microsoft/hve-core
Publicmirrored fromhttps://github.com/microsoft/hve-coreAvailable
docs/agents/github-backlog/why-backlog-manager.md
80lines · modecode
| 1 | --- |
| 2 | title: Why the Backlog Manager Works |
| 3 | description: Design principles and cognitive foundations behind the GitHub Backlog Manager workflow separation |
| 4 | sidebar_position: 2 |
| 5 | author: Microsoft |
| 6 | ms.date: 2026-05-20 |
| 7 | ms.topic: concept |
| 8 | keywords: |
| 9 | - github backlog manager |
| 10 | - workflow design |
| 11 | - github copilot |
| 12 | - backlog management |
| 13 | estimated_reading_time: 6 |
| 14 | --- |
| 15 | |
| 16 | Backlog management looks simple from the outside: read issues, assign labels, close duplicates. In practice, teams struggle with it because the work combines several cognitively different tasks into one undifferentiated session. The GitHub Backlog Manager addresses this by separating those tasks into focused workflows, each designed for one type of thinking. |
| 17 | |
| 18 | ## The Core Insight |
| 19 | |
| 20 | Discovering issues, classifying them, planning their execution, and applying changes require different mental models. Discovery is exploratory and divergent. Triage is analytical and convergent. Sprint planning is strategic and forward-looking. Execution is mechanical and precise. |
| 21 | |
| 22 | Combining these in a single pass forces constant context-switching between exploration, analysis, strategy, and action. The result is inconsistent labels, missed duplicates, and milestones that don't reflect actual priorities. |
| 23 | |
| 24 | The backlog manager solves this by giving each cognitive mode its own workflow, its own session, and its own output artifacts. You focus on one type of thinking at a time, and structured handoff files carry context forward without requiring you to hold it all in memory. |
| 25 | |
| 26 | ## How Each Workflow Helps |
| 27 | |
| 28 | Discovery narrows the aperture. Instead of staring at a full issue list, you define what you're looking for (your assignments, issues related to a branch, issues matching specific criteria) and get back a structured inventory. The analysis file captures what was found and why, so triage starts with organized input rather than raw data. |
| 29 | |
| 30 | Triage applies consistent classification. Working from discovery output rather than live issue lists means every issue gets evaluated against the same taxonomy in the same pass. Duplicate detection works better when issues are compared in batches rather than individually, because patterns only emerge when you see the full set. |
| 31 | |
| 32 | Sprint planning builds on classified data. With labels and duplicates resolved, milestone assignment becomes a mapping exercise rather than a judgment call. The workflow can reason about capacity and priority because triage has already done the classification work. |
| 33 | |
| 34 | Execution applies changes mechanically. By the time you reach execution, every change has been reviewed and approved in a handoff file. The workflow processes checkboxes, not decisions. This separation means bulk changes are safe because the decision-making happened in earlier phases with full context. |
| 35 | |
| 36 | ## Quality Comparison |
| 37 | |
| 38 | | Aspect | Manual Process | Managed Pipeline | |
| 39 | |----------------------|---------------------------------------------|-------------------------------------------------| |
| 40 | | Label consistency | Varies by who triages and when | Same taxonomy applied in every pass | |
| 41 | | Duplicate detection | Relies on memory and search skills | Systematic comparison across four dimensions | |
| 42 | | Milestone assignment | Often deferred or forgotten | Structured recommendations with capacity checks | |
| 43 | | Audit trail | Issue history only | Planning files, handoff logs, execution logs | |
| 44 | | Recovery from errors | Undo individual changes manually | Re-run execution; completed items are tracked | |
| 45 | | Time per issue | Decreases with fatigue during long sessions | Consistent because each workflow is short | |
| 46 | |
| 47 | ## Learning Curve |
| 48 | |
| 49 | The backlog manager is designed for progressive adoption: |
| 50 | |
| 51 | 1. Start with discovery alone to survey your backlog without changing anything |
| 52 | 2. Add triage when you want consistent labeling across issues |
| 53 | 3. Introduce sprint planning when milestones and priorities become important |
| 54 | 4. Use execution when you're comfortable with the handoff review process |
| 55 | |
| 56 | Each workflow is useful independently. You don't need to adopt the full pipeline to get value from individual workflows. |
| 57 | |
| 58 | > [!TIP] |
| 59 | > Most teams start with discovery and triage, adding sprint planning and execution as confidence grows. There is no requirement to use all four workflows together. |
| 60 | |
| 61 | ## Choosing Your Approach |
| 62 | |
| 63 | The backlog manager supports three autonomy levels. Choose based on your comfort with automated changes and the sensitivity of your repository: |
| 64 | |
| 65 | | Level | Discovery | Triage | Sprint Planning | Execution | |
| 66 | |---------|-----------|-----------|-----------------|-----------| |
| 67 | | Full | Automatic | Automatic | Automatic | Automatic | |
| 68 | | Partial | Automatic | Review | Automatic | Review | |
| 69 | | Manual | Automatic | Review | Review | Review | |
| 70 | |
| 71 | Full autonomy suits repositories where the cost of a mislabeled issue is low and velocity matters most. Manual control fits repositories where every change needs human approval. Partial autonomy balances speed with oversight by requiring review at the points where judgment matters most: classification and change application. |
| 72 | |
| 73 | The right level depends on your repository, not on the tool. Start with manual control and increase autonomy as you verify the workflow produces reliable results for your specific backlog. |
| 74 | |
| 75 | --- |
| 76 | |
| 77 | <!-- markdownlint-disable MD036 --> |
| 78 | *🤖 Crafted with precision by ✨Copilot following brilliant human instruction, |
| 79 | then carefully refined by our team of discerning human reviewers.* |
| 80 | <!-- markdownlint-enable MD036 --> |
| 81 | |