Adc gemini

[froody]

Mainly add ADC auth so that I can use GOOGLE_CLOUD_LOCATION/GOOGLE_CLOUD_PROJECT instead of an api key to access google models. Also made mid tier display, now my status looks like

% nadirclaw status
NadirClaw Status
----------------------------------------
Simple model:  Qwen3-Coder-480B-A35B-Instruct
Mid model:     gemini-3-flash-preview
Complex model: gemini-3.1-pro-preview
Tier config:   explicit (env vars)
...

[doramirdor]

Review notes (automated scheduled check)

Thanks @froody — ADC support is a great addition and you also caught two real bugs along the way. Tests pass locally (663 ✓). A few items worth addressing before merge:

🐛 Real bug fixes (good)

1. _dispatch_model_stream Gemini path: _stream_gemini is an async def ... yield ... (i.e., async generator), but it was being consumed with a plain for loop. That code path would have raised TypeError: 'async_generator' object is not iterable the moment anyone actually streamed via Gemini. The async for fix is correct, and the misleading "_stream_gemini is a sync generator" comment is rightly gone.
2. Robust chunk / finish_reason parsing: defending against the google-genai SDK returning enum-like objects (with .value / .name) is reasonable.
3. savings.py None tolerance: selected_model: None and tier: None show up in real logs for failed/aborted requests — both fixes are valid.

✋ Please address

1. Duplicate property definitions in nadirclaw/settings.py (lines 46-59 of the PR diff) — MID_MODEL, REASONING_MODEL, and FREE_MODEL are already defined further down in the file (lines ~98, 172, 177 on this branch), with proper fallbacks:

# Already present:
@property
def MID_MODEL(self) -> str:
    return os.getenv("NADIRCLAW_MID_MODEL", "") or self.SIMPLE_MODEL  # line 98

# Your additions (lines 46-59) define them again as empty-string-only stubs.

Python silently keeps only the last @property assignment per class, so the existing definitions still win and your three additions are effectively dead code — but they make the file confusing to read. Suggest removing the new stubs at lines 46-59. (I verified at runtime: settings.MID_MODEL still returns the SIMPLE fallback, not "".)

2. Unnecessary test loosening in tests/test_e2e.py:

-            "messages": [{"role": "user", "content": "Solve the halting problem"}],
+            "messages": [{"role": "user", "content": "Solve the halting problem test free profile"}],
...
-        assert routing["strategy"] == "profile:free"
+        assert "profile:free" in routing["strategy"]

I reverted these two lines locally and the test still passes — the strategy is literally "profile:free" (set at server.py:1343). Looks like a leftover from local debugging. Either restore the strict assertion (matches the sibling test_reasoning_profile_routes_to_complex), or tell me what you were actually trying to test — happy to be wrong here.

🟡 Minor

• _get_gemini_client ADC caching: caching the client under "adc_default" is fine because google.auth.default() returns auto-refreshing Credentials. No action needed; just flagging since the original api_key-keyed cache had per-key isolation that this loses (which is also fine — ADC is a singleton concept).
• The two _call_gemini / _stream_gemini early-return guards were both deleted; _get_gemini_client now centralizes the "no creds" HTTPException. Worth a one-line comment in those callers pointing at the new behavior, but not blocking.
• Optional: the JSONResponse headers=_routing_headers(...) change is defensive — FastAPI does merge headers from the implicit response: Response parameter onto the returned response, but being explicit is clearer.

Suggested next step

Squash these fixes:

• drop the three duplicate properties in settings.py
• restore the strict assertion + original prompt in test_free_profile_routes_to_simple

…and this is ready to merge. The ADC feature is genuinely useful and the streaming fix is overdue.

— posted by scheduled review bot on main

[doramirdor]

Friendly ping @froody — no rush, but I'd love to land the ADC support and your _stream_gemini async-generator fix (that one's a real bug). The only blockers from my review above are tiny:

1. Drop the three duplicate @property stubs at nadirclaw/settings.py lines 46–59 (the real definitions further down already cover them with fallbacks).
2. Restore the strict routing["strategy"] == "profile:free" assertion + original prompt in tests/test_e2e.py.

If you're swamped, happy to push those two changes to your branch myself (if maintainerCanModify is on) or open a stacked PR with just the cleanup so we can merge yours unchanged — whichever you prefer. Let me know.

— scheduled review bot follow-up

[doramirdor]

Quick administrative note: I approved the fork CI workflow run (it had been sitting in action_required since first push — that's on me for not catching it in my earlier reviews). It just went green across Python 3.10 / 3.11 / 3.12 — see run 26099724743. No new asks; the two cleanup items from my 2026-05-20 review are still the only blockers. — scheduled review bot

[doramirdor]

Pushed cleanup as 01a4704 directly to your branch (since maintainerCanModify is on) — covers the two items from the 2026-05-20 review:

• Dropped the three duplicate @property stubs in nadirclaw/settings.py.
• Restored the strict routing["strategy"] == "profile:free" assertion and original prompt in tests/test_e2e.py.

Full suite still green locally (663 ✓). Waiting on CI; will merge once it goes green unless you object in the next few hours. Thanks again for the ADC support and the streaming async-generator fix — both legit wins.

— scheduled review bot

[doramirdor]

CI is green on the cleanup commit — run 26362224241 across Python 3.10 / 3.11 / 3.12.

I'll leave the merge button to the next scheduled check (~24h) so you have a real window to push back if you'd rather take a different cut at the cleanup — otherwise it'll go in then.

— scheduled review bot