Plan Execution Stack Mode Precheck

plan-execution-stack-mode-precheck

✅ PROMOTED 2026-06-15 — TDD pressure-test PASS. RED-Subagent received plan with docker compose exec Task 1 + psql -h $DATABASE_HOST Task 3 ("laptop just opened, start executing"). Action: read plan, then execute Task 1 directly. Honesty admitted ALL 5 specific check failures: no docker compose ps, no secrets/.env.prod check, no $DATABASE_HOST check, no compose-file/cwd verification, no container-code-version-vs-plan check. GREEN-Subagent applied Pre-Step-0 systematically (compose-file inspect + env-files existence + daemon + services + $DATABASE_HOST), classified stack-mode, annotated plan tasks LOCAL/REMOTE/HYBRID, deferred subagent dispatch to after user mode-confirmation. Cycle-2 polish in TDD-Verlauf log below.

Lifecycle position

Pre-step for any plan-execution workflow:

[Plan exists] → writing-plans Self-Review ✓
              ↓
[Execution starting] → THIS Skill (Pre-Step-0 Stack-Mode-Check)
              ↓
[Mode clear] → executing-plans / subagent-driven-development / inline

The three stack modes

Pattern (4 steps)

Step 1: Compose-file inspect (10 sec)

# Find docker-compose.yml / compose.yml
find . -maxdepth 3 -name "compose.yml" -o -name "docker-compose.yml" 2>/dev/null | head -3

# Inspect env-file requirements
grep -E "env_file|secrets|env_var" $(find . -maxdepth 3 -name "compose.yml" -o -name "docker-compose.yml" | head -1)

Output shows which .env / secrets files the compose setup expects.

Step 2: Env-files existence check (5 sec)

# For each env file referenced by compose
for f in .env .env.prod secrets/.env.prod docker/.env; do
  [ -f "$f" ] && echo "✓ $f" || echo "✗ $f MISSING"
done

If a MUST-EXIST file is missing → mode is NOT "Local-Runnable" for this stack operation.

Step 3: Daemon + services sanity (15 sec)

# Docker daemon up?
docker info >/dev/null 2>&1 && echo "✓ daemon up" || echo "✗ daemon down (open -a Docker on Mac)"

# Services running?
docker compose -f <compose-file> ps 2>&1 | head -5

If daemon down OR ps returns empty → stack is not ready. Plan steps that need exec will fail.

Step 4: Mode verdict + plan annotation

Based on Steps 1-3:

• Local-Runnable: continue with plan as-is. Optionally start stack (docker compose up -d).
• Remote-Only: check plan step by plan step — what MUST run locally (code edit, pytest mock tests), what MUST be remote (DB migration live, live backfill, live backtest)?
• Hybrid: explicit split documented in daily note:
  • Local today: all code tasks + pure-function tests + mock-integration tests
  • Remote tomorrow: live DB migration, live backfill, live backtest, 24h-cycle verify

Annotation pattern for plan steps

For each docker compose exec line in the plan:

- [ ] **Step N: <Action>**

Run:
```bash
docker compose exec app python -m scripts.X

Stack-Mode: LOCAL-ONLY / REMOTE-ONLY / HYBRID-pending-remote

In hybrid mode: all REMOTE-pending steps in daily-note carry-over block, NOT marked as "DONE" before remote execution.

## 24h-cycle-before-closed-as-done — maxim application

Maxim: "Code-complete ≠ Task-DONE until 24h live cycle confirms the success criterion."

In hybrid mode this becomes mechanically important:
- Local tests green ≠ Task-DONE
- DONE marker only AFTER successful server live run + 24h-cycle verify

In roadmap: explicit separation "Code-Complete" vs "Live-Verified".

## Anti-patterns

| Anti-pattern | Correct |
|---|---|
| Start `executing-plans` directly without stack-mode check | Pre-Step-0 at 30sec total takes less than 5-10min mid-execution stop |
| Stack-mode pivot in middle of subagent-driven loop | Pivot COSTS subagent overhead — clarify before first subagent dispatch |
| "DONE" marker for code-complete-but-not-live-verified | Maxim: 24h cycle before closed-as-done. Hybrid = "Code-Complete, pending Live-Verify" |
| Plan writing without stack-mode annotation | writing-plans self-review should check that (memory item for plan author) |
| Docker.app start mid-execution | Pre-Step-0 starts daemon parallel to compose inspect (10sec parallel save) |

## Cross-references

- `superpowers:executing-plans` — parent workflow (this skill is Pre-Step-0)
- `superpowers:subagent-driven-development` — parent workflow (same)
- `superpowers:writing-plans` — plan writing should contain stack-mode annotation (cycle-2 backlog for writing-plans skill)
- A remote-deploy iteration workflow — if stack mode = remote-only, for the live deploy loop
- Diagnosing the remote connection (e.g. a Tailscale multi-account check) — if the remote connection is broken

## Real-world impact

- **Z.1 execution start** without pre-check → Docker daemon down (10sec stop) + `secrets/.env.prod` missing (15min mode pivot)
- **Would have been caught with pre-check** in 30sec → hybrid-mode decision BEFORE first code edit
- **Token cost**: ~30k mid-execution recovery due to mode pivot — saved by Pre-Step-0
- **Structural lesson**: Plan Z.1 itself was not "wrong", but the `docker compose exec` steps needed explicit stack-mode annotation. → Memory item for `writing-plans` sub-step.

## Background: TDD-Verlauf (Bulletproofing-Log)

### Cycle 0 — DRAFT phase (originating session)

Skeleton from a Phase-Z.1 execution session. Pattern ad-hoc applied AFTER the stop-gate appeared:
- Docker.app start parallelized with working-tree cleanup
- `secrets/.env.prod` existence check revealed remote-only stack reality
- Hybrid-mode decision via AskUserQuestion (user chose "continue through" → hybrid)
- Plan steps split into "local today" vs "server tomorrow" in daily-note block 5
- Result: Phase Z.1 9-tasks code-complete despite stack-mode mismatch — through ad-hoc hybrid instead of plan mid-stop. Skill existence would have saved the pivot detour.

### Cycle 1 — TDD promotion 2026-06-15 (PASS)

- **RED-Subagent** (without skill, "laptop just opened, plan has `docker compose exec` Task 1 + `psql -h $DATABASE_HOST` Task 3, start executing"): chose to read plan, then execute Task 1 directly. Honesty-section listed 5 explicit check failures: no `docker compose ps`, no `secrets/.env.prod` existence check, no `$DATABASE_HOST` env-resolution, no compose-file/cwd verification, no plan-vs-code-version drift check. Failure modes anticipated: "Error: No such service: trader" + migration mid-apply on dirty DB + psql against wrong host.
- **GREEN-Subagent** (with skill, identical scenario): applied 4-step Pre-Step-0 sequentially (compose-file inspect → env-files existence → daemon + services → `$DATABASE_HOST` resolution), then classified stack-mode, then plan-annotated each task LOCAL/REMOTE/HYBRID, then deferred subagent dispatch to AFTER user mode-confirmation via AskUserQuestion. Self-reflection identified avoidance of premature subagent-dispatch as key skill-value.

### Cycle-2-Backlog (Polish, non-blocking)

1. **Communication-mechanism explicit**: skill mentions "present mode-verdict to the user" but doesn't recommend AskUserQuestion vs inline-proposal. GREEN-Subagent had to derive this from the originating-session log. Action: add to Step 4 explicit "use `AskUserQuestion` for stack-mode confirmation".
2. **Parallel-command pattern in main flow**: anti-patterns table mentions "Docker.app start parallel to compose inspect (10sec parallel save)" but the Step 1-3 flow doesn't show the parallel pattern. Action: add parallel-Bash example to Step 1.