P0-4 verification: live-API smoke harness + Anthropic.SDK binding-gap docs

- Add tests/AnthropicVerification/ throwaway harness that hits the live
  Anthropic API for the three model IDs cited in Appendix B / Chapter 4.2.4
  (claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001). All three
  passed on 2026-05-02 with funded key.
- Pin harness to Microsoft.Extensions.AI 10.3.0 via VersionOverride; required
  because Anthropic.SDK 5.10.0 is compiled against M.E.AI 10.3 and crashes at
  runtime against the central pin (10.5.0) with a MissingMethodException on
  HostedMcpServerTool.AuthorizationToken. Per-project override does NOT work
  on samples that depend on Microsoft.Agents.AI 1.3 (which itself requires
  M.E.AI 10.5 and trips CS1705) -- documented in the sample's README.
- Record the 2026-05-02 verification pass and three remaining mitigation
  paths in docs/verification-log.md.
- Add a Known issues section to the repo README and a callout to the
  04.2.4-anthropic-agents README so readers aren't surprised by the runtime
  failure on the print-tagged code.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: NinjaRocks

README.md

@@ -135,6 +135,12 @@ If you intend to run any sample unattended (CI, scheduled jobs, scripts), set pe
 
 CI runs on every push and pull request, plus a weekly Monday build to catch transitive package drift. The matrix builds every sample in the `samples/` tree against the pinned versions in `Directory.Packages.props`. A green badge above means the print-tagged code still compiles end-to-end. If the badge is red, check the latest [Actions run](https://github.com/CodeShayk/generative-ai-dotnet-samples/actions/workflows/ci.yml) for the breaking package or API change before assuming the samples are wrong.
 
+CI catches build-level drift but not all runtime regressions. For the model-ID surface that ships in Appendix B, run the live-API smoke harness at [`tests/AnthropicVerification/`](tests/AnthropicVerification/) -- it iterates each cited Claude model ID against the live Anthropic API and reports `PASS`/`FAIL` per model. Each verification pass is recorded in [`docs/verification-log.md`](docs/verification-log.md).
+
+## Known issues
+
+- **`Anthropic.SDK` 5.10 ↔ `Microsoft.Extensions.AI` 10.5 binding gap.** `Anthropic.SDK` 5.10.0 is compiled against `Microsoft.Extensions.AI.Abstractions` 10.3.0; the central pin moved to 10.5.0 (required by `Microsoft.Agents.AI` 1.3), which reshapes `HostedMcpServerTool.AuthorizationToken`. Calling Claude through Anthropic.SDK's `IChatClient` bridge therefore throws `MissingMethodException` at runtime. This affects `samples/ch04-agent-framework/04.2.4-anthropic-agents` -- it builds clean but does not run as-shipped. The throwaway harness at `tests/AnthropicVerification/` works because it has no `Microsoft.Agents.AI` dependency and pins M.E.AI to 10.3.0 locally via `VersionOverride`; the same override does **not** work on the chapter sample because Agents.AI 1.3 itself rejects 10.3 with `CS1705`. Mitigation paths and pre-print decision are tracked in `docs/verification-log.md` (entry `2026-05-02`). Likely cleared once `Anthropic.SDK` 5.11+ ships rebuilt against M.E.AI 10.5+.
+
 ## Versioning and tags
 
 | Tag | Meaning |

docs/verification-log.md

@@ -91,3 +91,45 @@ Copy this block to start a new entry. Date format `YYYY-MM-DD`.
 ---
 
 *New entries go below this line, most recent first.*
+
+---
+
+## 2026-05-02 -- P0-4 Anthropic live-API smoke
+
+**Status:** Yellow -- model IDs and URLs verified green; one real defect (M.E.AI ↔ Anthropic.SDK runtime binding gap) blocks the existing Anthropic sample and must be cleared before print.
+
+**Test runs in this entry:** initial pass blocked on zero credit balance; re-run later the same day with funded key returned all three model IDs `OK`.
+
+### Packages
+- [x] `Anthropic.SDK` -- v5.10.0 (latest stable on NuGet); central pin unchanged.
+- [x] `Microsoft.Extensions.AI` -- v10.5.0 central pin; harness overrides to v10.3.0 to match Anthropic.SDK's compile-time target (see Issues).
+
+### Code samples
+- [x] Verification harness at `tests/AnthropicVerification/` builds clean with M.E.AI 10.3.0 override.
+- [x] Live-API model-ID smoke through the harness -- **green for all three IDs** (`PASS  claude-opus-4-7: OK`; `PASS  claude-sonnet-4-6: OK`; `PASS  claude-haiku-4-5-20251001: OK`).
+- [ ] Existing `samples/ch04-agent-framework/04.2.4-anthropic-agents` -- builds clean but **fails at runtime** with `MissingMethodException` from `Anthropic.SDK.Messaging.ChatClientHelper.CreateMessageParameters` -> `Microsoft.Extensions.AI.FunctionInvokingChatClient.GetResponseAsync` -> `Microsoft.Agents.AI.ChatClientAgent.RunCoreAsync`. Re-confirmed with funded key on 2026-05-02; the defect is independent of credits (see Issues).
+
+### URLs
+- [x] `https://docs.anthropic.com/` (cited at `Manuscript/Chapter-04.md:3609`) -- HTTP 200.
+- [x] `https://platform.claude.com/docs/en/about-claude/models/overview` (cited at `Manuscript/Appendix-B-Model-Quick-Reference.md:36`) -- HTTP 200.
+- [x] Sanity: `https://platform.claude.com/`, `https://docs.anthropic.com/en/docs/about-claude/models/overview` -- both HTTP 200.
+- [ ] Microsoft Learn / Azure / NuGet links -- deferred; not in the P0-4 scope.
+
+### Anthropic API surface
+- [x] Model IDs in `Manuscript/Appendix-B-Model-Quick-Reference.md:32-34` and `Manuscript/Chapter-04.md:557` (`claude-opus-4-7`, `claude-sonnet-4-6`, `claude-haiku-4-5-20251001`) -- all three callable; harness round-trip returned the expected `OK` response.
+- [x] `Anthropic.SDK` package surface used in chapter examples (`new AnthropicClient(key).Messages` as `IChatClient`) -- wire-compatible with the live API; full request/response round-trip confirmed when paired with M.E.AI 10.3.0.
+
+### Issues found / actions taken
+- **Defect (blocks print): `samples/ch04-agent-framework/04.2.4-anthropic-agents` does not run against the repo's central pin.** Throws `MissingMethodException: Method not found: 'System.String Microsoft.Extensions.AI.HostedMcpServerTool.get_AuthorizationToken()'.` from `Anthropic.SDK.Messaging.ChatClientHelper.CreateMessageParameters`. Cause: `Anthropic.SDK` 5.10.0's nuspec declares `Microsoft.Extensions.AI.Abstractions >= 10.3.0` but the assembly is compiled against 10.3.0 specifically; M.E.AI 10.5.0 changed the shape of `HostedMcpServerTool.AuthorizationToken` (property-getter no longer present). CI is green because builds succeed against the *abstractions* package surface; runtime binding does not.
+    - **Tried: per-project `VersionOverride` to 10.3.0 (initially proposed as option 3).** Verified to *not work* for this sample: `Microsoft.Agents.AI` 1.3.0 is itself compiled against `Microsoft.Extensions.AI.Abstractions` 10.5.0, so downgrading abstractions to 10.3.0 raises `CS1705 (uses higher version than referenced assembly)` at compile time. Override reverted; sample is back to clean build, broken runtime.
+    - **The harness avoids the conflict only because it has no `Microsoft.Agents.AI` dependency** -- pure Anthropic.SDK + M.E.AI. Override pin 10.3.0 stays on the harness as the runtime mitigation there.
+    - **Remaining viable paths (pre-print decision):**
+        1. Wait for `Anthropic.SDK` 5.11+ rebuilt against M.E.AI 10.5+. Cleanest, but not on NuGet as of 2026-05-02 (latest stable: 5.10.0). Watch <https://www.nuget.org/packages/Anthropic.SDK>.
+        2. Replace Anthropic.SDK's IChatClient bridge with a small custom `IChatClient` that calls `AnthropicClient.Messages.GetClaudeMessageAsync` directly. ~50 lines in the sample; bypasses the broken `ChatClientHelper`. Changes the chapter narrative slightly (the sample now ships a thin adapter rather than relying on the SDK's built-in adapter).
+        3. Switch the sample to the OpenAI SDK pointed at Anthropic's OpenAI-compatibility endpoint. Strongest provider-neutral story, but rewrites the chapter section.
+- **CI gap:** the build matrix needs at least one runtime-execution lane for samples that have non-cloud-key prerequisites (currently 04.2.4 needs ANTHROPIC_API_KEY so cannot run unattended; consider adding a stub-key smoke that exercises the M.E.AI binding via `IChatClient` invocation up to the wire-send point).
+- Verification harness lives at `tests/AnthropicVerification/` (throwaway, not in `AI in .Net.sln`). Re-run: `ANTHROPIC_API_KEY=<key> dotnet run --project tests/AnthropicVerification`.
+
+### Next-pass to-dos
+- [ ] Pick one of paths 1-3 above for the chapter sample. If path 1 (wait for SDK 5.11), open a tracking issue and re-check NuGet weekly. If path 2 (custom IChatClient), implement the bridge, re-run, confirm green. If path 3 (OpenAI-compat endpoint), revise the chapter prose for §4.2.4.
+- [ ] Run the broader URL audit (Microsoft Learn, Azure, NuGet) and fold into the next weekly entry.

samples/ch04-agent-framework/04.2.4-anthropic-agents/README.md

@@ -2,7 +2,9 @@
 
 Companion code for **Generative AI in .NET**, Chapter 4 section 4.2.4 ("Anthropic Agents").
 
-Wraps `Anthropic.SDK`'s message client as `IChatClient` via `.AsIChatClient(modelId)`, then hands it to a standard `ChatClientAgent`. The exact same agent API works against OpenAI, Azure OpenAI, Ollama, etc.
+Uses `new AnthropicClient(apiKey).Messages` -- which already implements `IChatClient` -- and hands it to a standard `ChatClientAgent`. The same agent API works against OpenAI, Azure OpenAI, Ollama, etc.
+
+> **Known issue (2026-05-02): this sample does not run as-shipped against the repo's central package pins.** It builds clean, but at runtime `ChatClientAgent.RunAsync(...)` throws `MissingMethodException: Method not found: 'System.String Microsoft.Extensions.AI.HostedMcpServerTool.get_AuthorizationToken()'.`. Root cause: `Anthropic.SDK` 5.10.0 was compiled against `Microsoft.Extensions.AI.Abstractions` 10.3.0; the repo pins 10.5.0 (required by `Microsoft.Agents.AI` 1.3.0), which reshapes that property. Per-project `VersionOverride` to 10.3.0 does not help here -- Agents.AI 1.3 itself rejects 10.3 with `CS1705` at compile time. **Mitigation paths under evaluation** (see [`docs/verification-log.md`](../../../docs/verification-log.md), entry `2026-05-02`): (1) wait for `Anthropic.SDK` 5.11+ rebuilt against 10.5+, (2) ship a thin custom `IChatClient` adapter that calls `AnthropicClient.Messages.GetClaudeMessageAsync` directly, or (3) re-target the sample at Anthropic's OpenAI-compatibility endpoint via the OpenAI SDK. The chapter prose still describes the intended pattern; once the underlying packages reconcile, the sample runs unchanged.
 
 ## Run it
 
@@ -21,3 +23,7 @@ Override the model with `ANTHROPIC_MODEL` (defaults to `claude-haiku-4-5-2025100
 ## Prerequisites
 
 - .NET 9 SDK and an Anthropic API key.
+
+## Verifying model IDs are still callable
+
+A throwaway harness at [`tests/AnthropicVerification/`](../../../tests/AnthropicVerification/) hits the live API with each model ID listed in Appendix B. The harness has no `Microsoft.Agents.AI` dependency, so it can pin to M.E.AI 10.3.0 and run end-to-end -- use it during the pre-print verification passes to confirm Anthropic has not rotated the cited identifiers.

tests/AnthropicVerification/AnthropicVerification.csproj

@@ -0,0 +1,25 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <OutputType>Exe</OutputType>
+    <TargetFramework>net9.0</TargetFramework>
+    <Nullable>enable</Nullable>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <LangVersion>latest</LangVersion>
+    <RootNamespace>AnthropicVerification</RootNamespace>
+    <AssemblyName>AnthropicVerification</AssemblyName>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageReference Include="Anthropic.SDK" />
+    <!--
+      Anthropic.SDK 5.10.0 was compiled against Microsoft.Extensions.AI.Abstractions 10.3.0;
+      the repo's central pin (10.5.0) reshapes HostedMcpServerTool.AuthorizationToken and
+      causes a runtime MissingMethodException. Pin this throwaway harness to the SDK's
+      compile-time targets via VersionOverride.
+    -->
+    <PackageReference Include="Microsoft.Extensions.AI" VersionOverride="10.3.0" />
+    <PackageReference Include="Microsoft.Extensions.AI.Abstractions" VersionOverride="10.3.0" />
+  </ItemGroup>
+
+</Project>

tests/AnthropicVerification/Program.cs

@@ -0,0 +1,37 @@
+using Anthropic.SDK;
+using Microsoft.Extensions.AI;
+
+var apiKey = Environment.GetEnvironmentVariable("ANTHROPIC_API_KEY");
+if (string.IsNullOrWhiteSpace(apiKey))
+{
+    Console.Error.WriteLine("ANTHROPIC_API_KEY not set.");
+    return 1;
+}
+
+string[] modelsToVerify =
+[
+    "claude-opus-4-7",
+    "claude-sonnet-4-6",
+    "claude-haiku-4-5-20251001",
+];
+
+IChatClient chat = new AnthropicClient(apiKey).Messages;
+
+var any = false;
+foreach (var model in modelsToVerify)
+{
+    try
+    {
+        var response = await chat.GetResponseAsync(
+            "Reply with exactly: OK",
+            new ChatOptions { ModelId = model });
+        Console.WriteLine($"PASS  {model}: {response.Text?.Trim()}");
+    }
+    catch (Exception ex)
+    {
+        any = true;
+        Console.WriteLine($"FAIL  {model}: {ex.GetType().Name}: {ex.Message}");
+    }
+}
+
+return any ? 2 : 0;

tests/AnthropicVerification/README.md

@@ -0,0 +1,45 @@
+# Anthropic Live-API Smoke Harness
+
+Throwaway test project used to verify, against the live Anthropic API, that the Claude model identifiers cited in the manuscript are still callable. Lives outside `AI in .Net.sln` on purpose -- it is not part of the shipped sample set.
+
+## What it does
+
+Iterates the three model IDs cited in `Manuscript/Appendix-B-Model-Quick-Reference.md` and `Manuscript/Chapter-04.md`, sends each one a fixed `"Reply with exactly: OK"` prompt via `IChatClient`, and prints `PASS`/`FAIL` per model. Exits 0 if all three respond, 2 if any fail, 1 if `ANTHROPIC_API_KEY` is unset.
+
+## Why it pins `Microsoft.Extensions.AI` to 10.3.0
+
+`Anthropic.SDK` 5.10.0 (latest stable at time of writing) is compiled against `Microsoft.Extensions.AI.Abstractions` 10.3.0. The repo's central pin is 10.5.0, which reshapes `HostedMcpServerTool.AuthorizationToken` -- runtime calls through the SDK's `IChatClient` bridge throw `MissingMethodException`. The harness sidesteps this by overriding the pin locally:
+
+```xml
+<PackageReference Include="Microsoft.Extensions.AI" VersionOverride="10.3.0" />
+<PackageReference Include="Microsoft.Extensions.AI.Abstractions" VersionOverride="10.3.0" />
+```
+
+The override works here because the harness has no `Microsoft.Agents.AI` dependency. The chapter sample at `samples/ch04-agent-framework/04.2.4-anthropic-agents/` cannot use the same override (Agents.AI 1.3 is itself compiled against M.E.AI 10.5 and refuses to bind to 10.3 with `CS1705`); see that sample's README and `docs/verification-log.md` for the workaround status.
+
+When `Anthropic.SDK` 5.11+ ships rebuilt against M.E.AI 10.5+, drop both overrides.
+
+## Run it
+
+```bash
+cd code/generative-ai-dotnet-samples
+ANTHROPIC_API_KEY='sk-ant-...' dotnet run --project tests/AnthropicVerification
+```
+
+Expected output:
+
+```
+PASS  claude-opus-4-7: OK
+PASS  claude-sonnet-4-6: OK
+PASS  claude-haiku-4-5-20251001: OK
+```
+
+If a model returns `FAIL`, update `Manuscript/Appendix-B-Model-Quick-Reference.md` to the current Anthropic-published identifier and re-run.
+
+## When to re-run
+
+- During the four pre-print weekly verification passes.
+- Whenever Anthropic announces a new Claude release (model IDs include date suffixes that may roll forward).
+- After bumping `Anthropic.SDK` or `Microsoft.Extensions.AI*` versions.
+
+Record results in `docs/verification-log.md`.