A chat turn that fails before streaming never surfaces an error — the UI hangs to a generic connection timeout

[nocstah]

Summary

When a chat turn fails before the provider produces any stream (e.g. the provider probe fails), no error event reaches the frontend — the user stares at an idle chat until the client's own ~2-minute timeout shows "No response from the agent after 2 minutes. Try again or check your connection."

Problem

What happened vs expected: web_chat logged run_chat_task failed … [claude-code] \claude` CLI not installedandreport_errorfired at11:42:46— the backend knew the turn was dead within milliseconds. The UI was never told: nochat_error-style socket event was emitted for the turn (an external_transfer_pendingegress event did go out, but nothing terminal). Expected: an immediate, actionable error bubble ("Claude Code CLI not found — …") the momentrun_chat_task` fails.

The "check your connection" copy is also misleading — the failure was local and permanent, not connectivity.

Steps to reproduce:

1. Configure the claude-code provider and make the CLI unresolvable (e.g. launch from Finder with the CLI only in ~/.local/bin — see the companion issue — or temporarily rename the binary).
2. Send a message.
3. Backend log shows run_chat_task failed instantly; the chat UI shows a spinner/typing state for ~2 minutes, then a generic connection error.

Version/platform: v0.63.14, macOS (Darwin 25.3.0), desktop bundle.

Solution (optional)

Emit a terminal per-turn error event on the socket surface from the run_chat_task failure path (src/openhuman/web_chat/ops.rs — the same place that logs the WRN), carrying the provider's error string; frontend renders it as the turn's result. The 2-minute client timeout then remains a true last resort for transport loss only.

Acceptance criteria

• Repro gone — a provider failure surfaces in the chat within ~1s with the provider's actual error message.
• Regression safety — a test covers "run_chat_task error ⇒ terminal error event emitted for that turn".
• Diff coverage ≥ 80% — the fix PR meets the changed-lines coverage gate (Vitest + cargo-llvm-cov, enforced by .github/workflows/ci-lite.yml).

Related

Companion issue: Finder-launch CLI resolution (the failure that exposed this); its fix removes this particular trigger but not the gap.