Declarative dependency-aware task-graph with parallel execution and automatic conflict resolution for plugin sub-agents

[sophiaso]

Describe the feature or problem you'd like to solve

Declarative dependency-aware task-graph with parallel execution and automatic conflict resolution for plugin sub-agents

Proposed solution

This is a follow-up to @github/copilot-cli/issues/3568 (which was closed and I don't have permission to reopen).

Add a built-in, declarative task-graph schema to Copilot CLI plugins, allowing plugin authors to define tasks, their dependencies/order, and conflict resolution strategies. Copilot CLI would then:

• Parse the dependency graph
• Run parallelizable tasks concurrently and sequential ones in order
• Handle conflicts between parallel tasks (e.g., serialize, merge, or rerun as needed)
• Return structured per-task results to support conditional workflows

This would reduce boilerplate, simplify orchestrator plugins, and improve reliability for complex, multi-step scenarios.

Example prompts or workflows

A scenario for .NET repo upgrades:

1. Upgrade .NET (must finish first)
2. Fix vulnerability A, B, C (run in parallel, but could conflict on shared files)
3. Deploy (wait for all fixes to succeed)

Sample YAML task-graph schema:

tasks:
  - id: upgrade
    agent: dotnet-upgrade-agent
    prompt: "Upgrade the project from net6 to net8"
  - id: fix-a
    agent: security-agent
    prompt: "Fix vulnerability CVE-A"
    dependsOn: [upgrade]
  - id: fix-b
    agent: security-agent
    prompt: "Fix vulnerability CVE-B"
    dependsOn: [upgrade]
  - id: fix-c
    agent: security-agent
    prompt: "Fix vulnerability CVE-C"
    dependsOn: [upgrade]
  - id: deploy
    agent: deploy-agent
    prompt: "Deploy the updated app to Azure"
    dependsOn: [fix-a, fix-b, fix-c]
on_conflict: serialize   # e.g. serialize | merge | rerun-affected

Benefits:

• Plugin authors describe dependencies instead of coding their own dispatch engine.
• Copilot CLI resolves parallelism and conflicts for them.
• Structured per-task results returned (pass/fail, output, logs, etc.).
• Enables orchestrator-style plugins for batch refactor, multi-repo actions, etc.

Additional context

Current plugin task tool exposes parallel dispatch, but still requires imperative handling of ordering, conflict resolution and result management. Is a declarative, dependency-aware task-graph (with automatic parallelization + conflict resolution + structured results) something you'd consider?