Add multi-agent pipeline: issue-scanner, issue-fixer, pr-verification

[YunchuWang]

Summary

Adds a three-agent Copilot pipeline modeled after the existing system in durabletask-js, adapted for the .NET SDK.

Architecture

Three collaborative agents are chained in a single GitHub Actions workflow via file-based handoff (JSON context files between sequential steps):

┌─────────────────────┐     /tmp/selected-issue.json     ┌─────────────────────┐     /tmp/pr-info.json     ┌─────────────────────┐
│   Issue Scanner     │ ──────────────────────────────▶  │   Issue Fixer       │ ────────────────────────▶ │   PR Verification   │
│   (Agent 1)         │                                   │   (Agent 2)         │                           │   (Agent 3)         │
└─────────────────────┘                                   └─────────────────────┘                           └─────────────────────┘

Agent 1: Issue Scanner (.github/agents/issue-scanner.agent.md)

• Fetches the 20 most recent GitHub issues
• Triages and classifies each issue into categories:
  • triage/actionable, triage/needs-human-verification, triage/known-blocker, triage/requires-redesign, triage/needs-info, triage/already-fixed, triage/too-large, triage/external-dependency, triage/duplicate, triage/feature-request
• Posts a triage comment on each issue with classification reasoning
• Selects the single best actionable issue for automated fixing
• Writes the issue context to /tmp/selected-issue.json for Agent 2
• If no issue qualifies, writes {\"found\": false} and the pipeline stops

Agent 2: Issue Fixer (.github/agents/issue-fixer.agent.md)

• Reads the selected issue context from Agent 1
• Deep-analyzes the codebase (source files, callers, existing tests, adjacent SDKs)
• Implements a correct, minimal fix following all repository conventions:
  • Copyright headers, XML docs, this. member access, Async suffix, sealed private classes
• Adds comprehensive tests:
  • Unit tests (xUnit + FluentAssertions + Moq)
  • Integration tests when behavioral changes are involved
• Runs the full test suite to verify
• Opens a PR linked to the issue with detailed documentation
• Writes the PR context to /tmp/pr-info.json for Agent 3

Agent 3: PR Verification (.github/agents/pr-verification.agent.md)

• Reads the PR context from Agent 2
• Checks out the PR branch and rebuilds
• Creates a standalone C# console app that reproduces the customer scenario
• Runs it against the DTS emulator (started in Docker as part of the workflow)
• Posts a structured verification report to the linked GitHub issue
• Updates PR labels (pending-verification → sample-verification-added)
• Pushes the verification sample to a verification/pr-<N> branch

GitHub Actions Workflow (.github/workflows/auto-issue-fix.yaml)

• Schedule: Daily at 09:00 UTC (+ manual trigger)
• Chaining mechanism: Sequential steps with file-based data passing — each agent writes a JSON handoff file that the workflow reads and injects into the next agent's prompt
• Deduplication: Collects existing copilot-finds PRs/issues before running to prevent duplicate work
• DTS Emulator: Started in Docker for Agent 3's verification step
• Labels: Auto-creates copilot-finds, pending-verification, sample-verification-added labels
• Summary: Posts a pipeline summary to the GitHub Actions step summary

Design Decisions

1. File-based handoff over built-in handoffs: GitHub Copilot CLI agents don't support built-in handoffs in CI context. Data is passed via JSON files written to /tmp/ and injected into prompts.
2. Sequential steps in one workflow: More reliable than event-triggered multi-workflow chains — guarantees ordering and shared build context.
3. Conservative triage: Agent 1 is intentionally conservative — it only selects issues with high confidence of correct automated fixing.
4. Idempotent verification: Agent 3 checks for existing verification before acting, preventing duplicate work across runs.
5. Separate verification branches: Verification samples are pushed to verification/pr-<N> branches, keeping main and PR branches clean.

Reference

This system is modeled after the existing multi-agent setup in durabletask-js:

• daily-code-review.agent.md → split into issue-scanner + issue-fixer for better separation of concerns
• pr-verification.agent.md → adapted for C#/.NET patterns (xUnit, dotnet build/test, DTS emulator on port 4001)
• daily-code-review.yaml → evolved into auto-issue-fix.yaml with three-agent chaining

Checklist

• Agent files follow the chatagent format from the JS repo
• Workflow uses proper GitHub Actions syntax
• .NET-specific build/test commands match existing validate-build.yml
• All agents respect repository coding conventions
• Deduplication prevents duplicate work
• DTS emulator integration for PR verification
• Pipeline summary for visibility

[Copilot]

Pull request overview

Adds a scheduled GitHub Actions workflow plus three Copilot “chatagent” prompt files to run a chained automation pipeline (issue triage → automated fix PR → emulator-based verification) for the DurableTask .NET SDK.

Changes:

• Introduces a daily/manual GitHub Actions workflow to orchestrate three sequential Copilot CLI agent runs with file-based JSON handoff in /tmp/.
• Adds agent prompt definitions for issue-scanner, issue-fixer, and pr-verification, including guidance for creating and pushing verification samples.
• Adds repo label bootstrapping and DTS emulator startup/cleanup as part of the workflow.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 6 comments.

File	Description
.github/workflows/auto-issue-fix.yaml	New scheduled workflow chaining three agents, dedup context collection, label creation, and DTS emulator orchestration.
.github/agents/issue-scanner.agent.md	New issue triage agent prompt and required JSON handoff contract.
.github/agents/issue-fixer.agent.md	New fix implementation agent prompt, testing requirements, PR creation, and PR handoff contract.
.github/agents/pr-verification.agent.md	New verification agent prompt describing sample creation, emulator execution, evidence posting, and label updates.

---

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 9 comments.

---

You can also share your feedback on Copilot code review. Take the survey.