feat(ai-project-generator): add DeepSeek as a selectable LLM provider

[parth0025]

Summary

Adds DeepSeek as a third LLM provider for the AI-Powered Project Creation feature, alongside the existing OpenAI and Anthropic providers. Selectable via LLM_PROVIDER="deepseek". Also fixes a latent bug in the AI task time estimator that affected all thinking models.

Changes

New provider — llmProvider/deepseekProvider.js

• DeepSeek exposes an OpenAI-compatible Chat Completions API, so this mirrors openaiProvider.js (same body shape, Bearer auth, response_format: json_object, choices[0].message + usage.*).
• Configurable base URL via DEEPSEEK_BASE_URL (default https://api.deepseek.com).
• Omits temperature for deepseek-reasoner (thinking mode); sends it for deepseek-chat / deepseek-v4-flash / deepseek-v4-pro.
• Full error mapping → friendly codes: 402 + quota-429 → LLM_QUOTA_EXCEEDED, 401/403 → LLM_AUTH_FAILED, 400 → LLM_BAD_REQUEST, 502/503 → LLM_UNAVAILABLE, timeout → LLM_TIMEOUT.
• Output clamp 65536 — above the orchestrator's 32000 default, under DeepSeek V4's 384K ceiling — so plan requests pass through unclamped.

Registry — llmProvider/index.js

• Registered deepseek in SUPPORTED, added to the fallback chain and isAnyProviderConfigured().

Config — config.js + .env.example

• Exposes DEEPSEEK_API_KEY / DEEPSEEK_MODEL; documents model options, timeout, and base-URL overrides.

Estimator fix — EstimatedTime/aiTaskEstimator.js

• Raised the estimate call's maxTokens 1024 → 8192 and timeout 60s → 120s.
• Why: thinking models (deepseek-reasoner, deepseek-v4-pro, OpenAI o-series) spend their output budget on hidden reasoning before the visible JSON. At 1024 the answer came back empty (finish_reason="length") → parseMinutes returned null → no estimate was ever saved. This is why DeepSeek v4-pro plans had zero task estimates while GPT-4.1 worked.

Verification

Check	Result
Provider selection via LLM_PROVIDER="deepseek"	✅
Per-model temperature handling (reasoner omits, others send)	✅
Output clamp pass-through (32000 unclamped; >ceiling → 65536)	✅
Live DeepSeek auth (401) + quota (402) error mapping	✅
Estimator sends 8192 + persists end-to-end	✅
gpt-4.1 unchanged — live call: 108 tokens, finish=stop, est=85	✅
Bonus: same fix repairs OpenAI o-series estimator path	✅

Existing functionality

• OpenAI and Anthropic providers are untouched and verified unchanged (gpt-4.1 produces identical results — the higher maxTokens cap is never hit, so zero cost/behavior change).

Notes for testing

• DeepSeek's deepseek-chat / deepseek-reasoner are aliases that now resolve to V4-Flash; deepseek-v4-flash / deepseek-v4-pro are the canonical names.
• Thinking models (reasoner, v4-pro) are slower and bill reasoning tokens; deepseek-chat/v4-flash are the cheapest/fastest for estimates.

🤖 Generated with Claude Code