docs: sharpen verification, debug-loop, and subagent guidance

Capture preferred away/E2E working style: real smoke over fake-done, defined debug-loop limit, wallclock note in handoff.

Co-authored-by: Cursor <cursoragent@cursor.com>

AGENTS.md

@@ -134,6 +134,12 @@ Fränky's bar is **pragmatic, not TDD-by-default** (`USER.md`: tests score 2/5).
 - Follow project `SOUL.md` if stricter (e.g. pytest before handoff)
 - Do not block small fixes on missing test infrastructure
 
+### Verification (pragmatic)
+
+- If the project has a runnable stack (Compose, CLI, service): prefer a **real smoke / happy-path** before handoff — not “code looks fine”
+- Waiting on slow builds or pipeline jobs is OK and preferred over fake-green
+- Unit/integration tests still apply when a suite exists; smoke does not replace them, and they do not replace smoke when infra is the product
+
 ---
 
 ## Commits & Issues
@@ -159,9 +165,10 @@ Scratchpad — use for blocked tasks, session park, overnight-run status, handof
 | Task | Subagent |
 |------|----------|
 | Unknown codebase or many files to scan | **Explore** |
+| Greenfield / large feature implementation | **generalPurpose** (parent keeps orchestrating deploy/ntfy) |
 | Browser / E2E / web UI | **Playwright** |
 | Broad research | **generalPurpose** or **Explore** |
-| Shell / git batch ops | **shell** |
+| Shell / git / sync / remote compose | **shell** |
 
 **Explore** for any project type when orientation is unclear — not only web. **Playwright** only when there is a web UI to test.
 
@@ -200,9 +207,15 @@ See `INFRASTRUCTURE.md` → **Project-local overrides**.
 ## Error Handling
 
 - Root cause first — do not guess
+- **Debug loop:** same failed hypothesis/fix tried **more than 2–3 times** → change strategy or escalate. A single deliberate course correction (wrong package → correct image, rebuild) is not a loop — note it and continue
+- Do not paper over broken tool/dependency versions with hacks when the real fix is a rebuild or version pin
 - Escalate on: debug loops, missing uninstallable software
 - Long-running jobs: no arbitrary timeout unless Fränky or SOUL says so; name `progress.log` when applicable
 
+### Away / long-task handoff
+
+After Unattended work (or any task that took wall-clock minutes): in chat and/or ntfy, **one short line on where the time went** (e.g. image build, E2E wait, one root-cause rebuild) — no essay.
+
 ---
 
 ## Notifications

BOUNDARIES.md

@@ -50,7 +50,7 @@ Rules that apply in **every** project unless explicitly overridden in a project-
 - **No hallucination** — if you don't know, say so; read the file, run the command, check docs
 - **No outdated advice** — flag when knowledge may be stale; verify against project code/version
 - **Stop and escalate** on:
-  - Debug loops (same error, same failed fix repeated)
+  - Debug loops — same failed hypothesis/fix **>2–3 times** (one deliberate course correction is fine)
   - Required software missing and not installable in context
   - Conflicting instructions you cannot resolve
   - **Plan Mode / read-only mode** blocks file edits — see below

USER.md

@@ -72,8 +72,9 @@ Who the human is and how they like to work. **Operational rules live in `AGENTS.
 | Small diffs over big refactors | 2 |
 | Tests are non-negotiable | 2 |
 | Docs only when necessary | 4 |
+| Real smoke before handoff > fake-done | 4 |
 
-**Tension:** Sometimes I want the AI to finish while I'm away — but I get angry when it doesn't work.
+**Tension:** Sometimes I want the AI to finish while I'm away — but I get angry when it doesn't work. Prefer a slow real smoke over a fast “looks fine”.
 
 ## Ideal Agent