feat(core): extended thinking + bigger output budget#83
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Findings
-
[Blocker] 缺少 DCO
Signed-off-by，会违反仓库硬约束并阻塞合并。证据：PR 当前提交b630bd80eef797b0c9b9dcc5e9dd0b56f6b37e0c的 commit message 未包含Signed-off-by:行。
Suggested fix:git commit --amend -s --no-edit git push --force-with-lease
-
[Major] 将
maxTokens无条件固定为32000，会把所有模型都提升到同一输出上限，存在超过模型实际max_output_tokens导致 4xx 失败的回归风险。证据packages/core/src/index.ts:278（无条件设置）以及默认模型目录含较小/异构模型packages/shared/src/config.ts:97packages/shared/src/config.ts:98packages/shared/src/config.ts:106packages/shared/src/config.ts:107。
Suggested fix:function maxTokensForModel(model: ModelRef): number | undefined { if (model.provider === 'anthropic' && /claude-(?:opus|sonnet)-4/i.test(model.modelId)) { return 32000; } return undefined; // 其他模型走 provider 默认，避免超限 } const maxTokens = maxTokensForModel(input.model); result = await completeWithRetry(input.model, input.messages, { apiKey: input.apiKey, ...(maxTokens !== undefined ? { maxTokens } : {}), ...(reasoning !== undefined ? { reasoning } : {}), }, ...);
Summary
- Review mode: initial
- 检查范围覆盖本次最新 diff（3 files）。发现 2 个问题（1 Blocker, 1 Major）。
docs/VISION.md与docs/PRINCIPLES.md：Not found in repo/docs。
Testing
- Not run (automation)
open-codesign Bot
| apiKey: input.apiKey, | ||
| ...(input.baseUrl !== undefined ? { baseUrl: input.baseUrl } : {}), | ||
| ...(input.signal !== undefined ? { signal: input.signal } : {}), | ||
| maxTokens: MAX_OUTPUT_TOKENS, |
There was a problem hiding this comment.
这里变成了对所有 provider/model 的硬编码上限。对于默认配置里的异构模型（例如 、），固定 32k 可能超过模型输出上限并触发 4xx。建议按模型能力分配上限，或仅对 Claude 4 开启 32k，其他模型保持 provider 默认。
Lift the per-call output budget to 32k tokens (was pi-ai's default of ~1/3 the context window, ≈10k for Opus 4) and ask Claude 4.x models to "think before answering" via pi-ai's reasoning='high' knob, which the Anthropic adapter maps to extended thinking. Non-Anthropic / pre-4 models keep their default reasoning behaviour but still benefit from the larger output budget.
hqhq1025
force-pushed
the
wt/loop-feat-thinking-bigger-tokens
branch
from
b630bd8 to
7bf6e61
Compare
There was a problem hiding this comment.
Findings
-
[Blocker] 缺少 DCO
Signed-off-by，会违反仓库硬约束并阻塞合并。证据：当前提交7bf6e6108de2cb6168b19c7d1ff9cbe1082c7df8的 commit message 未包含Signed-off-by:行。
Suggested fix:git commit --amend -s --no-edit git push --force-with-lease
-
[Major]
maxTokens被无条件固定为32000，会把所有 provider/model 都推到同一输出上限；对较小输出上限模型会触发 provider 4xx 并导致生成失败。证据packages/core/src/index.ts:291，且该行为被测试固化packages/core/src/generate.test.ts:132。
Suggested fix:function maxTokensForModel(model: ModelRef): number | undefined { if (model.provider === 'anthropic' && /claude-(?:opus|sonnet)-4/i.test(model.modelId)) { return 32000; } return undefined; // 其他模型走 provider 默认，避免超限 } const maxTokens = maxTokensForModel(input.model); result = await completeWithRetry(input.model, input.messages, { apiKey: input.apiKey, ...(maxTokens !== undefined ? { maxTokens } : {}), ...(reasoning !== undefined ? { reasoning } : {}), }, ...);
Summary
- Review mode: follow-up after new commits
- 本次最新 diff（3 files）仍有 2 个问题（1 Blocker, 1 Major）。
docs/VISION.md与docs/PRINCIPLES.md: Not found in repo/docs.
Testing
- Not run (automation)
open-codesign Bot
| apiKey: input.apiKey, | ||
| ...(input.baseUrl !== undefined ? { baseUrl: input.baseUrl } : {}), | ||
| ...(input.signal !== undefined ? { signal: input.signal } : {}), | ||
| maxTokens: MAX_OUTPUT_TOKENS, |
There was a problem hiding this comment.
这里对所有模型无条件传 maxTokens: 32000。这会把 OpenAI/Groq/部分 OpenRouter 小模型也抬到 32k，上游若不支持会直接 4xx。建议仅对已确认支持的 Claude 4.x 设 32k，其它模型保持 provider 默认。
1 participant
让模型有 10k thinking budget + 32k 输出预算，单次生成质量大幅提升（代价：时间翻 2-3 倍）。仅 Claude 4 系列启用 thinking。