fix(loops): sandbox live-path — sandbox 0.4.3 (SSE reconnect) + poll session-id + fail-loud infra errors

[drewstone]

What

The in-box agentic batch path (runLoop → sandbox → benchmark) was losing its SSE prompt stream on long quiet turns (clone/build/test), surfacing as Stream dropped without terminal event → Replay endpoint 404. Root cause: the bench resolved @tangle-network/sandbox 0.4.2, which drops the stream on a long tail. 0.4.3's reconnect-backoff survives it.

Verified

A commit0 in-box rollout (clone real repo → opencode implements → repo pytest judges) streamed >520s with zero drops on 0.4.3, completed end-to-end, wrote real RunRecords. On 0.4.2 the same path dropped at ~256s.

Changes

• bench: @tangle-network/sandbox ^0.4.0 → ^0.4.3.
• experiment.ts: surface the underlying iter0.error (an [infra-cause] line) instead of hiding every failure behind a generic INFRA-ERROR label — a swallowed-error fail-loud gap that cost real debugging time.
• sandbox-lineage.ts poll-mode: follow the session id dispatchPrompt actually assigned (DispatchedSession.sessionId), not the client-minted one. Polling the wrong id 404s the session-events endpoint.

Residual (not fixed here)

Poll-mode (dispatchPrompt + session().result()) is still gated on a platform session-events endpoint for dispatched sessions (result()→_sessionEvents→404 on all SDK versions). SSE is the working batch path; poll-mode remains opt-in and currently unusable until the platform serves events for dispatched sessions.

[tangletools]

⚠️ Review Incomplete — 2b23c33a

At least one required reviewer lane failed closed. No approval or request-changes review was published. Trigger a fresh review on the current PR head.

_{tangletools · 2026-06-07T00:46:41Z}

[tangletools]

✅ No Blockers — 2b23c33a

Readiness 92/100 · Confidence 80/100 · 1 finding (1 low)

deepseek: Correctness 92 · Security 92 · Testing 92 · Architecture 92

Full multi-shot audit completed 4/4 planned shots over 4 changed files. Global verifier still owns final merge decision.

🟡 LOW No test coverage for the platform-session-remap scenario — src/runtime/sandbox-lineage.ts

The poll-mode test at tests/loops/sandbox-lineage.test.ts:167-221 uses a fake dispatchPrompt that returns o?.sessionId ?? 'minted' — always echoing the supplied ID. The core fix (using the platform-assigned sessionId) has no test where dispatchPrompt returns a different ID than the one supplied. Adding a test with a divergent sessionId would cover the exact bug this PR fixes. Impact: low — the fix is straightforward and the existing test proves dispatchPrompt fires and the result path works end-to-end.

---

_{tangletools · 2026-06-07T01:09:23Z · trace}

[tangletools]

✅ Approved — 1 non-blocking finding — 2b23c33a

Full multi-shot audit completed 4/4 planned shots over 4 changed files. Global verifier still owns final merge decision.

Full immutable report for this review: trace

Summary comment for this run: full summary

---

_{tangletools · 2026-06-07T01:09:23Z · immutable trace}