Move req_llm to 1.24.0 and update the other outdated dependencies

.dialyzer_ignore.exs

@@ -57,8 +57,8 @@
   # defensive clause/guard: ReqLLM.Response types `usage` as map() on the
   # struct, but its schema defaults the field to nil and Response.usage/1 is
   # `map() | nil`, so the nil clause is reachable at runtime.
-  {"lib/imp/clients/req_llm.ex", :guard_fail, 1357},
-  {"lib/imp/clients/req_llm.ex", :pattern_match_cov, {1385, 8}},
+  {"lib/imp/clients/req_llm.ex", :guard_fail, 1414},
+  {"lib/imp/clients/req_llm.ex", :pattern_match_cov, {1442, 8}},
   {"lib/imp/clients/req_llm.ex", :pattern_match_cov, {147, 7}},
   # defensive error clause on an always-ok internal call
   {"lib/imp/clients/training.ex", :pattern_match, {1215, 13}},
@@ -116,9 +116,6 @@
   {"lib/mix/tasks/imp.benchmark.parity.ex", :pattern_match, {715, 11}},
   # raise-only helper: invalid_snapshot!/2 exists to Mix.raise
   {"lib/mix/tasks/imp.public_api.ex", :no_return, {838, 8}},
-  # dependency code: this file ships inside the req_llm package, not this
-  # repo; LLMDB.Model.t/0 is a Zoi-generated spec dialyzer cannot see
-  {"lib/req_llm.ex", :unknown_type, {109, 24}},
 
   # --- Surfaced 2026-08-08 during ignore-file regeneration (imp-fkwy): 583
   # commits landed without CI, and these warnings accumulated unpinned.
@@ -221,7 +218,7 @@
 
   # Defensive fallbacks and MapSet opacity retained at the 0.3 cut. These are
   # individually pinned so a changed success type makes the gate ask again.
-  {"bench/imp/benchmark_truth/multimodal_runner.ex", :pattern_match_cov, {341, 16}},
+  {"bench/imp/benchmark_truth/multimodal_runner.ex", :pattern_match_cov, {340, 16}},
   {"lib/imp/adapter/chat.ex", :pattern_match_cov, {774, 8}},
   {"lib/imp/adapter/xml.ex", :pattern_match_cov, {675, 8}},
   {"lib/imp/mcp.ex", :pattern_match_cov, {372, 8}},

bench/imp/benchmark_truth/multimodal_manifest.ex

@@ -2,25 +2,16 @@ defmodule Imp.BenchmarkTruth.MultimodalManifest do
   @moduledoc false
 
   @payload_keys ~w(assets campaign_id claim_policy created_at limitations provider samples schema_version scoring signature)
-  @provider_keys ~w(api capabilities credential_env endpoint generation identity_evidence model name pricing profile req_llm_dependency req_llm_model)
+  @provider_keys ~w(api capabilities credential_env endpoint generation identity_evidence model name pricing profile req_llm_model)
   @required_generation_keys ~w(max_tokens timeout_ms)
   @optional_generation_keys ~w(seed temperature top_p)
   @pricing_keys ~w(as_of cached_input_nano_usd_per_token cached_input_usd_per_1m currency input_nano_usd_per_token input_usd_per_1m output_nano_usd_per_token output_usd_per_1m pricing_basis source)
-  @req_llm_dependency_keys ~w(package package_sha256 repository source source_revision version)
   @signature_keys ~w(input output output_schema prompt_contract)
   @scoring_keys ~w(family_thresholds normalization scorer)
   @claim_keys ~w(document_family image_family required_families)
   @asset_keys ~w(bytes mime_type path sha256)
   @sample_keys ~w(asset_ids delivery expected_capability family gold id prompt)
   @deliveries ~w(typed_image_data_uri typed_native_file)
-  @req_llm_dependency %{
-    "package" => "req_llm",
-    "package_sha256" => "266c0e06c47b4562f243dcdf41332342cbed2ec37064750edd725fb66bb6e914",
-    "repository" => "https://github.com/agentjido/req_llm",
-    "source" => "hexpm",
-    "source_revision" => "33840077c2f1332eb6dff2d268dff02393014da4",
-    "version" => "1.17.1"
-  }
   @provider_profiles %{
     "google-gemini-2.5-flash-generate-content" => %{
       "api" => "generateContent",
@@ -43,7 +34,6 @@ defmodule Imp.BenchmarkTruth.MultimodalManifest do
         "pricing_basis" => "standard_text_image_video",
         "source" => "https://ai.google.dev/gemini-api/docs/pricing"
       },
-      "req_llm_dependency" => @req_llm_dependency,
       "req_llm_model" => "google:gemini-2.5-flash"
     },
     "openai-gpt-4.1-mini-2025-04-14-responses" => %{
@@ -66,7 +56,6 @@ defmodule Imp.BenchmarkTruth.MultimodalManifest do
         "pricing_basis" => "standard",
         "source" => "https://developers.openai.com/api/docs/models/gpt-4.1-mini"
       },
-      "req_llm_dependency" => @req_llm_dependency,
       "req_llm_model" => "openai:gpt-4.1-mini-2025-04-14"
     }
   }
@@ -93,7 +82,7 @@ defmodule Imp.BenchmarkTruth.MultimodalManifest do
     root = Keyword.get(opts, :root, File.cwd!()) |> Path.expand()
     exact_keys!(payload, @payload_keys, "manifest payload")
 
-    require_equal!(payload["schema_version"], 2, "schema_version")
+    require_equal!(payload["schema_version"], 3, "schema_version")
     require_string!(payload["campaign_id"], "campaign_id")
     require_string!(payload["created_at"], "created_at")
 
@@ -144,17 +133,44 @@ defmodule Imp.BenchmarkTruth.MultimodalManifest do
     end)
   end
 
-  def runtime_dependency!(payload) when is_map(payload) do
-    dependency = payload["provider"]["req_llm_dependency"]
+  @doc """
+  Binds the loaded ReqLLM package into the manifest's provider, so the
+  checkpoint identity, request audits and artifact record the dependency that
+  actually serializes and sends each request. A checkpoint written under one
+  ReqLLM package cannot resume under another.
+  """
+  def bind_runtime_dependency!(%{payload: payload} = manifest) do
+    %{
+      manifest
+      | payload: put_in(payload, ["provider", "req_llm_dependency"], runtime_dependency!())
+    }
+  end
 
-    runtime_version =
+  @doc """
+  The loaded ReqLLM package: its version and the Hex package checksum from the
+  project's lock. Raises when nothing is loaded or when the loaded version is
+  not the locked one.
+  """
+  def runtime_dependency! do
+    loaded =
       case Application.spec(:req_llm, :vsn) do
         nil -> raise ArgumentError, "ReqLLM runtime dependency is not loaded"
         version -> to_string(version)
       end
 
-    require_equal!(runtime_version, dependency["version"], "ReqLLM runtime version")
-    dependency
+    case Mix.Dep.Lock.read()[:req_llm] do
+      {:hex, :req_llm, ^loaded, package_sha256, _managers, _deps, "hexpm", _outer} ->
+        %{
+          "package" => "req_llm",
+          "package_sha256" => package_sha256,
+          "source" => "hexpm",
+          "version" => loaded
+        }
+
+      locked ->
+        raise ArgumentError,
+              "loaded ReqLLM #{loaded} is not the Hex package locked in mix.lock: #{inspect(locked)}"
+    end
   end
 
   def profiles, do: Map.keys(@provider_profiles) |> Enum.sort()
@@ -217,9 +233,6 @@ defmodule Imp.BenchmarkTruth.MultimodalManifest do
         require_string!(pricing[key], "pricing.#{key}")
       end
     )
-
-    exact_keys!(provider["req_llm_dependency"], @req_llm_dependency_keys, "req_llm_dependency")
-    require_equal!(provider["req_llm_dependency"], @req_llm_dependency, "req_llm_dependency")
   end
 
   defp validate_signature!(signature) do

bench/imp/benchmark_truth/multimodal_runner.ex

@@ -15,8 +15,7 @@ defmodule Imp.BenchmarkTruth.MultimodalRunner do
 
     root = Keyword.get(opts, :root, File.cwd!()) |> Path.expand()
     manifest_path = Keyword.get(opts, :manifest, Path.join(root, @default_manifest))
-    manifest = Manifest.load!(manifest_path, root: root)
-    Manifest.runtime_dependency!(manifest.payload)
+    manifest = manifest_path |> Manifest.load!(root: root) |> Manifest.bind_runtime_dependency!()
     max_concurrency = validate_concurrency!(Keyword.get(opts, :max_concurrency, 2))
 
     case mode do

bench/imp/benchmark_truth/optimize_anything/campaign.ex

@@ -166,12 +166,11 @@ defmodule Imp.BenchmarkTruth.OptimizeAnything.Campaign do
     do: raise(ArgumentError, "campaign options must be a keyword list, got: #{inspect(opts)}")
 
   @doc false
-  def handle_usage(_event, _measurements, _metadata, {_agent, owner})
-      when owner != self(),
-      do: :ok
-
-  def handle_usage(event, measurements, metadata, {agent, _owner}),
-    do: handle_usage(event, measurements, metadata, agent)
+  def handle_usage(event, measurements, metadata, {agent, owner}) when is_pid(owner) do
+    if Imp.Telemetry.emitted_for?(owner),
+      do: handle_usage(event, measurements, metadata, agent),
+      else: :ok
+  end
 
   def handle_usage(_event, measurements, _metadata, agent) do
     Agent.update(agent, fn audit ->

benchmarks/authorities.json

@@ -149,12 +149,12 @@
     "req_llm": {
       "role": "beam_runtime_dependency",
       "repository": "https://github.com/agentjido/req_llm",
-      "version": "1.17.1",
-      "git_ref": "refs/tags/v1.17.1",
-      "commit": "33840077c2f1332eb6dff2d268dff02393014da4",
+      "version": "1.24.0",
+      "git_ref": "refs/tags/v1.24.0",
+      "commit": "fd9e079fddf253e9b719b2d2c6920f4306592809",
       "source_hashes": {
-        "hex_package": "266c0e06c47b4562f243dcdf41332342cbed2ec37064750edd725fb66bb6e914",
-        "hex_registry": "169d77a77bac5cee7655a7a9158ebea1c2a7857670f4104b29516a25a2d24f1e"
+        "hex_package": "cdc5c5cd7f38c0e17cf7a949ced99b6760dcba0d860cd4d066fea1aeba422929",
+        "hex_registry": "8ab6bda68e28afc3adf070cc6616d51831b42b8ff144116a9f104999af4c763f"
       }
     }
   },

benchmarks/data/multimodal/manifest.json

@@ -81,14 +81,6 @@
         "source": "https://ai.google.dev/gemini-api/docs/pricing"
       },
       "profile": "google-gemini-2.5-flash-generate-content",
-      "req_llm_dependency": {
-        "package": "req_llm",
-        "package_sha256": "266c0e06c47b4562f243dcdf41332342cbed2ec37064750edd725fb66bb6e914",
-        "repository": "https://github.com/agentjido/req_llm",
-        "source": "hexpm",
-        "source_revision": "33840077c2f1332eb6dff2d268dff02393014da4",
-        "version": "1.17.1"
-      },
       "req_llm_model": "google:gemini-2.5-flash"
     },
     "samples": [
@@ -159,7 +151,7 @@
         "prompt": "Use both pages of the attached synthetic PDF. Among projects with review code R7 on page 2, which owner on page 1 has the highest Units value? Return exactly one JSON object with one field and no markdown: {\"answer\": \"owner\"}"
       }
     ],
-    "schema_version": 2,
+    "schema_version": 3,
     "scoring": {
       "family_thresholds": {
         "image": 0.75,
@@ -180,5 +172,5 @@
       "prompt_contract": "One pinned attachment followed by one pinned question; return only a single-field JSON object named answer."
     }
   },
-  "payload_sha256": "4b4c80497789d74dec60b2b26236f6eee2f1f9b29046f07ef95b3fa3faf8e50f"
+  "payload_sha256": "54bf6a6ca6bc66c42f927387409c5bc0ab6010a8b7b1c1fc7ef67cbbae0b930e"
 }

benchmarks/data/multimodal/openai-responses-manifest.json

@@ -79,14 +79,6 @@
         "source": "https://developers.openai.com/api/docs/models/gpt-4.1-mini"
       },
       "profile": "openai-gpt-4.1-mini-2025-04-14-responses",
-      "req_llm_dependency": {
-        "package": "req_llm",
-        "package_sha256": "266c0e06c47b4562f243dcdf41332342cbed2ec37064750edd725fb66bb6e914",
-        "repository": "https://github.com/agentjido/req_llm",
-        "source": "hexpm",
-        "source_revision": "33840077c2f1332eb6dff2d268dff02393014da4",
-        "version": "1.17.1"
-      },
       "req_llm_model": "openai:gpt-4.1-mini-2025-04-14"
     },
     "samples": [
@@ -157,7 +149,7 @@
         "prompt": "Use both pages of the attached synthetic PDF. Among projects with review code R7 on page 2, which owner on page 1 has the highest Units value? Return exactly one JSON object with one field and no markdown: {\"answer\": \"owner\"}"
       }
     ],
-    "schema_version": 2,
+    "schema_version": 3,
     "scoring": {
       "family_thresholds": {
         "image": 0.75,
@@ -178,5 +170,5 @@
       "prompt_contract": "One pinned attachment followed by one pinned question; return only a single-field JSON object named answer."
     }
   },
-  "payload_sha256": "04daa155d3f97edfff62e329dfdca1248855f22b2271f7e954228432f1ae39d8"
+  "payload_sha256": "7e0f46b93d76a0f1703e6edce0276436ac686fc5dd2ad57a0e7a9826776e3a09"
 }

docs/differentials/MULTIMODAL_FIDELITY.md

@@ -39,8 +39,12 @@ The provider manifests are:
 
 Their strict, checksummed schema pins every sample ID, prompt, expected output,
 family, delivery mode, asset byte count, MIME type, asset SHA-256, provider,
-endpoint, API, exact model, generation options, ReqLLM dependency, and pricing.
-Runtime ReqLLM version drift fails before dispatch.
+endpoint, API, exact model, generation options, and pricing. The ReqLLM
+dependency is not part of the manifest: the runner reads the loaded ReqLLM
+version and its Hex package checksum from `mix.lock`, refuses to run when the
+two disagree, and binds that dependency into the checkpoint identity, the
+request audits and the artifact. A checkpoint cannot resume under a different
+ReqLLM package.
 
 All assets are synthetic and repository-owned. Images cover shape counting,
 spatial relation, and OCR. The native-document family sends the original
@@ -57,7 +61,7 @@ observed. For every request, the redacted audit contains:
 - sanitized endpoint, API, HTTP method, serialized model, and body SHA-256;
 - ordered serialized part types;
 - each part's MIME type, decoded byte count, and content SHA-256;
-- ReqLLM package source, version, package hash, repository, and source revision;
+- ReqLLM package source, version, and package hash;
 - ReqLLM request ID and detected transport.
 
 The audit never persists data URIs, base64 payloads, file bytes, prompt text, or

docs/differentials/RESEARCH_LANDSCAPE.md

@@ -221,7 +221,7 @@ semantics, or produces useful held-out improvement.
 | [Ax](https://github.com/ax-llm/ax) | Strongest independent implementation comparator for a typed TypeScript interpretation of DSPy-style programming. | Current product-semantic audit: npm `24.0.4`, registry `gitHead` `a366e49759bd596c8217eca91dfdc9dd8382835d`. Historical executable differential: `23.0.0` at `eb5835e54ba0c5b2fbac380daed1cb87faeefd5e`. Use for API and behavioral comparison, not as scientific authority. |
 | [BAML](https://github.com/BoundaryML/baml) | Comparator for compiler diagnostics, generated typed clients, and partial structured streaming. | Study its contracts; do not add a separate Imp language unless Elixir modules and macros are demonstrably insufficient. |
 | [AdalFlow](https://github.com/SylphAI-Inc/AdalFlow), [TextGrad](https://github.com/zou-group/textgrad), and [SAMMO](https://github.com/microsoft/sammo) | Comparators for explicit parameter graphs, textual feedback, and structure-aware prompt transformations. | Borrow mechanisms only after pinning code and paper protocols independently. |
-| [ReqLLM](https://github.com/agentjido/req_llm) | Preferred BEAM provider substrate. | Release `v1.17.1` resolves to `33840077c2f1332eb6dff2d268dff02393014da4`. Integrate its provider, multimodal, tool, stream, usage, error, and telemetry contracts instead of rebuilding them. |
+| [ReqLLM](https://github.com/agentjido/req_llm) | Preferred BEAM provider substrate. | Release `v1.24.0` resolves to `fd9e079fddf253e9b719b2d2c6920f4306592809`. Integrate its provider, multimodal, tool, stream, usage, error, and telemetry contracts instead of rebuilding them. |
 | [Jido](https://github.com/agentjido/jido) and [Jido AI](https://github.com/agentjido/jido_ai) | Optional deployment and long-running-agent complements. | Study immutable state, explicit effects, supervision, and signals without making their agent model mandatory. |
 | [LangGraph](https://github.com/langchain-ai/langgraph), [Pydantic AI](https://github.com/pydantic/pydantic-ai), [MLflow](https://mlflow.org/docs/latest/genai/prompt-registry/optimize-prompts), and [Promptfoo](https://github.com/promptfoo/promptfoo) | Production references for checkpoints, durable execution, registries, eval matrices, and adversarial testing. | Treat as operational comparators, not optimizer parity authorities. |