| 1 | # Multi-provider live validation / 多供应商真实端点验证 |
| 2 | |
| 3 | Date / 日期: 2026-09-05. Local development checkout; no commit, publication or |
| 4 | issue closure is implied. / 当前本地开发版本，未提交、发布或关闭 issue。 |
| 5 | |
| 6 | ## Scope and method / 范围与方法 |
| 7 | |
| 8 | The user authorized live requests against LongCat, official DeepSeek, Zhipu |
| 9 | Coding Plan CN, and OpenCode Go. Each process receives only its selected |
| 10 | provider's credential, with disposable HOME/REASONIX_HOME directories. Keys |
| 11 | are never stored in the checkout or report. Model-facing tools consist only of |
| 12 | fixed markers, read-only argument/error fixtures, or a built-in writer confined |
| 13 | to a disposable directory. There is no shell, MCP or credential-reading tool. |
| 14 | |
| 15 | 用户授权使用四家供应商的真实凭据；每个测试进程只收到该供应商的密钥，使用 |
| 16 | 独立临时目录。模型工具仅为标记、只读参数/错误测试或限定临时目录的文件写入， |
| 17 | 不提供 shell、MCP 或凭据读取能力。请求串行执行，有每用例期限和请求上限。 |
| 18 | |
| 19 | Provider adapters and the real Agent loop execute the requests. A local proxy |
| 20 | can cut an actual upstream SSE response, remove reasoning, inject one HTTP 503, |
| 21 | or cancel just before commit. These are controlled faults applied to real |
| 22 | responses, not claims about naturally occurring outages. Ordinary retry delays |
| 23 | are shortened by the test clock; elapsed values do not measure production |
| 24 | 2/4/8-second backoff. Discarded attempts without final usage remain unknown. |
| 25 | |
| 26 | 实测经过真实适配器与 Agent 循环；故障代理作用于真实响应，不向公共服务注入 |
| 27 | 故障。测试时钟缩短退避，耗时不代表生产环境的恢复时延；缺少终态用量的请求 |
| 28 | 保持“未知”，不能根据报告中的 token 总数推算准确账单。 |
| 29 | |
| 30 | ### Protocol routing / 协议路由 |
| 31 | |
| 32 | | Provider / 供应商 | Routes / 路由 | Contract / 契约 | |
| 33 | | --- | --- | --- | |
| 34 | | LongCat | `/openai/v1/chat/completions`, `/anthropic/v1/messages` | LongCat-2.0; Bearer authentication for both / 两者均使用 Bearer | |
| 35 | | Zhipu Coding Plan CN / 智谱 | `/api/coding/paas/v4/chat/completions`, `/api/anthropic/v1/messages` | Seven GLM models; CN plan endpoint, not the international Z.AI route / 七个 GLM 模型，使用中国区套餐端点 | |
| 36 | | Official DeepSeek / 官方 | `/chat/completions`, `/responses`, `/anthropic/v1/messages` | Flash/Pro coverage in the earlier report; this extension adds the vision SKU's text/tool and image paths / 既有报告覆盖 Flash、Pro，本轮增加视觉 SKU 的文本工具与图像输入 | |
| 37 | | OpenCode Go | `/zen/go/v1/chat/completions`, `/zen/go/v1/messages`, `/zen/go/v1/responses` | Model-specific routing; Messages uses `x-api-key`, Chat/Responses use Bearer / 按模型选择路由与鉴权；另测 DeepSeek 自定义 Messages/Responses | |
| 38 | |
| 39 | Routing references: [LongCat API](https://longcat.chat/platform/docs/APIDocs.html), |
| 40 | [Zhipu Coding Plan](https://docs.bigmodel.cn/cn/coding-plan/using5-1), |
| 41 | [OpenCode Go endpoints](https://dev.opencode.ai/docs/go/). |
| 42 | The model catalogs were fetched live, but a successful catalog request does not |
| 43 | prove inference permission or remaining quota. / 模型目录已实时获取；目录返回 |
| 44 | 200 不代表生成请求有权限或额度。 |
| 45 | |
| 46 | ## Credential and quota gates / 凭据与额度边界 |
| 47 | |
| 48 | LongCat's direct generation routes returned HTTP 402, insufficient token quota. |
| 49 | The initial batch made 14 rejected requests, executing no model tool; these do |
| 50 | not count as recovery coverage. The harness now stops subsequent scenarios for |
| 51 | a route after a confirmed quota gate. LongCat-2.0 on OpenCode Go is a separate |
| 52 | route and did complete its baseline and recovery tests. |
| 53 | |
| 54 | LongCat 官方生成接口仍因 token 额度不足被阻塞；初始批次 14 次拒绝请求没有 |
| 55 | 执行模型工具，不能算作恢复覆盖。测试器已增加额度门槛后的停止规则。Go 上 |
| 56 | LongCat-2.0 的通过不能替代官方接口验证。 |
| 57 | |
| 58 | OpenCode Go initially returned CreditsError/HTTP 401. After the user renewed |
| 59 | the subscription, three correctly authenticated direct protocol probes returned |
| 60 | 200 and the funded model matrix passed. No billing settings were changed by |
| 61 | Reasonix or by these tests. / 用户重新订阅后，三种正确鉴权的原始请求均返回 |
| 62 | 200，后续模型矩阵通过；测试没有修改账户计费设置。 |
| 63 | |
| 64 | ## Initial matrices / 首轮矩阵 |
| 65 | |
| 66 | | Matrix / 矩阵 | Cases / 用例 | Client HTTP / 客户端请求 | Upstream HTTP / 真实上游 | Outcome / 结果 | |
| 67 | | --- | ---: | ---: | ---: | --- | |
| 68 | | Zhipu baseline, 7 models × 2 protocols / 智谱基础往返 | 14 | 28 | 28 | All completed; one tool each / 全部完成，每例一个工具 | |
| 69 | | OpenCode Go funded baseline, 27 models / Go 重新订阅后基础往返 | 31 | 62 | 62 | All completed; no retries / 全部完成，无重试 | |
| 70 | | Official DeepSeek vision SKU text/tool baseline / 官方视觉型号文本工具往返 | 3 | 6 | 6 | All three protocols completed / 三种协议均完成 | |
| 71 | | Zhipu recovery + continuity, 3 models × 2 protocols / 智谱恢复与连续会话 | 36 | 102 | 96 | 12/12 injected recovery cases completed; 6 cancellations safe / 12 个恢复用例完成，6 个取消未执行工具 | |
| 72 | | Official vision SKU recovery + continuity / 官方视觉型号恢复与连续会话 | 18 | 52 | 49 | 7/7 injected recovery cases completed; 3 cancellations safe / 7 个恢复用例完成，3 个取消未执行工具 | |
| 73 | | OpenCode Go recovery + continuity, 8 models / Go 恢复与连续会话 | 50 | 148 | 139 | 19/20 injected recovery cases completed; two Kimi tool-omission observations / 19 个恢复用例完成，Kimi 有两次未调用工具观察 | |
| 74 | |
| 75 | Go's 50 cases comprise 45 passes, two failures and three unexercised |
| 76 | missing-reasoning probes. One Kimi K3 failure happened after an injected cut; |
| 77 | another omitted the tool on the first response, so the intended second-request |
| 78 | 503 was never injected. They must not both be counted as transport-recovery |
| 79 | failures, nor counted as completed tasks. Diagnosis and repeat samples are |
| 80 | reported separately. The low-effort Kimi sample differs from the preset's max |
| 81 | setting; both levels are tested without changing the user's configuration. |
| 82 | |
| 83 | Go 的两次 Kimi 观察分别为断流重试后未调用工具、第一轮就未调用工具（没有机会 |
| 84 | 注入后续 503）。两者不能都记为网络恢复失败，也不能记为完成任务。低档位样本 |
| 85 | 与预设的 max 档位不同，另行对照，不修改用户设置。未触发故障的用例不算通过。 |
| 86 | |
| 87 | The Zhipu recovery matrix initially flagged three Anthropic cache-prefix checks. |
| 88 | All model turns and tool counts completed; the failure was the overly strict |
| 89 | cache-marker comparison described below. Corrected continuity replays are |
| 90 | reported separately. Initial permanent-read-error tests similarly exercised an |
| 91 | unintended host transient-read retry and are not counted as model duplicates. |
| 92 | |
| 93 | 智谱首轮有三个 Anthropic 缓存断言报错，但所有模型轮与工具次数符合预期； |
| 94 | 缓存标记检查校正后的复测另列。永久读取错误的旧测试也不混入模型重复调用统计。 |
| 95 | |
| 96 | Recorded input/output totals for the three recovery matrices are respectively |
| 97 | 20,227/1,455 (Zhipu), 19,588/1,303 (official vision SKU), and 44,792/3,640 (Go). |
| 98 | They include estimates for incomplete attempts; 18, 9 and 29 cases respectively |
| 99 | contain unknown usage. Successful recovery-case elapsed medians are 10.093 s, |
| 100 | 2.028 s and 4.055 s with shortened test waits. Each successful injected recovery |
| 101 | used one extra model request. No manual continuation was used to rescue a failed |
| 102 | sample; incomplete samples remain failures. These are selected-case results, |
| 103 | not production reliability percentages or accurate bills. |
| 104 | |
| 105 | 三个恢复矩阵记录的输入/输出 token 如上，包含中断估算；未知用量用例分别为 |
| 106 | 18、9、29。成功恢复用例中位耗时分别约 10.093、2.028、4.055 秒（测试退避缩短）。 |
| 107 | 每个成功恢复用例增加一次模型请求；没有通过人工继续挽救失败样本。不能将这些 |
| 108 | 受控结果解释为线上成功率或准确费用。表中请求数不包括目录查询、旧额度探测、 |
| 109 | 参数/文件/搜索/图像专门测试及后续对照，不能当作整个测试任务的请求总数。 |
| 110 | |
| 111 | ## Follow-up results / 复测与专项结果 |
| 112 | |
| 113 | | Check / 检查 | Coverage / 覆盖 | Outcome / 结果 | |
| 114 | | --- | --- | --- | |
| 115 | | Zhipu corrected Anthropic continuity / 智谱连续会话 | Three models, three user turns each, save/reload / 三模型各三轮并保存重载 | 18 HTTP requests, nine tools, all passed / 全部通过 | |
| 116 | | Canonical Unicode/newline arguments and permanent tool errors / 参数与永久工具错误 | Six Zhipu, five Go, three official vision protocol cases / 共十四组合 | Each made two requests and executed alpha/beta/unavailable once / 各两次请求，三个工具标记各执行一次 | |
| 117 | | File side effect followed by interrupted result persistence / 写入后结果持久化中断 | Six Zhipu, five Go, three official vision protocol cases / 共十四组合 | All passed: one durable intent and one actual disk write per case / 每例一个持久意图、一次真实写入，恢复不重复写 | |
| 118 | | Actual generated image input / 真实图像输入 | Official vision SKU over three protocols; Go vision SKU Chat / 四组合 | Left red/right blue correctly identified; one request per sample / 全部识别正确，每样本一次请求；Go 另有一个通过的重复样本 | |
| 119 | | Go native DeepSeek Flash search / Go 原生搜索 | Messages + Responses | Messages: ten sources; Responses: search/echo completed but no structured sources / 后者搜索及工具往返完成，来源不可用 | |
| 120 | | Actual server replay rejection / 服务端真实回放拒绝 | Go Flash/Pro × three protocols / 六组合 | Chat/Responses four recoveries; Messages two opaque 400 stops / 四例自动恢复，两例因无错误详情停止 | |
| 121 | |
| 122 | The file fixtures cancel after the write has actually happened, restore the |
| 123 | checkpoint written before its result, and verify the desired target before |
| 124 | continuing. Their one-write assertion refers to actual writes, not just model |
| 125 | call counts. It does not cover shell/MCP/remote/editor writes. |
| 126 | |
| 127 | 文件测试确实在写入完成、结果尚未持久化时中断，再恢复旧检查点并核验目标； |
| 128 | “没有重复写入”统计实际磁盘写入，不只是模型调用次数。未覆盖 shell、MCP、 |
| 129 | 远程或编辑器缓冲区写入。 |
| 130 | |
| 131 | ### Kimi and Go protocol diagnostics / Kimi 与 Go 协议诊断 |
| 132 | |
| 133 | The additional low-effort Kimi matrix repeated baseline, cut, follow-up 503 and |
| 134 | missing reasoning three times each: 10/12 completed, 27 HTTP requests, ten tool |
| 135 | executions and five retry events. The two omissions had no tool-call field on |
| 136 | the original wire. All three cut samples recovered; all three actual removed- |
| 137 | reasoning samples completed after the test stripper was corrected to include |
| 138 | Chat's `reasoning` alias. The max-effort control repeated baseline/cut three |
| 139 | times each: 5/6 completed, 14 requests, five tools, three retries. The failed |
| 140 | baseline also had no tool call on the wire. Changing effort alone is not a fix. |
| 141 | |
| 142 | Kimi 的补充 low 档位 12 例完成 10 例，max 档位 6 例完成 5 例；失败样本原始 |
| 143 | 响应也没有工具调用，因此不是适配器漏解析。两组各三个断流样本均恢复。 |
| 144 | reasoning 缺失测试补齐字段别名后，三个实际移除样本均完成。 |
| 145 | |
| 146 | A prompt-only experiment, inspired by OpenCode's explicit action requirements, |
| 147 | added a stable instruction to execute requested tools and never invent results. |
| 148 | At max effort, three baseline and three cut/retry samples all completed: 15 HTTP |
| 149 | requests, six physical tool calls and three retries. The original max control |
| 150 | completed five of six. This small sequential sample is promising, not a |
| 151 | statistically established improvement or an OpenCode runtime comparison. The |
| 152 | experiment does not change Reasonix's production prompt or thinking settings. |
| 153 | |
| 154 | 借鉴 OpenCode 行动提示的短提示实验在 max 档位六例全完成；原提示对照为五例。 |
| 155 | 这是小样本顺序实验，不能据此声称稳定收益，生产提示未改。另一个 Kimi 带参数 |
| 156 | 测试三次中两次通过，一次 Unicode/换行值不符合测试要求；每次都调用了三个 |
| 157 | 工具各一次，六次请求，无网络重试。该测试衡量首次参数正确性，并明确禁止 |
| 158 | 重复调用，不能当作一般参数自动纠错能力的结论。 |
| 159 | |
| 160 | The Go replay-rejection batch made 16 requests with six physical tool executions. |
| 161 | All four Chat/Responses cases followed 200 → 400 → 200 without replaying the |
| 162 | tool. Both Messages cases followed 200 → 400 with a body containing only the |
| 163 | model name, retaining the completed tool. The test knows the injected cause; |
| 164 | production clients cannot infer that cause from the opaque body. This remains |
| 165 | an automatic-recovery limitation, not a reason to retry arbitrary 400s. |
| 166 | |
| 167 | Go 回放拒绝共 16 次请求、六次工具执行；Messages 的不透明 400 仍是自动恢复 |
| 168 | 边界。测试器知道故障来源，不代表生产客户端可以猜测真实请求的错误原因。 |
| 169 | |
| 170 | Two follow-up Responses search samples were inspected at the raw terminal response; the last also verified unchanged raw search-item replay into the next request. |
| 171 | Search actions had queries but no `sources`; the following message had empty |
| 172 | `annotations`. Search execution and continuation passed, but source availability |
| 173 | was not demonstrated. The final harness separately checks raw search-item replay |
| 174 | and marks absent structured sources as unproven, without inferring URLs from |
| 175 | model prose. Initial source-count assertion failures remain recorded. |
| 176 | |
| 177 | Responses 搜索缺少来源是原始终态响应可见的事实；最终测试分别验证搜索 item |
| 178 | 回放与来源可用性，来源缺失记录为未证明，保留最初严格断言失败的证据。 |
| 179 | |
| 180 | The OpenCode source comparison and recommended changes are documented in |
| 181 | [OpenCode recovery comparison](OPENCODE_RECOVERY_COMPARISON.md). These are source |
| 182 | findings and Reasonix experiments, not an executed OpenCode-vs-Reasonix benchmark. |
| 183 | |
| 184 | ## Confirmed fixes / 已确认修复 |
| 185 | |
| 186 | 1. **Quota classification across HTTP statuses.** OpenCode Go returned HTTP 401 |
| 187 | with `CreditsError`/insufficient balance, while LongCat returned HTTP 402 |
| 188 | with `too_many_requests`/insufficient token quota. Exhausted allowance now |
| 189 | takes precedence over authentication and transient rate-limit labels. Both |
| 190 | managed and standalone requests stop after one attempt. The localized error |
| 191 | describes allowance, not an invalid key. Raw billing URLs are excluded; |
| 192 | status-based `APIError` consumers retain compatibility through unwrapping. |
| 193 | / 余额与套餐额度错误优先于 401、限流类标签，立即停止；不再提示用户更换 |
| 194 | 实际有效的密钥，不输出私有账单链接，并保留既有状态码错误接口兼容性。 |
| 195 | 2. **OpenCode transport identity.** All three built-in adapters supply the |
| 196 | client identity and `x-opencode-session`. A bound conversation path maps to |
| 197 | an opaque stable hash across resume; ephemeral/child sessions have their own |
| 198 | identity. Standalone calls use a stable client identity. Explicit custom |
| 199 | headers are preserved. Other vendor routes receive none of these additions. |
| 200 | / 三种适配器补齐客户端与会话标识；恢复保持稳定，子会话独立，不发送本地 |
| 201 | 路径。只调整官方 Go 的 HTTP 头，不改变正常消息、工具 schema 或思考设置。 |
| 202 | |
| 203 | Cache impact: service-side routing may change when the new header is first |
| 204 | introduced. Conversation/tool bytes remain unchanged. Cache-hit differences |
| 205 | between providers or headers are not an A/B causal benchmark. |
| 206 | |
| 207 | 缓存影响：首次引入会话头可能改变服务端缓存分组；正常会话正文与工具序列化 |
| 208 | 保持不变。供应商之间、请求头变化前后的命中数据不构成严格 A/B 因果对照。 |
| 209 | |
| 210 | ## Test harness distinctions / 测试口径校正 |
| 211 | |
| 212 | - An Anthropic tail `cache_control: {type: ephemeral}` marker normally moves to |
| 213 | the newly appended tail. The continuity check allows only this exact change |
| 214 | on the previous tail block. Text, tool arguments/results, reasoning and |
| 215 | signatures must remain identical; no general recursive field removal is used. |
| 216 | / 只允许旧尾部的标准临时缓存标记移动，不能借此掩盖历史正文或证明变化。 |
| 217 | - A synthetic read error containing “unavailable” initially activated the |
| 218 | existing one-time read-only transient retry. The permanent-error fixture now |
| 219 | implements `RetryableToolError() == false`. This distinguishes physical read |
| 220 | retries from a model replaying a completed tool. / 最初测试错误文本触发了既有 |
| 221 | 只读重试；永久错误使用显式类型后再验证，不将其误报为模型重复执行。 |
| 222 | - An initial label prompt double-escaped its JSON text, making a literal |
| 223 | backslash-n interpretation possible. The final argument tests generate the |
| 224 | JSON literal with `json.Marshal` and require the decoded Unicode/newline |
| 225 | value. The ambiguous sample is not attributed to the provider adapter. |
| 226 | / 参数提示最初有双重转义歧义；最终由程序生成标准 JSON，再验证解码值， |
| 227 | 不把该歧义归咎于适配器。 |
| 228 | - The first Go Messages probe incorrectly used Bearer authentication. It was |
| 229 | corrected to `x-api-key`, matching the adapter's actual preset. Historical |
| 230 | quota/auth probes are separated from funded inference results. / 初版测试的 |
| 231 | Messages 鉴权已纠正；历史拒绝请求不计为有效推理覆盖。 |
| 232 | |
| 233 | ## Evidence and boundaries / 证据与边界 |
| 234 | |
| 235 | See [the preceding official DeepSeek report](RECOVERY_VALIDATION.md) for its |
| 236 | 74 raw replay-contract probes, 54-case recovery matrix, independent search, |
| 237 | cache and write-recovery results. Those observations remain distinct from this |
| 238 | multi-provider extension. Native Claude, arbitrary relays, natural outages, |
| 239 | remote/editor-buffer write recovery and native Windows UI are not proven by |
| 240 | these endpoint tests. A synthetic cancellation does not measure user-interface |
| 241 | stop latency. The generated two-color image proves only this basic image-input fixture, not general visual reasoning. |
| 242 | |
| 243 | 既有 DeepSeek 报告与本轮结果分别保留。这里不能证明原生 Claude、任意中转、 |
| 244 | 自然故障、远程/编辑器通道写入恢复或 Windows 原生界面均已通过。测试取消不 |
| 245 | 等于测量用户界面停止时延；双色图片测试只证明该基础图像样本，不代表通用视觉能力。 |
| 246 | |
| 247 | ## Local verification / 本地验证 |
| 248 | |
| 249 | - Root full suite: `go test -p 1 ./... -timeout 180s` — passed. |
| 250 | - Desktop module full suite — passed separately. |
| 251 | - `make lint` — passed, without a new lint baseline waiver. |
| 252 | - Focused provider/agent/controller race tests — passed. |
| 253 | - Opaque Go HTTP 400 classification regression — passed. |
| 254 | - `git diff --check` — passed. |
| 255 | |
| 256 | 根模块、Desktop 独立模块、lint、涉及请求头/会话身份/额度/恢复的 race 回归均 |
| 257 | 通过。没有新增界面功能的原生 UI 验证结论，也没有执行提交、推送或发布。 |
| 258 | |
| 259 | Raw sanitized local logs are in `/tmp/reasonix-live-multiprovider/`; preceding |
| 260 | DeepSeek logs are in `/tmp/reasonix-live-official/`. They are ephemeral local |
| 261 | evidence, not published artifacts. No request/response reasoning bodies or |
| 262 | credentials are included in this report. |
| 263 | |
| 264 | Credential scan checked 295 modified/untracked source and local evidence files |
| 265 | against all four exact supplied keys: zero matches. The in-memory credential |
| 266 | supervisor was then stopped and temporary test binaries removed; sanitized logs |
| 267 | remain for diagnosis. / 四把密钥的精确值扫描覆盖 295 个源码和证据文件，零匹配； |
| 268 | 随后退出持有密钥的测试进程并清理测试二进制，保留脱敏日志。 |
| 269 |