菜单
Procedures for filing an upstream issue against microsoft/agent-framework. Use when the toolkit's analysis cannot map a symptom to a known registry pattern — i.e., this might be a real upstream bug. Pairs with the `MafDraftIssue` MCP tool, which assembles the issue body.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Detects new MAF releases, extracts breaking changes, and keeps the toolkit's registry/matrix/guide current. Pipeline: (1) deterministic data extraction via GitHub Actions + Python helpers + dotnet-inspect, opens a PR with raw data and TODOs on a per-version branch (never commits direct to main), (2) optionally, a maintainer manually runs a sibling workflow that opens a GitHub issue for GitHub Copilot Coding Agent to fill the TODOs, (3) Copilot opens a PR with the fills. This skill documents the whole loop end-to-end.
The fix-everything playbook for the maf-remediate loop: per-rule canonical fix + how to tell a real finding from a false positive (by confidence tier). Use when driving a codebase to a clean MAF health grade — triage heuristic findings before changing code.
Machine-readable registry of known CS0618 obsolete API warnings for MAF migrations. Each entry maps an obsolete method signature to its replacement, with the exact fix pattern. Prefer the MCP tool MafRegistryLookup (by ID) or MafApiSafety (by symbol) for one-shot queries; this skill is for understanding the registry schema and contributing new entries. Load this skill when cs0618-hunter has identified a CS0618 warning and you need the deterministic fix.
Maintainer runbook for keeping maf-autopilot current with new MAF releases. Use when a new MAF version has shipped and .maf-version is stale. Covers the full loop: pre-flight check → trigger watcher → monitor PR → fill TODOs → review → merge. Distinct from maf-release-watcher (which describes what the automation does internally); this skill tells YOU what to do as the operator.
Cross-framework migration map: Semantic Kernel (C#) → Microsoft Agent Framework. Load this skill when porting an SK app to MAF (NOT a MAF version bump). It is the smart index over docs/migration/from-semantic-kernel-to-maf.md — it tells you each SK construct's MAF target and its migration strategy (🌉 bridge / 🔁 rewrite / 🏗 re-architect), so you migrate construct-by-construct, side-by-side, non-destructively.
Validates fan-out / fan-in workflow topology — detects the silent fan-in starvation pattern (handlers returning void / non-generic Task that produce no downstream message). Canonical rules + procedure for what the compiler cannot catch and the build cannot surface.
| name | maf-issue-reporter |
| description | Procedures for filing an upstream issue against microsoft/agent-framework. Use when the toolkit's analysis cannot map a symptom to a known registry pattern — i.e., this might be a real upstream bug. Pairs with the `MafDraftIssue` MCP tool, which assembles the issue body. |
| version | 1.0 |
You should file an upstream issue only when all of the following are true:
maf://registry). If MafDraftIssue returns "Registry hits" — apply the deterministic fix first; don't file.maf://constraints doesn't already cover the pattern (e.g., session state in instance fields is a known anti-pattern, not a bug).If any of those is false → don't file. Apply the existing fix, or refine the repro.
Step 1. Run MafDraftIssue(repoPath, symptom, snippet?, expected?, actual?) to assemble the body. The tool detects:
.maf-version).Microsoft.Agents.AI* / Microsoft.Extensions.AI* <PackageReference> and its pinned version.It outputs a markdown body ready to paste.
Step 2. Review the body. Strip:
C:\Users\you\company-secret\…).Step 3. Post. Two routes:
microsoft/agent-framework titled <title> with this body." The GitHub MCP server's create_issue tool handles the API call. Never use a personal access token from inside maf-autopilot directly — that's why we don't post; the github-mcp-server (or a manual paste) keeps the trust boundary right.Step 4. After posting, add the issue URL to your team's local notes. If the issue is accepted and a fix lands, the maf-release-watcher workflow will surface the new MAF version automatically — file a follow-up to add the registry entry once the fix releases.
MafValidateFanOut / MAF001 Roslyn analyzer / MAF130-FAN-IN-001 registry entry. Not a bug.DefaultAzureCredential not working in production — the MAF002 analyzer warns; the MAF-AP-SEC-001 scanner rule flags. This is configuration, not a bug.Instructions is being ignored" — check that it's inside ChatClientAgentOptions.ChatOptions (registry entry MAF-AP-AGENT-001). Top-level Instructions is silently dropped in 1.3.0.✅ "AddFanInBarrierEdge argument order does not match the docs" — this turned out to be a known issue but the upstream surface bench-fix landed because someone reported it with a clean repro.
✅ "ChatClientAgent.RunAsync returns successfully but AgentResponse.Text is empty when the underlying chat client returns a tool call" — novel pattern; report it.
❌ "MAF is broken." (No repro, no version, no specifics.)
❌ "new DefaultAzureCredential() doesn't authenticate in production." (Already known — analyzer / scanner / constraint covers it.)
The MafDraftIssue tool never posts anything. We don't take a GitHub token, we don't make HTTP calls. The user reviews + routes. This is deliberate — posting to a public org-owned repo is a "modifies shared system" action that needs an explicit human gate.
If you want hands-off filing, install the official github-mcp-server and let it own the API call. It's a separate trust boundary with its own auth model.