dsh-github-copilot 0.4.2-alpha.10 → 0.4.2-alpha.12

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/AGENTS.md CHANGED
@@ -159,6 +159,15 @@ metadata remains unproven. Strict ephemeral candidate counts record one fixed
159
159
  first-rejection reason per candidate and distinguish native off preparation
160
160
  failure; no candidate IDs, raw metadata, prompts or new persistent population.
161
161
  No-fit warnings contain fixed codes only. Legacy absent evidence stays unknown.
162
+
163
+ Auxiliary context projection reserves current/nearest user text and, for exact
164
+ continuations, a substantive anchor within twelve preceding user messages plus
165
+ its user chain. Keep 12 rows/1600 characters per row/8000 serialized characters.
166
+ Task text limitations, absent anchors and task nontext content block downshift;
167
+ incidental history omissions alone do not. Retained text is not semantic
168
+ completeness. Preserve genuine classifier unknown, off/no-tools/one-call
169
+ behavior and native Chat input. Optional strict projection/returned-demand
170
+ evidence is content-free and ephemeral, never a new persistent population.
162
171
 
163
172
  - Explicit replay recovery (#279, #289) is default-off and bound to the initiating native request signal, Session, model and current account proof. Only exact verified scope failures capture bounded full-item fingerprints; no replay bytes reach Client or storage. Public Session running/error snapshots trigger read-only notices on open/settled turns, never polling or durable recovery events. Two-stage confirmation chooses next matching turn (all steps/retries until turn/end) or bounded Session consent, omitting matching old encrypted items and their item summaries only on outgoing requests, never durable history or new reasoning. Preserve get/setEnabled descriptor identities and legacy session-consent semantics; authorize is additive. Active-turn writes, ambient-scope calls and automatic retries are forbidden. Expiry, disposal and proof discontinuity revoke evidence. See `docs/model-compatibility-acceptance.md`.
164
173
 
@@ -359,9 +368,14 @@ Completed-turn Usage diagnostics (#259) are read-only plugin conversation data a
359
368
  Auto task routing (#258) uses current `model_picker_category` facts, never model
360
369
  names, effort or context capacity as a quality category. Hard eligibility/input
361
370
  fit precede task/preference policy; unknown category and uncertain task evidence
362
- remain explicit. Within the first fitting category, apply positive cost weights
363
- (ordinary 1, marked 0.2) and finite continuity (previous model multiplier 1.5),
364
- never unconditional retention or extreme-only cost gating. Capture actual reasons once
371
+ remain explicit. Within the first fitting category, high-cost-v3 retains every
372
+ fitting candidate with positive weight. Ordinary weight 1 receives finite
373
+ previous-model multiplier 1.5; marked base 0.2 decays to 0.02 only for the previous
374
+ model. One-turn recency resets after a different model, not a hidden streak
375
+ counter or task-failure inference. Preserve v1/v2 historical codecs/cohorts and
376
+ their original explanation. Fit is not task
377
+ success; never switch categories for cost, infer incapability or add escalation.
378
+ Capture actual reasons once
365
379
  per turn without durable decision events or changing Usage.
366
380
  `autoSemanticAssessment` is default-on by explicit user rollout request (#267),
367
381
  only for locally unknown demand; preserve explicit false opt-out. One bounded concrete
package/CHANGELOG.md CHANGED
@@ -1,6 +1,35 @@
1
1
  # Changelog
2
2
 
3
- ## 0.4.2-alpha.10 (prepared)
3
+ ## 0.4.2-alpha.12 (prepared)
4
+
5
+ - Restore positive opportunity for all fitting same-category candidates instead
6
+ of v2's hard high-cost deferral. Marked base weight remains 0.2; the previous
7
+ marked model decays tenfold to 0.02, with ordinary continuity unchanged.
8
+ No hidden streak counter, permanent penalty, automatic escalation or task
9
+ success inference. Single/all-marked pools and admitted turns stay usable.
10
+ - Explain high-cost-v3 one-turn recency decay and preserve original v1/v2
11
+ evidence, summaries and diagnostic cohorts. Sequential synthetic simulations
12
+ show fewer repeat selections, not measured production quality or savings.
13
+
14
+ ## 0.4.2-alpha.11 (published)
15
+
16
+ - Prefer fitting unmarked models inside the selected supplier category and defer
17
+ high-cost peers. Marked-only fallback stays usable without a continuity bonus.
18
+ Explain the cost tier and preserve v1 historical evidence; no category
19
+ weakening, task-success inference, automatic escalation or in-turn switch.
20
+ - Preserve a bounded substantive task anchor and its user continuation chain
21
+ ahead of verbose historical output in the auxiliary classifier projection.
22
+ Separate incidental omissions from limited task text; missing anchors,
23
+ truncated task text and task attachments still prevent simple/routine
24
+ downshift. Actual Chat input and native context admission stay untouched.
25
+ - Remove the fixed unknown example from the classifier instruction and explain
26
+ that historical omission alone is not evidence of unknown demand. Keep genuine
27
+ unknown, reasoning off, no tools, one call/no retry and the existing bounds.
28
+ - Expose strict content-free projection facts and the classifier's returned
29
+ demand in the existing selection explanation. These are diagnostics, not
30
+ measured classification accuracy or proof of reduced live unknown rates.
31
+
32
+ ## 0.4.2-alpha.10 (published)
4
33
 
5
34
  - Accept account-advertised protocol-native reasoning disable labels (`none`
6
35
  for Responses/Completions, `disabled` for Anthropic) as well as `off` for
package/README.md CHANGED
@@ -8,7 +8,7 @@
8
8
 
9
9
  GitHub Copilot account discovery, contextual Auto routing and hosted search for DSH. The plugin reuses DSH's published native adapter, retaining a profile-wide account default with independent Session next-turn choices; it does not patch Core or maintain a second model catalog.
10
10
 
11
- **Stable release: [`0.4.1`](https://github.com/cloga/dsh-github-copilot/releases/tag/v0.4.1); prepared prerelease: `0.4.2-alpha.10`.** `0.4.2-alpha.1` was not published: its release preflight failed, and its immutable tag will not be reused. Supported host: official DSH / Windows Desktop `0.2.0-rc.2`. The stable version is also published on npm's `latest` dist-tag; prereleases remain on their channel tags. Stability-channel promotion does not expand the qualified host baseline or prove live account, model, search or Desktop health. Earlier DSH pins are historical evidence, not supported installation targets. Published, installed in a profile and loaded by a running Host are separate states.
11
+ **Stable release: [`0.4.1`](https://github.com/cloga/dsh-github-copilot/releases/tag/v0.4.1); prepared prerelease: `0.4.2-alpha.12`.** `0.4.2-alpha.1` was not published: its release preflight failed, and its immutable tag will not be reused. Supported host: official DSH / Windows Desktop `0.2.0-rc.2`. The stable version is also published on npm's `latest` dist-tag; prereleases remain on their channel tags. Stability-channel promotion does not expand the qualified host baseline or prove live account, model, search or Desktop health. Earlier DSH pins are historical evidence, not supported installation targets. Published, installed in a profile and loaded by a running Host are separate states.
12
12
 
13
13
  **Upgrade from 0.4.0:** that first stable release can interrupt concurrent managed requests with `COPILOT_PREVIEW_CREDENTIAL_CHANGED` during normal same-account native OAuth renewal. In 0.4.1, a proven plugin-owned refresh with unchanged account, endpoint and entitlements preserves already-dispatched HTTP streams; subsequent requests use freshly validated auth and metadata. Unknown/external credential notifications, revocation, cancellation and provider rejections remain explicit—not a promise of invisible recovery or automatic retry. See [the refresh boundary](./docs/copilot-accounts.md#native-oauth-refresh-and-dispatched-requests-385).
14
14
 
@@ -16,9 +16,12 @@ GitHub Copilot account discovery, contextual Auto routing and hosted search for
16
16
 
17
17
  **High-cost Auto:** in **Manage → Model preferences**, mark costly models with
18
18
  **High cost**. Cost and task difficulty are independent: within the first
19
- fitting category, ordinary models have weight `1`, marked models `0.2`, and the
20
- previous model gets a finite `1.5×` bonus. All candidates keep a nonzero chance;
21
- marking does not change fixed choices or admitted turns. This is a declared
19
+ fitting category, every fitting candidate retains a nonzero chance. Ordinary
20
+ weight `1` gets a finite continuity bonus (`1.5×` for the previous model).
21
+ High-cost base weight is `0.2`; if that model was used in the previous turn,
22
+ its weight decays tenfold to `0.02` instead of receiving a bonus.
23
+ This one-turn recency penalty is a declared trial, not calibrated quality.
24
+ Marking does not change fixed choices or admitted turns. This is a declared
22
25
  policy, not a price/quality ranking or proven savings.
23
26
 
24
27
  The existing turn selection dialog exposes **Auto allocation observations**:
@@ -126,10 +129,10 @@ Supply any launcher patches with repeated `--patch /absolute/file`. Require `sup
126
129
  For a **standalone named profile**:
127
130
 
128
131
  ```sh
129
- dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.10/dsh-github-copilot-0.4.2-alpha.10.tgz
132
+ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.12/dsh-github-copilot-0.4.2-alpha.12.tgz
130
133
  ```
131
134
 
132
- For **Desktop**, its native package manager accepts an exact npm spec, for example `dsh-github-copilot@0.4.2-alpha.10`; stable versions publish on npm's `latest` dist-tag and prereleases on their channel tags. The official rc.2 **Desktop-bundled CLI** also supports reserved-profile plugin management; a global/generic `dsh` shim is not equivalent. Qualify the installed entry before use. See [Desktop CLI qualification](./docs/npm-distribution.md#desktop-bundled-cli-on-official-rc2) and the distinct [standalone offline procedure](./docs/npm-distribution.md#controlled-offline-cli-maintenance-for-standalone-profiles). Neither path authorizes registry-policy bypass, peer patches or configuration deletion.
135
+ For **Desktop**, its native package manager accepts an exact npm spec, for example `dsh-github-copilot@0.4.2-alpha.12`; stable versions publish on npm's `latest` dist-tag and prereleases on their channel tags. The official rc.2 **Desktop-bundled CLI** also supports reserved-profile plugin management; a global/generic `dsh` shim is not equivalent. Qualify the installed entry before use. See [Desktop CLI qualification](./docs/npm-distribution.md#desktop-bundled-cli-on-official-rc2) and the distinct [standalone offline procedure](./docs/npm-distribution.md#controlled-offline-cli-maintenance-for-standalone-profiles). Neither path authorizes registry-policy bypass, peer patches or configuration deletion.
133
136
 
134
137
  After an approved reload/restart:
135
138
 
@@ -162,7 +165,7 @@ All three Auto preferences use the same account-verified, non-excluded eligible
162
165
  | Complex | Powerful | Powerful | Powerful |
163
166
  | Unknown | Versatile | Versatile | Powerful |
164
167
 
165
- Categories come from authenticated supplier metadata, not model names, context capacity or measured quality rankings. Within the first fitting category, positive cost weights and the finite continuity bonus described above choose the model; a suitable previous model is not unconditionally retained. Category fallback is explained explicitly. One real answer model stays fixed throughout each admitted turn's steps and retries; compaction preserves its independently resolved native summary route.
168
+ Categories come from authenticated supplier metadata, not model names, context capacity or measured quality rankings. Within the first fitting category, positive cost weights and marked recency decay choose the model. The penalty applies to the previous high-cost model, not other marked peers, and resets to base weight after a different model's turn. Fit does not prove task success, and Auto does not add escalation or choose a weaker category for cost. Category fallback is explained explicitly. One real answer model stays fixed throughout each admitted turn's steps and retries; compaction preserves its independently resolved native summary route.
166
169
 
167
170
  Semantic assessment is **enabled by default only for locally unknown tasks**. It makes at most one bounded auxiliary call with a **30-second default end-to-end deadline** and **128-token output budget**. Set `github-copilot.autoSemanticAssessment: false` to opt out. The native plugin configuration exposes `autoSemanticAssessmentTimeoutMs` (1000–120000 ms; `8000` restores the old budget) and `autoSemanticAssessmentModel` (empty for automatic selection, or an exact current-account model ID). They apply to subsequent assessments, not admitted retries. An unavailable/excluded/unsuitable explicit classifier reports unavailable, never silently switches. Automatic selection is metadata-driven, not measured optimal; the new default is a trial, not proof of past timeout causes. See [assessment settings](./docs/auto-task-routing.md#using-auto).
168
171
 
@@ -172,6 +175,14 @@ The deadline is checked against a monotonic clock at preparation, native request
172
175
 
173
176
  The auxiliary classifier always requests the protocol's explicit disabled-reasoning mode (`none` for Responses/Completions, `disabled` for Anthropic) and receives no tools. A candidate is not dispatched unless both account metadata and the isolated public adapter model expose that capability; otherwise assessment fails closed. Missing or empty effort metadata is not proof that a model never reasons. The classifier does not inherit configured Chat reasoning effort, and its off mapping is not added to shared account models. Synthetic tests verify the outgoing payload, not supplier-side implementation. One call, bounded text input and 128 requested output tokens constrain overhead without promising a hard credit cap; ordinary Chat reasoning/tools remain unchanged.
174
177
 
178
+ Classification preserves the current request and a bounded substantive continuation
179
+ anchor before verbose historical output. Incidental historical omission alone no
180
+ longer forces a valid simple/routine result to unknown; missing anchors, limited
181
+ task text and nontext task dependencies still block downshift. The prompt does not
182
+ pre-fill an unknown answer. Assessment details distinguish projection limitations
183
+ from the classifier's returned demand. Unknown remains valid; retained text and
184
+ synthetic regressions do not prove calibrated accuracy or a reduced live unknown rate.
185
+
175
186
  Request-local strict selection evidence includes candidate counts and one fixed first-rejection reason per model: exclusion, configured-ID mismatch, category, text input, unproven disable capability or budget. No model IDs are added to these counts. No-fit warnings use fixed codes only, and native off preparation failure is distinguished from selection failure. This evidence is ephemeral, not a new persisted diagnostic population; legacy records without these fields remain unknown.
176
187
 
177
188
  The reply's selection disclosure shows captured reasons and optional content-free auxiliary milestones. `uncertain` and `insufficient-evidence` do not mean the conversation is empty. **Selection unknown** means no retained record; **Selection unavailable → Retry** means the read failed. Retry rereads that turn, not inference. Evidence is bounded and Host-local, so restart/eviction can lose it.
@@ -311,14 +322,14 @@ Commit attribution uses `Assisted-by` with the actual tool, never the model prov
311
322
  GitHub Releases and npm distribute the same original verified tarball. Pin a version; verify Release SHA-256 or npm `dist.integrity`. Never repack an immutable release or move/reuse its tag.
312
323
 
313
324
  ```sh
314
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.10/dsh-github-copilot-0.4.2-alpha.10.tgz
315
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.10/SHA256SUMS
325
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.12/dsh-github-copilot-0.4.2-alpha.12.tgz
326
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.12/SHA256SUMS
316
327
  sha256sum --check SHA256SUMS
317
328
  ```
318
329
 
319
330
  ```powershell
320
331
  $expected = (Get-Content .\SHA256SUMS).Split()[0]
321
- $actual = (Get-FileHash .\dsh-github-copilot-0.4.2-alpha.10.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
332
+ $actual = (Get-FileHash .\dsh-github-copilot-0.4.2-alpha.12.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
322
333
  if ($actual -cne $expected) { throw 'Release checksum mismatch' }
323
334
  ```
324
335
 
package/README.zh.md CHANGED
@@ -8,15 +8,17 @@
8
8
 
9
9
  为 DSH 提供 GitHub Copilot 账号模型发现、上下文 Auto 选模和 hosted search。插件复用 DSH 公开的原生适配器,保留 profile 全局默认账号,并支持 Session 后续 turn 指定独立账号;不修改 Core,也不维护第二套模型目录。
10
10
 
11
- **稳定版:[`0.4.1`](https://github.com/cloga/dsh-github-copilot/releases/tag/v0.4.1);待发布预览版:`0.4.2-alpha.10`。** `0.4.2-alpha.1` 的发布预检失败,因此未发布;其不可变 tag 不会复用。支持宿主:官方 DSH / Windows Desktop `0.2.0-rc.2`。稳定版也发布到 npm 的 `latest` dist-tag;预发布版本仍使用各自渠道 tag。稳定渠道不扩大已验证的宿主基线,也不证明实时账号、模型、搜索或 Desktop 健康状态。较早 DSH pin 仅是历史证据,不是当前安装目标。已发布、已安装到 profile、已被运行中的 Host 加载,是三个不同状态。
11
+ **稳定版:[`0.4.1`](https://github.com/cloga/dsh-github-copilot/releases/tag/v0.4.1);待发布预览版:`0.4.2-alpha.12`。** `0.4.2-alpha.1` 的发布预检失败,因此未发布;其不可变 tag 不会复用。支持宿主:官方 DSH / Windows Desktop `0.2.0-rc.2`。稳定版也发布到 npm 的 `latest` dist-tag;预发布版本仍使用各自渠道 tag。稳定渠道不扩大已验证的宿主基线,也不证明实时账号、模型、搜索或 Desktop 健康状态。较早 DSH pin 仅是历史证据,不是当前安装目标。已发布、已安装到 profile、已被运行中的 Host 加载,是三个不同状态。
12
12
 
13
13
  **从 0.4.0 升级:**首个正式版在同账号正常原生 OAuth 刷新时,可能以 `COPILOT_PREVIEW_CREDENTIAL_CHANGED` 中止并行托管请求。0.4.1 仅在插件拥有的原生刷新被精确验证,且账号、端点与权限连续时,保留已发出的 HTTP 流;后续请求重新验证认证与模型元数据。未知/外部凭据通知、撤销、取消与供应商拒绝仍明确报错,不承诺所有刷新无感,也不自动重试。参见[刷新边界](./docs/copilot-accounts.md#native-oauth-refresh-and-dispatched-requests-385)。
14
14
 
15
15
  ## 你可以做什么
16
16
 
17
17
  **高成本 Auto:**在 **Manage → Model preferences** 中用 **High cost** 标记
18
- 模型。成本与任务难度正交:第一个可容纳输入的同类候选池内,普通模型权重为 `1`,
19
- 高成本模型为 `0.2`,上一轮模型仅获得 `1.5×` 的有限加权。每个合格候选都有非零机会;
18
+ 模型。成本与任务难度正交:第一个可容纳输入的分类内,所有可容纳候选保留非零机会。
19
+ 普通权重 `1` 可获得上一轮 `1.5×` 有限加权;高成本基础权重 `0.2`,
20
+ 若上一轮用了该模型,本轮权重十倍衰减到 `0.02`,而不是获得连续性奖励。
21
+ 这是一轮近期惩罚的声明试行参数,不是经过校准的质量判断;
20
22
  标记不改变手动选择或已接纳的 turn。这是路由政策,不是价格/质量排名,也不证明节省费用。
21
23
 
22
24
  现有 turn 选择说明内提供 **Auto allocation observations**,显示当前 Session
@@ -106,10 +108,10 @@ node package/scripts/check-search-composition.mjs --profile-dir /absolute/profil
106
108
  **独立具名 profile**:
107
109
 
108
110
  ```sh
109
- dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.10/dsh-github-copilot-0.4.2-alpha.10.tgz
111
+ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.12/dsh-github-copilot-0.4.2-alpha.12.tgz
110
112
  ```
111
113
 
112
- **Desktop** 原生包管理器接受精确 npm spec,例如 `dsh-github-copilot@0.4.2-alpha.10`;稳定版发布到 npm `latest` dist-tag,预发布版使用对应渠道 tag。官方 rc.2 **Desktop 随附的专属 CLI**也支持管理保留 profile;全局/普通 `dsh` shim 不等价。使用前核验实际安装入口,参见[Desktop CLI 核验](./docs/npm-distribution.md#desktop-bundled-cli-on-official-rc2)及独立的[具名 profile 离线流程](./docs/npm-distribution.md#controlled-offline-cli-maintenance-for-standalone-profiles)。两种路径都不授权绕过包源政策、修改 peer 或删除配置。
114
+ **Desktop** 原生包管理器接受精确 npm spec,例如 `dsh-github-copilot@0.4.2-alpha.12`;稳定版发布到 npm `latest` dist-tag,预发布版使用对应渠道 tag。官方 rc.2 **Desktop 随附的专属 CLI**也支持管理保留 profile;全局/普通 `dsh` shim 不等价。使用前核验实际安装入口,参见[Desktop CLI 核验](./docs/npm-distribution.md#desktop-bundled-cli-on-official-rc2)及独立的[具名 profile 离线流程](./docs/npm-distribution.md#controlled-offline-cli-maintenance-for-standalone-profiles)。两种路径都不授权绕过包源政策、修改 peer 或删除配置。
113
115
 
114
116
  经批准 reload/restart 后:
115
117
 
@@ -142,7 +144,7 @@ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/release
142
144
  | 复杂 | Powerful | Powerful | Powerful |
143
145
  | 未知 | Versatile | Versatile | Powerful |
144
146
 
145
- 分类来自认证后的供应方元数据,不按模型名、上下文容量或臆测质量排名。在第一个可容纳输入的分类内,按前述正成本权重和有限连续性加权选择;上一模型合适也不会无条件保留。分类回退明确说明。同一已准入 turn 的工具步骤和重试保持同一真实回答模型;压缩保留独立解析的原生摘要路由。
147
+ 分类来自认证后的供应方元数据,不按模型名、上下文容量或臆测质量排名。在第一个可容纳输入的分类内,按正成本权重和高成本近期衰减选择。惩罚只针对上一轮高成本模型,不影响其它已标记模型;换用其它模型一轮后恢复基础权重。可容纳不证明能解决任务,不自动升级求助,也不为成本降低分类。分类回退明确说明。同一已准入 turn 的工具步骤和重试保持同一真实回答模型;压缩保留独立解析的原生摘要路由。
146
148
 
147
149
  语义判断**默认开启,仅用于本地需求未知的任务**。最多一次辅助请求,**默认 30 秒端到端期限、128 tokens 输出预算**。可用 `github-copilot.autoSemanticAssessment: false` 关闭。原生插件配置提供 `autoSemanticAssessmentTimeoutMs`(1000–120000 毫秒,`8000` 恢复旧预算)和 `autoSemanticAssessmentModel`(空字符串自动选择,或当前账号的精确模型 ID)。只影响后续评估,不改变已准入 turn 的重试。指定分类模型不存在/被排除/不合适时明确报告 unavailable,不偷偷换模型。自动选择依据元数据,不代表实测最优;30 秒是试验默认值,不证明过去主要因超时。参见[评估配置](./docs/auto-task-routing.md#using-auto)。
148
150
 
@@ -154,6 +156,11 @@ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/release
154
156
 
155
157
  辅助分类器始终请求协议明确的关闭推理模式(Responses/Completions 为 `none`,Anthropic 为 `disabled`),且不传工具。只有账号元数据和隔离后的公开适配器模型都支持此能力时才发送请求;否则关闭分类请求并按不可用处理。分类器不继承配置的 Chat reasoning effort,关闭映射也不会加到共享账号模型。合成测试验证出站载荷,不证明供应商内部实现。单次调用、有界文本输入和 128 tokens 请求输出预算限制额外开销,但不承诺额度硬上限;普通 Chat 的思考与工具保持不变。
156
158
 
159
+ 分类输入优先保留当前请求与有界的实质续接任务目标,再填入近期执行内容。仅省略旧历史
160
+ 不再强制把有效的 simple/routine 改成 unknown;目标缺失、任务文本受限或任务依赖非文本
161
+ 时仍阻止降级。提示词不预填 unknown 答案。判断详情区分文本投影限制与分类器实际返回值。
162
+ unknown 仍是有效结果;文本保留与合成回归不证明准确率已校准,也不证明真实 unknown 比例已下降。
163
+
157
164
  回复的选模说明展示捕获的理由和可选的不含内容的辅助里程碑。`uncertain`、`insufficient-evidence` 不等于会话没有上下文。**Selection unknown**表示记录未保留,**Selection unavailable → Retry**表示读取失败;Retry 只重读该 turn,不重新推理。证据有界且仅限 Host 生命周期,重启/淘汰可能丢失。
158
165
 
159
166
  失败/不完整轮次的原生 Usage 没有模型行时,沿用现有 ⓘ 打开**本轮模型与选择记录**:本轮成功消息来源显示**已记录模型**并提示失败尝试归属不完整;可靠的本轮已记录请求配置仅显示**请求模型**,不证明发送、执行或计费。缺失证据显示**未知**,不从当前选择、全局默认或相邻轮次倒推。原选模说明折叠在下方;原生完整模型记录不重复展示,没有关闭消息的轮次没有现有入口。原生 Usage 与重试保持不变。[证据边界](./docs/automatic-model-routing.md#attribution-and-explanation)。
@@ -291,14 +298,14 @@ Agent 从 `node scripts/agent.mjs describe --json`、`doctor --json`及`plan <ta
291
298
  GitHub Releases 和 npm 分发同一原始已校验 tarball。固定版本并核验 Release SHA-256 或 npm `dist.integrity`;不重打包不可变 Release、不移动/复用 tag。
292
299
 
293
300
  ```sh
294
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.10/dsh-github-copilot-0.4.2-alpha.10.tgz
295
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.10/SHA256SUMS
301
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.12/dsh-github-copilot-0.4.2-alpha.12.tgz
302
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.12/SHA256SUMS
296
303
  sha256sum --check SHA256SUMS
297
304
  ```
298
305
 
299
306
  ```powershell
300
307
  $expected = (Get-Content .\SHA256SUMS).Split()[0]
301
- $actual = (Get-FileHash .\dsh-github-copilot-0.4.2-alpha.10.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
308
+ $actual = (Get-FileHash .\dsh-github-copilot-0.4.2-alpha.12.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
302
309
  if ($actual -cne $expected) { throw 'Release checksum mismatch' }
303
310
  ```
304
311
 
@@ -8,7 +8,7 @@
8
8
  },
9
9
  "package": {
10
10
  "name": "dsh-github-copilot",
11
- "version": "0.4.2-alpha.10"
11
+ "version": "0.4.2-alpha.12"
12
12
  },
13
13
  "evidence": {
14
14
  "kind": "source-and-synthetic-test-inventory",
@@ -1341,6 +1341,9 @@
1341
1341
  "required": true,
1342
1342
  "sourceMarkers": [
1343
1343
  {"file": "src/auto-allocation.ts", "marker": "high-cost-v1"},
1344
+ {"file": "src/auto-allocation.ts", "marker": "high-cost-v2"},
1345
+ {"file": "src/auto-allocation.ts", "marker": "high-cost-v3"},
1346
+ {"file": "src/auto-allocation.ts", "marker": "deferredHighCostCount"},
1344
1347
  {"file": "src/config.ts", "marker": "highCostModelIds"},
1345
1348
  {"file": "src/model-preferences-card.ts", "marker": "High cost"},
1346
1349
  {"file": "src/turn-selection-host.ts", "marker": "allocationSummary"},
@@ -1349,7 +1352,11 @@
1349
1352
  {"file": "docs/auto-high-cost.md", "marker": "separate acceptance"}
1350
1353
  ],
1351
1354
  "tests": [
1352
- {"file": "tests/auto-allocation.spec.ts", "name": "gives every candidate opportunity without an absolute previous-model shortcut"},
1355
+ {"file": "tests/auto-allocation.spec.ts", "name": "simulates 20000 turns per previous state with positive cost-decayed opportunity"},
1356
+ {"file": "tests/auto-allocation.spec.ts", "name": "decays a marked previous model tenfold without eliminating its opportunity"},
1357
+ {"file": "tests/auto-allocation.spec.ts", "name": "keeps v2 hard-deferral evidence readable without relabeling it as v3"},
1358
+ {"file": "tests/auto-allocation.spec.ts", "name": "keeps v1 historical weights readable without relabeling them as v2"},
1359
+ {"file": "tests/auto-allocation-evidence.spec.ts", "name": "keeps legacy-only summaries v1 and separates mixed-policy decisions explicitly"},
1353
1360
  {"file": "tests/auto-allocation-evidence.spec.ts", "name": "counts each turn once, bounds retention and never recovers missing evidence after removal"},
1354
1361
  {"file": "tests/auto-model-routing.spec.ts", "name": "keeps costly Fast models eligible and evidence opt-out independent of the selected route"},
1355
1362
  {"file": "tests/authorization-controller.spec.ts", "name": "saves orthogonal high-cost preferences narrowly without credential or Session reads"},
@@ -42,28 +42,56 @@ unchanged official Core fixture, not that the affected live Host read the mark
42
42
  before the historical Astra decision. Do not hardcode a model, change live
43
43
  preferences or rewrite a frozen explanation to manufacture a repair.
44
44
 
45
- ## Routing policy v1
45
+ ## Routing policy v3 (#412)
46
46
 
47
47
  1. Account ownership, metadata validity, exclusion, input modality and input
48
48
  headroom are hard requirements.
49
49
  2. Assess each new turn using existing contextual local/semantic evidence.
50
50
  Preserve existing demand categories and supplier-category fallback.
51
- 3. Inside the first fitting category only, give ordinary candidates weight 1
52
- and marked candidates weight 0.2. Multiply an eligible previous model's
53
- weight by 1.5. Continuity is finite, never an early unconditional return.
51
+ 3. Inside the first fitting category only, retain all fitting candidates.
52
+ Ordinary candidates have weight 1 and an eligible previous ordinary model
53
+ receives multiplier 1.5. Marked candidates have base weight 0.2; the previous
54
+ marked model decays tenfold to 0.02 instead of receiving a continuity bonus.
54
55
  4. Select from cumulative positive weights using a stable Agent/turn seed.
55
56
  Capture the selected ID, candidate weights and conditional expected shares.
56
57
  No additional model calls are made to give candidates a chance.
57
58
  5. When no model fits, retain the largest-input-budget native recovery path.
58
59
  Cost cannot justify choosing a model that cannot fit.
59
60
 
60
- These initial parameters are a routing policy to evaluate, not calibrated
61
- quality or billing facts. A marked previous model has weight 0.3, lower than an
62
- ordinary peer's 1. An ordinary previous model has 1.5. A lone fitting marked
63
- model is selected; an all-marked pool remains usable. A marked Fast model can
64
- answer a simple task: it has a smaller share, not an extreme-only gate.
65
- Every eligible final-pool candidate has positive opportunity; finite workloads
66
- need not visit every model. Stable seeds do not prove independent random trials.
61
+ This corrects v2's over-strong hard deferral after the user clarified that high
62
+ cost should have rare opportunity and strongly decay after selection, not become
63
+ permanently unavailable. The 10x repeat penalty is a declared trial, not a
64
+ calibrated optimum. It uses existing previous-route evidence only; there is no
65
+ hidden multi-turn counter, persistent penalty, time cooldown or failed-task guess.
66
+ If the same marked model repeats, it remains at 0.02 rather than recursively
67
+ shrinking. A different model's turn resets it to 0.2. Other marked models retain
68
+ base weight. A lone fitting marked model is selected and an all-marked pool
69
+ remains usable. Fit and supplier
70
+ category cannot prove peers can solve the task; no automatic escalation, extra
71
+ calls, in-turn switch or lower-category substitution is introduced.
72
+ Every fitting final-pool candidate has positive opportunity. With one ordinary
73
+ peer, a marked model has share 2/17 (11.8%) when the ordinary model was previous,
74
+ then 1/51 (2.0%) when it was itself previous. The independent-draw approximation
75
+ for a predeclared pair is 2/867 (0.23%), not the probability of any streak in a
76
+ long Session or a guarantee against repeats. Stable seeds do not prove
77
+ independent random trials. Conditional shares are not observed billing.
78
+
79
+ Historical v1 retains mixed-pool weights and marked continuity; v2 retains hard
80
+ cost-tier deferral and its deferred count. Both remain readable unchanged.
81
+ Persistent cohorts retain original policy versions.
82
+ The Session summary separates incompatible policy cohorts and reports omitted
83
+ decisions instead of relabeling them or merging different opportunities.
84
+
85
+ Read-only October 11 alpha.10 aggregates showed 3 Astra selections against
86
+ 2.868 summed expected selections across 18 opportunities, without Session or
87
+ sequence identity. The requested local Session's bounded native metadata showed
88
+ five Sol turns followed by two Astra turns; its last 178 messages belonged to
89
+ one frozen turn. This does not establish a distribution violation. A conditional
90
+ two-peer seeded reproduction matches the two Astra choices, but the exact
91
+ historical pool was not retained. V2 implements the user's stricter preference,
92
+ not a claimed repair to a proven random-number defect. V3 supersedes v2's hard
93
+ exclusion with positive recency-decayed opportunity, without a live performance
94
+ claim.
67
95
 
68
96
  The auxiliary semantic classifier is a separate call, not another main-turn
69
97
  allocation. It prefers an unmarked eligible Lightweight classifier; if only
@@ -80,7 +108,8 @@ Synthetic acceptance covers ordinary/marked/all-marked/single pools, every
80
108
  supplier category, exclusions, capacity, category changes, stable seeds,
81
109
  preference changes and admitted-turn freezing. Distribution checks use large
82
110
  declared seed sets and compare actual allocation with conditional expected
83
- shares, including a marked previous model.
111
+ shares, including marked previous models, all-marked/single pools, sequential
112
+ previous-route updates and backward-compatible v1/v2 evidence.
84
113
 
85
114
  Runtime evidence must distinguish a selected decision from native attempted,
86
115
  delivered, finished, failed and cancelled calls. Count a decision once per
@@ -2,21 +2,21 @@
2
2
  "schemaVersion": 1,
3
3
  "feature": "autorouting",
4
4
  "status": "insufficient-sample",
5
- "policyVersion": "high-cost-v1",
5
+ "policyVersion": "high-cost-v3",
6
6
  "observation": {
7
7
  "start": "2026-10-08T00:00:00Z",
8
8
  "end": "2026-10-09T23:59:59Z",
9
9
  "sampleCount": 63,
10
- "limitations": "Retained prior aggregate review: producer alpha.2, daily UTC buckets, completeHistory=false, restarts=1. No Session/turn IDs, per-candidate reasons or model capability snapshot. A later user-reported turn says unavailable/not-selected and no adapter start, not timeout; the actual account candidate pool remains unavailable. A read-only local alpha.9 installed-manifest check and diagnostic inventory do not prove loaded turn version or live model metadata. No raw export uploaded."
10
+ "limitations": "The declared 63-sample window is the prior alpha.2 review, not the newer cohort. Separate alpha.10 read-only profile aggregates on October 11 show Astra 3 selections across 18 conditional opportunities, summed expected selections 2.868. No Session IDs or sequence in aggregates. Bounded native metadata from the requested local Session shows five Sol turns followed by two Astra turns; the last has 178 steps under one frozen allocation. Tail scan is limited, not full history. Actual per-turn pool/weights are not durably retained. Two user-reported valid unknown classifier results are not a population unknown rate. No content, credentials or raw export uploaded."
11
11
  },
12
- "previousReview": "Prior allocation review #332/#333, post-delivery review #334, fixed-outcome diagnostics #398, budget/config trial #400 and classifier disable #405.",
13
- "findings": "Prior 63 decisions: 44 local/unknown, 4 semantic/unknown, 15 semantic/complex. Missing diagnostics prevented causal attribution. Alpha.9 requires exact advertised off before native preparation. Regression confirms a supplier-advertised protocol-native none/disabled candidate is rejected solely by this spelling gate. Missing/empty metadata remains uncertain, not proof of no reasoning. Actual user-account rejection cause is not established.",
14
- "decision": "User requests classifier availability repair and better root-cause diagnostics. Accept explicit account-advertised off or protocol-native none for Responses/Completions and disabled for Anthropic, with one shared classifier-only predicate and published prepared off support. Do not manufacture capabilities from missing/empty metadata or names. Keep ordinary Chat, one call/no retry, no tools, 128 output tokens, timeouts, account proof and exclusions unchanged. Capture bounded strict request-local candidate first-rejection counts and native-off preparation failure; fixed no-fit Host warning codes only, no content, identities or raw errors and no new persistent diagnostic population. Preserve legacy missing fields as unknown. Do not claim live resolution before actual candidate evidence.",
15
- "validation": "Regression-first classifier tests failed for protocol-native disable labels and absent candidate observations before implementation. Complete pnpm verify rerun passed: 2757 Vitest tests with 2 expected skips, all 340 tooling tests, source/test types, Agent/baseline checks, clean build and built package smoke. Coverage retains exact off wire tests as well as native none/disabled, and verifies candidate counts/preparation failure in the existing bilingual explanation without fabricating legacy evidence. Initial tooling run caught a review-document evidence omission (No Session/turn IDs); restored that truthful limitation rather than weakening tests. Exact alpha.10 tarball verification is required after final documentation alignment. Required CI remains pending. No actual supplier metadata or inference was obtained; no Desktop restart, profile installation or credential read is part of this repair.",
12
+ "previousReview": "Prior #332/#333/#334 and alpha.11 review delivered in #411. User clarified on October 11 that v2 hard deferral is too strong and requested rare opportunity with recency decay (#412).",
13
+ "findings": "Prior 63 decisions cannot attribute later incidents. Valid unknown is not a timeout. Projection regressions establish avoidable anchor loss/incidental-omission downgrade, not the actual task root cause. The reported Astra streak is two allocations, not 178 independent choices. Under an explicitly assumed two-model category pool, production Session/turn seeds reproduce both Astra choices at conditional shares 11.8% then 23.1%; this is conditional reproduction, not recovered decision evidence. A 20000-turn synthetic fixed-previous v1 scan yields 4563 Astra hits and 1014 adjacent pairs, showing runs are possible, not proving actual distribution conformity. No observed aggregate inflation is established. V1 intentionally grants marked continuity; the user now requests a stricter cost preference, not a statistically proven RNG repair.",
14
+ "decision": "Supersede v2 hard deferral with high-cost-v3 positive mixed-pool weights in the first fitting supplier category. Ordinary weight 1 retains previous multiplier 1.5; marked base 0.2 becomes 0.02 only when that exact model is previous. This 10x one-turn penalty is a declared uncalibrated trial; it resets after a different model and never compounds or adds hidden streak/cooldown state. Lone/all-marked pools remain usable. Preserve v1/v2 strict decoding, historical explanations and separate cohorts; no relabeling. Fit is not task success; no automatic escalation, retry, category weakening or in-turn switch. Retain alpha.11 classifier projection and all native guards. No new persistent population or measured quality/cost claim.",
15
+ "validation": "Five regression-first cost-decay tests failed before implementation. Full pnpm verify passed: 2779 Vitest tests, 2 expected skips, 340 tooling tests, source/test types, build and package smoke. Identical synthetic Session/turn seeds for 100000 sequential draws, updating previous selection each turn: v1 13371 marked selections and 3041 adjacent pairs; v3 10822 selections and 210 pairs. Conditional expected selections were 13277.26 and 10703.73 respectively. This declared two-model synthetic pool proves behavior, not independent production draws or task quality. Fixed-previous 20000-seed sets preserve opportunity/shares; legacy v1/v2 codecs/cohorts remain. Native saved-mark/restart fixture covers base and previous-header decay, executed by required CI. Pack/tarball verification passed; final archive must be rechecked after this review update. Required CI pending before merge. No real model calls, credential reads, profile changes or Desktop restart.",
16
16
  "followUp": {
17
17
  "owner": "Maintainer or Agent assigned to post-delivery review issue #334",
18
- "trigger": "Review 200 new decisions with recorded reasons, or any further classifier capability/policy change; first inspect a subsequent no-fit turn's request-local counts.",
19
- "metric": "Fixed candidate first-rejection counts, native preparation failure, selected/no-fit assessment outcome and auxiliary milestones with producer/version coverage. Missing legacy evidence remains unknown; do not infer Session populations, billing, supplier latency or accuracy from aggregates.",
20
- "action": "Confirm whether capability, category, exclusions, configuration or budget dominates actual no-fit decisions. If metadata remains unproven, retain diagnostic failure rather than guessing; change only on reviewed public semantics. Continue the linked post-delivery review #334. No automatic upload or live benchmark."
18
+ "trigger": "Review 200 new decisions with recorded reasons, or any further classifier capability/context/policy change; first inspect a successful unknown turn's projection and decoded-demand fields.",
19
+ "metric": "By producer/policy, compare selected/no-fit and candidate opportunities versus conditional expected selections, splitting marked previous/base weights. Adjacent-turn rates require observed sequence, never infer them from daily aggregates. Retain valid unknown/projection limitations separately. Legacy absence is unknown; task quality and cost require independent evidence.",
20
+ "action": "Review whether positive high-cost opportunity and 10x repeat decay avoid over-strong exclusion without unwanted repeated selection; keep candidate changes and v1/v2/v3 separate. Check labeled task outcomes before interpreting lower high-cost usage as success or savings. Retain genuine uncertainty and linked #334. No automatic upload or live benchmark."
21
21
  }
22
22
  }
@@ -140,17 +140,31 @@ existing managed native adapter, never an Auto ID and never `ctx.llm` routing
140
140
  recursion. It has no tools, bounded text-only assessment context, bounded JSON
141
141
  output, one deadline, cancellation and no new retry loop. Text from messages is
142
142
  untrusted data, not router instructions. Files/images/reasoning/replay bodies
143
- are not read; omitted context makes classification uncertain and cannot justify
144
- downshifting. Assessment input is a separate projection: the actual chat
143
+ are not read; limited task text cannot justify downshifting. Assessment input is a separate projection: the actual chat
145
144
  messages, system prompt, tools, attachments and history stay untouched.
146
145
 
147
146
  Only locally unknown tasks invoke the enabled assessment. Context packing keeps
148
- the current and nearest preceding user requests before recent output, then
149
- restores chronological order. At most twelve complete JSON rows, 1600 text
147
+ the current and nearest preceding user requests before recent output. An explicit
148
+ continuation also reserves the nearest substantive user anchor among the previous
149
+ twelve user messages and the intervening user chain, then restores chronological
150
+ order. No expired anchor, task summary or semantic task boundary is invented.
151
+ At most twelve complete JSON rows, 1600 text
150
152
  characters per row and 8000 serialized characters are retained; JSON escape
151
153
  expansion is included in this bound. Any truncation, skipped row or non-text
152
- content still marks omitted context and prevents a semantic simple/routine
153
- downshift. Prioritization does not prove omitted output irrelevant.
154
+ content still marks projection omission. A missing current request/continuation
155
+ anchor, a missing reserved user row, truncated reserved task text, or nontext
156
+ content in the reserved user chain separately marks `taskTextLimited` and prevents
157
+ a semantic simple/routine downshift. Incidental assistant/tool/history omissions
158
+ alone no longer force that downgrade. The classifier must still judge whether
159
+ the retained evidence establishes demand; retained text is not semantic
160
+ completeness and missing required details must remain unknown.
161
+
162
+ The instruction requests a schema without a pre-filled unknown answer and
163
+ distinguishes projected data omissions from missing relevant task evidence.
164
+ Unknown is permitted for genuinely indeterminate demand, not required merely
165
+ because a long conversation exceeds the projection bounds. No reasoning/tool
166
+ enablement, second inference, larger context/output budget or full-history
167
+ serialization is introduced.
154
168
 
155
169
  The strict result contains demand (`simple`, `routine`, `complex`, `unknown`)
156
170
  and finite evidence signals, not a model ID, raw reasoning or an uncalibrated
@@ -160,6 +174,13 @@ instead of falling back. Credential, metadata and account revocation remain
160
174
  fatal through existing lease/admission guards. Extra calls incur supplier
161
175
  charges and latency, separate from native chat Usage and without inferred cost.
162
176
 
177
+ Strict ephemeral explanation evidence optionally records retained row count
178
+ (0–12), the projection omission flag, task-text limitation, fixed continuation
179
+ anchor status and the classifier's decoded demand. No text, attachment values
180
+ or new persistent population is recorded. Legacy absent fields stay absent.
181
+ These fields distinguish a valid classifier `unknown` from plugin-side
182
+ simple/routine blocking; they do not establish accuracy or a live unknown rate.
183
+
163
184
  ## Auxiliary latency evidence and bounded output (#272)
164
185
 
165
186
  The assessment has one **configured end-to-end deadline, 30 seconds by default**, including account
@@ -264,9 +285,13 @@ explicit fallback only, not fabricated categories. Unknown/new category values
264
285
  remain unclassified with a bounded catalog diagnostic.
265
286
 
266
287
  Within the first fitting category, use stable positive weighted Session/turn
267
- allocation: ordinary weight 1, user-marked high-cost weight 0.2, eligible previous
268
- model multiplier 1.5. Cost is independent of task difficulty and supplier
269
- category. There is no absolute continuity shortcut or extreme-only gate.
288
+ allocation across all fitting candidates. Ordinary weight 1 has finite
289
+ previous-model multiplier 1.5; high-cost base weight 0.2 decays tenfold to 0.02
290
+ only for the previous model. Every candidate keeps positive opportunity; a lone
291
+ candidate still has share 1. A different model's turn resets the penalty, without
292
+ extra streak state. Cost is independent of task difficulty and supplier
293
+ category; input fit does not prove task success. Never weaken category or
294
+ automatically escalate to establish whether another model can solve the task.
270
295
  These are reviewed policy parameters, not measured quality or prices. Once
271
296
  admitted, all steps/retries share the decision. No supplier cache savings are
272
297
  claimed. See [high-cost requirements](./auto-high-cost.md) and
@@ -58,7 +58,8 @@ remains authoritative after final framing/attachment projection.
58
58
  Task preference uses supplier `model_picker_category`, not capacity, model
59
59
  names or reasoning effort as quality rankings. [Task policy](./auto-task-routing.md)
60
60
  owns the demand matrix, default semantic assessment, category fallback,
61
- finite continuity and stable positive-weight same-category allocation.
61
+ high-cost one-turn recency decay, ordinary finite continuity and stable
62
+ positive-weight same-category allocation.
62
63
  User-marked high cost is orthogonal to demand and category; see
63
64
  [high-cost requirements](./auto-high-cost.md). Future policy changes must read
64
65
  the [latest review](./auto-iteration-review.json) before implementation.