@slamb2k/mad-skills 2.0.48 → 2.0.50

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "mad-skills",
   "description": "AI-assisted planning, development and governance tools",
-  "version": "2.0.48",
+  "version": "2.0.50",
   "author": {
     "name": "slamb2k",
     "url": "https://github.com/slamb2k"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@slamb2k/mad-skills",
-  "version": "2.0.48",
+  "version": "2.0.50",
   "description": "Claude Code skills collection — full lifecycle development tools",
   "type": "module",
   "repository": {

package/skills/brace/SKILL.md

@@ -308,17 +308,23 @@ unknown). Platform detection and all CLI/REST commands are in
 1. Detect platform from `git remote get-url origin`
 2. **GitHub:** Check existing protection via `gh api`. If unprotected, ask via
    AskUserQuestion:
-   - "Yes, require PR reviews (Recommended)" — require 1 approval, block force push
+   - "Yes, require PR reviews (team project — Recommended)" — require 1 approval, block force push + deletion
+   - "Yes, protect branch without review requirement (solo project)" — block force push + deletion, no reviewer gate so `/ship` squash-merges without `--admin`
    - "Skip" — leave unprotected
 3. **Azure DevOps:** Extract org/project from remote URL. Check existing
    policies via `az repos` CLI or REST fallback. If no minimum reviewer
-   policy, ask via AskUserQuestion (same options as GitHub).
+   policy, ask via AskUserQuestion:
+   - "Yes, require PR reviews (team project — Recommended)" — 1 approver, creator votes do not count
+   - "Yes, PR required, author can self-approve (solo project)" — same approver policy but `creator-vote-counts=true` so the author's own vote satisfies the gate
+   - "Skip" — no policy applied
 4. **Unknown platform:** Skip and report.
 
-If user accepts, apply using the procedures in
-`references/branch-protection-steps.md`.
+Apply the variant matching the user's choice using the procedures in
+`references/branch-protection-steps.md` (see "Apply protection — team
+project" vs "Apply protection — solo project" for each platform).
 
-Include result in the final report under a "🔒 Branch protection" section.
+Include result in the final report under a "🔒 Branch protection" section,
+labelled with the chosen variant (team / solo / skipped).
 
 ---
 

package/skills/brace/references/branch-protection-steps.md

@@ -30,7 +30,10 @@ gh api repos/{owner}/{repo}/branches/{default_branch}/protection
 ```
 404 = unprotected.
 
-### Apply protection
+### Apply protection — team project (reviewer gate)
+
+Requires 1 approving review before merge. Use when multiple humans work on
+the repo.
 
 ```bash
 gh api repos/{owner}/{repo}/branches/{default_branch}/protection \
@@ -42,6 +45,34 @@ gh api repos/{owner}/{repo}/branches/{default_branch}/protection \
   -F allow_deletions=false
 ```
 
+### Apply protection — solo project (no reviewer gate)
+
+Keeps force-push and deletion prevention but drops the reviewer
+requirement. Use when you are the only contributor — this is what makes
+`/ship` squash-merge succeed without `gh pr merge --admin`.
+
+The only difference from the team variant is the absence of
+`-f required_pull_request_reviews=...`:
+
+```bash
+gh api repos/{owner}/{repo}/branches/{default_branch}/protection \
+  -X PUT \
+  -f enforce_admins=false \
+  -f restrictions=null \
+  -f required_status_checks=null \
+  -F allow_force_pushes=false \
+  -F allow_deletions=false
+```
+
+**Migrating an existing team-project repo to solo:** if a reviewer policy
+is already in place, remove just that sub-rule without tearing down the
+rest of the protection:
+
+```bash
+gh api --method DELETE \
+  repos/{owner}/{repo}/branches/{default_branch}/protection/required_pull_request_reviews
+```
+
 ---
 
 ## Azure DevOps
@@ -94,7 +125,10 @@ curl -s -H "$AUTH" \
   | jq "[.value[] | select(.settings.scope[]?.refName == \"refs/heads/$default_branch\" and .settings.scope[]?.repositoryId == \"$REPO_ID\")]"
 ```
 
-### Create minimum reviewer policy
+### Create minimum reviewer policy — team project (reviewer gate)
+
+Use when multiple humans review each other's PRs. The PR author's own
+vote does not count toward the minimum approver count.
 
 **CLI:**
 ```bash
@@ -133,3 +167,24 @@ curl -s -X POST -H "$AUTH" -H "Content-Type: application/json" \
     }
   }"
 ```
+
+### Create minimum reviewer policy — solo project (author self-approve)
+
+Use when you are the only contributor. A PR is still required before
+merge, but `--creator-vote-counts true` means the author's own vote
+satisfies the 1-approver gate — no second human needed.
+
+**CLI:**
+```bash
+az repos policy approver-count create \
+  --org "$AZDO_ORG_URL" --project "$AZDO_PROJECT" \
+  --repository-id "$REPO_ID" --branch "$default_branch" \
+  --minimum-approver-count 1 \
+  --creator-vote-counts true \
+  --allow-downvotes false \
+  --reset-on-source-push true \
+  --blocking true --enabled true
+```
+
+**REST fallback:** same body as the team variant, but with
+`"creatorVoteCounts": true`.

package/skills/dock/references/pipeline-templates.md

@@ -115,9 +115,18 @@ jobs:
 
       - name: Smoke test
         run: |
-          # Wait for deployment to be ready, then hit health endpoint
-          sleep 10
-          curl -f "$DEV_URL/healthz" || exit 1
+          # Poll the health endpoint until ready (bounded deadline so a
+          # genuinely broken deploy still fails fast). Do not use a blind
+          # `sleep N && curl` — it flakes on slow cold starts and wastes
+          # time on fast ones.
+          deadline=$((SECONDS + 120))
+          until curl -fs "$DEV_URL/healthz" >/dev/null; do
+            if [ $SECONDS -ge $deadline ]; then
+              echo "Health check failed to pass within 120s" >&2
+              exit 1
+            fi
+            sleep 2
+          done
 
   # ── Promote to Staging (on release tag) ───────────────────────
   promote-staging:
@@ -173,8 +182,15 @@ jobs:
 
       - name: Post-deploy smoke test
         run: |
-          sleep 10
-          curl -f "$PROD_URL/healthz" || exit 1
+          # Poll until healthy, bounded deadline. Never `sleep N && curl`.
+          deadline=$((SECONDS + 120))
+          until curl -fs "$PROD_URL/healthz" >/dev/null; do
+            if [ $SECONDS -ge $deadline ]; then
+              echo "Production health check failed to pass within 120s" >&2
+              exit 1
+            fi
+            sleep 2
+          done
 ```
 
 ### Reusable deploy step patterns

package/skills/launch/SKILL.md

@@ -0,0 +1,355 @@
+---
+name: launch
+description: "Run the full OMC idea-to-merged-PR pipeline — cancel + deep-interview + ralplan + autopilot + mad-skills:ship — in a single invocation. Explicit-only; this skill never auto-activates. Only run when the user literally types /launch. Do not invoke on phrases like \"launch this\", \"ship it\", \"full pipeline\", or similar — none of those should trigger this skill."
+argument-hint: "<rough idea for the feature>"
+allowed-tools: Bash, Read, Skill
+disable-model-invocation: true
+---
+
+# Launch — Idea → Merged PR, end-to-end
+
+When this skill is invoked, IMMEDIATELY output the banner below before doing anything else.
+Pick ONE tagline at random — vary your choice each time.
+CRITICAL: Reproduce the banner EXACTLY character-for-character. The first line of the art has 4 leading characters (one invisible braille-blank + 3 spaces) — you MUST preserve them.
+
+```
+{tagline}
+
+⠀   ██╗██╗      █████╗ ██╗   ██╗███╗   ██╗ ██████╗██╗  ██╗
+   ██╔╝██║     ██╔══██╗██║   ██║████╗  ██║██╔════╝██║  ██║
+  ██╔╝ ██║     ███████║██║   ██║██╔██╗ ██║██║     ███████║
+ ██╔╝  ██║     ██╔══██║██║   ██║██║╚██╗██║██║     ██╔══██║
+██╔╝   ███████╗██║  ██║╚██████╔╝██║ ╚████║╚██████╗██║  ██║
+╚═╝    ╚══════╝╚═╝  ╚═╝ ╚═════╝ ╚═╝  ╚═══╝ ╚═════╝╚═╝  ╚═╝
+```
+
+Taglines:
+- 🚀 From idea to orbit!
+- 🛫 Cleared for takeoff!
+- 🎯 Zero-touch launch inbound!
+- 🎆 3, 2, 1… liftoff!
+- 🛰️ Launch sequence initiated!
+- 🏁 One command, full launch!
+- 🤖 Full auto, full send!
+- ⚙️ The whole launchpad!
+
+---
+
+## Output Formatting
+
+After the banner, display parsed input:
+
+```
+┌─ Input ────────────────────────────────────────
+│  {Field}:  {value}
+│  Flags:    {parsed flags or "none"}
+└────────────────────────────────────────────────
+```
+
+Pre-flight results:
+
+```
+── Pre-flight ───────────────────────────────────
+  ✅ {dep}           {version or "found"}
+  ⚠️ {dep}           not found → {fallback detail}
+  ❌ {dep}           missing → stopping
+──────────────────────────────────────────────────
+```
+
+Stage headers: `━━ {N} · {Name} ━━━━━━━━━━━━━━━━━━━━━━━━━`
+
+Status icons: ✅ done · ❌ failed · ⚠️ degraded · ⏳ working · ⏭️ skipped
+
+---
+
+## Purpose
+
+`/launch` is a thin orchestrator over five existing skills. Its job is to run them in order and keep going between them — nothing more. Each inner skill owns its own stage-level concerns (state, hooks, reviewers, regression). Treat this wrapper as dumb: if any inner stage fails, stop and report — do not attempt retry logic here.
+
+The collaboration window is **Stage 2 (deep-interview)**. Everything after is autonomous; any mid-pipeline pauses come from the inner skills surfacing `AskUserQuestion` prompts (e.g. an autopilot Phase 4 reviewer needs a call), not from this wrapper.
+
+## Flags
+
+Parse optional flags from the request:
+
+- `--skip-cancel`: skip the Stage 1 stale-state cleanup (use when you *know* the session is clean)
+- `--skip-interview`: skip Stage 2 if a usable spec already exists at `.omc/specs/deep-interview-*.md`
+- `--pr-only`: stop after autopilot (Stage 4); don't run ship
+- `--no-ship`: alias for `--pr-only`
+- `--critic=architect|critic|codex`: forwarded to ralplan for the consensus reviewer
+
+---
+
+## Pre-flight
+
+Before Stage 1, check all dependencies in this table:
+
+| Dependency | Type | Check | Required | Resolution | Detail |
+|---|---|---|---|---|---|
+| git | cli | `git --version` | yes | stop | Install from https://git-scm.com |
+| oh-my-claudecode | plugin | `ls "$HOME/.claude/plugins/cache/omc/oh-my-claudecode"/*/skills/cancel/SKILL.md 2>/dev/null` | yes | stop | OMC plugin not installed. Install via `/plugin add oh-my-claudecode` or follow the OMC installation docs. Without it, deep-interview / ralplan / autopilot / cancel are unavailable. |
+| oh-my-claudecode:cancel | skill | `ls "$HOME/.claude/plugins/cache/omc/oh-my-claudecode"/*/skills/cancel/SKILL.md 2>/dev/null` | yes | stop | OMC cancel skill missing — reinstall OMC |
+| oh-my-claudecode:deep-interview | skill | `ls "$HOME/.claude/plugins/cache/omc/oh-my-claudecode"/*/skills/deep-interview/SKILL.md 2>/dev/null` | yes | stop | OMC deep-interview skill missing — reinstall OMC |
+| oh-my-claudecode:ralplan | skill | `ls "$HOME/.claude/plugins/cache/omc/oh-my-claudecode"/*/skills/ralplan/SKILL.md 2>/dev/null` | yes | stop | OMC ralplan skill missing — reinstall OMC |
+| oh-my-claudecode:autopilot | skill | `ls "$HOME/.claude/plugins/cache/omc/oh-my-claudecode"/*/skills/autopilot/SKILL.md 2>/dev/null` | yes | stop | OMC autopilot skill missing — reinstall OMC |
+| mad-skills:ship | skill | `ls "$HOME/.claude/plugins/marketplaces/slamb2k/skills/ship/SKILL.md" 2>/dev/null` | yes | stop | mad-skills:ship not installed. Skip with `--pr-only` to stop after autopilot. Install with: `npx skills add slamb2k/mad-skills --skill ship` |
+
+For each row, in order:
+
+1. Skip rows that don't apply (e.g. the `mad-skills:ship` row when `--pr-only` is set — the skill isn't required in that path)
+2. Run the Check command
+3. If found: continue silently
+4. If missing: apply Resolution strategy
+   - **stop**: notify user with Detail, halt execution
+5. After all checks: summarize what's available and what's degraded
+
+The most important check is the **oh-my-claudecode plugin** — without it, four of the five stages are impossible and the skill must abort. Do not try to proceed with only some OMC skills installed; if the cache directory exists but individual skills are missing, the OMC installation is corrupt and needs reinstalling.
+
+**Additional state check (soft warning, not a blocker):** after the dependency table, run `git status --porcelain` once. If there is any output, warn the user that the working tree has uncommitted changes. `deep-interview` and `ralplan` don't touch source, but `autopilot` will. Let the user decide whether to stash / commit / proceed.
+
+---
+
+## Plan Resolution
+
+Capture the user's argument as the **IDEA** (the rough feature description). This is what deep-interview will use as its starting input.
+
+Display in the Input box as:
+
+```
+│  Idea:   {first 80 chars of idea}
+│  Flags:  {parsed flags or "none"}
+```
+
+---
+
+## Stage 1 · Cancel stale state
+
+Unless `--skip-cancel` was set, invoke the OMC cancel skill to wipe any orphaned state:
+
+```
+Skill("oh-my-claudecode:cancel", "--force")
+```
+
+**Why force:** `/launch` is meant to be a clean-slate starting point. If previous OMC sessions left orphaned `awaiting_confirmation` entries (this is a known rough edge — e.g. the word "autopilot" in a prompt can trigger autopilot state without user consent), those will pollute later stages. `--force` wipes everything so the downstream stages start from zero.
+
+After the cancel skill returns, confirm no modes are active before proceeding. Emit a short one-line status:
+
+```
+  ✅ state cleared — no active OMC modes
+```
+
+If cancel reports lingering state it couldn't clear, stop and report. Don't press on.
+
+---
+
+## Stage 2 · Deep-interview (requirements Q&A)
+
+Unless `--skip-interview` was set AND a pre-existing spec is found, invoke:
+
+```
+Skill("oh-my-claudecode:deep-interview", "{IDEA}")
+```
+
+Let deep-interview run to completion. It will:
+
+- Conduct Socratic Q&A via `AskUserQuestion` — **these pauses surface to the user**, not the wrapper
+- Continue until its ambiguity gate passes (≤ 20%)
+- Write the final spec to `.omc/specs/deep-interview-{slug}.md`
+
+**If `--skip-interview` was set:**
+
+- Look for any file matching `.omc/specs/deep-interview-*.md`
+- If found, use the most recently modified one and note its path in the stage output
+- If not found, warn the user and fall back to running deep-interview anyway
+
+When the stage completes, emit:
+
+```
+  ✅ spec written to .omc/specs/deep-interview-{slug}.md
+```
+
+---
+
+## Stage 3 · Ralplan (plan + consensus review)
+
+Invoke:
+
+```
+Skill("oh-my-claudecode:ralplan", "--direct{critic-flag-if-any}")
+```
+
+`--direct` tells ralplan to skip its own interview and consume the deep-interview spec directly. If the user passed `--critic=architect|critic|codex`, forward that flag to ralplan.
+
+Ralplan will run Planner → Architect → Critic in sequence, looping until Critic approves. When it completes, the consensus plan lives at `.omc/plans/ralplan-{slug}.md`.
+
+Emit:
+
+```
+  ✅ consensus plan at .omc/plans/ralplan-{slug}.md
+  ✅ Critic verdict: APPROVED
+```
+
+If Critic keeps rejecting past its internal loop budget, ralplan will stop and report. In that case, stop this wrapper too — don't press into autopilot with a rejected plan.
+
+---
+
+## Stage 4 · Autopilot (execution + QA + validation)
+
+Invoke:
+
+```
+Skill("oh-my-claudecode:autopilot")
+```
+
+No arguments. Autopilot will detect the consensus plan at `.omc/plans/ralplan-{slug}.md` and **skip its own Phase 0 + Phase 1** (expansion + planning), going straight to Phase 2 (Execution). It runs:
+
+- Phase 2: parallel executor agents implement the plan
+- Phase 3: QA cycling (up to 5 iterations) — build, lint, test, fix
+- Phase 4: multi-perspective validation (Architect + Security-reviewer + Code-reviewer in parallel; all must approve)
+- Phase 5: state cleanup
+
+**Critical:** autopilot's output is *verified local code*, not a merged PR. Autopilot's Phase 5 is only state-file cleanup — it does not push, open a PR, watch CI, or merge. Ship is a separate stage.
+
+Emit:
+
+```
+  ✅ all phases complete
+  ✅ tests pass, build green, validators approved
+```
+
+If autopilot's Phase 3 QA cycle repeats the same error 3 times, or Phase 4 validation fails 3 rounds, autopilot stops and reports. In that case stop this wrapper too — a human needs to review.
+
+---
+
+## Stage 5 · Ship (push → PR → merge → sync)
+
+**Skip if `--pr-only` or `--no-ship` was set.**
+
+Invoke:
+
+```
+Skill("mad-skills:ship", "{one-line summary of approach from ralplan} — Files: {files from autopilot IMPL_REPORT}")
+```
+
+`mad-skills:ship` handles the PR lifecycle:
+
+- Pushes the branch (creates one if needed)
+- Creates PR via `gh pr create` (or AzDO equivalent)
+- Watches CI — fixes failures in a separate sub-agent up to 2 attempts
+- Squash-merges on green
+- Syncs local back to default branch
+
+Emit:
+
+```
+  ✅ PR #{number} merged ({merge_commit})
+  ✅ local synced to {default_branch}
+```
+
+If ship's CI watch exhausts its fix attempts, ship stops and displays its own failure banner. In that case, stop — do not attempt additional retry here.
+
+---
+
+## Waiting patterns
+
+Whenever this skill (or its inner invocations) needs to wait on a file or condition, use a polling loop with an explicit deadline — **never** `sleep N && <cmd>`:
+
+```bash
+deadline=$((SECONDS + 600))
+until [ -s "$path" ] || [ $SECONDS -ge $deadline ]; do sleep 2; done
+```
+
+This keeps waits bounded and cache-friendly, and lets the loop exit early when the expected file appears.
+
+---
+
+## Final Report
+
+```
+┌─ Launch · Report ──────────────────────────────
+│
+│  ✅ Launch complete
+│
+│  💡 Idea:       {first line of idea}
+│  📄 Spec:       .omc/specs/deep-interview-{slug}.md
+│  🗺️ Plan:       .omc/plans/ralplan-{slug}.md
+│
+│  📝 Stages
+│     1. State cleared             ✅
+│     2. Requirements interview    ✅  ({Q&A rounds} rounds)
+│     3. Plan + Critic consensus   ✅  ({critic-approval-rounds} rounds)
+│     4. Execution + QA + review   ✅  ({QA cycles} QA cycles, {validator} approved)
+│     5. PR + merge                ✅  ({pr_url})
+│
+│  📊 Code
+│     Files changed:  {count}
+│     Tests:          {passed}/{total}
+│
+│  🔗 Links
+│     PR:    {pr_url}
+│     Merge: {merge_commit}
+│
+│  ⚡ Next
+│     {anything surfaced by debrief from autopilot or ship}
+│
+└─────────────────────────────────────────────────
+```
+
+If the run stopped short (`--pr-only`, `--no-ship`, or a stage failure), adjust the stages list accordingly — mark later stages with ⏭️ (skipped) or ❌ (failed) and surface what's left to do manually in the `⚡ Next` section.
+
+### Pipeline Summary (always emit)
+
+```
+┌─ Pipeline Summary ─────────────────────────────
+│
+│  {icon} Cancel         {"no-op" or "cleared N sessions"}
+│  {icon} Deep-interview {".omc/specs/... or reused spec"}
+│  {icon} Ralplan        {critic verdict}
+│  {icon} Autopilot      {"green / N files changed"}
+│  {icon} Ship           {pr_url → merge_commit}
+│
+└─────────────────────────────────────────────────
+```
+
+---
+
+## Failure Handling
+
+`/launch` is deliberately dumb about retries. Each inner skill already has its own retry/verification logic — trying to add more at this layer just fights with them. If any stage stops with a failure:
+
+1. Display the failure banner below (not the success report)
+2. List which stages completed, which failed, and which were skipped
+3. Stop. Do not advance to later stages. Do not invoke `/sync` or any cleanup.
+
+```
+┌─ Launch · FAILED ───────────────────────────────
+│
+│  ❌ Pipeline stopped at Stage {N}: {stage name}
+│
+│  Reason:           {specific failure reason from the inner skill}
+│  Stages completed: {list}
+│  Stages skipped:   {list}
+│
+│  ⚠️  Work may exist locally but is NOT merged.
+│     Review the inner skill's output for next steps.
+│
+└──────────────────────────────────────────────────
+```
+
+Rules on failure:
+
+- Do not emit a "Pipeline Summary" with a ✅
+- Do not suggest a cancel/sync/retry here — the inner skill that failed owns that
+- Do not invoke any recovery skill automatically
+
+---
+
+## Trigger Behaviour
+
+This skill has `disable-model-invocation: true`. That means:
+
+- It will **never** auto-activate on keyword hits, even if the user says things like *"launch this feature"*, *"ship it end to end"*, *"run the full pipeline"*, or mentions any of the inner skill names.
+- It runs **only** when the user literally types `/launch` as a slash command.
+- Another skill can still invoke it explicitly via `Skill("launch", "...")` if desired.
+
+This is deliberate: the inner skills (especially autopilot) have aggressive keyword triggers of their own, and a wrapper that also auto-triggered would compound the false-positive risk. Keep this file's `disable-model-invocation` flag set.

package/skills/manifest.json

@@ -1,12 +1,12 @@
 {
-  "generated": "2026-04-17T05:36:15.739Z",
-  "count": 10,
+  "generated": "2026-04-20T02:56:39.184Z",
+  "count": 11,
   "skills": [
     {
       "name": "brace",
       "directory": "brace",
       "description": "'Initialize any project directory with a standard scaffold for AI-assisted development. Creates specs/ and context/ directories, a project CLAUDE.md with development workflow and guardrails, .gitignore, and branch protection. Recommends claude-mem for persistent memory. Idempotent — safe to run on existing projects. Triggers: \"init project\", \"setup brace\", \"brace\", \"initialize\", \"bootstrap\", \"scaffold\".'",
-      "lines": 424,
+      "lines": 430,
       "hasScripts": false,
       "hasReferences": true,
       "hasAssets": true,
@@ -52,6 +52,16 @@
       "hasAssets": false,
       "hasTests": true
     },
+    {
+      "name": "launch",
+      "directory": "launch",
+      "description": "\"Run the full OMC idea-to-merged-PR pipeline — cancel + deep-interview + ralplan + autopilot + mad-skills:ship — in a single invocation. Explicit-only; this skill never auto-activates. Only run when the user literally types /launch. Do not invoke on phrases like \\\"launch this\\\", \\\"ship it\\\", \\\"full pipeline\\\", or similar — none of those should trigger this skill.\"",
+      "lines": 356,
+      "hasScripts": false,
+      "hasReferences": false,
+      "hasAssets": false,
+      "hasTests": false
+    },
     {
       "name": "prime",
       "directory": "prime",

package/skills/ship/scripts/merge.sh

@@ -94,30 +94,32 @@ if [ "$AZDO_MODE" = "cli" ]; then
     sleep 15
   done
 
-  # Complete the PR
-  if az repos pr update --id "$PR_NUMBER" --status completed \
-    --org "$AZDO_ORG_URL" \
-    --squash "$SQUASH_FLAG" \
-    --delete-source-branch "$DELETE_FLAG" 2>/dev/null; then
-    STATUS="success"
-    MERGE_COMMIT=$(git rev-parse --short HEAD 2>/dev/null)
-    BRANCH_DELETED=$DELETE_BRANCH
-  else
-    # Retry once after 30s (policies may still be evaluating)
-    sleep 30
+  # Complete the PR — retry on a bounded deadline; policies may still be
+  # evaluating after the initial poll finished. First attempt runs
+  # immediately, then we poll every 5s up to a 30s total budget.
+  # Avoid `sleep 30 && retry` (blind wait) — this loop exits early on success.
+  MERGE_DEADLINE=$((SECONDS + 30))
+  MERGE_OK=false
+  while :; do
     if az repos pr update --id "$PR_NUMBER" --status completed \
       --org "$AZDO_ORG_URL" \
       --squash "$SQUASH_FLAG" \
       --delete-source-branch "$DELETE_FLAG" 2>/dev/null; then
-      STATUS="success"
-      MERGE_COMMIT=$(git rev-parse --short HEAD 2>/dev/null)
-      BRANCH_DELETED=$DELETE_BRANCH
-    else
-      STATUS="failed"
-      ERRORS="Merge failed after retry"
-      MERGE_COMMIT=""; BRANCH_DELETED=false
-      emit_report; exit 2
+      MERGE_OK=true
+      break
     fi
+    [ $SECONDS -ge $MERGE_DEADLINE ] && break
+    sleep 5
+  done
+  if $MERGE_OK; then
+    STATUS="success"
+    MERGE_COMMIT=$(git rev-parse --short HEAD 2>/dev/null)
+    BRANCH_DELETED=$DELETE_BRANCH
+  else
+    STATUS="failed"
+    ERRORS="Merge failed after retry"
+    MERGE_COMMIT=""; BRANCH_DELETED=false
+    emit_report; exit 2
   fi
   emit_report
   exit 0