Repository navigation
zcode 用的是gpt-5.6-sol 上下文设置了272000，识别错了上下文空间了 #344
Description
相关日志
2026-08-29T11:53:38.257Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:53:38.297Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] processTurn: 372 msgs, renderTags=text-only, 96 text tagged, 0 tool tagged (should be 0 with text-only)
2026-08-29T11:53:38.297Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] nudge idle: usage=23% (208054/922000), growth=21792/22500 (ref=186262, interval=50000), pendingT1=135045/50000 model=gpt-5.6-sol, reason="growth 21792 < floor 22500, ready: T1 135045, blocked: T1 (cadence)"
2026-08-29T11:53:38.302Z [info] forward POST → https:///v1/chat/completions
2026-08-29T11:53:46.248Z [info] [update] checking npm registry for billion-context (last check 50s ago)…
2026-08-29T11:53:46.913Z [info] [update] current=0.1.63 latest=0.1.63 (up to date)
2026-08-29T11:53:49.172Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:53:49.209Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] processTurn: 380 msgs, renderTags=text-only, 96 text tagged, 0 tool tagged (should be 0 with text-only)
2026-08-29T11:53:49.209Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] nudge INJECT T1: usage=23% (212143/922000), growth=25881/22500 (ref=186262, interval=50000), pendingT1=137864/50000 model=gpt-5.6-sol, reason="T1 effective 137864 >= 50000, growth 25881, usage 23%"
2026-08-29T11:53:49.213Z [info] forward POST → https:///v1/chat/completions
2026-08-29T11:54:53.527Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:54:53.552Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] processTurn: 388 msgs, renderTags=text-only, 96 text tagged, 0 tool tagged (should be 0 with text-only)
2026-08-29T11:54:53.552Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] nudge idle: usage=24% (217405/922000), growth=5262/22500 (ref=212143, interval=50000), pendingT1=137864/50000 model=gpt-5.6-sol, reason="growth 5262 < floor 22500, ready: T1 137864, blocked: T1 (cadence)"
2026-08-29T11:54:53.556Z [info] forward POST → https:///v1/chat/completions
2026-08-29T11:55:32.105Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:55:32.109Z [warn] unrecognized path /api/v1/event/report — not a known protocol (/chat/completions, /v1/messages, /responses, /responses/compact); forwarding unchanged
2026-08-29T11:55:32.109Z [info] forward POST → https:///api/v1/event/report
2026-08-29T11:55:32.199Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:55:32.204Z [warn] unrecognized path /api/v1/snapshot/upload-credential?workspace_id=abcf6ff711a7 — not a known protocol (/chat/completions, /v1/messages, /responses, /responses/compact); forwarding unchanged
2026-08-29T11:55:32.204Z [info] forward GET → https:///api/v1/snapshot/upload-credential
2026-08-29T11:55:32.317Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:55:32.344Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] processTurn: 373 msgs, renderTags=text-only, 97 text tagged, 0 tool tagged (should be 0 with text-only)
2026-08-29T11:55:32.344Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] nudge idle: usage=22% (219425/988782), growth=7282/22500 (ref=212143, interval=50000), pendingT1=132397/50000 model=gpt-5.6-sol, reason="growth 7282 < floor 22500, ready: T1 132397, blocked: T1 (cadence)"
2026-08-29T11:55:32.348Z [info] forward POST → https:///v1/chat/completions
2026-08-29T11:55:52.827Z [info] [eb37a6f8-e344-4562-85e0-061fe856f968] [acp-usage] round 1 input=209684 cached=207360 (cache hit 99%)
2026-08-29T11:55:52.926Z [info] mitm :443 tunnel established (TLS terminated locally)
2026-08-29T11:55:52.928Z [warn] unrecognized path /api/v1/event/report — not a known protocol (/chat/completions, /v1/messages, /responses, /responses/compact); forwarding unchanged
Activity
ranxianglei commentedon Aug 29, 2026
[bot] 找到根因了，是 bili 的窗口优先级 bug，已修复（PR: https://github.com/ranxianglei/billion-context/pull/345）。
根因：
src/server.ts里 native 窗口解析顺序是beta ?? plugin ?? launcher ?? registry(peek) ?? 用户配置 ?? 内置表—— #212 那次改动本意是"registry 优先于内置静态表"，但把 models.dev registry 放到了用户显式声明之上。你的私有 relay host 不在 registry 的 host 映射表里，于是 lookup 退化为扫描所有 provider，采到了官方openai/gpt-5.6-sol的 1,050,000，你设的 272000 被完全忽略。日志里的两个分母正好印证（nudge 分母 = 窗口 − 该请求的 max_tokens 输出预留）：
922000= 1,050,000 − 128,000988782= 1,050,000 − 61,218修复：把优先级换回文档承诺的顺序（
config.ts/CONFIGURATION.md都写着"per-route declaration always wins"）：显式声明 > registry > 内置表。没声明时 registry 仍然生效（行为不变）。验证：新增回归测试
tests/configured-window-outranks-registry.test.ts（用 issue 里的精确数值：声明 272K vs registry 1,050K vs 内置表 400K），master 上第一个断言会失败（得到 1,050,000），修复后通过。typecheck 干净，825/825 测试全过，build 成功。一个确认：272000 是配在 bili 的 config（
providers["mitm://<host>"].models["gpt-5.6-sol"].context）吧？如果是配在 ZCode 自己的设置里，bili 是读不到的（MITM 只透传，不解析客户端配置），那就需要在 bili config 里加这条声明。合并 PR 后升级 bili 即可生效。Merge pull request #345 from ranxianglei/2026-08-29_configured-window…
fix: per-route model context declaration outranks the models.dev regi…
refactor: drop the dead SSE rewriters that kept causing mis-triage
feat: output-headroom fallback to model output when the request carri…
docs: point structure-aware compression at acp-kernel#344
fix: cap tier-gated registry windows at the standard window without t…