[codex] fix TUI fallback providers for dflash empty responses

[OmarB97]

What changed

• Load fallback_providers / legacy fallback_model config in TUI gateway session creation.
• Pass the resolved fallback chain into AIAgent, matching oneshot CLI behavior.
• Preserve the full fallback chain for TUI background turns instead of only carrying _fallback_model.
• Add regression coverage for TUI session construction and background agent kwargs.

Root cause

TUI sessions ignored the configured fallback chain when constructing AIAgent. In local taro/dflash sessions, dflash can occasionally return an empty post-tool response. Hermes then retried the same context but had no fallback available in TUI, despite fallbacks being present in config, so the user saw a terminal No reply result.

Impact

TUI behavior now matches oneshot CLI: if the primary model repeatedly returns empty output, the configured fallback chain can activate instead of ending the turn with no recovery path.

Validation

• python3 -m pytest tests/test_tui_gateway_server.py -q -k "make_agent or background_agent_kwargs"
• python3 -m pytest tests/run_agent/test_run_agent.py -q -k empty_response
• python3 -m pytest tests/run_agent/test_provider_fallback.py tests/cli/test_cli_init.py -q
• Live taro import canary confirmed _make_agent() passes the configured fallback chain.
• Live hermes-phone tmux session was restarted after deploying the patched gateway file to the installed taro checkout.

Notes

The full local tests/test_tui_gateway_server.py -q run had one unrelated environment-dependent browser-launch expectation failure on this Mac; the fallback-focused tests passed.

Fork mirror: OmarB97#47

[alt-glitch]

Related: #28753 (bug report — TUI sessions don't propagate fallback_model/fallback_providers to AIAgent) and PR #18310 (open, same fix). This PR supersedes #18310 with additional dflash empty-response handling and background-agent kwargs coverage.

[OmarB97]

Review Summary: APPROVE ✓

Focused, correct fix for the TUI dflash empty-response regression. The change is minimal (3 lines of production code + 2 targeted tests) and addresses the root cause: background agents in the TUI gateway were receiving only the first fallback provider instead of the full chain.

What the PR does

tui_gateway/server.py — 2 changes:

1. _make_agent: Now calls get_fallback_chain(cfg) and passes the full list as fallback_model. Previously, the config-level fallback chain was never wired into the agent constructor.
2. _background_agent_kwargs: Prefers agent._fallback_chain (full list) over agent._fallback_model (single dict legacy attr). Falls back to _fallback_model when the chain is empty or absent.

tests/test_tui_gateway_server.py — 2 new tests:

• test_make_agent_passes_fallback_chain_from_config: Verifies the config → _make_agent → AIAgent path passes the full chain.
• test_background_agent_kwargs_preserves_full_fallback_chain: Verifies the agent → _background_agent_kwargs → runner path prefers the chain.

Correctness analysis

Verified against init_agent (agent/agent_init.py:885-897): the fallback_model parameter already supports both list and dict — it filters valid entries and sets _fallback_chain. The PR leverages this existing capability correctly.

Edge cases handled:

• Empty chain ([]) → falsy → falls back to _fallback_model or None ✓
• None chain → or [] → list([]) → [] → falsy → same fallback ✓
• list() creates shallow copy to prevent mutation ✓

Verification

• Both new tests pass ✓
• Existing adjacent tests (test_make_agent_handles_null_agent_config, test_background_agent_kwargs_handles_null_agent_config) still pass ✓
• No other callers of _background_agent_kwargs or _make_agent affected ✓

Non-blocking suggestions

1. Type hint: fallback_model: Dict[str, Any] in both AIAgent.__init__ and init_agent should be Dict[str, Any] | List[Dict[str, Any]] | None. Pre-existing inaccuracy — init_agent already handles lists, but the hint hasn't been updated. Consider fixing in a follow-up.

2. Test coverage: No test for the case where _fallback_chain is explicitly set to [] (empty list) and _fallback_model is also None — the result would be None. This is correct behavior but worth a dedicated test assertion for clarity.

[OmarB97]

Obsolete: upstream 4b073d0 independently implemented the identical TUI fallback-provider fix with strictly more coverage (_load_fallback_model + _agent_fallback_model incl. the empty-chain nuance, 102 test lines). Nothing here is lost on main. Closing.