feat: real cross-backend integration tests (fresh + resume) + pi harness

[saucam]

Follows the codex integration work. Fake fixtures can't prove a backend actually works, or that a conversation survives a hop to a different backend — that's the exact blind spot that let the codex sandbox bug ship green. This adds opt-in integration tests that drive the real backends.

backends.integration.test.ts — the headline promise, live

CODEOID_INTEGRATION=1, backends from CODEOID_INTEGRATION_BACKENDS (default claude,codex), through the real SessionManager verbs:

• FRESH — each enabled backend takes a brand-new turn and replies.
• RESUME — a conversation started on backend S is session.fork-ed onto a different backend T; T must answer a question about a fact only established during S's turn. This is codeoid's "branch claude, continue on codex" promise, and it exercises the exact seedFromHistory path fork/switch rely on.

Verified locally against claude (Agent SDK) + codex (@openai/codex@0.144.1):

FRESH: claude ✓   FRESH: codex ✓
RESUME: claude → codex ✓   RESUME: codex → claude ✓     (5 pass, ~43s)

Both resume directions carry the fact — so history flows across backends regardless of which one produced it.

provider-pi.integration.test.ts — the pi counterpart

CODEOID_PI_INTEGRATION=1. The pi analogue of provider-codex.integration.test.ts, driving the real PiProvider against the actual pi binary (bundled @earendil-works/pi-coding-agent, system pi, or providers.pi.command):

• P1 — a fresh text turn runs to turn_done with a reply, no error.
• P2 — pi runs a shell tool end-to-end and its output reaches codeoid.

Honest caveat: pi needs a provider key/login to take a turn, and I have no pi auth on any box I can reach, so I could not execute P1/P2 here — they're written + gated and run wherever pi is authed. pi's turn mechanics remain covered offline by the extensive provider-pi.test.ts (fake-pi), and pi is also picked up automatically by the backends matrix above once you add it to CODEOID_INTEGRATION_BACKENDS on a pi-authed host.

Safety

Both files skip cleanly with no opt-in — CI stays hermetic (no network, no keys). Offline suite: 1347 pass / 9 skip / 0 fail, tsc + biome clean.

🤖 Generated with Claude Code

[github-actions]

🤖 Gemini code review

Introduces robust cross-backend and Pi provider integration tests that run against real binaries and SDKs when opted in via environment variables. The tests cover fresh turns and conversational resume/handoff between different backends. The implementation is clean and carefully gated to ensure CI remains hermetic.

Findings: 🔴 0 · 🟠 0 · 🟡 0 · 🟢 1

---

_{Tokens spent · ⬆️ Input: 5,259 · ⬇️ Output: 156 · Σ Total: 10,102
Total may be higher due to thinking token counts.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.51%. Comparing base (ccf28d3) to head (1a39384).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #165   +/-   ##
=======================================
  Coverage   84.51%   84.51%           
=======================================
  Files         104      104           
  Lines       18319    18319           
=======================================
  Hits        15482    15482           
  Misses       2837     2837           

Flag	Coverage Δ
daemon	84.51% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.