@slamb2k/mad-skills 2.0.37 → 2.0.38

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "mad-skills",
   "description": "AI-assisted planning, development and governance tools",
-  "version": "2.0.37",
+  "version": "2.0.38",
   "author": {
     "name": "slamb2k",
     "url": "https://github.com/slamb2k"

package/README.md

@@ -402,7 +402,7 @@ Three methods are available. The table below shows what each delivers:
 | | Plugin | npx skills | npm package |
 |---|---|---|---|
 | Skills (slash commands) | ✅ all 10 | ✅ all 10 | — |
-| Agents (e.g. ship-analyzer) | ✅ | ❌ | — |
+| Bundled scripts (sync, CI, merge) | ✅ | ✅ | — |
 | Session hooks (session-guard) | ✅ | ❌ | — |
 | Cross-agent (Cursor, Cline, etc.) | ❌ Claude Code only | ✅ | — |
 | Selective skill install | ❌ | ✅ | — |
@@ -410,7 +410,7 @@ Three methods are available. The table below shows what each delivers:
 
 ### Plugin (recommended)
 
-Installs skills, agents, and session hooks from the GitHub repo into
+Installs skills and session hooks from the GitHub repo into
 `~/.claude/plugins/`. Updates automatically. Claude Code only.
 
 **Step 1 — Register the marketplace (one-time):**
@@ -448,9 +448,8 @@ npx skills add slamb2k/mad-skills -g -y              # All skills, global
 npx skills add slamb2k/mad-skills --skill ship -g -y  # Specific skill
 ```
 
-Installs skills into `~/.claude/skills/` (and `~/.agents/skills/` for other agents). **Does not install agents or hooks.** This means:
+Installs skills into `~/.claude/skills/` (and `~/.agents/skills/` for other agents). **Does not install hooks.** This means:
 
-- `/build` falls back to `general-purpose` agent for the ship stage instead of the optimised `ship-analyzer` agent
 - The session-guard hook (CLAUDE.md staleness detection, git validation) is not active
 
 Use this method when you need cross-agent compatibility (Cursor, Cline, Amp, etc.) or want to install individual skills.
@@ -514,7 +513,7 @@ mad-skills/
 │   ├── build-manifests.js   # Generate skills/manifest.json
 │   └── package-skills.js    # Package .skill archives
 ├── hooks/                   # Session hooks + plugin hook config
-├── agents/                  # Agent definitions (ship-analyzer)
+├── agents/                  # Agent definitions (reserved for future use)
 ├── tests/results/           # Eval output
 ├── archive/                 # Legacy skills (v1.x)
 ├── .claude-plugin/          # Plugin metadata

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@slamb2k/mad-skills",
-  "version": "2.0.37",
+  "version": "2.0.38",
   "description": "Claude Code skills collection — full lifecycle development tools",
   "type": "module",
   "repository": {

package/skills/manifest.json

@@ -1,5 +1,5 @@
 {
-  "generated": "2026-03-17T02:08:37.271Z",
+  "generated": "2026-03-17T03:05:22.402Z",
   "count": 10,
   "skills": [
     {
@@ -76,8 +76,8 @@
       "name": "ship",
       "directory": "ship",
       "description": "\"Ship changes through the full PR lifecycle. Use after completing feature work to commit, push, create PR, wait for checks, and merge. Handles the entire workflow: syncs with main, creates feature branch if needed, groups commits logically with semantic messages, creates detailed PR, monitors CI, fixes issues, squash merges, and cleans up. Invoke when work is ready to ship.\"",
-      "lines": 428,
-      "hasScripts": false,
+      "lines": 418,
+      "hasScripts": true,
       "hasReferences": true,
       "hasAssets": false,
       "hasTests": true
@@ -96,8 +96,8 @@
       "name": "sync",
       "directory": "sync",
       "description": "Sync local repository with origin/main. Use before starting new work, after completing a PR, or when needing latest upstream changes. Safely stashes uncommitted changes, fetches and pulls origin/main, restores stash, and cleans up stale local branches (merged or with deleted remotes). Invoke when switching contexts or preparing for new feature work.",
-      "lines": 244,
-      "hasScripts": false,
+      "lines": 164,
+      "hasScripts": true,
       "hasReferences": false,
       "hasAssets": false,
       "hasTests": true

package/skills/ship/SKILL.md

@@ -59,9 +59,10 @@ Status icons: ✅ done · ❌ failed · ⚠️ degraded · ⏳ working · ⏭️
 
 ---
 
-Ship changes through the complete PR lifecycle. Every stage runs in a subagent
-to isolate context from the primary conversation. Prompts for each stage are
-in `references/stage-prompts.md`.
+Ship changes through the complete PR lifecycle. Deterministic stages (sync, CI
+polling, merge) run as bash scripts for speed and reliability. Only stages that
+require reasoning (commit/PR authoring, CI fix analysis) use LLM subagents.
+Stage prompts are in `references/stage-prompts.md`.
 
 ## Flags
 
@@ -112,7 +113,6 @@ Before starting, check all dependencies in this table. The table contains
 | git | cli | `git --version` | yes | stop | Install from https://git-scm.com |
 | gh | cli | `gh --version` | yes | url | https://cli.github.com |
 | az devops | cli | `az devops -h 2>/dev/null` | no | fallback | Falls back to REST API with PAT; see AzDO tooling below |
-| ship-analyzer | agent | `~/.claude/agents/ship-analyzer.md` or `~/.claude/plugins/marketplaces/slamb2k/agents/ship-analyzer.md` | no | fallback | Uses general-purpose agent |
 
 **Platform-conditional rules:**
 - **`gh`**: Only required when `PLATFORM == github`. Skip for AzDO repos.
@@ -129,6 +129,7 @@ For each applicable row, in order:
    - **ask**: notify user, offer to run command in Detail, continue either way (or halt if required)
    - **fallback**: notify user with Detail, continue with degraded behavior
 5. After all checks: summarize what's available and what's degraded
+6. Always show the commit/PR agent: `✅ commit agent    general-purpose`
 
 ### AzDO Tooling Detection
 
@@ -226,38 +227,33 @@ substituted into all stage prompts as `{REMOTE}` and `{DEFAULT_BRANCH}`.
 
 ## Stage 1: Sync
 
-Launch **Bash subagent** (haiku — simple git commands):
+Run the sync script directly (no LLM needed):
 
-```
-Task(
-  subagent_type: "Bash",
-  model: "haiku",
-  description: "Sync with default branch",
-  prompt: "Follow ~/.claude/skills/sync/SKILL.md or ~/.claude/plugins/marketplaces/slamb2k/skills/sync/SKILL.md subagent prompt. Return SYNC_REPORT."
-)
+```bash
+SKILL_ROOT="<resolved plugin root>"
+bash "$SKILL_ROOT/skills/sync/scripts/sync.sh" "{REMOTE}" "{DEFAULT_BRANCH}"
 ```
 
-Parse SYNC_REPORT. Extract `remote` and `default_branch`. Abort if sync failed.
+Parse SYNC_REPORT from output markers. Extract `remote` and `default_branch`.
+Abort if exit code is 1 (fatal).
 
 ---
 
 ## Stage 2: Commit, Push & Create PR
 
 This stage needs to **read and understand code** to write good commit messages
-and PR descriptions. Use a code-aware subagent.
+and PR descriptions — it's one of the few stages that requires an LLM.
 
-Launch **ship-analyzer subagent** (reads diffs + source files):
+Launch **general-purpose subagent** (reads diffs + source files):
 
 ```
 Task(
-  subagent_type: "ship-analyzer",
+  subagent_type: "general-purpose",
   description: "Analyze, commit, push, and create PR",
   prompt: <read from references/stage-prompts.md#stage-2>
 )
 ```
 
-> **Fallback:** If `ship-analyzer` is not available, use `subagent_type: "general-purpose"`.
-
 Substitute `{USER_INTENT}`, `{FILES_TO_INCLUDE}`, `{FILES_TO_EXCLUDE}`,
 `{REMOTE}`, `{DEFAULT_BRANCH}`, `{PLATFORM}`, `{AZDO_MODE}`, `{AZDO_ORG}`,
 `{AZDO_PROJECT}`, `{AZDO_ORG_URL}`, `{PAT}` into the prompt.
@@ -283,23 +279,19 @@ immediately. This stage loops: watch → detect failure → fix → push → wat
 
 ### Watch
 
-Launch **Bash subagent** (haiku — polling):
+Run the CI watch script directly (no LLM needed — just polling):
 
-```
-Task(
-  subagent_type: "Bash",
-  model: "haiku",
-  description: "Monitor CI checks",
-  prompt: <read from references/stage-prompts.md#stage-3>
-)
+```bash
+SKILL_ROOT="<resolved plugin root>"
+bash "$SKILL_ROOT/skills/ship/scripts/ci-watch.sh" \
+  "{PLATFORM}" "{PR_NUMBER}" "{BRANCH}" \
+  --azdo-mode="{AZDO_MODE}" --azdo-org-url="{AZDO_ORG_URL}" \
+  --azdo-project="{AZDO_PROJECT}" --azdo-project-url-safe="{AZDO_PROJECT_URL_SAFE}"
 ```
 
-Substitute `{PR_NUMBER}`, `{BRANCH}`, `{PLATFORM}`, `{AZDO_MODE}`,
-`{AZDO_ORG}`, `{AZDO_ORG_URL}`, `{AZDO_PROJECT}`, `{PAT}` into the prompt.
-
 Briefly inform the user: `⏳ Watching CI for PR #{pr_number}...`
 
-Parse CHECKS_REPORT when the subagent returns.
+Parse CHECKS_REPORT from output markers. Exit code 0=passed, 1=failed, 2=timeout.
 
 ### Fix (if needed)
 
@@ -347,29 +339,27 @@ defaults (override via `--no-squash` and `--keep-branch` flags only).
 
 ### 5a. Merge the PR
 
-Launch **Bash subagent** (haiku — simple git + platform CLI commands):
+Run the merge script directly (no LLM needed):
 
-```
-Task(
-  subagent_type: "Bash",
-  model: "haiku",
-  description: "Merge PR",
-  prompt: <read from references/stage-prompts.md#stage-5>
-)
+```bash
+SKILL_ROOT="<resolved plugin root>"
+bash "$SKILL_ROOT/skills/ship/scripts/merge.sh" \
+  "{PLATFORM}" "{PR_NUMBER}" \
+  {MERGE_FLAGS} {BRANCH_FLAGS} \
+  --azdo-mode="{AZDO_MODE}" --azdo-org-url="{AZDO_ORG_URL}" \
+  --azdo-project="{AZDO_PROJECT}" --azdo-project-url-safe="{AZDO_PROJECT_URL_SAFE}"
 ```
 
-Substitute `{PR_NUMBER}`, `{REMOTE}`, `{DEFAULT_BRANCH}`, `{PLATFORM}`,
-`{AZDO_MODE}`, `{AZDO_ORG_URL}`, `{AZDO_PROJECT}`, `{PAT}`, merge/branch flags.
-
-Parse LAND_REPORT.
+Parse LAND_REPORT from output markers. Exit code 0=merged, 1=failed.
 
 ### 5b. Sync local repo
 
-After the merge subagent succeeds, invoke `/sync` from the orchestrator to
-checkout the default branch, pull the merge commit, and **clean up stale
-branches** (prune gone remotes, delete merged branches). This replaces any
-manual `git checkout`/`git pull` — `/sync` handles everything including stash
-restore and branch cleanup.
+After the merge script succeeds, run the sync script to checkout the default
+branch, pull the merge commit, and **clean up stale branches**:
+
+```bash
+bash "$SKILL_ROOT/skills/sync/scripts/sync.sh" "{REMOTE}" "{DEFAULT_BRANCH}"
+```
 
 ---
 

package/skills/ship/references/stage-prompts.md

@@ -12,7 +12,10 @@ Use `{AZDO_PROJECT}` for CLI `--project` flags and display. Use `{AZDO_PROJECT_U
 
 ## Stage 2: Commit, Push & Create PR
 
-**Agent:** ship-analyzer (fallback: general-purpose)
+**Agent:** general-purpose
+
+This is the only shipping stage that requires an LLM — it reads and understands
+code to produce meaningful commit messages and PR descriptions.
 
 ```
 Ship the following changes. Analyze the diffs, create semantic commits,
@@ -31,6 +34,36 @@ FILES TO INCLUDE: {FILES_TO_INCLUDE}
 
 FILES TO EXCLUDE: {FILES_TO_EXCLUDE}
 
+## Principles
+
+1. **Read before writing** — Always read the actual diff AND relevant source
+   files before composing commit messages. Never guess at intent from filenames.
+2. **Semantic grouping** — Group related changes into logical commits. A
+   "logical group" shares a single purpose.
+3. **Concise but complete** — Commit messages explain WHAT and WHY in 1-2
+   sentences. PR descriptions give the full picture.
+4. **No attribution** — Never add Co-Authored-By or Generated-by lines.
+5. **Prioritize impact** — When changes span many files, read the most
+   impactful diffs first (source > tests > config).
+
+## Commit message format
+
+```text
+<type>(<scope>): <imperative description>
+
+<optional body: what changed and why, wrapped at 72 chars>
+```
+
+Types: feat, fix, refactor, docs, chore, test, perf
+
+Good examples:
+- feat(auth): replace pairing gate with channel allowlist
+- fix(memory): correct positional arg order in get_recent_commitments
+
+Bad examples:
+- update files (too vague)
+- feat: changes to auth system (no scope, vague)
+
 ## Steps
 
 1. **Analyze changes**
@@ -52,9 +85,9 @@ FILES TO EXCLUDE: {FILES_TO_EXCLUDE}
      )"
 
    Rules:
-   - Never add Co-Authored-By or attribution lines
    - Use HEREDOC for commit messages (ensures proper formatting)
-   - Types: feat, fix, refactor, docs, chore, test, perf
+   - Use `git add -p` when only some hunks in a file belong in a given commit
+   - Always verify the push succeeded before creating the PR
 
 4. **Push**
    git push -u {REMOTE} $BRANCH
@@ -149,152 +182,13 @@ SHIP_REPORT:
 
 ## Stage 3: Wait for CI
 
-**Agent:** Bash (haiku, background)
-
-```
-Monitor PR/pipeline status checks until complete. Fail fast — stop polling the
-moment any check fails so fixes can start immediately.
-
-Limit CHECKS_REPORT to 10 lines maximum.
-
-PLATFORM: {PLATFORM}
-PR_NUMBER: {PR_NUMBER}
-BRANCH: {BRANCH}
-
-## Steps
-
-**If PLATFORM == github:**
-
-1. **Wait for checks — fail fast on first failure**
-   gh pr checks {PR_NUMBER} --watch --fail-fast
-
-2. **Report final status**
-   gh pr checks {PR_NUMBER}
+**Replaced by script:** `scripts/ci-watch.sh`
 
-**If PLATFORM == azdo AND AZDO_MODE == cli:**
+The orchestrator runs this script directly via Bash — no LLM needed.
+See `skills/ship/scripts/ci-watch.sh` for the full implementation.
 
-1. **Wait for CI to start** (pipelines may take time to trigger after PR creation)
-   Poll until at least one run appears (max 2 minutes, check every 15s):
-   ```
-   RUNS_FOUND=false
-   for i in $(seq 1 8); do
-     RUN_COUNT=$(az pipelines runs list --branch "$BRANCH" --top 5 \
-       --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-       --query "length(@)" -o tsv 2>/dev/null)
-     if [ -n "$RUN_COUNT" ] && [ "$RUN_COUNT" != "0" ]; then
-       RUNS_FOUND=true
-       break
-     fi
-     sleep 15
-   done
-   ```
-   If no runs appear after 2 minutes, also check PR policy evaluations:
-   ```
-   az repos pr policy list --id {PR_NUMBER} \
-     --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-     --query "[].{name:configuration.type.displayName, status:status}" -o table
-   ```
-   If no policies exist either, report `no_checks`.
-
-2. **Wait for runs to complete — with fail-fast** (max 30 minutes, check every 15s)
-   ```
-   for i in $(seq 1 120); do
-     # Check for failures FIRST — stop immediately if any run has failed
-     FAILED=$(az pipelines runs list --branch "$BRANCH" --top 5 \
-       --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-       --query "[?result=='failed'] | length(@)" -o tsv 2>/dev/null)
-     if [ -n "$FAILED" ] && [ "$FAILED" != "0" ]; then
-       echo "CI failure detected — stopping wait"
-       break
-     fi
-
-     # Then check if anything is still running
-     IN_PROGRESS=$(az pipelines runs list --branch "$BRANCH" --top 5 \
-       --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-       --query "[?status=='inProgress'] | length(@)" -o tsv 2>/dev/null)
-     if [ "$IN_PROGRESS" = "0" ] || [ -z "$IN_PROGRESS" ]; then break; fi
-     sleep 15
-   done
-   ```
-
-3. **Determine final status**
-   Query both pipeline runs and PR policy evaluations:
-   ```
-   az pipelines runs list --branch "$BRANCH" --top 5 \
-     --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-     --query "[].{name:definition.name, status:status, result:result}" -o table
-   az repos pr policy list --id {PR_NUMBER} \
-     --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-     --query "[].{name:configuration.type.displayName, status:status}" -o table
-   ```
-
-   **Map results to CHECKS_REPORT status:**
-   - Any pipeline with `result == "failed"` → `some_failed`
-   - Any policy with `status == "rejected"` → `some_failed`
-   - All pipelines `result == "succeeded"` AND no rejected policies → `all_passed`
-   - Pipelines still `inProgress` after timeout → `pending`
-   - No pipelines and no policies found → `no_checks`
-
-   **IMPORTANT:** Do NOT report `all_passed` if any run has `result == "failed"`.
-   Do NOT continue waiting once a failure is detected — report `some_failed`
-   immediately with the failing check names.
-
-**If PLATFORM == azdo AND AZDO_MODE == rest:**
-
-1. **Wait for CI to start** (max 2 minutes, check every 15s)
-   ```
-   REPO_NAME=$(basename -s .git "$(git remote get-url origin)")
-   AUTH="Authorization: Basic $(echo -n ":{PAT}" | base64)"
-   BUILDS_URL="{AZDO_ORG_URL}/{AZDO_PROJECT_URL_SAFE}/_apis/build/builds?branchName=refs/heads/$BRANCH&\$top=5&api-version=7.0"
-
-   RUNS_FOUND=false
-   for i in $(seq 1 8); do
-     RUN_COUNT=$(curl -s -H "$AUTH" "$BUILDS_URL" | jq '.value | length')
-     if [ -n "$RUN_COUNT" ] && [ "$RUN_COUNT" != "0" ]; then
-       RUNS_FOUND=true
-       break
-     fi
-     sleep 15
-   done
-   ```
-   If no runs appear, check PR policy evaluations via REST:
-   ```
-   EVALS=$(curl -s -H "$AUTH" \
-     "{AZDO_ORG_URL}/{AZDO_PROJECT_URL_SAFE}/_apis/policy/evaluations?artifactId=vstfs:///CodeReview/CodeReviewId/{AZDO_PROJECT_URL_SAFE}/{PR_NUMBER}&api-version=7.0")
-   EVAL_COUNT=$(echo "$EVALS" | jq '.value | length')
-   ```
-   If no evaluations exist either, report `no_checks`.
-
-2. **Wait for runs to complete — with fail-fast** (max 30 minutes, check every 15s)
-   ```
-   for i in $(seq 1 120); do
-     BUILDS_JSON=$(curl -s -H "$AUTH" "$BUILDS_URL")
-
-     # Check for failures FIRST
-     FAILED=$(echo "$BUILDS_JSON" | jq '[.value[] | select(.result=="failed")] | length')
-     if [ "$FAILED" != "0" ]; then
-       echo "CI failure detected — stopping wait"
-       break
-     fi
-
-     # Check if anything still running
-     IN_PROGRESS=$(echo "$BUILDS_JSON" | jq '[.value[] | select(.status=="inProgress")] | length')
-     if [ "$IN_PROGRESS" = "0" ]; then break; fi
-     sleep 15
-   done
-   ```
-
-3. **Determine final status** — same mapping rules as CLI mode above.
-
-## Output Format
-
-CHECKS_REPORT:
-- status: all_passed|some_failed|pending|no_checks
-- checks:
-  - name: {check name}
-    status: success|failure|pending
-- failing_checks: {list of failed check names, or "none"}
-```
+Output is a structured `CHECKS_REPORT` between `CHECKS_REPORT_BEGIN`/`END` markers.
+Exit codes: 0=all_passed, 1=some_failed, 2=timeout, 3=tool error.
 
 ---
 
@@ -377,110 +271,10 @@ FIX_REPORT:
 
 ## Stage 5: Merge & Final Sync
 
-**Agent:** Bash (haiku)
+**Replaced by script:** `scripts/merge.sh`
 
-```
-Merge the pull request. Do NOT sync locally — the orchestrator handles that
-via /sync after this subagent returns.
-Do NOT ask the user for confirmation — proceed immediately.
-
-Limit LAND_REPORT to 10 lines maximum.
-
-PLATFORM: {PLATFORM}
-PR_NUMBER: {PR_NUMBER}
-MERGE_FLAGS: {--squash (default) | --merge if --no-squash}
-BRANCH_FLAGS: {--delete-branch (default) | omit if --keep-branch}
-
-## Steps
-
-1. **Merge**
-
-   **If PLATFORM == github:**
-     gh pr merge {PR_NUMBER} {MERGE_FLAGS} {BRANCH_FLAGS}
-
-   **If PLATFORM == azdo AND AZDO_MODE == cli:**
-     # 1. Verify all policies pass before attempting merge
-     REJECTED=$(az repos pr policy list --id {PR_NUMBER} \
-       --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-       --query "[?status=='rejected'] | length(@)" -o tsv 2>/dev/null)
-     if [ -n "$REJECTED" ] && [ "$REJECTED" != "0" ]; then
-       echo "ERROR: $REJECTED PR policies are rejected — cannot merge"
-       az repos pr policy list --id {PR_NUMBER} \
-         --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-         --query "[?status=='rejected'].{name:configuration.type.displayName, status:status}" -o table
-       exit 1
-     fi
-
-     # 2. Resolve merge strategy from flags
-     if echo "{MERGE_FLAGS}" | grep -q "\-\-merge"; then
-       SQUASH_FLAG="false"
-     else
-       SQUASH_FLAG="true"
-     fi
-     if echo "{BRANCH_FLAGS}" | grep -q "\-\-keep-branch"; then
-       DELETE_FLAG="false"
-     else
-       DELETE_FLAG="true"
-     fi
-
-     # 3. Complete the PR
-     az repos pr update --id {PR_NUMBER} --status completed \
-       --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-       --squash "$SQUASH_FLAG" \
-       --delete-source-branch "$DELETE_FLAG"
-     MERGE_RC=$?
-
-     # 4. If merge fails (e.g. policies still evaluating), wait and retry once
-     if [ $MERGE_RC -ne 0 ]; then
-       echo "Merge failed, waiting 30s for policies to finalize..."
-       sleep 30
-       az repos pr update --id {PR_NUMBER} --status completed \
-         --org "{AZDO_ORG_URL}" --project "{AZDO_PROJECT}" \
-         --squash "$SQUASH_FLAG" \
-         --delete-source-branch "$DELETE_FLAG"
-     fi
+The orchestrator runs this script directly via Bash — no LLM needed.
+See `skills/ship/scripts/merge.sh` for the full implementation.
 
-   **If PLATFORM == azdo AND AZDO_MODE == rest:**
-     AUTH="Authorization: Basic $(echo -n ":{PAT}" | base64)"
-     REPO_NAME=$(basename -s .git "$(git remote get-url origin)")
-     PR_API="{AZDO_ORG_URL}/{AZDO_PROJECT_URL_SAFE}/_apis/git/repositories/$REPO_NAME/pullrequests/{PR_NUMBER}"
-
-     # 1. Check for rejected policy evaluations
-     EVALS=$(curl -s -H "$AUTH" \
-       "{AZDO_ORG_URL}/{AZDO_PROJECT_URL_SAFE}/_apis/policy/evaluations?artifactId=vstfs:///CodeReview/CodeReviewId/{AZDO_PROJECT_URL_SAFE}/{PR_NUMBER}&api-version=7.0")
-     REJECTED=$(echo "$EVALS" | jq '[.value[] | select(.status=="rejected")] | length')
-     if [ "$REJECTED" != "0" ]; then
-       echo "ERROR: PR has rejected policy evaluations — cannot merge"
-       echo "$EVALS" | jq '.value[] | select(.status=="rejected") | {name: .configuration.type.displayName, status}'
-       exit 1
-     fi
-
-     # 2. Resolve merge strategy from flags
-     if echo "{MERGE_FLAGS}" | grep -q "\-\-merge"; then
-       MERGE_STRATEGY="noFastForward"
-     else
-       MERGE_STRATEGY="squash"
-     fi
-     if echo "{BRANCH_FLAGS}" | grep -q "\-\-keep-branch"; then
-       DELETE_BRANCH="false"
-     else
-       DELETE_BRANCH="true"
-     fi
-
-     # 3. Complete the PR
-     curl -s -X PATCH -H "$AUTH" -H "Content-Type: application/json" \
-       "$PR_API?api-version=7.0" \
-       -d "{\"status\": \"completed\", \"completionOptions\": {\"mergeStrategy\": \"$MERGE_STRATEGY\", \"deleteSourceBranch\": $DELETE_BRANCH}}"
-
-2. **Report**
-   MERGE_COMMIT=$(git rev-parse --short HEAD)
-
-## Output Format
-
-LAND_REPORT:
-- status: success|failed
-- merge_commit: {hash}
-- merge_type: squash|merge
-- branch_deleted: true|false
-- errors: {any errors}
-```
+Output is a structured `LAND_REPORT` between `LAND_REPORT_BEGIN`/`END` markers.
+Exit codes: 0=merged, 1=failed, 2=failed after retry.