dsh-github-copilot 0.4.1 → 0.4.2-alpha.11

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.
Files changed (51) hide show
  1. package/AGENTS.md +42 -6
  2. package/CHANGELOG.md +126 -0
  3. package/README.md +75 -25
  4. package/README.zh.md +55 -20
  5. package/deployment-baseline.json +30 -3
  6. package/docs/auto-high-cost.md +51 -20
  7. package/docs/auto-iteration-review.json +13 -13
  8. package/docs/auto-task-routing.md +118 -19
  9. package/docs/automatic-model-routing.md +2 -1
  10. package/docs/copilot-accounts.md +6 -0
  11. package/docs/copilot-compaction.md +1 -1
  12. package/docs/plugin-diagnostics.md +159 -6
  13. package/lib/client.js +814 -59
  14. package/lib/client.js.map +1 -1
  15. package/lib/diagnostics-analysis.js +4 -2
  16. package/lib/diagnostics-types-qvtuvM5E.js +661 -0
  17. package/lib/index.js +8775 -7934
  18. package/lib/manual-compaction-recovery.js +1 -1
  19. package/lib/remote.js +278 -6
  20. package/lib/routed-web.js +1 -1
  21. package/lib/types/auto-allocation-diagnostics.d.ts +126 -0
  22. package/lib/types/auto-allocation-evidence.d.ts +5 -1
  23. package/lib/types/auto-allocation.d.ts +7 -2
  24. package/lib/types/auto-assessment-evidence.d.ts +65 -0
  25. package/lib/types/auto-model-host.d.ts +1 -0
  26. package/lib/types/auto-model-routing.d.ts +1 -1
  27. package/lib/types/auto-task-assessment.d.ts +4 -2
  28. package/lib/types/auto-task-classifier.d.ts +2 -1
  29. package/lib/types/compaction-continuation.d.ts +1 -1
  30. package/lib/types/config.d.ts +18 -2
  31. package/lib/types/context-evidence.d.ts +4 -4
  32. package/lib/types/copilot-auth.d.ts +2 -1
  33. package/lib/types/copilot-usage-remote.d.ts +1 -1
  34. package/lib/types/diagnostics-analysis.d.ts +3 -2
  35. package/lib/types/diagnostics-card.d.ts +1 -0
  36. package/lib/types/diagnostics-collector.d.ts +14 -0
  37. package/lib/types/diagnostics-host.d.ts +114 -3
  38. package/lib/types/diagnostics-remote.d.ts +11 -9
  39. package/lib/types/diagnostics-types.d.ts +486 -30
  40. package/lib/types/native-settings-style.d.ts +1 -1
  41. package/lib/types/preview-provider.d.ts +8 -2
  42. package/lib/types/preview-route.d.ts +1 -1
  43. package/lib/types/remote.d.ts +1 -1
  44. package/lib/types/request-body-composition.d.ts +17 -0
  45. package/lib/types/request-body-evidence.d.ts +2 -1
  46. package/lib/types/request-diagnostics.d.ts +278 -0
  47. package/lib/types/session-continuation-types.d.ts +2 -2
  48. package/lib/types/turn-selection-host.d.ts +2 -1
  49. package/lib/types/turn-selection.d.ts +39 -2
  50. package/package.json +2 -2
  51. package/lib/diagnostics-types-FVf6zjob.js +0 -182
package/AGENTS.md CHANGED
@@ -52,7 +52,22 @@ projection may diagnose concrete failed-zero or invalid-sample incidents; ordina
52
52
  request progress and absent fresh sampling stay quiet. Never filter shared native
53
53
  usage, replace Core's meter, infer current occupancy/percentages or rewrite history.
54
54
  See [quota boundaries](./docs/copilot-usage.md).
55
- 12. Default-off local account/Checking/compaction diagnostics through public storage-domain and strict additive Remotes. Keep fixed-dimension Client/Host and logical/physical populations separate, with bounded handles/rows, 14-day retention, explicit loss/uncertainty and native commit evidence. Isolate the home-wide backend by public profile context, never account/Session identity; one writer per profile is required. No content, raw errors, identities/hashes, credential/history reads, ad-hoc files, Settings logs, upload, daily task or automatic repair. Enable/pause uses only the hidden volatile `diagnosticsEnabled` CAS leaf; clear is a separately confirmed epoch fence. Missing/corrupt storage remains unavailable, never healthy empty data. Publication does not enable collection. See [local diagnostics](./docs/plugin-diagnostics.md).
55
+ 12. Default-off local account/Checking/compaction diagnostics through public storage-domain and strict additive Remotes. Keep fixed-dimension Client/Host and logical/physical populations separate, with bounded handles/rows, 14-day retention, explicit loss/uncertainty and native commit evidence. Isolate the home-wide backend by public profile context, never account/Session identity; one writer per profile is required. No content, raw errors, identities/hashes, credential/history reads, ad-hoc files, Settings logs, upload, daily task or automatic repair. Aggregate and Auto controls remain independent. Separately consented `requestDiagnosticsEnabled` collects at most 128 content-free managed-adapter physical requests for 24 hours within the shared 2 MiB cap. Random stream groups/dispatch indices are not Session/turn identity or Core retry counts. Local body-write completion does not prove supplier receipt; abort attribution needs actual signal evidence, not EOF text. Never change transport/retries/Usage to obtain diagnostics. All controls use hidden volatile CAS leaves with exact readback; clear is a separately confirmed epoch fence. Missing/corrupt storage remains unavailable, never healthy empty data. Publication does not enable collection. See [local diagnostics](./docs/plugin-diagnostics.md).
56
+
57
+ Auto aggregate assessment outcomes are fixed codes only; keep legacy missing reasons
58
+ unrecorded and separate from new cohorts. Count decisions by selected/no-fit counts,
59
+ not repeated candidate opportunities. Never change assessment policy, deadline or
60
+ context safety merely to populate diagnostics; retain the shared storage/row bounds.
61
+
62
+ Credential-change observations (#402) use the existing independent request
63
+ consent: at most 128 fixed notification/commit rows and pending handles for 24
64
+ hours within the shared 2 MiB cap. Derive reasons only from existing refresh
65
+ checks, never new credential reads. No grant/record/identity/endpoint values or
66
+ hashes; random change IDs join verdict phases only. Live-wire counts are not
67
+ unique requests, retries or execution proof. Unknown source and missing legacy
68
+ evidence stay unknown. Preserve #385 qualification/revocation; clear, pause,
69
+ expiry and disposal fence later callbacks. Each diagnostic population, including
70
+ Auto, is independently collapsed by default; opening it never enables collection.
56
71
 
57
72
  ## Session-policy compaction replay consent (#349, #352)
58
73
 
@@ -121,14 +136,14 @@ Alpha.25 admits native `subagent/descriptor` v3, already v3 in the retained rc.1
121
136
  - `src/responses-reasoning.ts`, `src/responses-reasoning-text.ts`: selected-model effort mapping and public summary assembly.
122
137
  - `src/tool-schema-compat.ts`: Copilot-only prompt-assembly filter for unusable escalation arguments and action-specific Goal update schemas.
123
138
  - `src/plan.ts`: Copilot-only, fail-closed hosted-search candidate lifecycle.
124
- - `src/auto-assessment-evidence.ts`: bounded request-local semantic classifier milestones copied into strict ephemeral turn evidence; no prompts, replay, credentials or error bodies. Total includes preparation; adapter start and first text are not HTTP dispatch/byte timings. Keep one 8s monotonic deadline checked at preparation/native request/result boundaries and 128-token compact output; delayed timer settlement is not a real-time guarantee. Prefer eligible supplier-advertised reasoning-off classifiers, but request off only when supplier and public prepared-model effort agree; unsupported native controls stay unchanged. No guessed latency ranking, adaptive timeout, cooldown or cross-turn cache without reviewed observations.
139
+ - `src/auto-assessment-evidence.ts`: bounded request-local semantic classifier milestones copied into strict ephemeral turn evidence; no prompts, replay, credentials or error bodies. Total includes preparation; adapter start and first text are not HTTP dispatch/byte timings. Keep one configured monotonic deadline (30s default, 1000–120000ms) checked at preparation/native request/result boundaries and 128-token compact output; delayed timer settlement is not a real-time guarantee. The user-approved budget trial is not calibrated latency or proof of past timeout causes. Empty classifier setting retains metadata-driven automatic selection; a nonempty exact account ID preserves proof/exclusion/text/headroom guards, never a virtual Auto ID or silent alternate-model fallback. Both native volatile Config leaves affect future assessments and preserve admitted turns. Prefer eligible supplier-advertised reasoning-off classifiers, but request off only when supplier and public prepared-model effort agree; unsupported native controls stay unchanged. No guessed optimal/latency ranking, benchmark requests, adaptive timeout, cooldown or cross-turn cache without reviewed observations.
125
140
  - `src/probe.ts`: bounded native-search capability proof.
126
141
  - `src/wire.ts`, `src/wire-anthropic.ts`: inline hosted-search streaming.
127
142
  - `src/traditional-search.ts`: `github-copilot-hosted` `ctx.web` provider.
128
143
  - `src/serialize.ts`, `src/sse.ts`, `src/failure.ts`: protocol conversion and bounded error handling.
129
144
  - `src/response-error-body.ts`, `src/request-body-timeout.ts`, `src/request-upload-evidence.ts`: bounded clone-only HTTP error evidence and safe managed upload-timeout guidance. Request-scoped public Node diagnostic events supply numeric local body-write/header/ALPN evidence only after an exact verified `408/user_request_timeout`; absent events, multiple native requests and work limits stay explicit. Always dispose subscriptions when Fetch settles, without changing its arguments, Response/errors, dispatcher, proxy or retry owner. Local body-write completion is not upload duration, kernel ACK, supplier receipt or execution; its absence does not prove incomplete upload. Native failure classification/retry metadata stay unchanged; final JSON bytes are not tokens or a payload limit. A separate explicit opt-in compression setting is default-off and never activated by a 408; never infer overflow, auto-compact, trim history, switch models or add retries from a 408.
130
145
  - `src/responses-request-compression.ts`: asynchronous `node:zlib` gzip at `Z_BEST_SPEED` only for exact account-proven official Copilot HTTPS `/responses` JSON POSTs through the existing public Fetch seam. `github-copilot.responsesRequestCompression` defaults false; enforce the 256 KiB minimum, 32 MiB compression-work bound and both 5%/4 KiB savings thresholds. Skips retain the original request with a fixed reason, and cancellation/account/deadline fences run before dispatch. No custom Fetch, other protocol, explicit `auto`/WebSocket, DNS/proxy/dispatcher changes, context-admission or replay changes, resend after an encoding rejection, or new retry. Gzip is wire encoding, not compaction or proof of supplier receipt.
131
- - `src/request-body-evidence.ts`: verified-408-only, bounded numeric original-JSON span accounting through the existing fetch hook. Conversation/tool definitions/top-level system/residual partition total bytes; structural image and opaque replay spans are disjoint conversation subsets, never decoded contents. Unknown/limited evidence stays explicit. Fetch-to-header time excludes clone inspection and is not upload duration. With gzip, report original JSON composition and prepared HTTP-body bytes separately; neither establishes delivered bytes. No success-path analysis, retained body, automatic mitigation or supplier-limit claim.
146
+ - `src/request-body-evidence.ts`, `src/request-body-composition.ts`: bounded numeric original-JSON spans for verified 408s and separately consented request observations. Keep replay validation/recovery at 16 MiB; only diagnostic composition may use the 32 MiB cooperative scanner with 64 Ki-character yields, a two-second monotonic deadline and four scan slots. Do not build a full payload object graph or decode image/replay values. Conversation/tool definitions/top-level system/residual partition total bytes; structural image and opaque replay spans are disjoint conversation subsets. Unknown/limited/cancelled evidence stays explicit. Opted-in large scans start after Fetch invocation without delaying ordinary native delivery; retained numeric rows may update after settlement only behind consent/clear/retention/eviction fences. Retain source strings only for bounded in-flight counting, never in storage/logs. Fetch-to-header time excludes analysis and is not upload duration. With gzip, report original JSON and prepared HTTP-body bytes separately; neither establishes delivered bytes. No automatic mitigation, timeout/retry changes or supplier-limit claim.
132
147
  - `src/copilot-stream-liveness.ts`: request-local HTTP/SSE byte-idle observation, paired with the public native adapter's bounded assistant-chunk deadline. Default five-minute byte idle and ten-minute semantic silence; no fake assistant output, private watchdog access, new retry or Core patch. Consumer think time is excluded; explicit WebSocket/auto stays native-only. `chatStreamLiveness: false` restores native-only timing. `chatMaxRequestImageBytes` retains Core's 20 MiB default; smaller explicit settings authorize only native image offload, not hidden history trimming or a guessed supplier limit. Synthetic heartbeat evidence is not production heartbeat proof.
133
148
  - `tests/`: unit and integration evidence; mirror the source area being changed.
134
149
  - `deployment-baseline.json`: declared machine-readable compatibility and capability evidence inventory.
@@ -136,6 +151,23 @@ Alpha.25 admits native `subagent/descriptor` v3, already v3 in the retained rc.1
136
151
  - `lib/`: generated release output; never edit it.
137
152
 
138
153
  ## Non-negotiable invariants
154
+
155
+ Auxiliary classifier disable proof accepts explicit supplier `off` or the
156
+ selected protocol's native `none` (Responses/Completions) or `disabled`
157
+ (Anthropic), scoped to the isolated classifier model only. Missing/empty effort
158
+ metadata remains unproven. Strict ephemeral candidate counts record one fixed
159
+ first-rejection reason per candidate and distinguish native off preparation
160
+ failure; no candidate IDs, raw metadata, prompts or new persistent population.
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.
139
171
 
140
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`.
141
173
 
@@ -336,9 +368,13 @@ Completed-turn Usage diagnostics (#259) are read-only plugin conversation data a
336
368
  Auto task routing (#258) uses current `model_picker_category` facts, never model
337
369
  names, effort or context capacity as a quality category. Hard eligibility/input
338
370
  fit precede task/preference policy; unknown category and uncertain task evidence
339
- remain explicit. Within the first fitting category, apply positive cost weights
340
- (ordinary 1, marked 0.2) and finite continuity (previous model multiplier 1.5),
341
- never unconditional retention or extreme-only cost gating. Capture actual reasons once
371
+ remain explicit. Within the first fitting category, high-cost-v2 uses the
372
+ unmarked tier whenever it has fitting candidates; otherwise use the marked tier.
373
+ Ordinary weight 1 receives finite previous-model multiplier 1.5; marked weight
374
+ 0.2 never receives continuity. Preserve positive within-tier opportunity and v1
375
+ historical codecs/cohorts, and disclose deferred candidates. Fit is not task
376
+ success; never switch categories for cost, infer incapability or add escalation.
377
+ Capture actual reasons once
342
378
  per turn without durable decision events or changing Usage.
343
379
  `autoSemanticAssessment` is default-on by explicit user rollout request (#267),
344
380
  only for locally unknown demand; preserve explicit false opt-out. One bounded concrete
package/CHANGELOG.md CHANGED
@@ -1,5 +1,131 @@
1
1
  # Changelog
2
2
 
3
+ ## 0.4.2-alpha.11 (prepared)
4
+
5
+ - Prefer fitting unmarked models inside the selected supplier category and defer
6
+ high-cost peers. Marked-only fallback stays usable without a continuity bonus.
7
+ Explain the cost tier and preserve v1 historical evidence; no category
8
+ weakening, task-success inference, automatic escalation or in-turn switch.
9
+ - Preserve a bounded substantive task anchor and its user continuation chain
10
+ ahead of verbose historical output in the auxiliary classifier projection.
11
+ Separate incidental omissions from limited task text; missing anchors,
12
+ truncated task text and task attachments still prevent simple/routine
13
+ downshift. Actual Chat input and native context admission stay untouched.
14
+ - Remove the fixed unknown example from the classifier instruction and explain
15
+ that historical omission alone is not evidence of unknown demand. Keep genuine
16
+ unknown, reasoning off, no tools, one call/no retry and the existing bounds.
17
+ - Expose strict content-free projection facts and the classifier's returned
18
+ demand in the existing selection explanation. These are diagnostics, not
19
+ measured classification accuracy or proof of reduced live unknown rates.
20
+
21
+ ## 0.4.2-alpha.10 (published)
22
+
23
+ - Accept account-advertised protocol-native reasoning disable labels (`none`
24
+ for Responses/Completions, `disabled` for Anthropic) as well as `off` for
25
+ isolated Auto classifiers. Missing/empty capabilities remain unproven;
26
+ ordinary Chat mapping and single-call/no-tools behavior stay unchanged.
27
+ - Capture bounded first-rejection candidate counts and native-off preparation
28
+ failures in strict ephemeral selection evidence. No-fit logs use fixed codes
29
+ without IDs, content, credentials or raw errors. Historical missing fields
30
+ remain unknown; this is not live account root-cause proof.
31
+
32
+ ## 0.4.2-alpha.9 (published)
33
+
34
+ - Require the auxiliary Auto classifier to use advertised reasoning `off` and
35
+ the isolated public adapter's native disable mode. Serialize explicit
36
+ Responses/Completions `none` or Anthropic `disabled`; fail closed when either
37
+ capability is unavailable. Preserve ordinary Chat mappings and no-tools
38
+ behavior. Synthetic wire tests do not establish supplier-side implementation.
39
+
40
+ ## 0.4.2-alpha.8 (published)
41
+
42
+ - Keep every Local diagnostics population independently collapsed by default,
43
+ including Auto. Compact summaries retain status/counts, while collection,
44
+ scope and JSON controls remain separate. Opening a group does not collect data.
45
+ - Extend the independently consented request scope with bounded, content-free
46
+ credential-change qualification and commit evidence: fixed reasons and live
47
+ managed-wire populations only, with random operation IDs, 24-hour retention,
48
+ explicit loss/reopen uncertainty and pause/clear/disposal fences.
49
+ - Preserve the exact native-refresh exception and all unknown-change revocation,
50
+ transport, retry and Usage behavior. Unknown notification source stays unknown;
51
+ no credentials, identities, history or old incidents are reconstructed.
52
+
53
+ ## 0.4.2-alpha.7
54
+
55
+ - Raise the auxiliary Auto task-assessment default total budget from eight to
56
+ thirty seconds as a user-approved trial, with a native editable
57
+ `autoSemanticAssessmentTimeoutMs` setting bounded to 1000–120000 ms.
58
+ - Expose `autoSemanticAssessmentModel`: empty keeps the existing metadata-driven
59
+ automatic classifier; an exact current-account ID can select another eligible
60
+ concrete model. Missing, excluded or unsuitable explicit choices diagnose
61
+ unavailable without silently calling another classifier.
62
+ - Persist both volatile native settings across restart while preserving admitted
63
+ steps/retries. Keep cancellation, account proof, monotonic deadline checks,
64
+ 128-token output, context conservatism and one-call/no-retry ownership.
65
+ No empirical optimal-model search or claim that past unknowns were timeouts.
66
+ - Isolate auxiliary classifier reasoning from the route's configured Chat effort:
67
+ a high-effort profile no longer adds high reasoning to classification. Preserve
68
+ ordinary Chat behavior, no tools and native assembly; omitted reasoning or
69
+ native `off` is not proof of supplier-side disablement or a hard credit cap.
70
+
71
+ ## 0.4.2-alpha.6 (prepared)
72
+
73
+ - Retain fixed assessment outcomes in existing opted-in daily Auto allocation
74
+ aggregates and their diagnostic table: timeout, disabled, unavailable, invalid
75
+ result, failed, omitted context, and known/unknown local or semantic assessment.
76
+ - Preserve legacy missing reasons as unrecorded, keep outcome cohorts separate
77
+ through strict gateways and storage reopen, and distinguish task counts from
78
+ repeated candidate opportunities. No prompts, error bodies or Session IDs added.
79
+ - Leave classification, routing, eight-second assessment budget, conservative
80
+ context handling and collection consent unchanged. Historical unknown causes
81
+ remain unproven; review new outcome data before changing policy.
82
+
83
+ ## 0.4.2-alpha.5 (prepared)
84
+
85
+ - Measure large request JSON composition up to 32 MiB with a cooperative,
86
+ bounded span scanner instead of a full payload object graph. Preserve
87
+ disjoint image/replay byte subsets, native original spans and explicit
88
+ cancellation/time/work limits; replay validation keeps its 16 MiB boundary.
89
+ - Start opted-in large-body counting after Fetch invocation without holding
90
+ up ordinary response/stream delivery. Fence late numeric updates to their
91
+ own retained row across pause/re-enable, clear, teardown and eviction.
92
+ - Reuse bounded composition work in exact verified-408 guidance. Preserve
93
+ native bytes, response/error, retries, Usage, timeout and default-off gzip;
94
+ composition evidence does not establish a proxy or supplier fault.
95
+
96
+ ## 0.4.2-alpha.4 (prepared)
97
+
98
+ - Add independently opt-in, content-free managed-adapter physical request
99
+ observations to Local diagnostics: model/protocol, JSON/wire byte counts,
100
+ bounded composition, local transport timings, HTTP status and strict terminal
101
+ categories. Existing aggregate collection does not enable this scope.
102
+ - Retain at most 128 request rows for 24 hours within the shared 2 MiB snapshot.
103
+ Expose dropped/expired/pending/reopen gaps, independent pause and confirmed
104
+ clear, a native-style request table and explicit reviewed-JSON scope.
105
+ - Preserve native request bytes, retry decisions, stream results and accounting.
106
+ Body-write completion is not supplier receipt; missing terminal text is not
107
+ abort evidence. This improves investigation, not HTTP408/EOF/context recovery.
108
+ - Keep the offline account/compaction analyzer aggregate-only, with an explicit
109
+ gap when separately collected request observations are omitted.
110
+
111
+ ## 0.4.2-alpha.3 (prepared)
112
+
113
+ - Put daily Auto allocation aggregate rows in the primary Local diagnostics view, with UTC date, policy, model, decision cohort, opportunities, expected selections, selected counts and no-fit decisions. Keep cohorts distinct and state that these counts do not prove execution, quality or billing.
114
+ - Move account/Checking and compaction tables, collection controls, and data-scope/limits/JSON guidance into collapsed disclosures. Preserve status retry access, independent default-off switches and separately confirmed clear behavior.
115
+ - Match disclosure summaries and aggregate tables to native settings typography. Update bilingual usage guidance and add focused component regressions; no collection, CAS, storage, privacy or routing policy changes.
116
+
117
+ ## 0.4.2-alpha.1 (publication blocked: immutable tag metadata)
118
+
119
+ - Persist bounded daily Auto allocation aggregates in the existing profile-isolated local diagnostics storage domain. The Auto collection switch is independent and default-off; enabling it does not enable account/Checking or compaction collection.
120
+ - Retain only fixed policy/model/category/demand/assessment/high-cost/continuity dimensions and bounded opportunity, expected-share, selected and no-fit counts. Cap Auto strata at 1,024 and retain for 14 days under the existing snapshot-size limit; expose restart, eviction, drop, truncation and saturation evidence.
121
+ - Preserve turn routing and Host-lifetime explanations when observation is disabled or fails. No Session/turn IDs, conversation data, upload, automatic tuning or execution/quality/billing inference is added.
122
+ - Add collector, UI, Remote codec and storage-gateway regression coverage; synthetic tests do not establish live Desktop activation or Auto performance.
123
+
124
+ ## 0.4.2-alpha.2 (published)
125
+
126
+ - Carry forward the default-off local Auto allocation observations from the unpublished `0.4.2-alpha.1` candidate.
127
+ - Accept the repository's `(prepared)` changelog heading in the release preflight and verify exactly one version section before publishing.
128
+
3
129
  ## 0.4.1
4
130
 
5
131
  - Upgrade from 0.4.0: fix normal plugin-owned same-account native OAuth refresh aborting concurrent already-dispatched managed HTTP streams with `COPILOT_PREVIEW_CREDENTIAL_CHANGED`.
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). 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.11`.** `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
 
@@ -16,32 +16,68 @@ 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, use unmarked models whenever any fit; marked models are a
20
+ fallback tier only. Ordinary weight `1` gets a finite continuity bonus (`1.5×`
21
+ for the previous model); marked weight `0.2` never gets continuity. Participating candidates keep
22
+ a nonzero chance; deferred marked peers do not.
23
+ Marking does not change fixed choices or admitted turns. This is a declared
22
24
  policy, not a price/quality ranking or proven savings.
23
25
 
24
26
  The existing turn selection dialog exposes **Auto allocation observations**:
25
27
  local retained opportunities, expected shares and actual selections for this
26
- Session, with a read-only JSON export. No automatic upload or durable history
27
- is added; restart/eviction loses evidence. `autoAllocationEvidence: false`
28
- stops future candidate capture without changing routing; previously retained
29
- observations remain until disposal/eviction/restart. The auxiliary classifier
28
+ Session, with a read-only JSON export. This view remains Host-lifetime only.
29
+ For durable evaluation, default-off **Local diagnostics** also retains daily
30
+ aggregates by policy/model/category, demand and assessment source, high-cost
31
+ and continuity state, expected/actual selections and no-fit decisions for 14
32
+ days. Enable it independently through the plugin detail settings page; this
33
+ does not enable account/Checking or compaction collection. No Session/turn IDs,
34
+ conversation content, execution/outcome claims, automatic tuning or upload.
35
+ `autoAllocationEvidence: false` stops future candidate evidence in the
36
+ turn-selection view without changing routing. The auxiliary classifier
30
37
  prefers unmarked eligible Lightweight models separately. Advisor calls are not
31
- implemented. See [requirements](./docs/auto-high-cost.md) and the
38
+ implemented. See [requirements](./docs/auto-high-cost.md),
39
+ [local diagnostics](./docs/plugin-diagnostics.md) and the
32
40
  [required iteration review](./docs/evidence-driven-iteration.md).
33
41
 
34
42
  The [local diagnostics pilot](./docs/plugin-diagnostics.md) records bounded
35
- account/Checking and compaction aggregates, **default off**. After installing
43
+ account/Checking, compaction and Auto allocation aggregates, **default off**. After installing
36
44
  and loading the published build, open the plugin detail settings page →
37
45
  **Local diagnostics → Read status → Enable local collection**. This enables
38
- local persistence, not uploads or scheduled analysis. The package also has an
39
- explicit offline analyzer for a caller-selected persisted unit or reviewed
40
- aggregate view; it writes only to a new local report file and never scans,
41
- uploads or changes collection. Pause preserves evidence; Clear requires
42
- separate confirmation. Client/Host populations, uncovered paths and
43
- reporting/storage gaps remain explicit. Descriptive counts are not reconstructed
44
- history, failure rates or proof of an improvement.
46
+ local account/Checking and compaction persistence, not uploads or scheduled
47
+ analysis. Enable **Auto allocation observation** separately to collect only
48
+ Auto allocation aggregates. **Enable request observations** is a third, independent,
49
+ default-off control: up to 128 content-free managed-adapter physical requests
50
+ for 24 hours, with model, JSON/wire byte counts, timings, HTTP status and terminal
51
+ categories. No bodies, headers, credentials or Session/turn IDs are collected.
52
+ Local body-write completion does not prove supplier receipt; these diagnostics
53
+ do not fix timeouts or alter retry/Usage behavior.
54
+ All diagnostic groups, including Auto, are independently collapsed by default
55
+ inside the Local diagnostics region. Request consent also covers up to 128
56
+ credential-change notification/commit observations for 24 hours: fixed
57
+ qualification reasons and live-wire counts only. Unknown source remains unknown;
58
+ no past incident or account/Session attribution is reconstructed. See
59
+ [credential-change evidence](./docs/plugin-diagnostics.md#credential-change-decisions-402).
60
+ Large JSON bodies up to 32 MiB can receive bounded, cooperative composition
61
+ counts without a full payload parse; the approximately 20.4 MiB class no longer
62
+ necessarily hits the old 16 MiB diagnostic blind spot. Time/work/cancellation
63
+ limits remain explicit, and settled rows may acquire counts on a later read.
64
+ No gzip, history trimming or extra retry is enabled. See [request observations](./docs/plugin-diagnostics.md#content-free-physical-request-observations-392).
65
+ The package also has an explicit offline analyzer
66
+ for a caller-selected persisted unit or reviewed aggregate view; it writes only
67
+ to a new local report file and never scans, uploads or changes collection. Pause
68
+ preserves evidence; Clear requires separate confirmation. Client/Host
69
+ populations, uncovered paths and reporting/storage gaps remain explicit.
70
+ Descriptive counts are not reconstructed history, failure rates or proof of an
71
+ improvement. The diagnostics page opens on a daily Auto allocation table by UTC
72
+ date, policy, model and cohort, with opportunities, expected selections, selected
73
+ counts and no-fit decisions. Compare expected and selected counts only within
74
+ matching cohorts. New cohorts distinguish assessment timeouts, unavailable/invalid
75
+ results, omitted context, disabled assessment and genuinely unknown local/semantic
76
+ results; legacy reasons remain **not recorded**. Count tasks with Selected plus
77
+ No-fit, not candidate opportunities. This is diagnostic coverage, not a routing
78
+ change or proof of the supplier root cause. Account/Checking and compaction reports, collection controls,
79
+ and detailed scope, limits and JSON are in collapsed disclosures; clearing still
80
+ requires separate confirmation.
45
81
 
46
82
  To analyze an explicitly qualified file from a built package or source checkout:
47
83
 
@@ -92,10 +128,10 @@ Supply any launcher patches with repeated `--patch /absolute/file`. Require `sup
92
128
  For a **standalone named profile**:
93
129
 
94
130
  ```sh
95
- dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.1/dsh-github-copilot-0.4.1.tgz
131
+ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.11/dsh-github-copilot-0.4.2-alpha.11.tgz
96
132
  ```
97
133
 
98
- For **Desktop**, its native package manager accepts the exact stable npm spec `dsh-github-copilot@0.4.1`; stable versions publish on npm's `latest` dist-tag. 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.
134
+ For **Desktop**, its native package manager accepts an exact npm spec, for example `dsh-github-copilot@0.4.2-alpha.11`; 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.
99
135
 
100
136
  After an approved reload/restart:
101
137
 
@@ -128,11 +164,25 @@ All three Auto preferences use the same account-verified, non-excluded eligible
128
164
  | Complex | Powerful | Powerful | Powerful |
129
165
  | Unknown | Versatile | Versatile | Powerful |
130
166
 
131
- 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.
167
+ Categories come from authenticated supplier metadata, not model names, context capacity or measured quality rankings. Within the first fitting category, the cost tier and positive weights described above choose the model; a suitable previous high-cost model is deferred when an unmarked peer fits. 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.
132
168
 
133
- Semantic assessment is **enabled by default only for locally unknown tasks**. It makes at most one bounded auxiliary call with an **8-second end-to-end deadline** and **128-token output budget**. Set `github-copilot.autoSemanticAssessment: false` to opt out. Timeout/invalid output leaves demand unknown and uses the preference fallback; it never means simple or triggers another classifier. Caller cancellation and account invalidation remain terminal. Auxiliary calls add latency and supplier charges outside native answer Usage.
169
+ 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).
134
170
 
135
- The deadline is checked against a monotonic clock at preparation, native request and result boundaries, even if the event loop delays the timer. Such delays can postpone settlement; they do not authorize late dispatch or late-result acceptance. Eligible Lightweight classifiers advertising reasoning `off` take precedence, with deterministic ID tie-breaking; the request uses `off` only if the public native model also supports it. This is not a measured speed ranking. Explanations distinguish the fitting category pool and previous-model continuity from semantic merit.
171
+ Timeout/invalid output leaves demand unknown and uses the preference fallback; it never means simple or triggers another classifier. Caller cancellation and account invalidation remain terminal. Auxiliary calls add latency and supplier charges outside native answer Usage.
172
+
173
+ The deadline is checked against a monotonic clock at preparation, native request and result boundaries, even if the event loop delays the timer. Such delays can postpone settlement; they do not authorize late dispatch or late-result acceptance. Eligible Lightweight classifiers must advertise an explicit reasoning-disable mode: `off`, or protocol-native `none` for Responses/Completions or `disabled` for Anthropic. Deterministic ID tie-breaking follows. This is not a measured speed ranking. Explanations distinguish the fitting category pool and previous-model continuity from semantic merit.
174
+
175
+ 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.
176
+
177
+ Classification preserves the current request and a bounded substantive continuation
178
+ anchor before verbose historical output. Incidental historical omission alone no
179
+ longer forces a valid simple/routine result to unknown; missing anchors, limited
180
+ task text and nontext task dependencies still block downshift. The prompt does not
181
+ pre-fill an unknown answer. Assessment details distinguish projection limitations
182
+ from the classifier's returned demand. Unknown remains valid; retained text and
183
+ synthetic regressions do not prove calibrated accuracy or a reduced live unknown rate.
184
+
185
+ 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.
136
186
 
137
187
  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.
138
188
 
@@ -271,14 +321,14 @@ Commit attribution uses `Assisted-by` with the actual tool, never the model prov
271
321
  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.
272
322
 
273
323
  ```sh
274
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.1/dsh-github-copilot-0.4.1.tgz
275
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.1/SHA256SUMS
324
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.11/dsh-github-copilot-0.4.2-alpha.11.tgz
325
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.11/SHA256SUMS
276
326
  sha256sum --check SHA256SUMS
277
327
  ```
278
328
 
279
329
  ```powershell
280
330
  $expected = (Get-Content .\SHA256SUMS).Split()[0]
281
- $actual = (Get-FileHash .\dsh-github-copilot-0.4.1.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
331
+ $actual = (Get-FileHash .\dsh-github-copilot-0.4.2-alpha.11.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
282
332
  if ($actual -cne $expected) { throw 'Release checksum mismatch' }
283
333
  ```
284
334
 
package/README.zh.md CHANGED
@@ -8,31 +8,55 @@
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)。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.11`。** `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,只要未标记候选可容纳输入,
19
+ 就暂缓高成本候选;否则使用高成本兜底层。普通权重 `1` 可获得上一轮 `1.5×` 有限加权,
20
+ 高成本权重 `0.2` 不获得连续性加权。参与层内候选有非零机会,暂缓候选没有本轮分配机会;
20
21
  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
22
 
22
23
  现有 turn 选择说明内提供 **Auto allocation observations**,显示当前 Session
23
- 保留的候选机会、预期份额与实际选择,并提供只读 JSON 导出。不自动上传,也不写入持久
24
- 历史;重启或淘汰会丢失观察。设置 `autoAllocationEvidence: false` 可停止后续候选
25
- 记录,不改变路由;此前观察保留至销毁、淘汰或重启。辅助分类器独立优先使用未标记的
26
- 合格 Lightweight 模型。顾问求助尚未实现。参见[完整需求](./docs/auto-high-cost.md)
24
+ 保留的候选机会、预期份额与实际选择,并提供只读 JSON 导出;此视图仍仅在 Host
25
+ 生命周期内保留。为了持久评估,默认关闭的 **Local diagnostics** 还会按策略/模型/类别、
26
+ 需求与评估来源、高成本/连续性状态汇总预期/实际选择及 no-fit 决策,保留 14 天。
27
+ 可在插件详情设置页单独启用,不会因此采集账号/Checking 或压缩汇总。不记录
28
+ Session/turn ID、对话内容,也不推断执行/结果,不自动调参或上传。
29
+ `autoAllocationEvidence: false` 会停止 turn 选择视图的后续候选证据,但不改变路由。
30
+ 辅助分类器独立优先使用未标记的合格 Lightweight 模型。顾问求助尚未实现。参见
31
+ [完整需求](./docs/auto-high-cost.md)、[本地诊断](./docs/plugin-diagnostics.md)
27
32
  与[强制迭代评审流程](./docs/evidence-driven-iteration.md)。
28
33
 
29
- [本地诊断试点](./docs/plugin-diagnostics.md)记录有界的账号/Checking 与压缩
34
+ [本地诊断试点](./docs/plugin-diagnostics.md)记录有界的账号/Checking、压缩与 Auto 分配
30
35
  汇总,**默认关闭**。安装并加载已发布版本后,进入插件详情设置页 →
31
36
  **Local diagnostics → Read status → Enable local collection**,开启本地持久化;
32
- 不上传,不启用定时分析。包内另提供显式离线分析器,只处理调用者指定的持久化单元
33
- 或经审阅的聚合视图,并且只新建本地报告文件;不会扫描、上传或更改采集设置。
34
- 暂停保留证据,清除需单独确认。Client/Host 分开统计,未覆盖路径、报告丢失与存储
35
- 故障明确显示。描述性计数不是历史重建、错误率,也不表示优化效果已被验证。
37
+ 仅用于账号/Checking 与压缩汇总,不上传或启用每日分析。另行启用
38
+ **Auto allocation observation** 才会单独收集 Auto 分配汇总。第三个独立、默认关闭的
39
+ **Enable request observations** 开关保留最多 128 条托管适配器的物理请求观测,保留 24 小时:
40
+ 仅包含模型、JSON/传输字节数、耗时、HTTP 状态和终止分类,不记录正文、headers、
41
+ 凭据或 Session/turn ID。本地写出完成不证明供应方收到了请求;诊断不会修复超时,
42
+ 也不改变重试或 Usage。最高 32 MiB 的大 JSON 请求可通过有界、协作式扫描获得组成计数,
43
+ 不完整解析 payload 对象图;约 20.4 MiB 的请求不再必然落入旧 16 MiB 诊断盲区。
44
+ 耗时/工作量/取消限制仍明确可见,已结束的记录可能在之后读取时才补齐计数。
45
+ 不自动启用 gzip、截断历史或增加重试,参见[请求观测](./docs/plugin-diagnostics.md#content-free-physical-request-observations-392)。
46
+ Local diagnostics 区域内各诊断组均默认独立折叠,Auto 也只按需展开。
47
+ 请求观测同意范围还包含最多 128 条凭据变更通知/提交观测,保留 24 小时:
48
+ 仅记录固定资格判定原因码和活动请求计数。未知来源仍是未知,不还原过去故障,
49
+ 也不推断账号或 Session 归属,参见[凭据变更证据](./docs/plugin-diagnostics.md#credential-change-decisions-402)。
50
+ 包内另提供显式离线分析器,
51
+ 只处理调用者指定的持久化单元或经审阅的聚合视图,并且只新建本地报告文件;不会扫描、
52
+ 上传或更改采集设置。暂停保留证据,清除需单独确认。Client/Host 分开统计,未覆盖路径、
53
+ 报告丢失与存储故障明确显示。描述性计数不是历史重建、错误率,也不表示优化效果已被验证。诊断页默认显示按 UTC 日期、
54
+ 策略、模型和 cohort 划分的 Auto 分配每日汇总表,包含候选机会、预期选择、实际选择及
55
+ no-fit 决策;只能在匹配 cohort 内比较预期与实际选择。新 cohort 区分评估超时、不可用/无效
56
+ 结果、上下文缺失、评估关闭及本地/语义判断确实未知;旧数据原因保持 **not recorded**,
57
+ 不倒推。按 Selected 加 No-fit 计任务数,不按候选机会计数。这是诊断覆盖,不改变选模策略,
58
+ 也不证明供应商根因。账号/Checking 和压缩报告、采集控制、
59
+ 详细范围/限制/JSON 位于折叠区;清除仍需单独确认。
36
60
 
37
61
  在已构建的包或源码 checkout 内,显式分析已核验的文件:
38
62
 
@@ -83,10 +107,10 @@ node package/scripts/check-search-composition.mjs --profile-dir /absolute/profil
83
107
  **独立具名 profile**:
84
108
 
85
109
  ```sh
86
- dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.1/dsh-github-copilot-0.4.1.tgz
110
+ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.11/dsh-github-copilot-0.4.2-alpha.11.tgz
87
111
  ```
88
112
 
89
- **Desktop** 原生包管理器接受精确稳定版 npm spec `dsh-github-copilot@0.4.1`;稳定版发布到 npm `latest` dist-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
+ **Desktop** 原生包管理器接受精确 npm spec,例如 `dsh-github-copilot@0.4.2-alpha.11`;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。
90
114
 
91
115
  经批准 reload/restart 后:
92
116
 
@@ -119,11 +143,22 @@ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/release
119
143
  | 复杂 | Powerful | Powerful | Powerful |
120
144
  | 未知 | Versatile | Versatile | Powerful |
121
145
 
122
- 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。
146
+ 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,按前述成本层和正权重选择;未标记同类候选可容纳输入时,暂缓上一轮高成本模型。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。
123
147
 
124
- 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,**8 秒端到端期限、128 tokens 输出预算**。可用 `github-copilot.autoSemanticAssessment: false` 关闭。超时/无效结果保留未知,再按偏好兜底,不解释为简单,也不换分类模型重试。用户取消和账号失效仍终止请求。辅助调用增加延迟与供应方费用,不计入原生回答 Usage。
148
+ 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)。
125
149
 
126
- 准备、原生请求和结果边界均用单调时钟检查期限,不依赖定时器及时执行。事件循环延迟可能推迟结束,但不授权超期发起请求或接受结果。合格 Lightweight 分类模型优先考虑声明 reasoning `off` 的候选,同级按 ID 确定顺序;只有公开原生模型也支持时才发送 `off`。这不是实测速率排名。解释分别说明该分类的可容纳候选数和上一模型连续性,不将保留上一模型说成语义优选。
150
+ 超时/无效结果保留未知,再按偏好兜底,不解释为简单,也不换分类模型重试。用户取消和账号失效仍终止请求。辅助调用增加延迟与供应方费用,不计入原生回答 Usage。
151
+
152
+ 准备、原生请求和结果边界均用单调时钟检查期限,不依赖定时器及时执行。事件循环延迟可能推迟结束,但不授权超期发起请求或接受结果。Lightweight 分类候选必须由当前账号元数据明确声明禁用 reasoning 的模式:`off`,或 Responses/Completions 的 `none`、Anthropic 的 `disabled`,再按确定性 ID 顺序选择。字段缺失或空列表不证明模型不会 reasoning。这不是实测速率排名。解释分别说明该分类的可容纳候选数和上一模型连续性,不将保留上一模型说成语义优选。
153
+
154
+ 请求内严格选模证据包含候选总数、可用数和每个模型的首个固定排除原因计数:被排除、指定 ID 不匹配、类别、文本输入、禁用能力未证实或预算。不在计数中新增模型 ID。无候选日志仅输出固定原因码,并区分原生适配器无法准备 off 的情况;不记录正文、凭据或原始错误。这是临时证据,不新增持久诊断采集;旧记录缺字段仍为未知。
155
+
156
+ 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。
157
+
158
+ 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
159
+ 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
160
+ 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。
161
+ 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。
127
162
 
128
163
  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。
129
164
 
@@ -262,14 +297,14 @@ Agent 从 `node scripts/agent.mjs describe --json`、`doctor --json`及`plan <ta
262
297
  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。
263
298
 
264
299
  ```sh
265
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.1/dsh-github-copilot-0.4.1.tgz
266
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.1/SHA256SUMS
300
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.11/dsh-github-copilot-0.4.2-alpha.11.tgz
301
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.2-alpha.11/SHA256SUMS
267
302
  sha256sum --check SHA256SUMS
268
303
  ```
269
304
 
270
305
  ```powershell
271
306
  $expected = (Get-Content .\SHA256SUMS).Split()[0]
272
- $actual = (Get-FileHash .\dsh-github-copilot-0.4.1.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
307
+ $actual = (Get-FileHash .\dsh-github-copilot-0.4.2-alpha.11.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
273
308
  if ($actual -cne $expected) { throw 'Release checksum mismatch' }
274
309
  ```
275
310