict keys / span below are stamped with it.
-        exp = exporter.get_exporter()
         if exp:
             eval_run.eval_run_id = exp.export_eval_start(eval_run)
 
@@ -201,13 +207,33 @@ class EvalHandler:
 
         # Capture evaluation config in span attributes
         span.attributes["eval.run_id"] = str(eval_run.eval_run_id)
-        span.attributes["eval.program_name"] = eval_run.program_name or ""
+        span.attributes["eval.module_class"] = eval_run.module_class or ""
         span.attributes["eval.metric_name"] = eval_run.metric_name or ""
         span.attributes["eval.dataset_size"] = str(eval_run.dataset_size)
 
         self._callback._call_id_to_span[call_id] = span
         context.push_span(span)
 
+    def _program_id(self, exp: Any) -> str | None:
+        """The program id the eval run carries, resolved the way a trace's is.
+
+        A program name the call declares wins over the callback's program id. The
+        exporter resolves the name now, on the caller's thread, because the run's
+        start is sent synchronously and must carry the id. The exporter keeps the
+        answer for the process, so a name its traces already resolved costs no
+        request. A name the exporter cannot resolve leaves the run without an id.
+        """
+        from . import callback
+
+        name = callback._declared_program_name()
+        if name is None:
+            return self._callback._resolve_program_id()
+        # Read with getattr because the member is optional: an in-memory collector
+        # that resolves no names has none.
+        resolve = getattr(exp, "resolve_program_name", None)
+        program_id = resolve(name) if callable(resolve) else None
+        return program_id if isinstance(program_id, str) else None
+
     def on_evaluate_end(
         self,
         call_id: str,
@@ -228,7 +254,7 @@ class EvalHandler:
                 span.status = models.TraceStatus.OK
             context.pop_span()
             if self._callback._should_stream_span(span):
-                exp = exporter.get_exporter()
+                exp = self._callback._current_exporter()
                 if exp:
                     exp.export_span(span)
 
@@ -373,7 +399,7 @@ class EvalHandler:
                 span.attributes["eval.score"] = str(overall_score)
 
         # Export the eval run
-        exp = exporter.get_exporter()
+        exp = self._callback._current_exporter()
         if exp:
             exp.export_eval_run(eval_run)
 

{cmpnd-0.68.0 → cmpnd-0.70.0}/cmpnd/models.py

@@ -1274,7 +1274,11 @@ class EvalRunBuilder:
     start_time: datetime | None = None
 
     # Eval Identity (defines comparability)
-    program_name: str | None = None
+    # The program the run evaluates: the id serve gave the program name the call
+    # declared, or the client's program id. None when the run names no program.
+    program_id: str | None = None
+    # The evaluated module's class, such as "ReAct".
+    module_class: str | None = None
     metric_name: str | None = None
     dataset_hash: str | None = None
     dataset_size: int = 0
@@ -1393,7 +1397,10 @@ class EvalRunBuilder:
         return {
             "eval_run_id": "new",
             "status": EvalRunStatus.RUNNING.value,
-            "program_name": self.program_name or "",
+            "program_id": self.program_id,
+            "module_class": self.module_class or "",
+            # Also under the key a serve from before module_class reads.
+            "program_name": self.module_class or "",
             "metric_name": self.metric_name or "",
             "dataset_hash": self.dataset_hash or "",
             "dataset_size": self.dataset_size,
@@ -1411,7 +1418,10 @@ class EvalRunBuilder:
         """Convert to API-compatible dictionary."""
         d = {
             "eval_run_id": str(self.eval_run_id) if self.eval_run_id else None,
-            "program_name": self.program_name or "",
+            "program_id": self.program_id,
+            "module_class": self.module_class or "",
+            # Also under the key a serve from before module_class reads.
+            "program_name": self.module_class or "",
             "metric_name": self.metric_name or "",
             "dataset_hash": self.dataset_hash or "",
             "dataset_size": self.dataset_size,

{cmpnd-0.68.0 → cmpnd-0.70.0}/docs/cli.md

@@ -458,6 +458,18 @@ program signature hash. `cmpnd programs optimized` commands and
 optimization result, not a program. Each group's `--help` names the column its
 IDs come from.
 
+`cmpnd evals list [--program NAME] [--module CLASS] [--metric] [--status]`
+lists eval runs, newest first. `--program` takes the program's name, current
+or retired, and a name no program answers to is an error that names it.
+`--module` takes the evaluated module's class, such as `ReAct`, and matches it
+exactly. The CLI sends them as `program_name` and `module_class`. Each eval
+run in `evals list`, `evals show` and `programs evals` names what it evaluated
+in three fields, as a trace does. `program_name` is the program's name, empty
+when the program has none or the run names no program. `program_id` is the
+program's `p_` id, and `module_class` is the evaluated module's class. A server
+that predates `module_class` sends the class in `program_name` and filters
+`--program` by class.
+
 `cmpnd deployments openapi <id>` prints the OpenAPI document a deployment
 serves, which an agent reads before it calls the deployment.
 

{cmpnd-0.68.0 → cmpnd-0.70.0}/docs/sdk-instrumentation.md

@@ -727,6 +727,8 @@ What it captures beyond the normal span flow:
 - An `EvalRunBuilder` (constructed at `eval_handler.py:79`) that groups all per-example traces under one eval ID.
 - The metric name (via `get_metric_name` from `python-sdk/cmpnd/helpers.py`) and inferred provider.
 - Links to the enclosing optimization run when one exists (`eval_handler.py:85`), so eval runs triggered inside `GEPA.compile` roll up under the optimization in the UI.
+- The program the run evaluates, as `program_id`. A program name the call declares wins: the exporter resolves it with `resolve_program_name`, the same once-per-process answer its traces use, on the caller's thread because the eval start is sent synchronously. Without a declared name, the run takes the callback's program id. A name serve refuses leaves the run without an id. The evaluated module's class goes as `module_class`, and again under `program_name`, the key a serve from before `module_class` reads.
+- The exporter it publishes through: the callback's own (`_current_exporter`), so a client's eval runs go to the client's account.
 - The dataset — on `on_evaluate_start` it triggers the `dataset_sync` pre-flight described above.
 
 Per-example results (scores, statuses, example hashes) are collected in `on_evaluate_end` and shipped as part of the `EvalRunBuilder`'s `to_api_dict()`, which the exporter routes to `/api/v1/eval_runs`.
@@ -803,7 +805,9 @@ name form. A program name wins over the config's id. `configure()`
 checks the id's shape and raises `ValueError` on a bad one, and a `Config` built directly
 is checked the same way.
 
-The id rides every body that can declare a program, beside `program_signature_hash`:
+The id rides every body that can declare a program. A deployment create names its program
+by the id alone and sends no `program_signature_hash`. The other bodies still carry the
+hash beside the id:
 
 | Body | Where | Built by |
 |---|---|---|

{cmpnd-0.68.0 → cmpnd-0.70.0}/pyproject.toml

@@ -5,7 +5,7 @@ name = "cmpnd"
 # it's new) — no git tags. Bump with `uv version --bump {minor,patch}` or by
 # hand. Runtime `cmpnd.__version__` reads installed metadata, so it tracks
 # this automatically. See docs/plans/2026-07-02-sdk-pypi-tag-free-publish.md.
-version = "0.68.0"
+version = "0.70.0"
 description = "DSPy observability and deployment SDK for cmpnd"
 readme = "README.md"
 requires-python = ">=3.11"

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/e2e/test_deploy_programs.py

@@ -219,7 +219,6 @@ class TestDeployBadImportRejected:
             create_resp = client.post(
                 f"{config.endpoint}/api/v1/deployments",
                 json={
-                    "program_signature_hash": 0,
                     "declared_content_length": len(zip_bytes),
                     "program_id": program_id_for(f"test_deploy_programs-{uuid.uuid4().hex[:8]}"),
                     # A worker that spends egress tokens needs the model a token names, so

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/e2e/test_optimize_loop.py

@@ -32,7 +32,6 @@ import dspy
 import pytest
 
 import cmpnd
-from tests import _identity_helpers
 from tests.e2e import committee_task, fake_lm
 from tests.e2e.helpers import named_program, platform_client, poll_until, program_id_for
 from tests.e2e.target_capabilities import needs_an_optimizer
@@ -190,7 +189,6 @@ class TestOptimizeGepaDeployDetail:
 
     def test_gepa_deploy_run_stamps_traces_and_lands_detail(self, deployed_source, program_id, server_client):
         program, trainset, valset, deployed = deployed_source
-        sig_hash = _identity_helpers.program_hash_for_sig(program.signature)
         before = time.time()
 
         platform_client().optimize_deployment(
@@ -210,10 +208,10 @@ class TestOptimizeGepaDeployDetail:
 
         # Correlate to THIS run through the optimization index, not by scanning
         # the trace stream: `deployed_source` is module-scoped, so every other test
-        # in this file emits traces under the SAME program_signature_hash. This run
-        # is the only one of this optimizer + signature started after `before`.
-        # (The optimization search is not filtered by program_signature_hash, so
-        # match it on the returned row, which surfaces it.)
+        # in this file emits traces under the SAME program. This run is the only
+        # one of this optimizer + program started after `before`. A run serve
+        # starts names its deployment's program_id. The optimization search does
+        # not filter by program, so match it on the returned row, which carries it.
         #
         # Poll, don't sleep-once: on serve the deploy run's detail + traces land
         # via a background completion watcher (5s poll cadence) AFTER optimize_deployment()
@@ -228,11 +226,11 @@ class TestOptimizeGepaDeployDetail:
                 order_direction="desc",
                 per_page=50,
             ).get("optimization_runs", [])
-            return [r for r in runs if str(r.get("program_signature_hash")) == str(sig_hash)]
+            return [r for r in runs if r.get("program_id") == program_id]
 
         mine = poll_until(_find_run, timeout=60.0, interval=3.0)
         assert mine, (
-            f"no GEPA optimization run for program_signature_hash={sig_hash} after {start_from} "
+            f"no GEPA optimization run for program_id={program_id} after {start_from} "
             "(serve completion watcher didn't land it within 60s)"
         )
         run_id = mine[0]["run_id"]
@@ -254,9 +252,7 @@ class TestOptimizeGepaDeployDetail:
         assert candidates, f"GEPA deploy run {run_id} landed a header but no candidate detail (CMP-323)"
 
         # The worker's tracker stamps the run id onto the run's traces (the collector
-        # stamps only deployment_id), and serve carries the int64
-        # program_signature_hash through the poll without float-rounding it, so the
-        # run row matches this signature. Its traces land under the deployment's program.
+        # stamps only deployment_id), and the traces land under the deployment's program.
         # Page wide (page_size=500, the API max) so this run's few traces aren't crowded
         # out of the default window by the other tests' runs — module-scoped
         # `deployed_source` means they all share this program —

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/test_authored.py

@@ -142,8 +142,9 @@ def test_bundle_pins_tools_without_importing_the_program_twice(monkeypatch):
     check.assert_called_once()
 
 
-def test_deploy_body_omits_unknown_hash_and_trace_limits():
+def test_deploy_body_omits_trace_limits_and_sends_no_hash():
     assert authored.ProgramFacts.minimal().deploy_body() == {}
+    assert "program_signature_hash" not in authored.ProgramFacts(module_class="Predict").deploy_body()
 
 
 def test_deploy_body_preserves_empty_signature_field_lists():

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/test_cli/test_evals.py

@@ -86,20 +86,34 @@ class TestEvalsList:
     # own spelling passes while the filter does nothing.
     def test_minimal_list(self, mock_client):
         mock_client.post.return_value = {"eval_runs": []}
-        args = argparse.Namespace(program_name=None, metric_name=None, status=None, limit=50, offset=0)
+        args = argparse.Namespace(
+            program_name=None, module_class=None, metric_name=None, status=None, limit=50, offset=0
+        )
         handle_list(mock_client, args)
         body = mock_client.post.call_args[1]["json"]
         assert body == {"page": 1, "page_size": 50}
 
     def test_with_filters(self, mock_client):
         mock_client.post.return_value = {"eval_runs": []}
-        args = argparse.Namespace(program_name="RAG", metric_name="accuracy", status="completed", limit=50, offset=0)
+        args = argparse.Namespace(
+            program_name="RAG", module_class=None, metric_name="accuracy", status="completed", limit=50, offset=0
+        )
         handle_list(mock_client, args)
         body = mock_client.post.call_args[1]["json"]
         assert body["program_name"] == "RAG"
         assert body["metric"] == "accuracy"
         assert body["status"] == "completed"
 
+    def test_program_names_the_program_and_module_names_the_class(self, mock_client):
+        # --program is the program's name; --module is the evaluated module's
+        # class, which the server matches exactly under module_class.
+        mock_client.post.return_value = {"eval_runs": []}
+        args = _build_parser().parse_args(["evals", "list", "--program", "capital-qa", "--module", "ReAct"])
+        handle_list(mock_client, args)
+        body = mock_client.post.call_args[1]["json"]
+        assert body["program_name"] == "capital-qa"
+        assert body["module_class"] == "ReAct"
+
     def test_deprecated_search_verb_dispatches_and_is_marked(self, mock_client):
         # The warning is the dispatcher's job (see test_main.py); the wiring is this one's.
         parser = _build_parser()

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/test_deploy.py

@@ -414,7 +414,8 @@ class TestDeploy:
         create_call = client.post.call_args_list[0]
         assert "ct_test_key" in str(create_call)
 
-    def test_sends_program_signature_hash(self, mock_httpx):
+    def test_sends_no_program_signature_hash(self, mock_httpx):
+        # A deployment names its program by id alone.
         client = self._setup_mocks(mock_httpx)
 
         with mock.patch.object(deployment, "time") as mock_time:
@@ -425,9 +426,7 @@ class TestDeploy:
 
         create_call = client.post.call_args_list[0]
         body = create_call.kwargs.get("json", {})
-        assert "program_signature_hash" in body
-        # CMP-331: carried as a quoted decimal string on the wire, not a number.
-        assert isinstance(body["program_signature_hash"], str)
+        assert "program_signature_hash" not in body
 
     def test_sends_the_programs_signature(self, mock_httpx):
         # A program's signature is known the moment the module is packaged, and the

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/test_deploy_program_id.py

@@ -108,7 +108,7 @@ class TestDeployBody:
             lambda: deployment.deploy_module_under(dspy.Predict("question -> answer"), None, 300, None, None, config)
         )
         assert "program_id" not in body
-        assert "program_signature_hash" in body
+        assert "program_signature_hash" not in body
 
     def test_an_undeclared_bundle_deploy_omits_the_key(self):
         _configure()

{cmpnd-0.68.0 → cmpnd-0.70.0}/tests/test_eval.py

@@ -199,7 +199,7 @@ class TestEvalLifecycle:
         eval_run = exporter.eval_runs[0]
 
         # Program metadata
-        assert eval_run.program_name == "TestProgram"
+        assert eval_run.module_class == "TestProgram"
         assert eval_run.metric_name == "exact_match"
         assert eval_run.dataset_size == 3