Instrument Observability

Instrument A365 Observability

Trigger phrases — any of these will activate this skill automatically:

• "instrument observability for this agent"
• "add a365 observability to this agent"
• "add observability to this agent"
• "set up tracing for this agent"
• "make this agent visible in microsoft defender"
• "enable agent 365 telemetry"
• "wire up opentelemetry for this agent"
• "add observability to this .net agent"
• "add observability to this node.js agent"
• "add a365 observability to this python agent"

Overview

This skill instruments Microsoft Agent 365 observability into an existing agent codebase without disrupting the agent's core logic. It:

1. Detects the agent type (.NET AgentFramework, Node.js, or Python)
2. Installs the correct A365 observability packages (core + hosting + optional extensions)
3. Wires observability in the entry point
4. Adds BaggageBuilder context or BaggageMiddleware to message handlers
5. Implements the agentic token resolver with caching
6. Adds manual instrumentation scopes (InvokeAgentScope, InferenceScope, ExecuteToolScope — required for store publishing)
7. Updates configuration files with observability settings
8. Validates the build passes

Store publishing requirement: The Agent 365 store validation requires InvokeAgentScope, InferenceScope, and ExecuteToolScope to be implemented. This skill wires them.

All changes are additive and idempotent — rerunning the skill is safe.

Phase 0: Load Detection Cache and Validate

Task-list display (applies throughout this skill). This skill creates tasks inline via **TaskCreate** — "..." markers at the start of each phase, and marks them complete at phase end. The user must see this progress visibly. Each TaskCreate line corresponds to one checklist item; exactly one item in_progress at a time.

• Claude Code: TaskCreate is in allowed-tools — calling it renders a native checklist UI; subsequent TaskUpdate calls flip statuses.
• VS Code Copilot Chat / GitHub Copilot CLI: allowed-tools is ignored — before Phase 0.1, scan this SKILL.md for all **TaskCreate** — "..." lines and emit a markdown checklist in chat upfront (- [ ] Load detection cache…, - [ ] Determine agent kind…, etc.); flip items to - [x] as each phase completes.

TaskCreate — "Load detection cache and validate with user"

Step 0.1 — Triage the workspace

Run in parallel:

• Glob **/*.csproj, package.json, requirements.txt, pyproject.toml, src/**/*.ts, **/*.cs, **/*.py → hasProjectFiles.
• Read .a365-workspace-detection.local.json → cacheState (fresh if detectedAt < 60 min, stale if older, missing if absent).

Decide:

Step 0.2 — Load from cache

🛑 STOP — .a365-workspace-detection.local.json MUST exist before this step. Read the file path .a365-workspace-detection.local.json in the working directory. If it does not exist, you arrived at Step 0.2 by skipping Step 0.1's triage routing. Do NOT proceed. Do NOT invent default cache values. Do NOT run any further phase (no npm install, no dotnet add package, no pip install, no file edits). Instead:

1. Tell the user verbatim: "I skipped the Step 0.1 triage and the detection cache wasn't written. Running a365-setup now to fix that, then I'll return here."
2. Read ${CLAUDE_PLUGIN_ROOT}/skills/a365-setup/SKILL.md and follow it to completion.
3. Re-verify the file now exists, then continue below.

The stop hook (validate-instrument-observability.js) will fail the session at end if the cache file is missing — this guard exists so the model halts immediately rather than instrumenting against unknown authMode.

Load from cache: agentStack, programmingLanguage, usesTeamsOrCopilot, agentType, authMode (if previously stored).

Present the loaded values in one message and wait for confirmation:

Here's what we detected about your agent:
  • Stack:    {agentStack}
  • Language: {programmingLanguage}

Reply **yes** to confirm, or describe any corrections.

TaskUpdate — Mark complete: "Load detection cache and validate with user"

Phase 0.5: Agent Kind and Authentication Mode

TaskCreate — "Determine agent kind and authentication mode"

Read ${CLAUDE_PLUGIN_ROOT}/shared/agent-detection.md — section "Agent Type and Auth Mode Detection" — and follow it exactly.

If agentType and authMode are already present in the detection cache (from a prior skill run in this session OR pre-populated by a parent skill like make-ai-teammate), the confirmation behavior depends on agentType:

• agentType = "ai-teammate" — skip the confirmation prompt entirely. The AI Teammate identity model is unambiguous (authMode = agentic-user, no obo/s2s decision exists), so a confirm prompt adds friction without catching drift. Proceed silently.
• agentType = "system-agent" — confirm the cached values with the user before proceeding, since the obo/s2s choice is meaningful and a stale value would silently route to the wrong token path.

Read authMode case-insensitively (S2S = s2s, OBO = obo); always write back the canonical lowercase value.

Store agentType (ai-teammate = AI Teammate, or system-agent = Agent (Non AI Teammate)) and authMode:

• AI Teammate: agentic-user (agent's own M365 identity — not the caller's token; auto-set, no question needed)
• Agent (Non AI Teammate): obo (On-Behalf-Of — signed-in user token) or s2s (Service Principal, no user token)

Update .a365-workspace-detection.local.json — merge agentType and authMode into the existing cache file, preserving all other fields (agentStack, programmingLanguage, usesTeamsOrCopilot, detectedAt). Use the Write tool to write the merged object back.

The authMode value drives Phases 3–5: OBO and S2S paths differ in entry point wiring (Phase 3), message handler pattern (Phase 4), and token resolver (Phase 5). Phases 2, 6, 7, and 8 are identical regardless of authMode.

TaskUpdate — Mark complete: "Determine agent type and authentication mode"

Phase 1: Detect Agent Type

TaskCreate — "Detect agent type and load reference patterns"

1. Read ${CLAUDE_PLUGIN_ROOT}/shared/agent-detection.md for detection heuristics.

2. Run detection following the rules in agent-detection.md:

   • Check for .NET AgentFramework indicators (Microsoft.Agent.*, AgentFramework) → .csproj
   • Check for Node.js indicators (package.json, @langchain, openai, @microsoft/agents-*)
   • Check for Python indicators (requirements.txt, pyproject.toml, .py files, microsoft-agents)
   • Determine package file (*.csproj, package.json, pyproject.toml/requirements.txt)
   • Determine entry point (Program.cs, index.ts/js, app.py / host_agent_server.py)
   • Determine message handler location

3. Load reference patterns:

   • If .NET: Read ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/references/dotnet-observability.md
   • If Node.js: Read ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/references/nodejs-observability.md
   • If Python: Read ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/references/python-observability.md

4. If agent type cannot be determined, write marker .a365setup-unknown-agent and exit early with clear error message.

5. Framework-support soft-warn matrix. Check (programmingLanguage, agentStack) from the detection cache and surface a warning when the stack lacks first-class auto-instrumentation in @microsoft/opentelemetry or microsoft-opentelemetry. The skill still proceeds — observability is the OTel SDK underneath, which works for any HTTP-based LLM — but the user should know they'll need to add manual InferenceScope.start wrappers around each LLM call.

   Lang	Stack	Action
   Node.js	LangChain, OpenAI Agents SDK, Claude SDK	✅ Auto-instrumented (Claude with custom shape — see Phase 5.5)
   Node.js	Semantic Kernel, Google ADK	⚠ Soft-warn — auto-instrumentation may not patch the LLM library; add manual InferenceScope.start around each LLM call
   Python	Agent Framework, OpenAI, Google ADK	✅ Auto-instrumented
   Python	LangChain, Claude SDK, CrewAI	⚠ Soft-warn — same as Node.js SK/ADK
   .NET	Agent Framework, Semantic Kernel	✅ Auto-instrumented via .UseOpenTelemetry() on IChatClient
   .NET	Azure AI Foundry	⚠ Soft-warn — best-effort wiring

   For soft-warn rows, surface verbatim: "Auto-instrumentation in <unified-distro-package> doesn't patch your LLM library directly. The skill will still wire useMicrosoftOpenTelemetry/UseMicrosoftOpenTelemetry (OTel SDK + A365 exporter), but you'll need to manually wrap each LLM call with InferenceScope.start(...) to capture gen_ai.* spans. See <language>-observability.md § 'InferenceScope — Manual Wrapping' for the pattern." Continue to Phase 2.

6. TaskUpdate — Mark complete and report detected agent type (+ any soft-warn) to user.

Phase 2: Install A365 Observability Packages

TaskCreate — "Install A365 observability packages"

All languages converge on a single unified distro that re-exports the legacy A365 observability + hosting types and auto-instruments common LLM SDKs:

Do not install legacy *.Observability.Runtime / -hosting / -extensions-* packages alongside the unified distro — the distro re-exports their types and mixing the two produces CS0433 duplicate-type errors (.NET) or duplicate spans (Node.js / Python). After install, verify the package appears in the manifest (*.csproj / package.json / requirements.txt or pyproject.toml). pip install does not update the dependency manifest — prefer uv add microsoft-opentelemetry (or poetry add ...) for Python.

Python — Google ADK gotcha: if pyproject.toml lists google-adk, uv sync will backtrack for minutes resolving the OTel graph. Pin OTel via [tool.uv] override-dependencies — see python-observability.md → "Google ADK projects — pin the OTel stack" for the exact block. Other Python stacks (AgentFramework, LangChain, OpenAI, Claude, Semantic Kernel) don't need this.

Full per-language package tables, version constraints, and the LangChain extras flag live in the references — see the "Required packages" section of:

• ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/references/dotnet-observability.md
• ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/references/nodejs-observability.md
• ${CLAUDE_PLUGIN_ROOT}/skills/instrument-observability/references/python-observability.md

TaskUpdate — Mark complete.

Phase 3: Wire Observability in Entry Point

TaskCreate — "Wire observability in entry point"

Pre-existing placeholders: As of CLI 1.1, a365 setup all auto-writes Agent365Observability placeholder sections to appsettings.json (.NET) or .env (Node.js/Python). Before creating config from scratch, check if placeholders already exist and fill in values rather than duplicating the section.

For .NET AgentFramework

1. Read the current entry point (Program.cs or detected file).

2. Edit — Add observability wiring following the reference pattern in dotnet-observability.md:

   • Add using Microsoft.OpenTelemetry; to Program.cs.
   • OBO / agentic-user path (AI Teammate AND Standard .NET agents): Call builder.UseMicrosoftOpenTelemetry(o => { ... }) with o.Exporters = ExportTarget.Agent365 | ExportTarget.Console (Dev) or ExportTarget.Agent365 (Production). The distro auto-registers IExporterTokenCache<AgenticTokenStruct> in DI — no AddAgenticTracingExporter() call needed. Leave o.Agent365.Exporter.UseS2SEndpoint at its default (false) — the exporter POSTs to /observability/ which the OBO token cache authenticates. Also set "EnableAgent365Exporter": true in appsettings.json to activate the backend exporter — the SDK defaults this to false when absent, so without it the exporter is wired but inert.
   • Required: IChatClient.UseOpenTelemetry() — when registering the IChatClient (e.g. Azure OpenAI), chain .AsBuilder().UseFunctionInvocation().UseOpenTelemetry(sourceName: null, cfg => cfg.EnableSensitiveData = true).Build(). This is what makes the AI SDK emit the gen_ai.inference and gen_ai.tool spans that InvokeAgentScope (Phase 5.5) anchors as children. Skipping this means no LLM spans appear in MAC, even with everything else wired — the InvokeAgent parent becomes a hollow span. EnableSensitiveData = true includes prompts/completions in span attributes (PII consideration — set to false for regulated data).
   • S2S path: First Write the two scaffold files from the reference doc — Observability/ObservabilityServiceExtensions.cs (DI extension with AddAgent365Observability() using ServiceTokenCache and conditional ObservabilityTokenService) and Observability/ObservabilityTokenService.cs (background service that acquires the Observability API token via the MSAL FMI 3-hop chain with .WithFmiPath() targeting scope api://9b975845-388f-4429-889e-eab1ef63949c/.default, supports MSI with client-secret fallback). Then call builder.Services.AddAgent365Observability(); and builder.UseMicrosoftOpenTelemetry(...) with token resolver reading from the ServiceTokenCache. Critical: Set o.Agent365.Exporter.UseS2SEndpoint = true in the options callback — without this, the exporter posts to the wrong path (/observability/ instead of /observabilityService/) and gets HTTP 401. See "Known Issues" section.
   • Optionally register adapter.Use(new BaggageTurnMiddleware()) (OBO path only) to auto-populate baggage on every request
   • Mark all new lines with: // A365 Observability — best-effort instrumentation (verify against official sample)

3. Preserve all existing code — only add new lines, never remove.

For Node.js

1. Read the current entry point (index.ts, app.ts, or detected file).

2. Edit — Add observability initialization following the reference pattern in nodejs-observability.md:

   • Import useMicrosoftOpenTelemetry, shutdownMicrosoftOpenTelemetry, configureA365Hosting, and AgenticTokenCacheInstance — all from @microsoft/opentelemetry (single package as of GA 1.0; do NOT import from the legacy -observability, -hosting, or -runtime packages).
   • OBO / agentic-user path: Call useMicrosoftOpenTelemetry({ a365: { enabled: true, enableObservabilityExporter: true, tokenResolver } }) before any LLM/framework imports. Both enabled: true AND enableObservabilityExporter: true are required in 1.0+ to actually export spans. Wire tokenResolver to AgenticTokenCacheInstance.getObservabilityToken(agentId, tenantId) ?? ''.
   • S2S path: First Write observability/token-cache.ts (in-memory token cache with cacheToken/getCachedToken/tokenResolver) and observability/observability-token-service.ts using the scaffold pattern from nodejs-observability.md (S2S section). This module acquires the Observability API token via MSAL FMI 3-hop chain (@azure/msal-node with fmiPath parameter, targeting scope api://9b975845-388f-4429-889e-eab1ef63949c/.default, supports MSI with client-secret fallback) and refreshes it every 50 min. Then call useMicrosoftOpenTelemetry({ a365: { enabled: true, enableObservabilityExporter: true, useS2SEndpoint: true, tokenResolver: a365TokenResolver } }). useS2SEndpoint: true is now a first-class option (1.0+); the old workaround with custom Agent365Exporter via spanProcessors and ENABLE_A365_OBSERVABILITY_EXPORTER=false is no longer needed for new instrumentation. If the old workaround is already present in an existing agent, leave it in place — do not delete code as part of this additive skill; flag it in the final summary as a candidate for cleanup if the user explicitly asks to migrate.
   • Both paths: Call configureA365Hosting(adapter, { enableBaggage: true }) once at startup to register BaggageMiddleware. This replaces manual adapter.use(new BaggageMiddleware()) and removes the need for BaggageBuilderUtils.fromTurnContext in handlers.
   • Both paths: Register SIGTERM/SIGINT handlers calling await shutdownMicrosoftOpenTelemetry() to flush pending spans on shutdown.
   • Auto-instrumentation note: Do NOT call OpenAIAgentsTraceInstrumentor.enable() or LangChainTraceInstrumentor.instrument() — these are auto-enabled in 1.0+ and manual calls cause duplicate spans. To opt out, set instrumentationOptions: { openaiAgents: { enabled: false } }.
   • Mark all new lines with: // A365 Observability — best-effort instrumentation (verify against official sample)

3. Preserve all existing code — only add new lines, never remove.

For Python

1. Read the current entry point (app.py, host_agent_server.py, or detected file).

2. Edit — Add observability configuration following the reference pattern in python-observability.md:

   • Import use_microsoft_opentelemetry from microsoft.opentelemetry (single unified package; do NOT import from the legacy microsoft_agents_a365.* namespace).
   • OBO / agentic-user path: Call use_microsoft_opentelemetry(enable_a365=True, a365_enable_observability_exporter=True, a365_token_resolver=...). Both enable_a365=True AND a365_enable_observability_exporter=True are required in 1.0+ to actually export spans. Wire a365_token_resolver to AgenticTokenCache().get_observability_token from microsoft.opentelemetry.a365.hosting.token_cache_helpers (or a custom resolver reading from token_cache.py).
   • S2S path: First Write observability/token_cache.py (in-memory token cache with cache_token/get_cached_token) and observability/observability_token_service.py using the scaffold pattern from python-observability.md (S2S section). This module acquires the Observability API token via a 3-hop FMI chain: direct HTTP POST with fmi_path for Hops 1+2 (MSAL Python does not properly serialize fmi_path — known limitation), then msal.ConfidentialClientApplication for Hop 3, targeting scope api://9b975845-388f-4429-889e-eab1ef63949c/.default, supports MSI with client-secret fallback, refreshes every 50 min via an asyncio background task. Then call use_microsoft_opentelemetry(enable_a365=True, a365_enable_observability_exporter=True, a365_use_s2s_endpoint=True, a365_token_resolver=...). a365_use_s2s_endpoint=True is now a first-class kwarg — no workaround needed. Schedule run_token_service() as an asyncio task and call acquire_initial_token() in your aiohttp lifespan startup. Also install msal, azure-identity, and httpx.
   • Both paths: Call ObservabilityHostingManager.configure(adapter.middleware_set, ObservabilityHostingOptions(enable_baggage=True)) once at startup to auto-populate baggage from TurnContext. Note: enable_baggage defaults to False — must be explicitly set to True.
   • Auto-instrumentation note: Do NOT call legacy *Instrumentor().instrument() methods for LangChain/OpenAI/SK/AgentFramework — these are auto-enabled in 1.0+ and manual calls cause duplicate spans.
   • Mark all new lines with: # A365 Observability — best-effort instrumentation (verify against official sample)

3. Preserve all existing code — only add new lines, never remove.

4. TaskUpdate — Mark complete.

Phase 4: Add BaggageBuilder Context to Message Handler

TaskCreate — "Add BaggageBuilder context to message handler"

Skip this phase if BaggageMiddleware was registered in Phase 3 — the middleware handles baggage propagation automatically for every request.

Auth mode note: All three authMode values use authHandlerName: "AGENTIC" in the code — the token exchange call is identical. The identity in traces is determined by Azure AD provisioning and the incoming token. Add an inline comment indicating which mode was chosen.

For .NET AgentFramework

1. Read the detected message handler file.

2. Edit — Follow the reference pattern in dotnet-observability.md (full code sample under "Agent Class — Message Handler (OBO Path)"):

   OBO path (obo / agentic-user) — applies to both AI Teammate agents and Standard .NET agents:

   • Inject IExporterTokenCache<AgenticTokenStruct> in the constructor (auto-registered by the distro — no AddAgenticTracingExporter() call needed).
   • Inject IConfiguration (for blueprint/observability config) and ILogger<MyAgent>.
   • Resolve agent ID for BOTH auth paths — agentic instance ID from the Activity for agentic turns, decoded from the auth token via Utility.ResolveAgentIdentity(context, authToken) for non-agentic turns (the SDK names the second parameter generically authToken — it accepts both OBO tokens and agentic-path tokens returned by UserAuthorization.GetTurnTokenAsync). Do NOT fall back to Guid.Empty.ToString() — that creates a synthetic identity the exporter cannot authenticate, polluting traces with "No token obtained. Skipping export for this identity." warnings.

     string? resolvedAgentId = null;
     if (turnContext.Activity.IsAgenticRequest())
     {
         resolvedAgentId = turnContext.Activity.GetAgenticInstanceId();
     }
     else if (!string.IsNullOrEmpty(authHandlerName))
     {
         try
         {
             var authToken = await UserAuthorization
                 .GetTurnTokenAsync(turnContext, authHandlerName, cancellationToken: cancellationToken)
                 .ConfigureAwait(false);
             if (!string.IsNullOrEmpty(authToken))
             {
                 resolvedAgentId = Utility.ResolveAgentIdentity(turnContext, authToken);
             }
         }
         catch (Exception ex)
         {
             _logger.LogDebug(ex, "Could not resolve agent id from auth token; A365 observability skipped for this turn.");
         }
     }
     
     var resolvedTenantId = turnContext.Activity.Conversation?.TenantId
                         ?? turnContext.Activity.Recipient?.TenantId;
     
     var hasObservabilityIdentity = !string.IsNullOrEmpty(resolvedAgentId)
                                 && !string.IsNullOrEmpty(resolvedTenantId);
     

     GetAgenticInstanceId() returns the agent's service principal object ID (the instance ID assigned by A365 for the Teams agentic identity). Utility.ResolveAgentIdentity(context, authToken) decodes the agent identity from a JWT — works for both OBO tokens and agentic-path tokens (SDK signature names the param generically authToken). Both paths produce the same kind of ID — what shows up in MAC Advanced Hunting.
   • Conditional baggage + token registration — only when hasObservabilityIdentity == true. Skip both calls cleanly when the identity can't be resolved:

     using IDisposable? baggageScope = hasObservabilityIdentity
         ? new BaggageBuilder()
             .TenantId(resolvedTenantId!)
             .AgentId(resolvedAgentId!)
             .Build()
         : null;
     
     if (hasObservabilityIdentity)
     {
         try
         {
             _agentTokenCache.RegisterObservability(
                 resolvedAgentId!,
                 resolvedTenantId!,
                 new AgenticTokenStruct(
                     userAuthorization: UserAuthorization,
                     turnContext: turnContext,
                     authHandlerName: authHandlerName ?? string.Empty),
                 EnvironmentUtils.GetObservabilityAuthenticationScope());
         }
         catch (Exception ex)
         {
             _logger.LogWarning(ex, "Failed to register observability token.");
         }
     }
     

     Note: Some SDK versions support object-initializer syntax instead. If the constructor form fails to compile, try property-initializer: new AgenticTokenStruct { UserAuthorization = ..., TurnContext = ..., AuthHandlerName = ... }.
   • The authHandlerName should resolve to the agentic auth handler name (from config AgentApplication:AgenticAuthHandlerName) when IsAgenticRequest() is true, OBO handler name (from AgentApplication:OboAuthHandlerName) otherwise.
   • Keep the Agent365Observability section in appsettings.json (EnableAgent365Exporter and base exporter settings are still required — Phase 6 handles these). For OBO, you do not need to hardcode per-agent IDs, tenant IDs, or S2S credentials in that section — the agent ID and tenant ID are resolved from the request at runtime on each turn.
   • The inline pattern shown above is preferred for new code (mirrors PR #308 in microsoft/Agent365-Samples). The older A365OtelWrapper.InvokeObservedAgentOperation(...) static-wrapper pattern at Agent365-samples/dotnet/agent-framework/sample-agent/telemetry/A365OtelWrapper.cs is functionally equivalent but uses a separate helper class.

   S2S path:

   • Inject Agent365ObservabilityContext (singleton registered by AddAgent365Observability()) in the constructor — not IExporterTokenCache<AgenticTokenStruct>
   • Baggage: Use new BaggageBuilder().FromTurnContext(turnContext).Build() as a separate using var baggageScope — FromTurnContext() is an extension on BaggageBuilder only; it does not exist on InvokeAgentScope or any scope type
   • Scope: Use InvokeAgentScope.Start(new Request(...), new InvokeAgentScopeDetails(endpoint: new Uri("...")), _obs.AgentDetails, callerDetails) as a separate using var scope — InvokeAgentScopeDetails has no parameterless constructor; always pass at least endpoint. CallerDetails with the blueprint sponsor's identity is required for S2S traces to appear in the portal
   • No per-turn RegisterObservability() call; no .FromTurnContext() chaining on the scope
   • Add inline comment: // A365 auth mode: S2S — FMI 3-hop chain via ObservabilityTokenService (scope: api://9b975845-388f-4429-889e-eab1ef63949c/.default)

   Mark all new lines with: // A365 Observability — best-effort instrumentation (verify against official sample)

3. Preserve all existing handler logic.

For Node.js

1. Read the detected message handler file.

2. Edit — Refresh the per-turn exporter token following the reference pattern in nodejs-observability.md:

   • Import AgenticTokenCacheInstance from @microsoft/opentelemetry (single unified package).
   • OBO paths only (obo / agentic-user): Resolve agentId and tenantId dynamically from TurnContext each turn (never from config), then refresh the exporter token (non-fatal, wrap in try/catch):

     const agentId  = turnContext.activity?.recipient?.agenticAppId ?? '';
     const tenantId = turnContext.activity?.recipient?.tenantId     ?? '';
     await AgenticTokenCacheInstance.RefreshObservabilityToken(
       agentId, tenantId, turnContext,
       agentApplication.authorization,   // ← the AgentApplication auth object, NOT an auth-handler name string
     );
     

     • obo (signed-in user): agentApplication.authorization exchanges the token as the signed-in user → traces attributed to the user
     • obo (agentic identity): agentApplication.authorization exchanges the token as the agentic user provisioned in Azure AD → traces attributed to the agent
     • Default observability scope is auto-applied (api://9b975845-388f-4429-889e-eab1ef63949c/.default) — no need to import getObservabilityAuthenticationScope (removed in 1.0).
     • Recommended pattern: Extract the agentId/tenantId resolution and token refresh into a preloadObservabilityToken(turnContext) helper function to keep the handler clean. See nodejs-observability.md for the full helper implementation.
   • S2S path: Do NOT call AgenticTokenCacheInstance.RefreshObservabilityToken — there is no user authorization token. The tokenResolver passed to useMicrosoftOpenTelemetry() (set up in Phase 3) handles authentication via the FMI 3-hop chain token service.
   • Baggage construction is done in Phase 5.5 (canonical pattern), NOT here. In Phase 5.5 the message handler builds BaggageBuilderUtils.fromTurnContext(new BaggageBuilder(), turnContext as any).build() and runs InvokeAgentScope.start(...) inside baggageScope.run(...). The configureA365Hosting(adapter, { enableBaggage: true }) middleware registered in Phase 3 is a fallback that auto-populates baggage outside the handler, but it does NOT cover the scopes you'll add in Phase 5.5 — those need the manual outer wrapping or they get filtered as Partitioned into 0 identity groups. In Phase 4 itself, just refresh the token; do NOT call InvokeAgentScope.start here.
   • Add inline comment: // A365 auth mode: {authMode} — see: https://learn.microsoft.com/en-us/entra/agent-id/agent-on-behalf-of-oauth-flow
   • Mark all new lines with: // A365 Observability — best-effort instrumentation (verify against official sample)

3. Preserve all existing handler logic.

For Python

1. Read the detected message handler file AND host_agent_server.py — the helper lives in the HOST file, not the agent class. The verified AF sample places _setup_observability_token in host_agent_server.py:130-156 so it has access to the AgentApplication instance and can be called by activity middleware. Per-turn baggage construction also lives in the handler/middleware layer in host_agent_server.py, NOT in agent.py.

2. Edit host_agent_server.py (the host file) — Refresh the per-turn exporter token following the reference pattern in python-observability.md:

   • Default observability scope is auto-applied by microsoft-opentelemetry 1.1+ — do not import get_observability_authentication_scope unless you need to override the default. If overriding, pass via a365_observability_scope_override to use_microsoft_opentelemetry. The exchange_token() call below omits scopes= and lets the auth handler resolve the default.
   • Import cache_agentic_token from token_cache (the custom module created in Phase 5) — or use AgenticTokenCache from the hosting helpers.
   • OBO paths only (obo / agentic-user): Resolve agent_id and tenant_id dynamically from context each turn (never from config), then exchange the OBO token (non-fatal, wrap in try/except):

     agent_id  = context.activity.recipient.agentic_app_id
     tenant_id = context.activity.recipient.tenant_id
     await self._setup_observability_token(context, tenant_id, agent_id)
     

     The _setup_observability_token helper exchanges and caches the token:

     async def _setup_observability_token(self, context, tenant_id, agent_id):
         exaau_token = await self.agent_app.auth.exchange_token(
             context,
             scopes=get_observability_authentication_scope(),
             auth_handler_id=self.auth_handler_name  # from config — NOT hardcoded "AGENTIC"
         )
         cache_agentic_token(tenant_id, agent_id, exaau_token.token)
     

     • auth_handler_name must come from config (e.g., AgentApplication:AgenticAuthHandlerName) — never hardcode "AGENTIC"; it is the registered auth handler name in your agent setup.
     • agentic-user (AI Teammate): the exchange returns a token for the agent's own Agentic User identity → traces attribute to the agent
     • obo (non-AI Teammate): the exchange returns whatever the configured auth handler resolves — typically the signed-in user, but it can also be the agent's own identity if the handler is configured that way
   • S2S path: Do NOT call _setup_observability_token — token comes from the background token service wired in Phase 3. The handler should NOT touch tokens.
   • Baggage: No manual baggage construction in the handler. Phase 3 registered ObservabilityHostingManager.configure(adapter.middleware_set, ObservabilityHostingOptions(enable_baggage=True)) which auto-populates baggage from TurnContext for every request. (Optional fallback if you skipped that: build manually with populate(builder, context) then with builder.build():.)
   • Add inline comment: # A365 auth mode: {authMode} — see: https://learn.microsoft.com/en-us/entra/agent-id/agent-on-behalf-of-oauth-flow
   • Mark all new lines with: # A365 Observability — best-effort instrumentation (verify against official sample)

3. Preserve all existing handler logic.

4. TaskUpdate — Mark complete.

Phase 5: Implement Agentic Token Resolver

TaskCreate — "Implement agentic token resolver with caching"

For AI Teammate agents and Standard agents on the OBO/agentic-user path, the built-in token cache handles caching automatically — no custom resolver needed. With the Microsoft.OpenTelemetry distro the cache is auto-registered by UseMicrosoftOpenTelemetry(...) (Phase 3). With the legacy individual packages it's registered explicitly via AddAgenticTracingExporter (.NET), AgenticTokenCacheInstance (Node.js), or AgenticTokenCache (Python). Skip to step 3 for these agents.

For .NET AgentFramework (hosting path)

1. The distro's builder.UseMicrosoftOpenTelemetry(...) call (Phase 3) auto-registers IExporterTokenCache<AgenticTokenStruct> in DI — no separate AddAgenticTracingExporter() call is needed. If you're on the legacy two-package wiring, AddAgenticTracingExporter() provides the same DI instance.

2. In the agent class, inject IExporterTokenCache<AgenticTokenStruct> in the constructor and call RegisterObservability(...) per turn (already done in Phase 4).

For .NET AgentFramework (S2S path)

The ObservabilityTokenService background service (created in Phase 3 via the scaffold) acquires and refreshes the Observability API token automatically via the FMI 3-hop chain (Blueprint → Agent Identity → Power Platform PFAT token) — no manual TokenResolver delegate needed.

1. Check if Observability/ObservabilityServiceExtensions.cs and Observability/ObservabilityTokenService.cs exist. If yes, skip — they were already created in Phase 3.

2. If absent (Phase 3 was skipped or re-running the skill on a partial state), create them now following the S2S scaffold patterns in dotnet-observability.md. These files provide AddAgent365Observability() (DI extension registering AddServiceTracingExporter, ObservabilityTokenService, and Agent365ObservabilityContext) and ObservabilityTokenService (background service that acquires the Observability API token via the FMI 3-hop chain and refreshes it every 50 minutes).

For Node.js (OBO path)

AgenticTokenCacheInstance from @microsoft/agents-a365-observability-hosting handles caching automatically. The useMicrosoftOpenTelemetry() call in Phase 3 wires it as the tokenResolver. No additional token resolver module is needed unless Use_Custom_Resolver=true is required (see reference doc for custom resolver pattern).

For Node.js (S2S path)

Check if observability/observability-token-service.ts exists. If yes, skip — it was created in Phase 3.

If absent (Phase 3 was skipped or re-running), create observability/token-cache.ts and observability/observability-token-service.ts now using the scaffold from nodejs-observability.md (S2S section). The token service uses MSAL (@azure/msal-node) with fmiPath to acquire tokens via the FMI 3-hop chain targeting scope api://9b975845-388f-4429-889e-eab1ef63949c/.default. Call startTokenService(config) at app startup and pass tokenResolver from the cache module to useMicrosoftOpenTelemetry().

For Python (OBO path)

The token_cache.py custom module (located at project root or observability/token_cache.py) provides cache_agentic_token and get_cached_agentic_token. The a365_token_resolver in use_microsoft_opentelemetry() (Phase 3) is wired to get_cached_agentic_token. The per-turn _setup_observability_token helper (Phase 4) calls cache_agentic_token after each OBO exchange. If token_cache.py is absent (e.g., this phase is reached before Phase 4 ran), create it now following the OBO token cache pattern in python-observability.md.

For Python (S2S path)

Check if observability/observability_token_service.py exists. If yes, skip — it was created in Phase 3.

If absent, create observability/token_cache.py and observability/observability_token_service.py now using the scaffold from python-observability.md (S2S section). The token service uses MSAL (msal.ConfidentialClientApplication) with fmi_path to acquire tokens via the FMI 3-hop chain targeting scope api://9b975845-388f-4429-889e-eab1ef63949c/.default. Call acquire_initial_token() for pre-warm, schedule run_token_service() as asyncio.create_task(), and pass token_cache.get_cached_token as the a365_token_resolver in use_microsoft_opentelemetry().

TaskUpdate — Mark complete.

Phase 5.5: Wire Manual Instrumentation Scopes

TaskCreate — "Wire InvokeAgentScope, InferenceScope, ExecuteToolScope"

Auto-instrumentation vs manual: Whether manual scopes are needed depends on authMode and whether auto-instrumentation framework extensions were installed in Phase 2.

Situation	InvokeAgentScope	InferenceScope	ExecuteToolScope
authMode = "s2s"	Required — add always	Required — add always	Required — add always
OBO + framework extension installed (Phase 2)	Required — add always	Skip — auto-instrumentation generates these	Only for local/custom tools not covered by the extension
OBO + no framework extension	Required — add always	Required — add always	Required — add always

"Autonomous" agents can run on either OBO or S2S — auth mode is the actual differentiator here, not whether the agent is autonomous.

Rule: Never skip InvokeAgentScope — it wraps the turn and is always required for traces to appear in the MAC portal. Auto-instrumentation extensions cover LLM calls (InferenceScope) and framework-managed tool calls (ExecuteToolScope), but they do not wrap the agent turn itself.

Determine which scopes to add:

• If authMode = "s2s": proceed directly — add all three scopes without prompting (required for S2S agents).
• If OBO and a framework extension was installed in Phase 2:
  • Add InvokeAgentScope always.
  • **Skip `In

…(truncated)