Skip to content

Commit bc12539

Browse files
authored
release: v0.7.26 巩固 effort 陷阱根治 + LLM 连通性测试 + 文档收尾 (#99)
* docs: 巩固模型引导/配置表同步 v0.7.26 resolveDreamEffort 新行为 - 思考型模型处理顺序:'被拒绝自动去掉重试一次(v0.7.16)' → 'resolveDreamEffort 自动换用支持档位或省略字段(v0.7.26+)' - 思考模型分类声明:'插件侧 fallback 无法绕开' → 'v0.7.26+ 已根治(发流前能力探测)' - 配置表 dreamReasoningEffort:同步 v0.7.26 自动换档行为 - README.en.md 三处同步 Co-authored-by: Anans-Ivresse <[email protected]> * release: v0.7.26 巩固 effort 陷阱根治 + LLM 连通性测试 + 文档收尾 bump 0.7.25 → 0.7.26;CHANGELOG [0.7.26](随 PR #97 入库);cherry-pick 补入 PR #98 squash 丢失的引导/配置表同步 commit Co-authored-by: Anans-Ivresse <[email protected]>
1 parent 5517650 commit bc12539

4 files changed

Lines changed: 10 additions & 10 deletions

File tree

‎dsh-mneme/README.en.md‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -67,14 +67,14 @@ The default `32768` reserves headroom for reasoning models, where reasoning alon
6767
| Medium (10k–50k chars) | `65536` |
6868
| Large (>50k chars) | `131072` (cap) |
6969

70-
> With **reasoning models** (e.g. DeepSeek-R1-like), the model may spend the entire budget on reasoning and return an empty body (the log shows `no json array in llm output`). Resolution order: ① set `dreamReasoningEffort` to `low` to suppress reasoning overhead (when the provider rejects the parameter it is stripped automatically and retried once — the rejection reason lands in `llm_audit`); ② if the retry still returns an empty body under the model's default reasoning behavior, raise `dreamMaxTokens` (reasoning and body share this budget) or route `dreamProvider`/`dreamModel` to a non-reasoning model. The sleep side has the corresponding `sleepReasoningEffort`.
70+
> With **reasoning models** (e.g. DeepSeek-R1-like), the model may spend the entire budget on reasoning and return an empty body (the log shows `no json array in llm output`). Resolution order: ① set `dreamReasoningEffort` to `low` to suppress reasoning overhead (v0.7.26+: if the model doesn't support that tier, `resolveDreamEffort` auto-falls back to the model's default/first supported tier or omits the field — the rejection reason lands in `llm_audit`); ② if the retry still returns an empty body under the model's default reasoning behavior, raise `dreamMaxTokens` (reasoning and body share this budget) or route `dreamProvider`/`dreamModel` to a non-reasoning model. The sleep side has the corresponding `sleepReasoningEffort`.
7171
7272
**Consolidation model classification** (settings panel "consolidation model" = `dreamProvider`/`dreamModel`; sleep side: `sleepProvider`/`sleepModel`):
7373

7474
| Model kind | Examples | Notes |
7575
|-----------|----------|-------|
7676
| **Non-reasoning (recommended)** | glm-5-2-class | No reasoning declaration; even if an effort is configured and the harness rejects it, the fallback strips the field and the retry succeeds. Lowest risk of empty-body runs |
77-
| **Reasoning (test first)** | deepseek-v4-flash-ga and other v4-flash-ga family | Reasons by default and may burn the whole token budget on an empty body; some SKUs (e.g. v4-flash-ga) are additionally declared by the harness as accepting **no reasoning effort at all** — the no-effort retry still gets rejected through the harness `defaultEffort` (`UNSUPPORTED_REASONING_EFFORT`), which the plugin fallback cannot bypass. If you use one, set `dreamReasoningEffort` and test; switch to a non-reasoning model otherwise |
77+
| **Reasoning (test first)** | deepseek-v4-flash-ga and other v4-flash-ga family | Reasons by default and may burn the whole token budget on an empty body; some SKUs (e.g. v4-flash-ga) are additionally declared by the harness as accepting **no reasoning effort at all** (the no-effort retry gets rejected through the harness `defaultEffort`, `UNSUPPORTED_REASONING_EFFORT`). Fixed in v0.7.26+: `resolveDreamEffort` probes the model's supported effort tiers before streaming — an unsupported configured tier auto-falls back to the model's default/first supported tier, and the field is omitted for models that declare no reasoning capability. If you use one, set `dreamReasoningEffort` and test; switch to a non-reasoning model otherwise |
7878

7979
### Sleep Mode: System-Level Sleep 💤 (v0.4.0, opt-in)
8080

@@ -306,7 +306,7 @@ It works out of the box with the defaults. To adjust, override in `~/.dsh/profil
306306
| `dreamDelayMs` | `2000` | Asynchronous consolidation delay (debounce) |
307307
| `dreamProvider` / `dreamModel` | empty | Explicit dream LLM route — config wins over the agent's default model (config-first, v0.7.16); left empty, the agent's default model is used |
308308
| `dreamMaxTokens` | `32768` | Maximum tokens per dream LLM call (cap 131072; reasoning and body share this budget on reasoning models — raise it when the body comes back empty, see the tuning guide below) |
309-
| `dreamReasoningEffort` | `none` | Reasoning-effort passthrough for the dream LLM: `low` / `medium` / `high` / `none` (`none` = omit the field and use the model default; set `low` when a reasoning model exhausts its budget on reasoning and produces an empty body) |
309+
| `dreamReasoningEffort` | `none` | Reasoning-effort passthrough for the dream LLM: `low` / `medium` / `high` / `none` (`none` = omit the field and use the model default; set `low` when a reasoning model exhausts its budget on reasoning and produces an empty body; v0.7.26+ auto-falls back to the model's default/first supported tier when the configured tier is unsupported, and omits the field for models without reasoning capability) |
310310
| `apiToken` | empty | Optional API auth token; once set, write operations and key endpoints require `Authorization: Bearer <apiToken>` |
311311
| `embedProvider` | `openai` | Semantic backend: `openai` (default, v0.1-compatible) / `local` (ONNX offline) / `ollama` |
312312
| `localEmbedModel` | `Xenova/bge-small-zh-v1.5` | Local ONNX embedding model |

‎dsh-mneme/README.md‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -90,14 +90,14 @@ dsh web
9090
| 中等(1 万-5 万字) | `65536` |
9191
| 大型(5 万字以上) | `131072`(上限) |
9292

93-
> 若使用**思考型模型**(如 deepseek-v4-flash / DeepSeek-R1 类),模型可能把全部预算花在 reasoning 上导致正文为空(日志出现 `no json array in llm output`)。处理顺序:① 把 `dreamReasoningEffort` 设为 `low` 显式压低思考(被方舟拒绝该参数时自动去掉重试一次,拒绝原因会记入 llm_audit);② 重试走模型默认思考行为后正文仍为空的,调大 `dreamMaxTokens`(reasoning 与正文共享该预算)或配置 `dreamProvider`/`dreamModel` 指向非思考模型。sleep 侧对应 `sleepReasoningEffort`。
93+
> 若使用**思考型模型**(如 deepseek-v4-flash / DeepSeek-R1 类),模型可能把全部预算花在 reasoning 上导致正文为空(日志出现 `no json array in llm output`)。处理顺序:① 把 `dreamReasoningEffort` 设为 `low` 显式压低思考(v0.7.26+ 若模型不支持该档位,`resolveDreamEffort` 会自动换用模型支持的默认/首个档位或省略字段,拒绝原因会记入 llm_audit);② 重试走模型默认思考行为后正文仍为空的,调大 `dreamMaxTokens`(reasoning 与正文共享该预算)或配置 `dreamProvider`/`dreamModel` 指向非思考模型。sleep 侧对应 `sleepReasoningEffort`。
9494
9595
**巩固模型分类声明**(settings panel「巩固模型」= `dreamProvider`/`dreamModel`,睡眠侧对应 `sleepProvider`/`sleepModel`):
9696

9797
| 模型类别 | 例子 | 说明 |
9898
|---------|------|------|
9999
| **非思考模型(推荐)** | glm-5-2 类等 | 无 reasoning 声明;即使配了 effort 被 harness 拒绝,fallback 去掉字段重试即成功。空体风险最低 |
100-
| **思考模型(需实测)** | deepseek-v4-flash-ga 等 v4-flash-ga 系 | 默认开推理,可能烧光 token 预算返回空体;且部分型号(如 v4-flash-ga)在 harness 侧被声明为**不接受任何 reasoning effort** —— 去掉 effort 重试时 harness 的 `defaultEffort` 仍会顶上来再次拒绝(`UNSUPPORTED_REASONING_EFFORT`),插件侧 fallback 无法绕开。选用时建议配 `dreamReasoningEffort` 实测,不行就换非思考模型 |
100+
| **思考模型(需实测)** | deepseek-v4-flash-ga 等 v4-flash-ga 系 | 默认开推理,可能烧光 token 预算返回空体;部分型号(如 v4-flash-ga)在 harness 侧被声明为**不接受任何 reasoning effort**(去掉 effort 时 harness 的 `defaultEffort` 会顶上来再次拒绝)。v0.7.26+ 已根治:`resolveDreamEffort` 发流前探测模型支持的档位,不支持的配置档位自动换用模型默认/首个支持档位,声明无 reasoning 能力的型号则省略字段。选用时建议配 `dreamReasoningEffort` 实测,不行就换非思考模型 |
101101

102102
### Sleep Mode 系统级睡眠 💤(v0.4.0,opt-in)
103103

@@ -380,7 +380,7 @@ dsh web
380380
| `dreamDelayMs` | `2000` | 整理异步延迟(去抖) |
381381
| `dreamProvider` / `dreamModel` | 空 | dream 的 LLM 路由覆盖(显式配置优先于 agent 默认模型;留空则回退到 agent 默认模型) |
382382
| `dreamMaxTokens` | `32768` | dream LLM 调用最大 token 数(上限 131072;思考型模型的 reasoning 与正文共享该预算,正文为空时优先调大,见下方调优指南) |
383-
| `dreamReasoningEffort` | `none` | dream LLM 推理强度透传:`low` / `medium` / `high` / `none`(`none`=不传该字段,沿用模型默认;思考型模型(如 deepseek-v4-flash)想压低思考可设 `low`;被方舟拒绝该参数时 v0.7.16 起自动去掉重试一次) |
383+
| `dreamReasoningEffort` | `none` | dream LLM 推理强度透传:`low` / `medium` / `high` / `none`(`none`=不传该字段,沿用模型默认;思考型模型(如 deepseek-v4-flash)想压低思考可设 `low`;v0.7.26+ 模型不支持配置档位时自动换用其支持的默认/首个档位,无 reasoning 能力的型号省略字段) |
384384
| `apiToken` | 空 | 可选 API 鉴权 token;设置后写操作与密钥接口要求 `Authorization: Bearer <apiToken>` |
385385
| `embedProvider` | `openai` | 语义后端:`openai`(默认,兼容 v0.1)/ `local`(ONNX 离线)/ `ollama` |
386386
| `localEmbedModel` | `Xenova/bge-small-zh-v1.5` | 本地 ONNX embedding 模型 |

‎dsh-mneme/package-lock.json‎

Lines changed: 3 additions & 3 deletions
Some generated files are not rendered by default. Learn more about customizing how changed files appear on GitHub.

‎dsh-mneme/package.json‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
11
{
22
"name": "@modusensus/dsh-mneme",
33
"description": "Cross-session memory plugin for DeepSeek Harness with autoDream consolidation: SQLite store, Markdown mirrors, 7 model tools, automatic injection, session summarization, user profile/rules, custom slash commands, vector (semantic) search, and a Web GUI panel",
4-
"version": "0.7.25",
4+
"version": "0.7.26",
55
"license": "MIT",
66
"repository": {
77
"type": "git",

0 commit comments

Comments
 (0)