Work Item Workflow

Work Source Detection

Classify the work source before any other action. Resolve in this priority order:

1. GitHub Issue or Azure DevOps Work Item if a numeric ID is referenced anywhere in the user's message, the recent conversation, or the tasks.md being implemented (e.g., "implement issue #42", "#15", "work item 1234", "task 5678", or a tasks.md row that links to #1690). Any such reference activates board mode for the entire implementation, even if the user phrases the request as "implement the Terraform task" or "continue the work".
2. tasks.md only if no issue or work item ID is reachable from the current context. Skip this skill.

When in doubt between the two, choose board mode and read the work item. Reading is read-only and cheap; silently skipping board updates is the failure this skill exists to prevent.

Reading the Work Item

1. Identify platform: Read .memory/board-config.md or detect from remote (git config --get remote.origin.url)
2. Get current user:
   • GitHub: gh api user --jq '.login'
   • Azure DevOps: az account show --query user.name -o tsv or ask the user
3. Read work item via MCP: Extract title, description, assignee, acceptance criteria, state, and tags/labels

Required Checks

Execute in order. If any check blocks, STOP and inform the user.

1. Assignee check

When the work item has no assignee, do not skip this step. Execute the assignment before moving to the State update check:

• GitHub: Call github/issue_write with method: "update", the issue number, and assignees: ["<current-user-login>"] (the login obtained in step "Get current user").
• Azure DevOps: Call ado/wit_update_work_item to set the System.AssignedTo field to the current user.

If the assignment call fails (permission error, user lookup empty, API error), STOP and report the failure rather than silently proceeding to later checks. The downstream flow (state update, branch creation, PR linkage) assumes the work item is owned by the current user.

2. State update

State transitions are mandatory, not advisory. If the board MCP is unreachable, surface a blocking error rather than skipping the transition.

• Move the task to the active state:
  • GitHub: Add label status:in-progress
  • Azure DevOps: Change State per process template (invoke board-config skill for details; typical: New to Active)
• Check the parent User Story:
  • If the parent is not yet in the active state, update it as well
  • Inform: "User Story #[ID] moved to In Progress."
• Confirm the transition succeeded by re-reading the work item and verifying the new state. If the read shows the old state, retry once and then surface the failure to the user before any code is written.

3. Dependency check

• GitHub: Look for "Depends on: #NNN" in the body
• Azure DevOps: Look for links of type "Predecessor"
• For each unfinished dependency (open issue / state != Done|Closed):

  This task has unfinished dependencies:
  
  - #[ID]: [title] (state: [state])
  
  [V] View dependency details
  [O] Work on another task
  [I] Ignore and continue anyway (not recommended)

4. Priority check

• Search for available tasks (no assignee, same feature)
• If there are tasks with higher priority available:

  There are higher priority tasks available:
  
  - #[ID]: [title] (P1)
  
  Requested task: #[ID] (P2)
  
  [P] Pick the higher priority task (recommended)
  [C] Continue with the requested task

5. Capacity check

• Search for in-progress tasks assigned to the current user
• If the user already has 3 or more tasks in progress:

  You already have [N] tasks in progress:
  
  - #[ID]: [title]
  
  [F] Finish a task before picking up another (recommended)
  [C] Continue and pick up another task

tasks.md (when no issue/work item is mentioned)

1. Identify the current feature by checking the docs/features/ directory. If the user specified a feature, use it.
2. Read tasks.md for the task list and execution plan
3. Proceed with the implementation flow

Main Execution Rule

Developers should work on only one task at a time. Work happens at the task level, not at the feature, user story, or epic level directly.

Board Synchronization

When synchronizing with the board, the agent should:

1. List ALL issues/work items in the repository/project
2. Identify feature-related issues using multiple strategies:
   • By label/tag: feature:<name>, type:user-story, type:task
   • By title: contains the feature name or keywords
   • By content: mentions files or concepts from the plan
3. Classify found issues:
   • With copilot-generated label: Issues created by the agent (managed)
   • Without labels, related: Manual issues that appear related
   • Without labels, unrelated: Issues from other features/contexts
4. For unlabeled issues that appear related, ask the user:
   • Add labels and incorporate into the plan
   • Ignore (not from this feature)
   • Mark as duplicate of a planned task
5. Map current state: User Stories and Tasks (open, in progress, closed)
6. Present summary to user with comparison between board and plan

Cascade Closure

When synchronizing with the board, check and propose cascade closures:

Closure Rules

• Always ask before closing (do not close automatically)
• Show summary of what will be closed
• Validate that there are no orphan tasks/US before closing

Work Item Hierarchy

GitHub

• Use sub-issues for parent-child hierarchy
• For dependencies between tasks, use the "Dependencies" section in the body: Depends on: #123

Azure DevOps

• Parent-Child: use ado/wit_add_child_work_items
• Dependencies: use ado/wit_work_items_link with link type Predecessor/Successor