第三十三轮:模型接口加入上下文参数(context_window / max_output_tokens)

老板 2026-09-24 要求:「文档里有写明支持上下文多少吗?有的话就修改 ai 网关,
接口加入上下文参数」+「Agnes 先加上,opencode go 看下 v2/docs/console/go/ 文档,
没有的话就看接口里有没有,这个也要加入同步脚本里」。

数据来源(逐一核查过):
- Agnes:官方**每个文本模型详情页**的「限制与价格」表写明 `上下文窗口` / `最大输出`
  (定价页没有这两项)。3.0-flash = 512K/65,536、2.5-flash = 512K/65.5K、
  2.5-pro 与 2.5-pro-beta = 1M/65,536。图片/视频页无此表 → 留 0。
- OpenCode:官方文档(Zen 页、v2/docs/console/go/)**都不写**上下文;`/zen/go/v1/models`
  实测只回 id/object/created/owned_by(单模型详情 404、include 参数被忽略)。
  唯一机器可读来源是 **models.dev**(OpenCode 团队维护),其 `opencode-go` /
  `opencode` 两供应商每条模型都带 `limit.context` / `limit.output`。
  口径(老板拍板):只用于**更新网关已有模型**,不拿它当增删依据(models.dev 收录
  比官方 Go 文档多,含 omen-alpha 等未收录的);跨源不一致时取**保守下限**
  (minimax-m3 Go 1M / Zen 512K → 512,000;qwen3.6-plus → 262,144)。

改动:
- app/config.py:ModelInfo 新增 context_window / max_output_tokens(默认 0 = 未标注),
  解析 + to_dict 序列化一并带上。
- app/main.py:GET /models 与 /v1/models 输出这两个字段;别名继承目标模型的值。
- app/settings.py:「供应商与模型」面板往返不丢字段(默认 0 → 未配置的模型不会被
  补出 `context_window: 0` 污染 YAML);提交了就以此为准,没提交则沿用文件原文。
- data/gateway.yaml:37 个模型落盘(33 OpenCode + 4 Agnes 文本),文本级插入保住
  全部手写注释(注释行 162 → 236)。
- scripts/sync_models.py:新增 fetch_models_dev()(拉失败返回 None → 只跳过上下文
  同步、不当成漂移)、fetch_agnes_context()(逐模型页解析「限制与价格」表)、
  diff() 产出 ctx_drift、build_plan/apply_plan(_set_context 块内改值,保留行尾注释、
  幂等)、verify() 回读校验、render_plan/render 显示漂移;_DRIFT_KEYS 与
  plan_is_empty 认 ctx_set。
- tests:新增 3 个(models.dev 保守下限合并 + 抓失败返回 None;_set_context 幂等与
  保注释;/models 带出字段且未标注不瞎补)→ **201 passed / 0 failed**;ci_check.sh 全绿。

上线:supervisorctl restart api-gateway:api-gateway_00(pid 3492804)后实测
/v1/models 45 个模型、38 个带 context_window(含 wjd 别名继承 = 1,000,000/384,000),
图片模型正确为 0。
This commit is contained in:
api-gateway committed 2026-09-24 15:00:26 +08:00
1 parent 1c4bc9fb1e
commit 64054ad920
6 files changed
+428 -7

No files matched your search

+207 -6
View File
@@ -107,6 +107,50 @@ def fetch_opencode() -> dict[str, dict]:
return models
# ── models.dev(上下文窗口 / 最大输出)──────────────────────────
# 第三十三轮(2026-09-24 老板要求:接口加入上下文参数):
# OpenCode 官方文档(Zen / Go / v2)**都不写**上下文长度,`/zen/go/v1/models`
# 也只回 id/object/created/owned_by(实测:单模型详情 404、include 参数被忽略)。
# 唯一可用的机器可读来源是 models.dev(OpenCode 团队维护的模型目录),其
# `opencode-go` / `opencode` 两个供应商的每条模型都带 `limit.context` / `limit.output`。
# 口径(老板拍板):
# · 只用于**更新网关已有模型**的这两个字段,不拿它当增删依据
# (增删仍以官方 Go 文档表为准 —— models.dev 收录更多,含 omen-alpha 等文档未收的);
# · 同名模型在两个供应商下不一致时(如 minimax-m3:Go 1M / Zen 512K)取**保守下限**
# (宁可早压缩,不能超窗口);
# · models.dev 挂了/改版 → 只跳过上下文同步,不影响增删判定(返回 None)。
MD_DEV_URL = "https://models.dev/api.json"
def fetch_models_dev() -> dict[str, dict] | None:
"""返回 {model_id: {"context": int, "output": int}}(两供应商合并取保守下限)。
拉不到就返回 None —— 调用方据此跳过上下文同步,不当成漂移。
"""
try:
raw = _get(MD_DEV_URL, timeout=60)
data = json.loads(raw)
except Exception as exc: # noqa: BLE001
print(f"[warn] models.dev 拉取失败,跳过上下文同步:{exc}", file=sys.stderr)
return None
merged: dict[str, dict] = {}
for prov in ("opencode-go", "opencode"):
for mid, m in ((data.get(prov) or {}).get("models") or {}).items():
lim = (m or {}).get("limit") or {}
ctx, out = lim.get("context"), lim.get("output")
if not ctx:
continue
cur = merged.get(mid) or {"context": ctx, "output": out or 0}
cur["context"] = min(cur["context"], ctx) # 保守下限
if out:
cur["output"] = min(cur["output"] or out, out) if cur["output"] else out
merged[mid] = cur
if not merged:
print("[warn] models.dev 解析出 0 个模型,跳过上下文同步", file=sys.stderr)
return None
return merged
# ── Agnes ──────────────────────────────────────────────────────
_MODEL_ID_RE = re.compile(r"[a-z][a-z0-9.\-]*[0-9][a-z0-9.\-]*")
@@ -208,6 +252,62 @@ def fetch_agnes() -> dict[str, dict]:
return out
# ── Agnes 上下文窗口(模型详情页「限制与价格」表)──────────────────
# 第三十三轮(2026-09-24): Agnes 官方**每个文本模型页**都有一张「限制与价格」表,
# 写明 `上下文窗口`(如 `512K` / `1M tokens`)与 `最大输出`(如 `65,536 Token`)。
# 定价页没有这两项 —— 所以要去逐个模型页抓。图片/视频模型页无此表 → 抓不到 = 留空。
AGNES_MODEL_PAGES = {
"agnes-3.0-flash": "https://www.agnes-ai.cn/zh-Hans/docs/agnes-30-flash",
"agnes-2.5-flash": "https://www.agnes-ai.cn/zh-Hans/docs/agnes-25-flash",
"agnes-2.5-pro": "https://www.agnes-ai.cn/zh-Hans/docs/agnes-25-pro",
"agnes-2.5-pro-beta": "https://www.agnes-ai.cn/zh-Hans/docs/agnes-25-pro-beta",
}
def _parse_tokens(cell: str) -> int | None:
"""`512K` / `1M tokens` / `65,536 Token` / `65536` → 整数 token 数。"""
s = _strip_tags(cell).lower().replace(",", "").replace(" ", "")
m = re.search(r"([0-9.]+)\s*([km])?", s)
if not m:
return None
val = float(m.group(1))
unit = m.group(2)
if unit == "k":
val *= 1000
elif unit == "m":
val *= 1_000_000
return int(val)
def fetch_agnes_context() -> dict[str, dict]:
"""逐个 Agnes 文本模型页抓 `上下文窗口` / `最大输出` → {id: {context, output}}。
单页抓失败只跳过该模型(不抛)—— 上下文同步是增量能力,不该拖垮整个同步任务。
"""
out: dict[str, dict] = {}
for mid, url in AGNES_MODEL_PAGES.items():
try:
html = _get(url)
except Exception as exc: # noqa: BLE001
print(f"[warn] Agnes 模型页抓取失败 {mid}: {exc}", file=sys.stderr)
continue
ctx = out_tok = None
for row in re.findall(r"<tr>(.*?)</tr>", html, re.S):
cells = [_strip_tags(c) for c in re.findall(r"<t[dh][^>]*>(.*?)</t[dh]>", row, re.S)]
if len(cells) < 2:
continue
label = cells[0].replace(" ", "")
if "上下文窗口" in label:
ctx = _parse_tokens(cells[1])
elif "最大输出" in label:
out_tok = _parse_tokens(cells[1])
if ctx:
out[mid] = {"context": ctx, "output": out_tok or 0}
if not out:
print("[warn] Agnes 模型页一个上下文都没解析出来(页面可能改版)", file=sys.stderr)
return out
# ── 比对 ───────────────────────────────────────────────────────
def _references(cfg: dict) -> dict[str, str]:
"""被别名 / media 指向的模型 id → 引用原因。删这些会直接打断功能,故永不自动删。"""
@@ -251,7 +351,9 @@ def _split_removed(prov: str, cur: dict, doc: dict, refs: dict) -> tuple[list[st
return drop, keep
def diff(cfg: dict, oc_doc: dict, agnes_doc: dict) -> dict:
def diff(cfg: dict, oc_doc: dict, agnes_doc: dict,
md_dev: dict[str, dict] | None = None,
ag_ctx: dict[str, dict] | None = None) -> dict:
providers = cfg.get("providers", {})
pricing = (cfg.get("pricing") or {}).get("per_model") or {}
refs = _references(cfg)
@@ -270,6 +372,19 @@ def diff(cfg: dict, oc_doc: dict, agnes_doc: dict) -> dict:
for mid, meta in sorted(oc_doc.items())
if mid in oc_cur and meta["free"] != bool(oc_cur[mid].get("free"))]
oc_drop, oc_keep = _split_removed("opencode", oc_cur, oc_doc, refs)
# 第三十三轮: 上下文窗口漂移(models.dev 现值 vs 网关现值)。
# 只比对**网关已有模型**(不拿 models.dev 当增删依据);md_dev 为 None(拉取失败)
# 时留空 → 不产生漂移、不写盘。
oc_ctx_drift = []
for mid, m in sorted(oc_cur.items()):
want = (md_dev or {}).get(mid)
if not want or not want.get("context"):
continue
have_ctx = int(m.get("context_window") or 0)
have_out = int(m.get("max_output_tokens") or 0)
if have_ctx != want["context"] or (want.get("output") and have_out != want["output"]):
oc_ctx_drift.append({"id": mid, "cfg_ctx": have_ctx, "doc_ctx": want["context"],
"cfg_out": have_out, "doc_out": want.get("output") or have_out})
report["opencode"] = {
"doc_count": len(oc_doc), "cfg_count": len(oc_cur),
"added": [{"id": k, "name": v["name"], "protocol": v["protocol"], "free": v["free"]}
@@ -278,6 +393,7 @@ def diff(cfg: dict, oc_doc: dict, agnes_doc: dict) -> dict:
"protected": oc_keep,
"protocol_drift": proto_drift,
"free_drift": free_drift,
"ctx_drift": oc_ctx_drift,
}
# Agnes
@@ -291,8 +407,20 @@ def diff(cfg: dict, oc_doc: dict, agnes_doc: dict) -> dict:
if have != want:
price_drift.append({"id": mid, "cfg": have, "doc": want})
ag_drop, ag_keep = _split_removed("agnes", ag_cur, agnes_doc, refs)
# 第三十三轮: Agnes 上下文漂移(模型详情页 vs 网关现值),同 opencode 口径。
ag_ctx_drift = []
for mid, want in sorted((ag_ctx or {}).items()):
m = ag_cur.get(mid)
if not m or not want.get("context"):
continue
have_ctx = int(m.get("context_window") or 0)
have_out = int(m.get("max_output_tokens") or 0)
if have_ctx != want["context"] or (want.get("output") and have_out != want["output"]):
ag_ctx_drift.append({"id": mid, "cfg_ctx": have_ctx, "doc_ctx": want["context"],
"cfg_out": have_out, "doc_out": want.get("output") or have_out})
report["agnes"] = {
"doc_count": len(agnes_doc), "cfg_count": len(ag_cur),
"ctx_drift": ag_ctx_drift,
"added": [{"id": k, "section": v["section"], "type": v["type"],
"display_name": pretty_name(k), "input": v["input"], "output": v["output"]}
for k, v in sorted(agnes_doc.items()) if k not in ag_cur],
@@ -343,7 +471,12 @@ def render(report: dict) -> str:
if oc["free_drift"]:
lines.append(" 免费标记漂移 → 落盘时更新:")
lines += [f" ~ {d['id']} 文档={d['doc_free']} 网关={d['cfg_free']}" for d in oc["free_drift"]]
if not any(oc[k] for k in ("added", "removed", "protocol_drift", "free_drift")):
if oc.get("ctx_drift"):
lines.append(" 上下文窗口漂移(models.dev vs 网关)→ 落盘更新:")
lines += [f" ~ {d['id']} 网关={d['cfg_ctx']:,}/{d['cfg_out']:,}"
f" 来源={d['doc_ctx']:,}/{d['doc_out']:,}" for d in oc["ctx_drift"]]
if not any(oc[k] for k in ("added", "removed", "protocol_drift", "free_drift",
"ctx_drift")):
lines.append(" 无漂移 ✓")
lines.append("")
@@ -369,6 +502,10 @@ def render(report: dict) -> str:
label = {True: "免费", "limited": "限时免费", None: "(无标注)"}
lines += [f" ~ {d['id']} 网关={label.get(d['cfg'], d['cfg'])}"
f" 文档={label.get(d['doc'], d['doc'])}" for d in ag["free_drift"]]
if ag.get("ctx_drift"):
lines.append(" 上下文窗口漂移(Agnes 模型页 vs 网关)→ 落盘更新:")
lines += [f" ~ {d['id']} 网关={d['cfg_ctx']:,}/{d['cfg_out']:,}"
f" 文档={d['doc_ctx']:,}/{d['doc_out']:,}" for d in ag["ctx_drift"]]
vd = ag.get("video_default") or {}
if vd.get("cfg"):
mark = ("✓ 文档里还在" if vd.get("in_doc")
@@ -381,7 +518,8 @@ def render(report: dict) -> str:
return "\n".join(lines)
_DRIFT_KEYS = ("added", "removed", "protocol_drift", "free_drift", "price_drift")
_DRIFT_KEYS = ("added", "removed", "protocol_drift", "free_drift", "price_drift",
"ctx_drift")
def has_drift(report: dict) -> bool:
@@ -558,6 +696,37 @@ def _set_free(seg: list[str], want: bool, date: str) -> list[str]:
_AGNES_FREE_LINE_RE = re.compile(r"^\s*(free|free_limited):\s*")
_CTX_KEY_RE = re.compile(r"^\s*(context_window|max_output_tokens):")
def _set_context(seg: list[str], ctx: int, out_tok: int, date: str) -> list[str]:
"""块内设置 context_window / max_output_tokens(第三十三轮)。
· 已有该行 → 只改值,保留行尾注释;
· 没有 → 插在 `type:` 行之后(与手工插入的位置一致)。
只动这两行,其余原文(含手写注释)不动。
"""
want = {"context_window": ctx, "max_output_tokens": out_tok}
seen: set[str] = set()
for i, ln in enumerate(seg):
m = _CTX_KEY_RE.match(ln)
if not m:
continue
key = m.group(1)
seen.add(key)
tail = ln.split("#", 1)[1] if "#" in ln else f" # 同步 {date}(scripts/sync_models.py)"
seg[i] = f" {key}: {want[key]} #{tail}"
missing = [k for k in want if k not in seen]
if missing:
for i, ln in enumerate(seg):
if re.match(r'^\s{8}type:\s*"', ln):
extra = [f" {k}: {want[k]} # 同步 {date}(scripts/sync_models.py)\n"
for k in missing]
seg[i + 1:i + 1] = extra
break
return seg
def _set_agnes_free(seg: list[str], want: object, date: str) -> list[str]:
"""按文档写 Agnes 模型的免费标注(2026-09-24 老板要求 Agnes 也要标)。
@@ -589,7 +758,7 @@ def build_plan(report: dict, cfg: dict, oc_doc: dict, agnes_doc: dict) -> dict:
docs = {"opencode": oc_doc, "agnes": agnes_doc}
pricing = (cfg.get("pricing") or {}).get("per_model") or {}
plan: dict = {"add": {}, "remove": {}, "price_set": {}, "price_add": {},
"proto_add": {}, "free_set": {}, "agnes_free": {}}
"proto_add": {}, "free_set": {}, "agnes_free": {}, "ctx_set": {}}
for prov in SYNC_PROVIDERS:
plan["add"][prov] = {m["id"]: m for m in report[prov]["added"]}
plan["remove"][prov] = list(report[prov]["removed"])
@@ -597,6 +766,9 @@ def build_plan(report: dict, cfg: dict, oc_doc: dict, agnes_doc: dict) -> dict:
plan["free_set"] = {d["id"]: d["doc_free"] for d in report["opencode"]["free_drift"]}
# Agnes 的免费标注(True / "limited" / None)单独一条 plan,走 _set_agnes_free
plan["agnes_free"] = {d["id"]: d["doc"] for d in report["agnes"]["free_drift"]}
# 第三十三轮: 上下文窗口 / 最大输出(models.dev + Agnes 文档)→ 块内改值
plan["ctx_set"] = {d["id"]: {"context": d["doc_ctx"], "output": d["doc_out"]}
for d in report["opencode"]["ctx_drift"] + report["agnes"]["ctx_drift"]}
for mid, meta in sorted(agnes_doc.items()):
if meta["input"] is None: # 图片/视频模型文档没给单价
continue
@@ -621,7 +793,8 @@ def plan_is_empty(plan: dict) -> bool:
if plan["add"].get(prov) or plan["remove"].get(prov):
return False
return not any(plan.get(k) for k in
("price_set", "price_add", "proto_add", "free_set", "agnes_free"))
("price_set", "price_add", "proto_add", "free_set", "agnes_free",
"ctx_set"))
def render_plan(plan: dict, date: str) -> str:
@@ -638,6 +811,8 @@ def render_plan(plan: dict, date: str) -> str:
_flabel = {True: "免费", "limited": "限时免费", None: "撤掉标注"}
for mid, want in sorted(plan.get("agnes_free", {}).items()):
lines.append(f" 免费标注 agnes/{mid} → {_flabel.get(want, want)}")
for mid, spec in sorted(plan.get("ctx_set", {}).items()):
lines.append(f" 上下文 {mid} → {spec['context']:,} 输入 / {spec['output']:,} 输出")
for mid, spec in sorted(plan["price_set"].items()):
lines.append(f" 改价 {mid} {spec['old']} → {spec['new']}")
for mid, want in sorted(plan["price_add"].items()):
@@ -704,6 +879,16 @@ def apply_plan(text: str, plan: dict, date: str) -> str:
s, e = blocks[mid]
lines[s:e] = _set_agnes_free(lines[s:e], plan["agnes_free"][mid], date)
# 3c) 上下文窗口 / 最大输出(第三十三轮)——同一套倒序处理,字段插在块的 type: 之后。
if plan.get("ctx_set"):
for prov in SYNC_PROVIDERS:
blocks = _find_model_blocks(lines, prov)
mids = [m for m in plan["ctx_set"] if m in blocks]
for mid in sorted(mids, key=lambda m: blocks[m][0], reverse=True):
s, e = blocks[mid]
spec = plan["ctx_set"][mid]
lines[s:e] = _set_context(lines[s:e], spec["context"], spec["output"], date)
# 4) 定价:改价 / 补价 / 删掉被删模型的价
gone = {m for prov in SYNC_PROVIDERS for m in plan["remove"].get(prov) or []}
span = _pricing_span(lines)
@@ -779,6 +964,18 @@ def verify(new_text: str, cfg_before: dict, plan: dict) -> list[str]:
problems.append(f"免费标记 {mid} 没改成 {want}")
# Agnes 的免费标注:True(免费)/ "limited"(限时免费)/ None(该没有标注)
ag2 = {m["id"]: m for m in (cfg2.get("providers", {}).get("agnes") or {}).get("models", [])}
# 第三十三轮: 上下文窗口 / 最大输出真写进去了
for prov, table in (("opencode", oc2), ("agnes", ag2)):
for mid, spec in plan.get("ctx_set", {}).items():
m = table.get(mid)
if m is None:
continue # 该模型属于另一个 provider
got_ctx = int(m.get("context_window") or 0)
got_out = int(m.get("max_output_tokens") or 0)
if got_ctx != spec["context"] or got_out != spec["output"]:
problems.append(
f"上下文 {prov}/{mid} 没写成 {spec['context']}/{spec['output']}"
f"(实际 {got_ctx}/{got_out})")
for mid, want in plan.get("agnes_free", {}).items():
got = _cfg_free_state(ag2.get(mid, {}))
if got != want:
@@ -874,7 +1071,11 @@ def main() -> int:
_append_log(args.log, msg, 2)
return 2
report = diff(cfg, oc_doc, agnes_doc)
# 第三十三轮: 上下文来源(拉不到 = None/空 dict,只跳过上下文同步,不产生漂移)
md_dev = fetch_models_dev()
ag_ctx = fetch_agnes_context()
report = diff(cfg, oc_doc, agnes_doc, md_dev, ag_ctx)
plan = build_plan(report, cfg, oc_doc, agnes_doc)
drift = has_drift(report)
code = 1 if drift else 0