Improve vmr-codeflow-status skill: current-state-first analysis, force push & empty diff detection

[lewing]

This improves the vmr-codeflow-status Copilot CLI skill with enhancements discovered while investigating a stale codeflow PR (#124095).

Design Change: Current State First

Inspired by the code-review skill restructuring in #124229, the analysis now follows a "current state first, comments as history" pattern:

1. Assess current state from primary signals (PR state, diff size, force pushes, timeline) before reading any Maestro comments
2. Form an independent verdict (MERGED / CLOSED / NO-OP / IN PROGRESS / STALE / ACTIVE)
3. Then read comments as historical context to explain how we got to the current state
4. Recommendations are generated by the agent from a structured JSON summary, not hardcoded script logic

Previously, the script read Maestro conflict/staleness comments first and drew conclusions from them — even when those comments were stale (e.g., after someone had already force-pushed a resolution). Recommendations were a 130-line if/elseif chain that couldn't adapt to edge cases.

New Capabilities

1. Current State Assessment (Step 2)

Synthesizes primary signals into an immediate verdict before any comment parsing:

• ✅ MERGED — PR has been merged, no action needed
• ✖️ CLOSED — PR was closed without merging, Maestro should create a replacement
• 📭 NO-OP — empty diff, likely already resolved
• 🔄 IN PROGRESS — recent force push, awaiting update
• ⏳ STALE — no activity for >3 days
• ✅ ACTIVE — PR has content and recent activity

2. Force Push Detection

Queries PR timeline for head_ref_force_pushed events, showing who force-pushed and when. Uses --slurp for correct pagination handling.

3. Empty Diff Detection

Checks changedFiles/additions/deletions to flag 0-change PRs as NO-OP (regardless of whether a force push caused it).

4. Post-Action Staleness Analysis

Cross-references force push timestamps against conflict/staleness warnings. When a force push post-dates these warnings, the script correctly identifies the issues as potentially resolved.

5. JSON Summary + Agent-Generated Recommendations

The script's hardcoded recommendations (130+ lines of branching logic) have been replaced with:

• A [CODEFLOW_SUMMARY] JSON block emitted at the end of the script with all key facts
• isCodeflowPR boolean for explicit codeflow vs non-codeflow classification
• A "Generating Recommendations" section in SKILL.md that teaches the agent how to reason over the summary
• The agent now synthesizes contextual, nuanced recommendations instead of picking from a canned decision tree

6. Codeflow History (renamed section)

The former "Staleness & Conflict Check" is now "Codeflow History" with a framing line: "Maestro warnings (historical — see Current State for present status)". This signals that comments describe past events, not necessarily current state.

Testing

• Verified against PR [main] Source code updates from dotnet/dotnet #124095 (closed backflow with conflicts + force push) and PR Improve vmr-codeflow-status skill: current-state-first analysis, force push & empty diff detection #124231 (non-codeflow PR)
• Multi-model review round 1 (Claude Sonnet 4, GPT-5): consensus findings → added merged/closed state handling, empty-diff-only NO-OP, --slurp pagination fix
• Multi-model test round 2 (Claude Sonnet 4, GPT-5): functional test of JSON summary → fixed isUpToDate null handling, negative daysSinceUpdate, added isCodeflowPR field, clean exit codes, non-codeflow guidance in SKILL.md

[Copilot]

Pull request overview

Updates the vmr-codeflow-status Copilot CLI skill to better diagnose codeflow PR state transitions by detecting force-pushes, recognizing empty/no-op PR diffs, and refining recommendations when warnings have likely already been acted on.

Changes:

• Adds PR “empty diff” detection using changedFiles/additions/deletions from gh pr view.
• Queries the PR timeline for head_ref_force_pushed events and reports actor/time/SHA.
• Cross-references force-push timestamps against conflict/staleness warning timestamps to adjust recommendations (merge/close/force-trigger vs resolve-conflict).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
.github/skills/vmr-codeflow-status/scripts/Get-CodeflowStatus.ps1	Adds force-push timeline querying, empty-diff detection, and post-action staleness/conflict analysis to refine recommendations.
.github/skills/vmr-codeflow-status/SKILL.md	Documents the new force-push and empty-diff detection behavior and adds guidance for “Maestro may be stuck” scenarios.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

.github/skills/vmr-codeflow-status/scripts/Get-CodeflowStatus.ps1:1523

• The closing-brace comment # end of else (non-terminal state) appears to annotate the wrong else block (it actually closes the $issues.Count -eq 0 branch). Consider updating/removing the comment (and/or re-indenting the nested blocks) to avoid confusion when modifying this section later.

        Write-Host "    1. Wait — Maestro should auto-update the PR" -ForegroundColor White
        Write-Host "    2. Trigger manually — if auto-updates seem delayed" -ForegroundColor White
        if ($subscriptionId) {

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

.github/skills/vmr-codeflow-status/scripts/Get-CodeflowStatus.ps1:176

• In Get-CodeflowPRHealth, clearing HasStaleness depends on successfully fetching/parsing gh pr checks (the commit-date comparison is nested under the $checksJson block). If gh pr checks fails/returns nothing (auth/rate limit), staleness will remain flagged even when the PR has commits newer than the last staleness warning. Consider running the “last commit after last warning” check independently of gh pr checks (and independently of mergeability), using the already-fetched comments + a gh pr view --json commits call guarded with try/catch/exit-code checks.

        $checksJson = gh pr checks $PRNumber -R $Repo --json name,state 2>$null
        if ($LASTEXITCODE -eq 0 -and $checksJson) {
            try {
                $checks = ($checksJson -join "`n") | ConvertFrom-Json
                $codeflowCheck = @($checks | Where-Object { $_.name -match 'Codeflow verification' }) | Select-Object -First 1
                if (($codeflowCheck -and $codeflowCheck.state -eq 'SUCCESS') -or $isMergeable) {
                    # No merge conflict — either Codeflow verification passes or PR is mergeable
                    $hasConflict = $false
                    # For staleness, check if there are commits after the last staleness warning
                    if ($hasStaleness) {
                        $commitsJson = gh pr view $PRNumber -R $Repo --json commits --jq '.commits[-1].committedDate' 2>$null
                        if ($LASTEXITCODE -eq 0 -and $commitsJson) {
                            $lastCommitTime = ($commitsJson -join "").Trim()
                            $lastWarnTime = $null
                            foreach ($comment in $prDetail.comments) {
                                if ($comment.author.login -match '^dotnet-maestro' -and $comment.body -match 'codeflow cannot continue|the source repository has received code changes') {
                                    $warnDt = [DateTimeOffset]::Parse($comment.createdAt).UtcDateTime
                                    if (-not $lastWarnTime -or $warnDt -gt $lastWarnTime) {
                                        $lastWarnTime = $warnDt
                                    }
                                }
                            }
                            $commitDt = if ($lastCommitTime) { [DateTimeOffset]::Parse($lastCommitTime).UtcDateTime } else { $null }
                            if ($lastWarnTime -and $commitDt -and $commitDt -gt $lastWarnTime) {
                                $hasStaleness = $false
                            }
                        }