# Troubleshoot Observability

> Coordinate Azure observability troubleshooting for application or agent health, runtime failures, operational latency, telemetry, monitoring, alerts, evidence-backed fixes and recovery. Use for an end-to-end observability investigation or explicit instrumentation, collection or alerting guidance. Not general resource/app/agent creation, model deployment, RBAC, cost or coding. Conceptual questions need no journey; explicit specialist requests stay direct. Actions require applicable approval.

- Skill: `microsoft/troubleshoot-observability` (Agent Skill, multi-file: 25 files)
- Install (CLI): `npx get-skill add microsoft/troubleshoot-observability`
- Raw SKILL.md: https://api.skillmd.com/api/skills/microsoft/troubleshoot-observability/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- License: MIT
- Author: Microsoft (https://skillmd.com/u/microsoft)
- Updated: 2026-11-18
- Page: https://skillmd.com/skills/microsoft/troubleshoot-observability

---


# Azure Observability skill plugin

Copilot chooses the next useful step; available tools execute approved operations.
This is instruction-guided routing, not a diagnostics engine or permission sandbox.

## Current-request intent gate

On every user turn, before any journey header, diagram, discovery, onboarding
question or read, confirm a current operational observability request. "My API
is slow", an actual failure (including a failed deployment), or explicit
instrumentation/collection/alert design qualifies, including before deployment.
Conceptual questions get direct answers. Generic coding, creation, deployment
readiness, cost and unrelated topic switches get no journey, tools or onboarding.
Prior context is not sticky; clarify only genuinely ambiguous operational intent
with one minimal question. Explicit specialist requests stay direct.
Read [intent details](references/intent.md) only for an ambiguous admission decision.

## Standalone availability gate

For an admitted task, check availability before reading any cross-skill reference
or loading a sibling. Individual installation may omit them.
Use available core instructions in this bundle; their use below is conditional.
Unknown availability follows the missing-reference fallback.
Never download, install or attempt a mandatory missing sibling-file read.
Retain scope/window, evidence and budgets; report the precise handoff blocker and
smallest supported manual next step. Do not claim an absent specialist executed.
Canonical `azure-diagnostics` and `microsoft-foundry` are external dependencies:
discover them through the host under [dependency selection](references/canonical-dependencies.md),
never by assumed sibling paths. Missing external instructions do not block
the local evidence-only affected-path workflow or imply permission to install.
For agent/model/tool failures or AI-specific analysis, check the host's exposed
skills and load enabled, relevant `microsoft-foundry` instructions before
choosing a core fallback. Not bundled does not mean unavailable. Follow the
dependency selection checkpoint; unknown discovery is not confirmed absence.

## Select one path

Load only the applicable guidance below, not every reference.

| Current need | Instructions to select |
|---|---|
| Find telemetry or resolve missing target/source | Available `connect-observability` |
| Broad health | Available `assess-observability-health` |
| Completed ranked health, no finding selected | Ask for selection; never auto-select rank 1 or reassess |
| Named alert, failure, slow operation or selected finding | [Bounded affected-path investigation](references/investigation/README.md); host-discover canonical `azure-diagnostics` for service-specific expertise or `microsoft-foundry` for AI-specific expertise when needed |
| Missing scope/access or a live-read plan | [Scope and permissions](references/scope-permissions.md) |
| Specialist handoff or additional capability need | [Selection](references/selection.md), then only the relevant matching/setup/trace contract |
| Evidenced issue: what should I do? | [Next actions](references/next-actions.md): supported fix, mitigation, PR/config proposal or alert/collection prevention |
| Effective change: did it work? | Available `verify-observability-fix` with supplied comparable evidence |
| Recovery verdict: how do I keep watching? | [Monitoring follow-up](references/monitoring-follow-up.md) |
| Explicit setup/design | Applicable available specialist; no invented gap or implicit provisioning |

## Boundaries and presentation

Reuse confirmed source/schema, UTC window, selected findings and consumed/remaining
budgets. Supplied evidence needs no cloud, installation, scripts, jobs or saves.
Discover actual operations and access; no guessed tool names or topology.
Missing/denied/empty/stale/partial/failed/throttled evidence remains distinct;
insufficient evidence is INCONCLUSIVE, and correlation is not causation.
Reads, sensitive content, connections/sign-in, edits/persistence, jobs/traffic
and deployment/alert/infrastructure changes retain separate applicable approvals.
A "continue" retains only covered unexpired consent; honor stops and declines.

At journey entry, before a scope/approval question or specialist handoff, read
the [indicator](references/journey-indicator.md) and show its fenced Mermaid
stepper below actual **Using:** / **Stage:**. Telemetry mapping stays within
Detect or Investigate; it does not postpone presentation. Repeat only at meaningful
transitions. On reported rendering failure use its text fallback, not an SVG file.
Locally inspected guidance uses **Instructions read:** and **host activation
unverified**. Show **Discovered:** only for relevant capabilities actually found;
discovery/loading is not execution.
Apply [next actions](references/next-actions.md) at a substantive result or plan
review: explain missing source or the exact next action and separate approval.
Report planned tools separately from actual outcomes, skips and fallback.

Remediate means options and a reviewed plan, not automatic changes. Validate
means recovery comparison first, then optional monitoring; dashboards, quiet
alerts, deployment success and better evaluation scores are not recovery proof.
No dashboard or monitoring setup is required. Runtime/host behavior remains
unverified by offline instruction tests.

