twig-sdlc

End-to-End SDLC Workflow

Orchestrated SDLC pipeline powered by the conductor skill. Takes an ADO work item (Epic or Issue) or a natural language prompt and drives it through planning, implementation, code review, PR management, and close-out — all via multi-agent orchestration.

This workflow is long-running — typically 30-120+ minutes depending on scope. Always launch with conductor run ... --web (no --silent) in a hidden window — this captures full agent output in logs for post-mortem analysis while the web dashboard provides real-time monitoring. Do NOT use --web-bg — it does not work correctly; always use --web.

Always launch detached — use Start-Process -WindowStyle Hidden so conductor survives if the parent session drops and doesn't clutter the terminal. Non-detached async shells die when the Copilot CLI session reconnects or terminates.

Workflows

All workflows are registered in the twig conductor registry. Use short names — no absolute paths needed.

Workflow source: PolyphonyRequiem/twig-conductor-workflows Install: conductor registry add twig --source github://PolyphonyRequiem/twig-conductor-workflows Update: conductor registry update twig

Workflow Inputs

pr_owner and pr_repo_name are required. They tell the workflow which GitHub repository to target for PR creation, review, and merge. Omitting them causes the workflow to fail at the PR lifecycle stage. For twig, always pass pr_owner=PolyphonyRequiem and pr_repo_name=twig.

skip_plan_review is removed — use intent=resume instead (default behavior). For new runs from a prompt, intent=new is inferred automatically.

Workflow Metadata

All metadata is passed dynamically via --metadata / -m flags at invocation time. The workflow YAMLs contain no metadata — the invoking agent resolves all values.

All values must be resolved — templates with {braces} are skipped by the dashboard.

Quick Reference

Always run in a dedicated worktree — never on main or your working tree. Name the worktree twig2-<ID> based on the work item ID.

# 1. Create a worktree for the work item
git worktree add -b sdlc/<ID> ../twig2-<ID> main
cd ../twig2-<ID>

# 2. Restore dependencies (worktrees don't share NuGet packages)
dotnet restore

# 3. Copy .twig workspace and set context to the target work item
Copy-Item -Recurse ../twig2/.twig .twig
twig set <ID>
twig sync

# 4. Run the full SDLC — default intent is "resume"
conductor run twig-sdlc-full@twig --input work_item_id=<ID> --input pr_owner=PolyphonyRequiem --input pr_repo_name=twig -m tracker=ado -m project_url=https://dev.azure.com/dangreen-msft/Twig -m git_repo=C:\Users\dangreen\projects\twig2 -m workitem_id=<ID> -m worktree_name=twig2-<ID> -m cwd=C:\Users\dangreen\projects\twig2-<ID> --web

# For a brand new work item (no prior work):
conductor run twig-sdlc-full@twig --input work_item_id=<ID> --input intent=new --input pr_owner=PolyphonyRequiem --input pr_repo_name=twig -m tracker=ado -m project_url=https://dev.azure.com/dangreen-msft/Twig -m git_repo=C:\Users\dangreen\projects\twig2 -m workitem_id=<ID> -m worktree_name=twig2-<ID> -m cwd=C:\Users\dangreen\projects\twig2-<ID> --web

# To redo from scratch (deletes existing children/branches):
conductor run twig-sdlc-full@twig --input work_item_id=<ID> --input intent=redo --input pr_owner=PolyphonyRequiem --input pr_repo_name=twig -m tracker=ado -m project_url=https://dev.azure.com/dangreen-msft/Twig -m git_repo=C:\Users\dangreen\projects\twig2 -m workitem_id=<ID> -m worktree_name=twig2-<ID> -m cwd=C:\Users\dangreen\projects\twig2-<ID> --web

Launching Multiple Runs

When running multiple work items, use discrete worktrees and launch 10 seconds apart.

Option A: Direct launch with logging (recommended)

Simpler approach — launches conductor directly with Tee-Object for log capture. Dashboard URLs are written to conductor.log in each worktree.

# Create worktrees (NEVER on main — always a dedicated branch)
git worktree add -b sdlc/1673 ../twig2-1673 main
git worktree add -b sdlc/1782 ../twig2-1782 main

# Launch 10s apart with log capture
$ids = 1673, 1782
foreach ($id in $ids) {
    $wt = "C:\Users\dangreen\projects\twig2-$id"
    Start-Process -FilePath "pwsh" -ArgumentList "-NoProfile", "-Command",
        "cd $wt; conductor run twig-sdlc-full@twig --input work_item_id=$id --input pr_owner=PolyphonyRequiem --input pr_repo_name=twig -m tracker=ado -m project_url=https://dev.azure.com/dangreen-msft/Twig -m git_repo=C:\Users\dangreen\projects\twig2 -m workitem_id=$id -m worktree_name=twig2-$id -m cwd=$wt --web 2>&1 | Tee-Object -FilePath $wt\conductor.log" `
        -WindowStyle Hidden -PassThru | ForEach-Object { "Launched #$id — PID $($_.Id)" }
    if ($id -ne $ids[-1]) { Start-Sleep -Seconds 10 }
}

# Retrieve dashboard URLs after ~15s
Start-Sleep -Seconds 15
foreach ($id in $ids) {
    Get-Content "C:\Users\dangreen\projects\twig2-$id\conductor.log" |
        Select-String "Dashboard:" | Select-Object -First 1
}

Option B: Job Object wrapper (orphan cleanup)

Use tools/run-conductor.ps1 when orphaned processes are a concern (MCP servers, test runners surviving after conductor exits). The wrapper creates a Windows Job Object that kills all child processes when conductor exits.

⚠️ Argument quoting is critical — the -Arguments value must be wrapped in escaped quotes so Start-Process passes it as a single token to the script parameter.

$ids = 1673, 1782
foreach ($id in $ids) {
    $wt = "C:\Users\dangreen\projects\twig2-$id"
    $args = "run twig-sdlc-full@twig --input work_item_id=$id --input pr_owner=PolyphonyRequiem --input pr_repo_name=twig -m tracker=ado -m project_url=https://dev.azure.com/dangreen-msft/Twig -m git_repo=C:\Users\dangreen\projects\twig2 -m workitem_id=$id -m worktree_name=twig2-$id -m cwd=$wt --web"
    Start-Process -FilePath "pwsh" -ArgumentList "-NoProfile", "-File",
        "tools\run-conductor.ps1",
        "-WorkingDirectory", "`"$wt`"",
        "-Arguments", "`"$args`"" `
        -WindowStyle Hidden
    if ($id -ne $ids[-1]) { Start-Sleep -Seconds 10 }
}

The 10-second stagger avoids MCP server port collisions and rate-limit spikes.

Always share the dashboard URL — after launching, provide the user the web dashboard URL from terminal output.

Worktree Cleanup

After runs complete, clean up worktrees:

# Remove worktrees and branches
git worktree remove --force ../twig2-1673
git branch -D sdlc/1673

# If directories are locked, kill lingering processes first:
dotnet build-server shutdown
Get-CimInstance Win32_Process | Where-Object { $_.CommandLine -match 'twig2-\d+' } |
    ForEach-Object { Stop-Process -Id $_.ProcessId -Force }

Phases

Phase 0: Preflight Validation

Validates external dependencies before expensive LLM planning begins. Catches broken auth, missing tooling, and connectivity failures early — saving minutes of wasted Opus inference that would otherwise fail opaquely in a downstream agent.

Entry Points

Full preflight runs as the preflight_check script node in twig-sdlc-full.yaml. On failure, it routes to the preflight_gate human gate where the user can retry, proceed anyway (advisory-only failures), or abort.

Lite preflight runs as the preflight_lite script node in sub-workflows. This catches the most common failures (gh auth, git repo, ADO connectivity) without repeating the full 12-check suite. On failure, it routes to preflight_lite_gate.

Required vs. Advisory Checks

Checks are classified into two categories:

• **Required

Content truncated.