aep-wrap

Wrap

Post-merge archive and workspace cleanup. Run this on the integration branch ($BASE — main in single-branch mode, develop in two-branch mode; see git-ref → "Integration Branch") after the PR merges to archive the OpenSpec change and clean up the workspace.

Where this fits:

/aep-onboard → /aep-scaffold → [ /aep-design → /aep-launch → /aep-build → /aep-wrap ]
                                                        ▲ you are here

Session: Main session, post-merge Input: Merged PR notification Output: Archived OpenSpec change, cleaned up workspace

---

Phase 13: Archive & Cleanup on the Integration Branch

Run on the integration branch ($BASE) after the PR merges. Do not run from the workspace.

1. Fetch merged state and fast-forward the integration branch

# Resolve $BASE — see git-ref "Integration Branch" (override → develop → main)
BASE=$(git config --get aep.integration-branch 2>/dev/null || true)
[ -z "$BASE" ] && { git show-ref --verify --quiet refs/heads/develop \
  || git show-ref --verify --quiet refs/remotes/origin/develop; } && BASE=develop
BASE=${BASE:-main}

git fetch origin
git checkout "$BASE"
git pull --ff-only origin "$BASE"
git status

Update the local integration branch to include the merged workspace PR. --ff-only is intentional — if it fails because $BASE has unpushed local commits, push or rebase those first.

After this checkout you are on the integration branch; later steps recover its name with BASE=$(git branch --show-current) (HEAD persists across shells, so this is safe even in a fresh command).

Check for lost OpenSpec changes: If the dispatch commit included OpenSpec changes that are now missing (common if the dispatch commit wasn't pushed before launching), recover them from the original dispatch commit:

# Find the original dispatch commit
git log --oneline -n 20
# Restore OpenSpec files from the dispatch commit
git restore --source=<dispatch-commit-sha> -- openspec/

Verify the workspace is clean before archiving. If git status shows unexpected modified files (code in apps/, packages/, etc.), investigate before proceeding. Only openspec/ and product-context.yaml files should change during the wrap step.

2. Stop the dev server (from the workspace, if still running)

source .feature-workspaces/<name>/.dev-workflow/ports.env 2>/dev/null
lsof -ti :$SERVER_PORT | xargs kill 2>/dev/null
lsof -ti :$WEB_PORT | xargs kill 2>/dev/null

3. Run archive

/opsx:archive <change-name>

4. Commit and push the archive

BASE=$(git branch --show-current)   # integration branch, checked out in step 1
git add openspec/
git commit -m "chore: archive <change-name>"
git pull --ff-only origin "$BASE"
git push origin "$BASE"

5. Sync story status from workspace signals (Product-Cycle Mode Only)

Standalone mode: If product-context.yaml doesn't exist, skip this step and proceed to step 6.

If product-context.yaml exists and this feature was a dispatched story, read the workspace signals and cross-check against the actual PR state:

# Read completion data from workspace signals
cat .feature-workspaces/<name>/.dev-workflow/signals/status.json

# Cross-check: verify actual PR state (signals can be stale)
gh pr view <pr-number> --json state,mergedAt

Signal validation: Workspace agents don't always update their signal files after merge (they may show in_review even though the PR is already merged). Always cross-check the signal's story_status against the actual PR state via gh pr view. If the PR is merged but the signal says in_review, treat the story as completed.

From the signal file (corrected by PR state if needed), extract story_status, pr_url, cost_usd, completed_at, and failure_log (if present). Update the story in product-context.yaml:

# Update the matching story:
status: completed # from signal story_status
completed_at: <timestamp> # from signal completed_at
pr_url: <url> # from signal pr_url
cost_usd: <cost> # from signal cost_usd

If story_status is failed, update with failure data instead:

status: failed
failure_logs:
  - <structured failure_log from signal>

After updating the story, check if any pending stories should transition to ready (all dependencies now completed). Validate and commit all transitions atomically:

# Validate YAML before committing (see product-context references/yaml-guardrails.md)
npx js-yaml product-context.yaml > /dev/null && echo "YAML OK"

BASE=$(git branch --show-current)   # integration branch, checked out in step 1
git add product-context.yaml
git commit -m "chore: update story <id> status to completed"
git pull --ff-only origin "$BASE"
git push origin "$BASE"

Concurrency protocol: This is the only place where story completion status enters product-context.yaml. Workspace agents write to signals; /aep-wrap (running on the integration branch) reads signals and writes to YAML.

5.5. Archive lessons learned

Before forgetting the workspace, check for lessons captured during the build:

LESSONS=".feature-workspaces/<name>/.dev-workflow/lessons.md"
if [ -f "$LESSONS" ] && [ "$(wc -l < "$LESSONS")" -gt 12 ]; then
  # File has content beyond the template header
  mkdir -p lessons-learned
  cp "$LESSONS" "lessons-learned/<change-name>.md"
  BASE=$(git branch --show-current)   # integration branch, checked out in step 1
  git add lessons-learned/<change-name>.md
  git commit -m "docs: archive lessons from <change-name>"
  git pull --ff-only origin "$BASE"
  git push origin "$BASE"
fi

Why before worktree removal: Once git worktree remove runs, the worktree directory and its .dev-workflow/lessons.md are gone. This is the only chance to extract them.

If the lessons file contains only the template header (no Solutions, Errors, Missing, or Summary entries), skip — don't archive empty ceremony.

6. Tear down the worker + worktree (executor.teardown())

Stop the workspace's worker before removing the worktree — otherwise an OS-bound worker keeps running against a deleted directory, and these orphans accumulate across an autopilot run. The stop step is per launch mode (recorded as backend/agent_id in autopilot state, or evident from how you launched):

# Mode-specific worker stop (each is a no-op for the other modes):
#   native-bg-subagent → TaskStop(<bare-hex bg-subagent id>) (session-bound, no team)
#   claude-bg      → claude stop <agent_id>; claude rm <agent_id>
#   codex-subagent → close_agent(<agent_id>) if still running
#   codex-exec     → nothing to kill (the exec process exited with the build)
#   legacy         → tmux kill-session -t <name> 2>/dev/null || true

git worktree remove .feature-workspaces/<name> \
  || git worktree remove --force .feature-workspaces/<name>   # --force only if leftover files block removal
git worktree prune
git branch -d feat/<name>   # PR was merged → branch is reachable from the integration branch, safe to delete

If git branch -d warns the branch isn't fully merged (e.g., the PR was squash-merged so commit SHAs differ), force with git branch -D feat/<name> after confirming via gh pr view <number> --json state that the PR is MERGED.

---

Guardrails

• Never run /opsx:archive from a workspace — it writes to openspec/specs/ and causes conflicts when parallel workspaces are active. Archive always runs on the integration branch ($BASE).
• Phase 13 clean-workspace check — after git fetch && git pull --ff-only, run git status to verify no unexpected files are modified.
• Verify OpenSpec changes exist before archiving — if openspec/changes/<name>/ is missing, the dispatch commit may have been lost. Recover from the original dispatch commit using git restore --source=<sha> before running archive.
• Cross-check signals against PR state — workspace signals can be stale (e.g., showing in_review after PR is merged). Always verify via gh pr view before updating product-context.yaml.
• Read signals AND lessons before removing worktrees — once git worktree remove runs, the worktree directory, signal files, and lessons.md are gone. Extract all needed data (status, lessons) first.

---

Reflect and Advance (Product-Cycle Mode)

Standalone mode: If product-context.yaml doesn't exist, skip the layer gate check. You can still run /aep-reflect if you want to classify observations.

After archiving, check the product context:

Layer Gate Check

If product-context.yaml exists and this feature was a dispatched story:

# Check: was this the last story in the current layer?
# Read product-context.yaml and check if all stories in the active layer are completed

If all stories in the current layer are completed, run the two-phase layer gate — a gate is green only when the layer is covered, not when one journey passes.

Read skills/e2e-test/policy.md first (if the project has the e2e-test skill): it declares the applicable tiers (run only these — a none-target / CLI project has no Tier-2), the dogfood target (none / local / deployed:<url>), and the timing. With journey_timing: post-deploy, run the journey against the deployed:<url> target here (after merge/deploy) — that is what flips scripted_passed → passed.

1. Tier-1 (machinery). Run the project's scripted suite for this layer. If green, set layer_gates[layer].status: scripted_passed and record the test file under evidence.scripted.
2. Tier-2/3 (product) + regression. Skip this step entirely if dogfood_target == none (CLI/library — Tier-2 N/A; prove criteria via Tier-1/Tier-3). Otherwise run the matching BDD journey/journeys in skills/e2e-test/journeys/ (layer: N) via their tool-selection.md, plus any applicable API drivers, and replay prior-layer journeys — seeding the policy's target first (a deployed:<url> target needs SERVER_URL=<url> bash skills/e2e-test/scripts/seed.sh, not local). Record evidence — screenshots, API JSON, PASS/FAIL per Then, and the two coverage matrices — in docs/layer-gates/<layer>.md.
3. Check coverage. Confirm every layer acceptance criterion maps to ≥1 proving test (coverage.criteria_covered == criteria_total). Genuine gaps were already auto-closed during /aep-build Phase 6; a deliberate deferral must carry a WAIVER: <reason> line. Never flip to passed while criteria are silently uncovered.
4. Flip to passed only when all applicable tiers are green AND coverage is complete-or-waived AND the regression replay passed; set completed_at. If only Tier-1 passed, leave it scripted_passed.
5. Ask the human before advancing. Surface the coverage summary (criteria_covered / criteria_total, per-tier status, any waivers) and confirm with the user that the next layer's design should begin. The gate flip is automatic-on-evidence; the advance is a human decision — the next /aep-dispatch proceeds only after that confirmation.

Feedback Loop

Consider running /aep-reflect to classify observations from this feature and update the product context:

/aep-reflect

This closes the feedback loop — bugs, refinements, and discoveries get routed back to the right phase.

---

Next Step

Pick the next story from the dispatch queue:

/aep-dispatch

Or classify feedback from the feature you just shipped:

/aep-reflect