e2e — board-steward end-to-end test harness

A reusable harness that codifies the verification patterns proven during ARCH REDESIGN v2 (multi-board routing + reconciliation). Run it to regression-test those overhauls — or extend it for the next one — instead of hand-writing throwaway test scripts each time.

Source of the redesign it tests: ~/Desktop/WorkBoard/docs/ARCH_REDESIGN_V2.md.

The cardinal rule (why this exists)

A test must NEVER pollute the user's live board or global state. The harness enforces this structurally:

• Isolated state — BOARD_REGISTRY / BOARD_ASSIGNMENTS / BOARD_ACTIVE are pointed at temp files; BOARD_NO_SERVER=1 forces direct file writes for tests that don't need a server. The real ~/.board-steward/ is never touched.
• Throwaway boards — every board under test is a temp dir, seeded in-memory or from templates/board.json.
• Live-board guard — the live board's (:7891) card count is captured before the run and asserted unchanged after. If a test touches it, the suite fails.
• Cleanup — all temp dirs are removed at the end.

How to run

python3 ~/.claude/skills/e2e/e2e_workboard.py multiboard    # routing + $HOME disambiguation (no LLM, free)
python3 ~/.claude/skills/e2e/e2e_workboard.py recon         # only_discovered flag + gating + CLAUDECODE path (free)
python3 ~/.claude/skills/e2e/e2e_workboard.py recon-haiku   # real Haiku: IP→done / IP→backlog (~1 Haiku call)
python3 ~/.claude/skills/e2e/e2e_workboard.py all           # multiboard + recon (free tier)
python3 ~/.claude/skills/e2e/e2e_workboard.py all --haiku   # everything, including the Haiku E2E

Override the repo with BOARD_REPO=/path/to/WorkBoard (default ~/Desktop/WorkBoard).
Exit code 0 = all passed, 1 = a failure (printed at the bottom), 2 = bad args / repo not found.

Test catalog (what each group proves)

multiboard — the v2 multi-board work

• routing_isolation — a card added from inside project A lands on A, never B; last-active tracks the last-mutated board.
• disambiguation — at $HOME, the last-active board wins over a newer-mtime board; mtime is only the fallback (the wrong-board bug that started v2).

recon (free) — the v2 reconciliation work

• only_discovered_flag — reconcile_sweep(only_discovered=True) scopes to bootstrap-mined cards; False reconciles every non-done card (so SessionStart catches live cards).
• gates_short_circuit — --reconcile-only makes no Haiku call when there are no non-done cards (Gate A) or no recorded project activity (Gate B).
• claudecode_path — CLAUDECODE=1 writes a prose recon_pending.json (no Haiku); proves the spawn's env -u CLAUDECODE is load-bearing for the autonomous Haiku path.

recon-haiku (costs ~1 Haiku call) — the real end-to-end

• haiku_e2e — a real claude -p --model haiku pass moves a shipped In-Progress card → done and a "nvm/skip" card → backlog.

Extending for the NEXT overhaul

Add a function test_<group>_<name>(ctx) in e2e_workboard.py and register it in GROUPS. Use the context helpers — they keep the cardinal rule automatic:

• ctx.board(cards=[...]) → a throwaway project board.json (pass from_template=True for the real seed).
• ctx.card(num, column, title, tags=[...]) → a minimally-valid card dict (real cards always carry an id).
• ctx.assert_eq(name, got, want) / ctx.assert_true(name, cond, detail) → record pass/fail.
• State is already isolated in Ctx.__init__; the live-board guard runs automatically.

Always: isolate state, use throwaway boards, and never assume the live server's state. If a feature needs a running server, start it on an explicit free port (e.g. 7950) and verify :7891's count is unchanged.