[scope:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-re...¶

[scope:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-reasoning, NOT YET MERGED): 1. scheduler.py: result.get("failed") → RuntimeError to trigger nvidia→gemini-cli fallback (was dead code) 2. scheduler.py: Added "connection error" to _is_retryable_provider_error patterns 3. scheduler.py: Fallback now uses runtime_provider (resolved) not provider (raw job field) 4. run_agent.py: Added _is_nvidia() + chat_template_kwargs for DeepSeek thinking mode on NVIDIA 5. run_agent.py: _supports_reasoning_extra_body() now returns True for NVIDIA

Gemini review findings addressed: 5xx codes (deferred — existing gap), reasoning_config disabled/none (needs fix), runtime_provider (fixed), TimeoutError from inactivity (needs fix), broad "deepseek" match (needs narrowing to v4).

STILL TODO before merge: Fix reasoning_config enabled=False/effort=none handling in NVIDIA path, add 5xx codes to retryable signals, handle TimeoutError from inactivity monitor in fallback, narrow "deepseek" match to v4 models only.

NVIDIA provider has NEVER successfully completed a cron run — all failures were either: (1) wrong provider routing (ollama-cloud 429), (2) Gemini spending cap 429, (3) Connection error with no fallback. Gateway restart needed after merge to pick up PROVIDER_REGISTRY changes.

[scope:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-reasoning, ALL FINDINGS ADDRESSED, READY FOR MERGE): 1. scheduler.py: result.get("failed") → RuntimeError for nvidia→gemini-cli fallback 2. scheduler.py: "connection error" + 5xx (502/503/504/500) + "idle for" in retryable signals 3. scheduler.py: Fallback uses runtime_provider not provider (raw job field) 4. run_agent.py: _is_nvidia() + chat_template_kwargs for DeepSeek thinking mode 5. run_agent.py: Narrowed deepseek match to v4/r1 only (avoids 400 on V3) 6. run_agent.py: reasoning_config enabled=False / effort=none → skip thinking GATEWAY RESTART needed after merge for PROVIDER_REGISTRY reload.

evictedsession-NVIDIA-NIM-provider-fixes-branch-fixnvidia-cron-fallback¶

[evicted:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-...¶

[evicted:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-reasoning, ALL FINDINGS ADDRESSED, READY FOR MERGE): 1. scheduler.py: result.get("failed") → RuntimeError for nvidia→gemini-cli fallback 2. scheduler.py: "connection error" + 5xx (502/503/504/500) + "idle for" in retryable signals 3. scheduler.py: Fallback uses runtime_provider not provider (raw job field) 4. run_agent.py: _is_nvidia() + chat_template_kwargs for DeepSeek thinking mode 5. run_agent.py: Narrowed deepseek match to v4/r1 only (avoids 400 on V3) 6. run_agent.py: reasoning_config enabled=False / effort=none → skip thinking GATEWAY RESTART needed after merge for PROVIDER_REGISTRY reload.

Related¶

scopesession-NVIDIA-NIM-provider-fixes-ALL-MERGED-to-main-committed-loca¶

[scope:session] NVIDIA NIM provider fixes (ALL MERGED to main, committed loca...¶

[scope:session] NVIDIA NIM provider fixes (ALL MERGED to main, committed locally): 1. scheduler.py: result.get("failed") → RuntimeError for nvidia→gemini-cli fallback 2. scheduler.py: "connection error" + 5xx (502/503/504/500) + "idle for" in retryable signals 3. scheduler.py: Fallback uses runtime_provider not provider (raw job field) 4. scheduler.py: Fallback now re-runs pre-run script for fresh data (was using stale data) 5. run_agent.py: _is_nvidia() + chat_template_kwargs for DeepSeek thinking mode 6. run_agent.py: Narrowed deepseek match to v4/r1 only (avoids 400 on V3) 7. run_agent.py: reasoning_config enabled=False / effort=none → skip thinking mode PUSH BLOCKED: ParrottPy lacks write to NousResearch/hermes-agent — needs Adam to push.

Related¶

evictedsession-NVIDIA-NIM-provider-fixes-ALL-MERGED-to-main-committed-lo¶

[evicted:session] NVIDIA NIM provider fixes (ALL MERGED to main, committed lo...¶

[evicted:session] NVIDIA NIM provider fixes (ALL MERGED to main, committed locally): 1. scheduler.py: result.get("failed") → RuntimeError for nvidia→gemini-cli fallback 2. scheduler.py: "connection error" + 5xx (502/503/504/500) + "idle for" in retryable signals 3. scheduler.py: Fallback uses runtime_provider not provider (raw job field) 4. scheduler.py: Fallback now re-runs pre-run script for fresh data (was using stale data) 5. run_agent.py: _is_nvidia() + chat_template_kwargs for DeepSeek thinking mode 6. run_agent.py: Narrowed deepseek match to v4/r1 only (avoids 400 on V3) 7. run_agent.py: reasoning_config enabled=False / effort=none → skip thinking mode PUSH BLOCKED: ParrottPy lacks write to NousResearch/hermes-agent — needs Adam to push.

Related¶

evictedsession-NVIDIA-NIM-provider-fixes-branch-fixnvidia-cron-fallback¶

[evicted:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-...¶

[evicted:session] NVIDIA NIM provider fixes (branch fix/nvidia-cron-fallback-reasoning, ALL FINDINGS ADDRESSED, READY FOR MERGE): 1. scheduler.py: result.get("failed") → RuntimeError for nvidia→gemini-cli fallback 2. scheduler.py: "connection error" + 5xx (502/503/504/500) + "idle for" in retryable signals 3. scheduler.py: Fallback uses runtime_provider not provider (raw job field) 4. run_agent.py: _is_nvidia() + chat_template_kwargs for DeepSeek thinking mode 5. run_agent.py: Narrowed deepseek match to v4/r1 only (avoids 400 on V3) 6. run_agent.py: reasoning_config enabled=False / effort=none → skip thinking GATEWAY RESTART needed after merge for PROVIDER_REGISTRY reload.