[andyhsu6]

Why

Chapter generation was taking 5–9 minutes per chapter. Measurements on one real project showed the cost is entirely reasoning tokens, not input size:

input chars	reasoning tokens	first token	total
208,363	18,888	252s	282s
207,907	19,675	280s	316s
207,782	36,114	477s	514s
53,921 (context cut 4x)	33,546	476s	508s

Cutting the prompt by 74% left the reasoning count unchanged, so context work cannot fix this. The gateway does accept a thinking-effort level, which was never sent.

What

• ChapterGenerateRequest / BatchGenerateRequest: new reasoning_effort: low | medium | high, default medium. Explicit null keeps the gateway default (the field is then not injected — asserted by test).
• Plumbed AIService.generate_text_stream → providers → OpenAIClient payload. Anthropic/Gemini accept and ignore it; OpenAI forwards it, including the tool-call round.
• Both retry recursions forward the level (stream_options rejection, learned output ceiling). Without that a retry silently drops back to the gateway default and the chapter takes 5–8 minutes again — covered by a dedicated test.
• Style tail block gained a length-floor clause (93% of the target, zh/en), because lower effort returns shorter chapters.

Measurements (same chapter, same prompt, glm-5.3-flash)

level	n	reasoning tokens	total	output chars
default (not sent)	4	18,888–36,114	282–514s	3,380–4,402
low	4	100–303	43–53s	2,208–2,579
low + length floor	2	92–426	46–54s	2,708–2,909
medium + length floor	2	294–661	46–64s	2,726–3,363
high	1	9,747	209s	2,646

Style metrics stay inside the range of the drafts already delivered for this project (sentence length 22.4–25.5 vs 21–29; short-sentence ratio 0.29–0.37 vs 0.21–0.38). Lower effort does not make the short-sentence habit worse, but it does not fix it either — that remains a separate style issue.

finish_reason was stop in every sample, so the truncation gate is unaffected.

Test plan

• pytest tests/ -q in a clean checkout: 1684 passed, 0 failed (baseline 1669 + 15 new).
• New tests/test_reasoning_effort.py (15 cases): schema default/validation, payload injection, no field when unset, retry recursion keeps the level, length-floor text (zh/en/absent).
• Updated test_every_generation_path_appends_the_style_tail_block to match the new 3-argument call form (still guards all 5 wiring sites).
• End-to-end through the production call chain (AIService.generate_text_stream(reasoning_effort="medium"), real gateway, 213k-char prompt): 96.6s total, 1,875 reasoning tokens, 3,420 chars, finish_reason=stop.

refs #295