dsh-github-copilot 0.4.0-alpha.88 → 0.4.0-alpha.92

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/README.md CHANGED
@@ -1,486 +1,187 @@
1
1
  # dsh-github-copilot
2
2
 
3
3
  [![CI](https://github.com/cloga/dsh-github-copilot/actions/workflows/ci.yml/badge.svg)](https://github.com/cloga/dsh-github-copilot/actions/workflows/ci.yml)
4
- [![Release](https://github.com/cloga/dsh-github-copilot/actions/workflows/release.yml/badge.svg)](https://github.com/cloga/dsh-github-copilot/actions/workflows/release.yml)
5
- [![Latest release](https://img.shields.io/github/v/release/cloga/dsh-github-copilot)](https://github.com/cloga/dsh-github-copilot/releases/latest)
4
+ [![Release](https://img.shields.io/github/v/release/cloga/dsh-github-copilot?include_prereleases)](https://github.com/cloga/dsh-github-copilot/releases)
6
5
  [![License](https://img.shields.io/github/license/cloga/dsh-github-copilot)](./LICENSE)
7
6
 
8
7
  **English** | [简体中文](./README.zh.md)
9
8
 
10
- A focused DSH companion for GitHub Copilot sign-in, account-aware model profiles, Copilot-specific tool compatibility, and provider-hosted search. It reuses DSH's built-in `@deepseek-ai/dsh-llm-pi-ai`; it is not a second Copilot model adapter or catalog.
9
+ GitHub Copilot account discovery, contextual Auto routing and hosted search for DSH. The plugin reuses DSH's published native adapter and one Host-owned OAuth account; it does not patch Core or maintain a second model catalog.
11
10
 
12
- ## Current experience
11
+ **Current release: `0.4.0-alpha.92`. Supported host: official DSH / Windows Desktop `0.2.0-rc.2`.** Earlier DSH pins are historical evidence, not supported installation targets. Publication, profile installation and the version loaded by a running Host are separate states.
13
12
 
14
- Auto checks historical user/tool image blocks as well as the latest message.
15
- Managed requests reject an explicitly incompatible projected image MIME before
16
- network dispatch; omitted account MIME lists remain unverified, not guessed.
17
- Core still owns image conversion and history. See [image compatibility limits](./docs/image-input-compatibility.md).
13
+ ## What you can do
18
14
 
19
- - **Account and models:** sign in, inspect account-discovered models, refresh metadata, and manage the account from the native GitHub Copilot row under **Settings → Models**. **Manage → Model preferences** can exclude exact models from both the managed picker and Auto without changing Core's flat picker.
20
- - **Search routing:** choose the primary and fallback search providers under **Plugins → dsh-github-copilot → Details**. The page intentionally exposes only those user-facing choices; protocol probes, timeouts, allowlists, and transport switches are not presented as routine UI. It also keeps the rc.2 lifecycle warning visible before disable/remove/upgrade actions. If the updated routed service is not active yet, the card reports **Restart required** instead of presenting a broken form.
21
- - **Usage:** eligible Copilot sessions get an optional **Credits** control beside the composer, with provider-reported billing-cycle data and explicit unavailable/stale states. The quota request alone combines Node default and OS trust roots without disabling certificate verification or relying on Desktop launcher environment inheritance.
22
- - **Context evidence:** managed requests preserve native usage, including terminal zero on failure/cancellation, so context diagnostics cannot remove shared accounting evidence. A separate composer disclosure recognizes unreliable failed-zero samples and shows the last applicable historical input-plus-cache reading; it is not current occupancy, replaces no native `0%` meter and rewrites no history. See [context evidence boundaries](./docs/copilot-usage.md#historical-context-evidence).
23
- - **Missing Turn Usage:** a completed Copilot reply without a native total can show **Turn Usage incomplete**. Its read-only explanation distinguishes recorded local pre-dispatch budget blocks, same-attempt HTTP 408 diagnoses, other missing samples and incomplete history. Bounded step/record references and proven same-step retry completion explain what happened, not missing token amounts or billing; it never invents zero usage or a partial turn total. [Observation and boundaries](./docs/copilot-usage.md#missing-completed-turn-usage).
24
- - **Compatibility:** the plugin applies Copilot-only tool-schema, reasoning-presentation, request-budget, and account-model compatibility through published DSH APIs. It does not patch Core or maintain a second model catalog.
15
+ | Task | Where to start |
16
+ |---|---|
17
+ | Sign in and manage account models | **Settings → Models → GitHub Copilot → Sign in** |
18
+ | Exclude or restore individual models | **Manage → Model preferences** |
19
+ | Choose models automatically | Pick **Auto · Balance**, **Auto · Efficiency** or **Auto · Intelligence** |
20
+ | Follow a parent model in supported subagents | **Plugins → dsh-github-copilot → Details → Follow parent model** |
21
+ | Choose primary/fallback search providers | **Plugin details → Web search** |
22
+ | Read account Credits and context evidence | Copilot Session composer; native Turn Usage stays separate |
25
23
 
26
- ![Current Web search provider routing on the plugin details page](./docs/images/copilot-search-routing.png)
24
+ ![Account controls and model preferences from the published Client](./docs/images/copilot-model-preferences.png)
27
25
 
28
- The screenshot uses the current built Client with synthetic providers in an isolated browser fixture. It demonstrates the shipped configuration surface, not live credentials, account access, search success, or production Desktop activation.
29
-
30
- On official rc.2, routing is exposed through native live `Config` fields, not a manually registered settings namespace. Saving updates the two routing leaves together without remounting the plugin or changing ordinary safety settings. A legacy search-model override is reset only by its separate explicit action.
31
-
32
- ## Compatibility baselines and qualification targets
33
-
34
- Managed Copilot HTTP `408 / user_request_timeout` with the verified request-body timeout message reports `COPILOT_REQUEST_BODY_TIMEOUT`, the final JSON request's UTF-8 byte count, bounded numeric composition and fetch-to-response-header elapsed time when observable. Composition partitions conversation, tool definitions, top-level system/instructions and other/framing; image blocks and opaque replay are disjoint subsets of conversation. Remaining conversation includes text/tool history, framing and unknown fields, not a proven text-only count. Image blocks include URLs/references, not decoded image sizes. Timing is a round trip, **not upload duration**; bytes are not context tokens, a supplier payload limit or proof of the cause. Missing, unsupported or work-limited evidence stays explicit. Analysis runs only after the exact error is verified and never records request text, credentials or replay contents. Native retries and failure classification remain unchanged. After retries exhaust, these numbers can inform an **explicit** native image-offload budget change or compaction for compressible history; tool definitions and opaque replay may remain. Also check network/proxy health and [GitHub service status](https://www.githubstatus.com/). Compaction is not guaranteed to succeed or cure network/service faults; increasing a client timeout cannot change the server timeout. No automatic history trimming, settings change, model switch, compaction or new retry loop is introduced. Unknown errors retain native behavior.
35
-
36
- Managed HTTP/SSE chat now distinguishes **byte idle** from **assistant-output silence**. Native Core ignores SSE heartbeats when timing assistant chunks; synthetic unchanged-adapter tests reproduce a timeout despite incoming bytes. By default, five minutes without response headers/nonempty body bytes still fails, but real byte progress permits up to ten minutes without an assistant chunk. Endless heartbeats cannot postpone failure indefinitely, and consumer think time is excluded. `chatStreamLiveness: false` restores native-only timing; explicit WebSocket/`auto` transport remains native-only. This fixes the reproduced false-idle mechanism, not every retry or upstream HTTP 408, and does not establish heartbeat activity in a production incident.
37
-
38
- | DSH surface | Exact source pin | Models UI seam |
39
- |---|---|---|
40
- | Controlled Desktop `0.1.1-rc.2` baseline | Controlled Core commit [`a772dbb`](https://github.com/cloga/deepseek-harness/commit/a772dbbde82780bff2b9394427e9f0a24cafa1d5) on `cloga-pi-ai-model-api` | Dedicated **Settings → GitHub Copilot** section |
41
- | DSH `0.1.2-rc.1` | Tag commit [`a66e470`](https://github.com/deepseek-ai/deepseek-harness/commit/a66e4702047846cdaa10c66c9d3df3951f5ea70d) | **Settings → Models** provider card |
42
- | DSH `0.1.3-alpha.1` | Tag commit [`d347e70`](https://github.com/deepseek-ai/deepseek-harness/commit/d347e703908d0406b7a7ef80e3a0e594d86b2215) | **Settings → Models** provider card |
43
- | Official DSH `0.1.5-alpha.1` | Tag commit [`5dda764`](https://github.com/deepseek-ai/deepseek-harness/commit/5dda764ed3aa172535a7967b06ff95d9cbfe536a) | **Settings → Models** provider card |
44
- | Official DSH `0.1.5-alpha.2` | Tag commit [`b2e3b2a`](https://github.com/deepseek-ai/deepseek-harness/commit/b2e3b2a0125854567a4a5fcba75782e42fe84901) | **Settings → Models** provider card |
45
- | Official DSH `0.1.5-rc.1` | Tag commit [`183f08e`](https://github.com/deepseek-ai/deepseek-harness/commit/183f08e9c6dde7e36cd2318eaee70b0da08fb35e) | **Settings → Models** provider card |
46
- | Official DSH `0.1.5-rc.2` | Tag commit [`fb2c4b9`](https://github.com/deepseek-ai/deepseek-harness/commit/fb2c4b9e698e30edb738bca4cf0618587db7d203) | **Settings → Models** provider card |
47
- | Official DSH `0.1.6-alpha.1` | Tag commit [`0a15e36`](https://github.com/deepseek-ai/deepseek-harness/commit/0a15e36e7f82b6ed45af6fa9759f29b40dcd965d) | **Settings → Models** provider card |
48
- | Official DSH `0.1.6-alpha.2` | Tag `dsh-v0.1.6-alpha.2`, commit [`ddefc45`](https://github.com/deepseek-ai/deepseek-harness/commit/ddefc45fbc7f8e46dd73185e68295696d1297887) | **Settings → Models** provider card |
49
- | Official DSH `0.2.0-rc.1` (historical) | Tag `dsh-v0.2.0-rc.1`, commit [`4878cd`](https://github.com/deepseek-ai/deepseek-harness/commit/4878cdabd87d4041bdaff61d04c966883b9fd07a) | **Settings → Models** provider card |
50
- | Official DSH and Windows Desktop `0.2.0-rc.2` (current qualification target) | Tag `dsh-v0.2.0-rc.2`, commit [`639ed0`](https://github.com/deepseek-ai/deepseek-harness/commit/639ed015397290b3745d163aafe02ffee4aa3f84) | **Settings → Models** provider card |
51
-
52
- The table preserves historical source pins; this release admits and gates **only `0.2.0-rc.2`**. The prior pins are not current support claims. The published-artifact synthetic transport fixture and development dependencies target exact `0.2.0-rc.2` npm artifacts. Its unchanged tagged-source fixture runs without building or patching Core. The signed Desktop shared-package descriptor was audited for peer ownership, not plugin loading or live endpoints. Existing public Host, Client, and Remote seams are retained. Stock Core model-entry `api` support is not a prerequisite for the plugin-owned route. Exact peer and `engines.dsh` declarations are admission checks, not live compatibility proof. No Core patch is installed by this plugin.
53
-
54
- ### Alpha.11 compatibility correction (#105)
55
-
56
- Core `0.1.5-alpha.2` requires `ResolvedPiAiProviderProfile.modelErrors` and reads it unconditionally in `PiAiAdapter.modelOf`. The plugin now supplies its own empty diagnostic map: account descriptors have already passed validation, and rejected models never enter the provider. No upstream profile, catalog, prototype or dependency artifact is changed. Actual pinned-adapter tests cover resolution, preparation and synthetic streaming; the older pins retain the same plugin code.
57
-
58
- Legacy Anthropic inline requests containing an in-band `system` message delegate unchanged to Core **before probing** rather than demoting system authority to a user turn. This intentionally limits inline search on those requests. The legacy Responses inline wire retains its existing mapping of explicit system content to user input text; this is not system-message filtering. The managed route always uses native Core transport. See the [seam audit and evidence limits](./docs/agent-readiness.md#core-alpha2-compatibility-follow-up-105-planned-alpha11).
59
-
60
- ### Alpha.12 release prerequisite correction (#107)
61
-
62
- The first alpha.11 publication attempt stopped before packing: the tagged Session/Remote fixture required Core's `mime-types` dependency, but the release job installed only the pi-ai closure. CI and release now explicitly install the unchanged pinned Session Controller dependency closure before that fixture runs. The tests remain enabled; no Core source or live dependency is patched. Alpha.12 carries the same runtime compatibility fixes with a fresh release version.
63
-
64
- ### Alpha.17 provider-aware web search (#118)
65
-
66
- The bundle routes search through a plugin-owned policy configured in plugin details (with a Models-page fallback): `auto` prefers eligible native Copilot search and otherwise uses the configured default provider; `fixed` always uses that provider. A separately selected account model lets Volcengine and other non-Copilot chat Sessions use Copilot hosted search. The implementation uses public web-service composition without Core or preset edits. Source and synthetic tests are not proof of live Copilot search, release publication or local activation; see [routing evidence](docs/session-search-routing.md).
67
-
68
- ### Alpha.19 DSH 0.1.6 compatibility adaptation (#125)
69
-
70
- The exact `dsh-v0.1.6-alpha.1` source fixture now awaits serialized `agent/created` initialization before reading live Session projections. Static and runtime gates also verify the retained request-header/projection path instead of adding synchronous history reads; MCP SDK v2 resource cursors; the `dsh-ptc-runtime` and `dsh-workflow-ptc` contracts; empty model environment for isolated Node PTC; asynchronous cancellable Sandbox/Shell preparation; consumer-owned optional-plugin startup failures; attachment request caches under DSH cache while normalized attachment paths remain stable; and provider-owned Team task pagination. For image-budget recovery, the adapter's first `IMAGE_OFFLOAD_REQUIRED` result remains an error; the fixture records the Core `image/offload` projection and proves the retried Copilot request sends the mapped read-only normalized path as placeholder text without image bytes. The plugin does not import or own MCP, PTC, Workflow, Sandbox, Shell, or Team services. Upstream exposes no generic `HostGrant`/`hostGrants` API in this tag, and the plugin registers no such coupling. Copilot schema compatibility continues to remove unsupported escalation controls from `pwsh`, filesystem, and `run_code` schemas while preserving Team pagination fields. This compatibility version is prepared as a Draft and is not a release claim.
71
-
72
- ### Alpha.21 Desktop shared-package ownership fix (#125)
73
-
74
- The candidate manifest treats `@deepseek-ai/dsh-authorization` and `@deepseek-ai/schemastery` as required Host peers instead of private runtime dependencies. Development copies remain pinned for standalone build, unit, Host import, Client loader, and Remote codec verification. The packed-tarball gate audits every declared dependency and peer against a hash-pinned actual Desktop 0.1.5 runtime descriptor and the generated Desktop 0.1.6 package-set input; it rejects bundled, optional, incompatible, or newly unaudited shared ownership. The 0.1.6 package-set is authoritative input to descriptor generation but is not a materialized Desktop descriptor, live activation, OAuth, or model-call result. This change preserves the existing Settings → Models provider card, authentication flow, lifecycle adaptation, and image-offload behavior, and does not weaken Desktop validation.
75
-
76
- ### Alpha.22 Client React ownership fix (#132)
77
-
78
- The built Client intentionally requests React from DSH's browser `ModuleLoader` singleton. React is not a Desktop Host shared package and must not be installed as a required Node profile peer. Alpha.22 declares `dsh.client.external: [\"react\"]`, removes the root React peer, and retains React only as a development dependency. Packed verification now audits every required peer and Client external, confirms the real built Client requests exactly React from the loader, and preserves authorization/schemastery as required Host peers. This fixes the packaged Desktop startup error `requires missing react@^18.2.0` without enabling peer auto-install, bundling another React copy, or weakening Desktop graph validation. Actual packaged Electron loading remains a downstream acceptance gate.
79
-
80
- Alpha.24 fixed the Web search card's missing traced Remote dependency. At that time it lived in **Settings → Models** (`settings.models.footer`, list/root), with an old-Core Web search section fallback. It now prefers the plugin's detail page; search routing is stored under the existing `github-copilot` namespace via nested paths without requiring a separate section registration, and without affecting account controls or search safety guards. Actual packaged Desktop acceptance is still a separate gate.
81
-
82
- ### Alpha.25 official-first DSH 0.1.6-alpha.2 adaptation
83
-
84
- The historical baseline inventory includes `dsh-v0.1.6-alpha.2` at `ddefc45fbc7f8e46dd73185e68295696d1297887`; it is not admitted by this release. Strict Remote descriptors now provide alpha.2 `create()` factories while retaining a legacy `schema` bridge to the same strict parser. Dedicated executor projection admission uses native `subagent/descriptor` **v3**, already v3 in rc.1: the old v1 assumption was a plugin bug, not an upstream v1-to-v3 migration. Projection cache **`stateVersion: 2`** forces refolding, not history conversion. Unknown/v1/v2 descriptor histories fail closed and remain unmodified; review the old child and explicitly create a new child if needed, never fake a conversion by relabeling its descriptor.
85
-
86
- Source markers, local rc.1-backed focused tests and fifteen scoped exact-source runtime tests passed (alpha.2 contracts 8, Remote 1 and Session-context 6). Full local `pnpm verify` passed: 1373 Vitest tests with 2 expected skips, 176 tooling tests, typechecks, build and package smoke; pack/tarball verification also passed. The scoped run used a supplemental resolver with official TypeScript `6.0.3`, declared `mime-types@3.0.2` and `ws@8.21.0`, and shared Zod `^4.4.3`, without source/dependency patches; it is not full official-root-helper qualification. Broad frozen dependency installation remains blocked by the configured mirror returning HTTP 404 for `node-addon-require-builtin@0.1.6`. **CI qualification has not executed.** No published-artifact compatibility, live Desktop activation, OAuth or real model-call success is claimed. The [official-first matrix](./docs/official-first-016-alpha2.md) records exact official sources, support scope, retained gaps and retirement triggers rather than assuming missing parity from feature names.
87
-
88
- ### Alpha.26 official DSH 0.2.0-rc.1 adaptation
89
-
90
- The historical source target was official tag `dsh-v0.2.0-rc.1`, commit `4878cdabd87d4041bdaff61d04c966883b9fd07a`; the independently verified Windows Desktop release feed also reported `0.2.0-rc.1` (SHA-512 `hPdqeajEHTXYEUOkOX7q1FflGU9jyOkRVDNSPxIDWhNihWjiVal0gtuT5zvtBRlX29VhqfHL31q4owGNp4ee6A==`). Its unchanged-source runtime fixture covered the plugin's OAuth credential seam, account-owned model discovery and native adapter, reasoning/streaming/tool behavior, Web search, Remote codecs, Session context and strict schema compatibility; the published npm fixture independently checked package/class identity and synthetic transport. Those tests did not validate the 0.2.0 Desktop package set, Electron loading, live sign-in or profile upgrade.
91
-
92
- ### Alpha.39 official DSH 0.2.0-rc.2 adaptation (#170)
93
-
94
- Official tag `dsh-v0.2.0-rc.2` at `639ed015397290b3745d163aafe02ffee4aa3f84` is the sole admitted target. Its published adapter now depends on pi-ai `^0.87.1`; this plugin pins `0.87.1`, updates the native transcript-based request estimator and keeps the same public stream instance. The install preflight uses public `bundlePatchPaths` to compose ordered array patches from rc.2 bundles, still rejecting invalid manifests and unsupported routing. Signed Windows Desktop `dsh/desktop-runtime.json` (SHA-256 `29561160fb7825c3968708f5700564f4381a976985fc270ee5dcc42e5a0a3a54`) records 287 shared packages; declared peers match the rc.2 host graph. The [official-first rc.2 review](./docs/official-first-020-rc2.md) records public API decisions and remaining limits. Source, package, and Desktop peer evidence do not prove loaded plugin behavior, Copilot OAuth, or live model calls.
95
-
96
- Installation does not reconcile stale profile-local copies of DSH host peers. Before an upgrade, inspect the effective dependency resolution against the Desktop's shipped package versions; an older local `@deepseek-ai/dsh-authorization` can shadow the bundled service, fail Core peer validation and disable authorization-dependent account plugins. This package does not remove or replace such files. Use the separate approved Desktop repair procedure and verify effective module resolution before attributing a startup failure to this plugin. See [the 0.2.0 official-first review](./docs/official-first-020-rc1.md).
97
-
98
- ## Copilot usage beside the composer
99
-
100
- Alpha.33 corrects alpha.32's missing selector argument to the public Session hook, which could retire the registered usage entry after a render failure. Positive eligible-Session rendering must be verified separately from installed-version receipts or signed-out absence. The fix is plugin-only; no Core patch or automatic restart is included.
101
-
102
- For a Session using `github-copilot` or `github-copilot-preview`, an optional **Credits** control shows the signed-in account's reported billing-cycle usage and remaining allocation. It uses the public composer dock, leaves the native Context meter unchanged, and disappears for other providers. Open it for budget, reset time, freshness and refresh details. Unsupported Client slots disable only this optional control.
103
-
104
- Alpha.35 matches native secondary statistics typography and pill spacing. A host with a shared wrapping dock can place Credits directly after Cache hit and wrap entries when space runs out; older hosts retain their own layout. A reading wider than its allocated width is ellipsized, with its complete text retained for accessibility, the tooltip and independent details. The plugin does not move native DOM or replace statistics; [responsive layout ownership and evidence limits](./docs/copilot-usage.md#responsive-statistics-presentation-alpha35) remain explicit.
105
-
106
- Numbers are not token estimates: legacy accounts show **Premium requests**, a shared organization pool may provide used credits without any individual remaining balance, and absent data is **Unavailable**, never zero or unlimited. Failed refreshes may retain explicitly **last-known** same-account values; authentication/account invalidation removes them. Per-session credits are not displayed because the supported native adapter does not expose complete provider-reported credit accounting through the public plugin contract. They are not inferred from account deltas. Reset time is shown only when a valid supplier-reported next-reset date is later than the snapshot observation; missing, invalid, zero or elapsed reset metadata is omitted without discarding otherwise valid account amounts.
107
-
108
- The Host reads GitHub's internal quota endpoint using the existing canonical OAuth grant, with bounded requests and cache lifetimes; no new sign-in, credential store, model transport, settings write or model switch is added. This internal endpoint can change or deny access; enterprise custom endpoints are not guessed. The public enterprise/organization metrics reports are not used as a live personal balance service. See [data semantics, lifecycle and evidence limits](./docs/copilot-usage.md). Synthetic tests do not establish real account access, billing freshness or installed Desktop activation.
26
+ Actual published Client components with synthetic models in an isolated browser. These screenshots demonstrate presentation, not live sign-in, model availability, search or loaded Desktop state.
109
27
 
110
28
  ## Install and sign in
111
29
 
112
- The commands below target the package version `0.4.0-alpha.88`. Versioned URLs describe the intended release artifacts, not proof that publication or local activation has completed; use them only once that Release and its checksums are available. Install into the profile you use (replace `web` when targeting another profile):
30
+ Use the exact verified release, the intended profile and an authorized package source. No `copilot2api`, external gateway, pasted GitHub token, placeholder key or separate `dsh-web-search-provider` is required.
113
31
 
114
- Before installing/updating, unpack the **checksum-verified** archive into a temporary directory and run its read-only composition preflight (replace all paths with absolute paths for the intended profile):
32
+ Before installation, extract the checksum-verified archive and run its **read-only composition preflight** with absolute paths:
115
33
 
116
34
  ```sh
117
35
  node package/scripts/check-search-composition.mjs --profile-dir /absolute/profile --home /absolute/DSH_HOME --install-anchor /absolute/dsh/package.json
118
36
  ```
119
37
 
120
- Include any launcher patch files with repeated `--patch /absolute/file` arguments. Require `supported: true`; otherwise do not install the routing bundle. The preflight rejects custom, disabled, nested or already-isolated web-service layouts, reserved routing collisions, and a nonempty disposable `cordis.yml` root before any mutation. It parses through public Core APIs and never boots plugins, resolves credentials or rewrites configuration. DSH 0.2's normal `composeProfile()` also rewrites that root to `[]`, including on Desktop's resolved-profile path; preserve and resolve any nonempty root before attempting install/startup. **`dsh plugin add` does not automatically run this preflight.** It is a required installer/operator step, not a universal compatibility guarantee. Also check that effective profile dependency resolution does not let a stale profile-local DSH peer shadow the host's bundled `0.2.0-rc.2` service; the plugin installer does not reconcile old host modules.
38
+ Supply any launcher patches with repeated `--patch /absolute/file`. Require `supported: true`: the install command does not automatically run this preflight. Unsupported/custom Web composition and `NONEMPTY_DISPOSABLE_PROFILE_ROOT` are stop conditions. Preserve a nonempty root and review whether public composition reconstructs it before any separately approved normalization; never assume it can be cleared. Check effective shared-peer resolution so stale profile-local packages do not shadow Desktop's official peers.
121
39
 
122
- For a standalone named profile only, the supported CLI command is:
40
+ For a **standalone named profile**:
123
41
 
124
42
  ```sh
125
- dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.0-alpha.88/dsh-github-copilot-0.4.0-alpha.88.tgz
43
+ dsh plugin --profile web add https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.0-alpha.92/dsh-github-copilot-0.4.0-alpha.92.tgz
126
44
  ```
127
45
 
128
- For Desktop profiles, use Desktop's native package manager with `dsh-github-copilot@0.4.0-alpha.88` after the same preflight and only when registry access is permitted. Desktop reserves its `desktop` profile from CLI management. If registry access is blocked or unavailable, stop; the documented [controlled offline CLI procedure](./docs/npm-distribution.md#controlled-offline-cli-maintenance-for-standalone-profiles) is not a Desktop workaround.
129
-
130
- Then open the Models UI listed above, find **GitHub Copilot**, select **Sign in**, and complete the GitHub device-code flow. Plugin installation changes the selected profile; activation follows that profile's normal reload/restart policy.
131
-
132
- ### User authorization flow
133
-
134
- 1. Open **Settings → Models** and find **GitHub Copilot**. Account controls appear in an existing configured canonical `github-copilot` card, suppressing the separate footer controller; otherwise the footer fallback (or old-Core **Settings → GitHub Copilot**) remains usable. An already-signed-in account automatically ensures missing/idle/stale/error/loading metadata once on opening; fresh ready metadata makes no discovery request. No native provider or manual refresh is needed for the normal discovery flow.
135
- 2. Select **Sign in with GitHub**. The authorization area expands automatically with a prominent one-time code, **Open GitHub verification page**, **Copy code**, and **Cancel sign-in**. No extra **Manage** click is needed. Recognized Desktop v1 hosts use their existing same-window external-navigation handoff to the system browser instead of a popup; web hosts retain a new tab. If nothing opens, select and copy the displayed verification address into your browser. The plugin cannot confirm that the system browser opened.
136
- 3. Copy the code, open the verification link, and complete authorization in your own GitHub browser session. Copy success/failure is announced accessibly; manual copying remains available. Never paste a GitHub token into DSH.
137
- 4. DSH polls only while authorization is in flight. Successful explicit **Start sign-in** (the **Sign in with GitHub** button), including a UI account switch, forces exactly one bounded discovery after immediate or polled success. The code and verification link clear, the automatic authorization area closes, and the account shows **Signed in** and **Manage**. Manually opened details stay open; cancellation clears the old code. Failures show only a safe stage diagnostic: `COPILOT_AUTHORIZATION_BEGIN_FAILED` means authorization did not complete; `COPILOT_ROUTE_REPAIR_FAILED` means authorization completed but local route repair failed. A failed begin can also show a fixed **latest observed milestone** (`AUTHORIZATION_REQUESTED`, `INTERACTION_PROMPT_OBSERVED`, or `INTERACTION_NOTICE_OBSERVED`). It reports only progress observed through the public interaction callbacks, not the failing operation or a credential commit. The latter route-repair code retains authentication and directs you to **Manage → Repair model configuration**. These codes and milestones do not explain or resolve the underlying sign-in failure; raw provider details are never displayed or logged.
138
- 5. Choose an accepted model under **GitHub Copilot** (stable route ID `github-copilot-preview`). Normal opening/use maintains metadata without requiring **Refresh models**. **Manage → Model preferences** provides search, **All / Enabled / Excluded** filters, and immediate single-model **Exclude/Restore** actions. Excluded exact IDs disappear from the managed picker and new Auto pools; temporarily absent excluded IDs remain restorable. A selected model can be excluded: already admitted native turns keep their model through tool steps and retries, while new turns cannot use it. A fixed selection is not automatically replaced; select another model before its next turn. Exclusion-driven parent renders retain the mounted account controller and metadata without restarting status checks or model discovery. No exclusion rewrites history, changes selection, or mutates account entitlement. **Discovery details** contains metadata diagnostics separately from the model controls. Legacy configuration diagnostics and explicit repair remain available when needed, and the [migration guide](./docs/single-route-migration.md) remains in the documentation.
139
-
140
- If preferences cannot be read, available account rows remain visible but read-only. Unknown exclusion status is not shown as enabled. Named diagnostics distinguish unavailable/invalid settings; **Retry** reads Host status only, without refreshing OAuth or discovering models. Failed writes retain the last confirmed state. Individual saves return only strict model preferences, without scanning sessions or reading credentials/full account status. Exclusions use native live, hidden Config fields and path-level settings CAS, so saving does not remount the plugin. Native configuration queue, file-lock and Loader synchronization may still delay confirmation; no optimistic success is shown before persistence completes. Account model discovery alone does not prove that preference settings or model transport are working.
141
-
142
- ### Auto model routing
143
-
144
- **One setting for subagents:** in the plugin configuration page, turn on **Subagent models → Follow parent model** and Save. It applies to existing and new supported Copilot subagents and Team mates, without Session IDs or per-child setup. A fixed parent model takes effect on each child's next turn; Auto inherits the same preference but chooses using each child's own context. Running turns stay unchanged and child-owned manual selections still win. Default Off (`github-copilot.followParentModel: false`). Enabling replaces creation-time model choices; it never clears explicit child selection history. Legacy `parentModelFollow` bindings remain active independently and are not migrated. See [requirements, mockup and boundaries](./docs/parent-model-follow.md). Native Team creation labels are not rewritten.
145
-
146
- The managed GitHub Copilot group includes **Auto · Balance**, **Auto · Efficiency**, and **Auto · Intelligence** as flat virtual picker rows. They share one account-verified eligible pool after applying exact model exclusions and input-fit filtering. Routing now uses the account's **Powerful / Versatile / Lightweight** classification and conservative contextual task assessment, not reasoning effort or context capacity as a quality ranking. Even Intelligence uses Lightweight for an isolated greeting; short difficult reasoning stays Powerful. Unknown tasks prefer Versatile with Balance/Efficiency and Powerful with Intelligence. Suitable previous-model continuity wins before stable equal-weight category tie-breaking. Missing categories and fallback remain explicit. If no candidate fits, Auto preserves largest-input-budget native recovery; if none is eligible, it fails closed. One decision stays fixed across the turn's tool steps, retries and compaction. Images retain their independent verified admission. Search, actual model/Usage and histories remain native-owned. See the [task policy and limits](./docs/auto-task-routing.md) and [parent-inheritance boundary](./docs/automatic-model-routing.md#requested-parent-to-child-selection-inheritance). Core's picker remains flat.
147
-
148
- Local rules also recognize isolated fenced JSON/CSV conversions and line sorting as routine, and exact continuations of a recent explicit difficult task as complex. New tasks break continuation inheritance. `uncertain` means demand is unresolved; insufficient local evidence is not proof that the conversation has no context.
149
-
150
- Semantic assessment is **enabled by default**, using one bounded auxiliary model call **only when local demand is unknown**. Set `github-copilot.autoSemanticAssessment: false` to opt out; existing explicit false settings remain respected and are not migrated. Its bounded input prioritizes current and preceding user requests without dismissing omitted context. It adds latency and supplier charges not included in native chat Usage. Failure or omitted assessment context never means the task is simple; it retains conservative local evidence. This user-authorized rollout is not labeled-corpus calibration or GitHub's private Auto service.
151
-
152
- Auto's auxiliary semantic assessment uses a single 8-second deadline and 128-token compact JSON output budget. Its existing selection disclosure can show the classifier ID, total elapsed/budget, adapter-start and first-text milestones, native finish and result validation. A timeout leaves task demand unknown and uses the preference-policy fallback; it is not a main-answer timeout or evidence the task is complex. These content-free, ephemeral observations are not HTTP timings or native Usage, and do not yet prove a speed improvement. Reasoning-off is requested only when current supplier metadata and the public prepared model both support it; otherwise native policy stays intact. See [assessment boundaries](docs/auto-task-routing.md#auxiliary-latency-evidence-and-bounded-output-272).
153
-
154
- Completed Copilot replies show **Auto (preference)** and an expandable recorded reason after native Usage/time, or **Manual** only with captured explicit fixed-selection evidence. A successful read with no retained evidence shows **Selection unknown**, never a guess from today's picker. Failed reads instead show **Selection unavailable → Retry**; retry only reads the same turn's record, not another model request. Explicit Session/turn arguments remain stable inside a session-bound Chat, avoiding the former native gateway argument-count failure on fresh master/lead replies. There is no repeated model name or separate Model details button: actual model display stays in native Usage, whose missing-route behavior is unchanged. New selection records are bounded and Host-local; restart, disposal or eviction loses them. Compatible historical Auto records remain readable. The public assistant-actions slot supports internal wrapping only, not changing the native fixed-height row.
155
-
156
- The information button explains this turn's captured task/category choice, fallback and continuity/tie-breaking. Supporting assessment facts sit in a disclosure; older records without details explicitly say that the reason was not retained.
157
-
158
- The selection footer mounts independently of optional conversation projections. Native completed-turn evidence lets retained Auto/Manual records remain visible if those projections are unavailable; incomplete attribution stays explicit in the information disclosure. This does not restore lost selection records or prove that an installed update is loaded in Desktop.
159
-
160
- Alpha.79 also fixes the Host half of selection reads: the retained decision store is now exposed through the public Remote service binding and method marker, not a plain provided object rejected by the native gateway. Actual Client-to-Host roundtrips cover both source discovery and strict descriptors. This complements alpha.76's Client argument fix; it does not recover lost records or change Auto model selection.
161
-
162
- Auto input fitting uses Core's native TokenMeter for every entered message, including reasoning and tool history, plus its current-surface/tool-envelope floor. Missing or invalid native measurement fails Auto explicitly rather than silently selecting from an undercount. This prevents the known estimation error but does not rescue history whose summary input already exceeds every eligible model's budget; see [compaction limits](./docs/copilot-compaction.md#existing-oversized-history-and-recovery-limits).
163
-
164
- Alpha.63 fixes a missing footer on official Desktop by supporting its native ChatNodeStore rather than requiring a JavaScript Map. It does not change Auto selection or recover reasons already lost on Host restart.
165
-
166
- **History loading error mentioning `github-copilot/auto-model-decision`:** alpha.54 prevents new incompatible events; upgrading or restarting does not repair existing logs. The [explicit recovery command](./docs/automatic-model-routing.md#recovering-affected-histories) validates a detached copy with the official codec and retains an exact original backup. Applying it to a live Session requires separate approval and stopped writers; no automatic history migration occurs.
167
-
168
- If a conversation disappeared after using the original alpha.54 recovery utility, use the corrected source-checkout utility: it restores the independent zstd header frame required for Desktop discovery, without deleting or changing logical history rows. Its report includes `reframedHeader`; logical event validation alone was insufficient in the original utility.
169
-
170
- When a virtual Auto preference is inherited as the default in a new Session, the first resolved request records that exact preference as the Session's durable selection while its request header still records the real model. Later turns retain the preference instead of locking to that header; explicitly selecting a real model switches to the fixed route. The existing `auto` ID remains Balance. Existing Sessions with a concrete recorded choice are not migrated by a change to the global default.
171
-
172
- One shared account-state owner survives transfer while another eligible surface remains mounted. Last-surface unmount or nonoverlapping declaration replacement stops polling; a later mount reads status and separately ensures metadata if needed. Status reads and details toggles themselves remain network-free, but opening Models can discover missing/stale signed-in metadata. Background credential/reset notifications clear Client state and read status rather than forcing discovery on every token event; the next open/use ensures metadata.
173
-
174
- Concurrent status retries join the compact account's pending read. An obsolete read must settle before that controller reads for a new lifetime; Remote has no cancellation contract, so a permanently hung read still needs connection recovery. Authorization polling backs off from 500 ms to 1 s and then 2 s (32 reads in the first minute with immediate replies). Status errors stop polling and expose an explicit retry; sign-in, cancellation and other mutations are never automatically replayed. Credential invalidation and the bounded initial metadata check retain their existing behavior. These are plugin request-pressure safeguards, not a fix for incompatible Agent presets or Core-wide request scheduling.
175
-
176
- The public provider-card slot is additive: it cannot replace Core's **Edit/Delete** controls. The native editor remains available, but normal plugin discovery needs no manual model definitions. Embedding the controls does not merge `github-copilot` with `github-copilot-preview`, remove configuration, rewrite history or change model selection. If controls are missing from both eligible card and fallback, verify the active profile and loaded Host/Client version.
177
-
178
- The subdued **Updated … ago** timestamp uses the last successful model snapshot, not the time the page opened. Its tooltip and accessible label provide the complete local date/time and time zone. Display text updates locally as time passes without fetching status or models. Refresh/failure retains the previous successful time while that account's display evidence remains valid; sign-out/account invalidation clears it. Missing or invalid timestamps are hidden; future clock values use an absolute date instead of a misleading age. Manage does not repeat the timestamp.
179
-
180
- ### Agent and automation flow
181
-
182
- Agents should treat the browser authorization as a human handoff, not as a token-acquisition task:
183
-
184
- 1. Install the pinned release into the requested profile and restart/reload that profile when required.
185
- 2. Direct the user to **Settings → Models → GitHub Copilot → Sign in with GitHub**.
186
- 3. Tell the user to open the displayed verification URL and enter the displayed one-time code. Do not ask for, read, copy, log, or persist the user's GitHub token.
187
- 4. Wait for the user to complete the browser step. Do not repeatedly start new authorization attempts while one is in flight.
188
- 5. Confirm **Signed in** and inspect the automatic discovery result before asking the user to choose a model. Already-signed-in Models opening ensures missing/stale metadata automatically; fresh ready cache makes no request. Use visible **Retry** for errors or **Manage → Refresh models** for an intentional forced update, not routine setup. Status alone does not discover, and login, metadata and successful model calls remain separate evidence.
189
- 6. Use **Sign out** only when the user explicitly asks to disconnect the account. It deletes the Copilot credential record but preserves route settings.
190
-
191
- GitHub Releases and npm are the default distribution channels for each new version, using the same verified tarball. Pin the version and verify the evidence for the channel used: Release `SHA256SUMS`, and npm `dist.integrity` when installing from npm. Install into Desktop profiles only through Desktop's native package manager; the reserved `desktop` profile is not managed by the DSH CLI. After npm publication is verified, the native manager accepts `dsh-github-copilot@0.4.0-alpha.88`, not a URL or file. Controlled offline CLI maintenance is documented for standalone named profiles only, never Desktop-managed profiles. Follow the mandatory search-composition preflight, backup, single-writer and post-install checks; do not bypass a corporate registry ban, disable TLS verification or restart without separate approval. See [distribution and publication requirements](./docs/npm-distribution.md).
192
-
193
- **Desktop package-update and Web-lifecycle boundaries:** the native manager's `restart-required` result for an in-place installed-package upgrade is authoritative and distinct from plugin-bundle removal/re-add behavior. Any disable, removal, or upgrade involving service recomposition (`cordis.patch.yml`, such as the isolated routed web search layer) may leave dependent entries pending until Desktop is fully exited and cold-restarted. If the native manager reports only entries such as `pending (waiting for service: web)` while this plugin already shows **Off**, the requested disabled state has been persisted; cold-restart Desktop instead of repeatedly toggling or reinstalling the plugin. Other errors are not covered by this guidance and must still be investigated.
194
-
195
- A pinned public Loader characterization records that, with the exact bundle IDs and an active provider, staged first-add can leave global `ctx.web` and an active consumer unavailable even after the routed row is enabled. A separate complete cold-loaded Web-routing composition works initially but loses global `ctx.web` and its consumer on removal and remains unavailable after re-add/reconciliation. An earlier temporary harness used alias IDs and no active provider, so its successful final add was not equivalent to the exact bundle fixture. These are synthetic Loader observations; the native warning above is operational evidence, not proof that every pending diagnostic is harmless or that restart recovery has been exercised by the fixture. Do not use bundle removal/re-add as a hot-upgrade workaround. See [the rc.2 lifecycle characterization and bounded public-API finding](./docs/web-lifecycle-rc2.md).
196
-
197
- **Host TLS trust boundary:** this plugin adds system roots only to its quota request; it does not change Desktop process trust, sign-in, models, search or other requests. A correctly environmented Desktop launch previously restored quota, while a recovery launch without the inherited process variable caused it to regress despite HKCU retaining `NODE_USE_SYSTEM_CA=1`; normal relaunch restored it. The quota fix avoids this launcher dependency after an authorized upgrade and reload. No live Desktop profile, trust store or process is changed by the package alone.
198
-
199
- No `copilot2api` process, external gateway, placeholder API key, pasted GitHub token, or separate `dsh-web-search-provider` installation is required.
46
+ For **Desktop**, its native package manager accepts `dsh-github-copilot@0.4.0-alpha.92`. Official rc.2's **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.
200
47
 
201
- ## What this package owns
48
+ After an approved reload/restart:
202
49
 
203
- - A conditional authorization-service fallback for profiles such as rc.2 that omit Core's service.
204
- - Account controls embedded in an existing canonical Models provider card, with a shared-state footer/old-Core section fallback, Client-safe Remote descriptors, and Host authorization controller.
205
- - Strict normalization of pi-ai's provider-owned Copilot OAuth grant.
206
- - Preservation of intentional canonical-route absence, compatibility repair of existing legacy profiles, and restoration of verified old overrides. It never automatically removes user profiles or replaces Core's model list from another pi catalog.
207
- - A data-driven account-model route using pi `0.85.1`, authenticated Copilot metadata, and the published native DSH adapter. New model IDs do not require a model-specific code patch when their advertised protocol and capabilities are supported.
208
- - Direct provider-hosted search through inline agent-loop interception and a Responses-only `ctx.web` provider.
50
+ 1. Open **Settings → Models → GitHub Copilot** and select **Sign in with GitHub**. No native provider or manual model definition is needed.
51
+ 2. Copy the displayed one-time code and complete GitHub's device flow in your own browser. Desktop uses its system-browser handoff; a selectable verification URL remains available if it does not open.
52
+ 3. Wait for **Signed in** and account discovery, then choose a model. Opening Models and normal use ensure missing/stale metadata; **Manage → Refresh models** is an intentional forced refresh, not routine setup.
209
53
 
210
- DSH Core continues to own model selection, sandboxing, tools, attachments, and other providers. `@deepseek-ai/dsh-llm-pi-ai` owns the Copilot adapter, catalog, OAuth method and grant format, token exchange, refresh, and normal model transport. Credentials remain Host-only.
54
+ Sign-out requires an explicit action and removes only the Copilot credential record, not route settings. Upgrades preserve existing native Copilot profiles; two groups can remain until [explicit single-route migration](./docs/single-route-migration.md). Installation never migrates conversations or defaults.
211
55
 
212
- ## Retired planner / executor experience
56
+ **Desktop lifecycle:** disable/remove/upgrade can require a full cold restart for Web service recomposition. If the plugin is Off and the native manager reports only `pending (waiting for service: web)`, do not repeatedly toggle or reinstall. Obtain restart approval and follow [rc.2 lifecycle guidance](./docs/web-lifecycle-rc2.md).
213
57
 
214
- Since `0.4.0-alpha.34`, **Model roles**, its legacy **Copilot · Model roles** fallback, the planning/execution selectors and the dedicated-session creation button are retired (#158). Use ordinary Sessions and Core-owned model selection/subagents; this plugin adds no replacement role UI or automatic model mapping.
58
+ ## Auto and model preferences
215
59
 
216
- Parent-model → subagent-model rules and their native settings UI are separate, still-pending work in [Core PR #95](https://github.com/cloga/deepseek-harness/pull/95). Removing this plugin UI does not make those rules available on older or currently installed Core versions, and does not depend on that PR shipping.
60
+ All three Auto preferences use the same account-verified, non-excluded eligible pool. Hard input/image capability checks run before soft task preferences.
217
61
 
218
- Existing dedicated histories and captured policies remain supported through the retained compatibility runtime, without conversion or model substitution. Saved legacy settings are not deleted or migrated. Legacy Remote clients receive `DUAL_MODEL_RETIRED` for settings writes and new roots; recovery of an already-created matching request remains available. See [retirement and compatibility boundaries](./docs/dual-model.md). Publication and runtime activation are separate from this source change.
62
+ | Task demand | Efficiency | Balance | Intelligence |
63
+ |---|---|---|---|
64
+ | Proven simple | Lightweight | Lightweight | Lightweight |
65
+ | Routine | Lightweight | Versatile | Powerful |
66
+ | Complex | Powerful | Powerful | Powerful |
67
+ | Unknown | Versatile | Versatile | Powerful |
219
68
 
220
- The plugin's Web search-provider dropdowns use paired application-theme surface/text colors, including unavailable options, so dark mode does not leave pale labels on a white popup. Hosts without those theme tokens use readable system-color fallbacks. Appearance changes do not save or replace search selections; Core owns its model picker.
69
+ Categories come from authenticated supplier metadata, not model names, context capacity or measured quality rankings. A suitable previous model is retained; otherwise the plugin uses stable equal-weight allocation within the eligible category. Category fallback is explained explicitly. One real model stays fixed throughout each admitted turn's steps, retries and compaction.
221
70
 
222
- ## Shared account, independent sessions (V3)
71
+ 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.
223
72
 
224
- One Host-owned Copilot account supplies many account-discovered models. Each explicitly selected or history-backed Session keeps its own model context: search for Session A uses the captured effective request-header/config of initiating Session A (or explicit request `GenerateOptions`), not Session B's choice or a future global default C. Search plans are cached per owner so different-model A/B requests do not reuse or cancel each other's plans. Account metadata remains shared; capability/probe and credential checks still apply.
73
+ 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.
225
74
 
226
- Cold Chat picker and `/model` `listModels()` calls ensure the shared managed source without opening Settings first. An actual managed search also performs a non-forcing shared ensure before deriving route/model facts. These paths use the same 24-hour maximum TTL and failure cooldown below; no new global current-model or search-status card is added.
75
+ **Model exclusions** remove exact IDs from the managed picker and new Auto pools. Selected models are excludable: an admitted turn keeps its model, but a new turn/direct request cannot use an excluded ID. A fixed selection is never silently replaced. Unknown preference settings leave controls read-only, not falsely enabled. Saves use narrow native CAS without credentials, discovery or Session scans.
227
76
 
228
- **Native selection limit:** public Core `session.selectModel` also saves the future global default. This plugin does not replace that behavior: selecting one Session does not rewrite other selected/history-backed Sessions, but empty Sessions without a selection may still inherit the changed default. Do not promise every unselected Session is immutable.
77
+ See [task assessment](./docs/auto-task-routing.md) and [routing contracts, exclusions and history recovery](./docs/automatic-model-routing.md).
229
78
 
230
- **Native Add is warning-only:** the new Copilot provider draft warns that saving a native profile adds another real model group, not a second account. Published additive APIs cannot veto native Add or disable **Save**. The native editor remains; this warning is not absolute prevention or enforcement of a single route. Actual one-managed-route deployment requires the separately approved [Ops migration](./docs/single-route-migration.md) after release, not hidden groups or automatic configuration removal.
79
+ ## Subagents and search
231
80
 
232
- ### Read-only Ops migration readiness (alpha.9)
81
+ ![Current parent-following and Web search controls](./docs/images/copilot-search-routing.png)
233
82
 
234
- No-argument Remote `githubCopilot.migrationStatus()` provides fresh live evidence for a separately authorized maintenance operation. Generic `session/list` data can be stale and plugin inventory alone does not identify the loaded version. This separate strict result reports the loaded plugin build's `plugin.name`/`plugin.version`, `protocolVersion: 1`, `observedAt`, capability flags (`agentsList`, `sessionProjections`, `settingsCas`, `providerRegistry`, `defaultSelection`) and completeness flags for sessions/default/routes. Missing capability or incomplete required selection/route evidence is unknown, not permission to migrate. An idle Agent's `activeRequestSelection: null` is expected.
83
+ **Follow parent model** defaults Off and applies profile-wide to supported native children/Team mates of managed Copilot parents. Fixed parents apply on each child's next turn; Auto parents pass the exact preference while each child evaluates its own context. Child-owned explicit selections win. Running turns, roots, forks, other providers and dedicated legacy policies are not rewritten. Existing legacy bindings remain independent when the broad switch is Off. [Parent-following contract](./docs/parent-model-follow.md).
235
84
 
236
- For each live Agent Session it reports `effectiveSelection` and `selectionSource`: pending model projection first, then recorded request-header config, then the current default only for a genuinely empty Session with known projection state. Running Agents also have a separate `activeRequestSelection`: the latest recorded request header, **not proof of an in-flight LLM call**. Route flags distinguish effective native configuration (`nativeConfigured`) from actual native/managed registration. No credential data or full settings/history is returned.
85
+ **Web search** has two ordinary controls: primary provider and final fallback. Auto follows the initiating Chat provider when a corresponding search registration exists; fixed choices stay fixed when Chat changes. At most one distinct final fallback is tried, with possible charges disclosed. Cancellation/account-proof loss never authorizes a fallback.
237
86
 
238
- The call does not invoke authorization status or model discovery, access credentials/network, or mutate settings/Sessions. It adds no normal UI or global current-model/search-status card. The seven ordinary authorization Remotes and their codec are unchanged; this eighth Remote has its own `GitHubCopilotMigrationStatus` codec.
87
+ Copilot search requires current account/protocol evidence and capability proof. Fixed/fallback Copilot preserves a nonempty legacy `searchModel`; otherwise it considers at most three account-owned Responses candidates. The final user query is sent once, never replayed across those candidates. A saved choice or successful Chat request does not prove hosted search. [Search routing and composition](./docs/session-search-routing.md).
88
+ The legacy `github-copilot.searchFallback` setting applies only to its separate
89
+ canonical inline path; it is not a third routed fallback. Fetch is unchanged.
239
90
 
240
- **Limits:** `historyScope: live-agents-only` excludes cold stored histories; the operator must acknowledge that those conversations may need explicit model selection later. Build identity and structural capability self-reports do not attest all Desktop/Core bytes. This is not an atomic cross-namespace snapshot; recheck evidence immediately before CAS. The planned `tools/migrate-copilot-managed-route.ps1` in `cloga/dsh-windows-ops` is a separate, post-release **config-only** maintenance command: v1 does not write Session/default selections, inspect cold history, install the plugin, restart DSH, or certify a full Desktop baseline. Its publication/installation and any live migration remain separate evidence, not a claim made by this document.
91
+ ## Usage, request limits and recovery
241
92
 
242
- ## Authorization and route behavior
93
+ **Credits** shows validated account billing-cycle data, not context tokens or Session cost. Stale, pooled and unavailable readings keep their real meanings. The quota request alone uses Node plus system CA roots with TLS verification enabled; it does not change trust for sign-in, models, search or Desktop globally.
243
94
 
244
- `llm-pi-ai` registers the OAuth method; the authorization service orchestrates the interaction; this package contributes the UI/Remote controller and route reconciliation. Core supplies authorization on rc.1. On rc.2 profiles that omit it, this package mounts its runtime dependency and reuses any provider already present.
95
+ Native context occupancy and Turn Usage remain Core-owned. The plugin forwards usage unchanged, including zero samples on failure/cancellation. Its separate context disclosure reports historical evidence, never current occupancy or a replacement `0%` meter. **Turn Usage incomplete** explains missing samples/lifecycle and recorded local blocks; it invents neither zero usage nor partial totals. Cancellation can retain usage, but does not guarantee a final supplier receipt. [Usage boundaries](./docs/copilot-usage.md).
245
96
 
246
- New installations without a canonical profile use the account-discovered route displayed as **GitHub Copilot**, with the unchanged actual ID `github-copilot-preview`. Account controls use an existing configured canonical provider card when mounted; otherwise the footer/old-Core section provides the same flow. Exactly one bounded `/models` discovery follows this UI user's successful Start sign-in, and **Refresh models** remains available for explicit updates. Neither UI placement nor discovery changes a session, default, history, credential ownership or route configuration automatically.
97
+ Managed request admission preserves truthful input/output limits and native transactions. Automatic pressure uses the initiating Agent's actual bound compaction service; a preset without an engine cannot borrow global recovery. Already oversized manual summaries require separately selected [manual recovery](./docs/manual-compaction-recovery.md), not a promise that preventive admission can rescue them.
247
98
 
248
- An existing Core-owned `github-copilot` profile is preserved, so upgrades can still show two real routes until the user completes [explicit single-route migration](./docs/single-route-migration.md). This is not a UI filter or facade. After actual removal of the reviewed legacy profile, both the composer picker and `/model` receive only the managed Copilot group. Old conversations remain stored unchanged, but a paused conversation still selecting the removed canonical route needs an explicit managed model selection when resumed.
99
+ For continuing steps with a committed managed route and no pending model change, known input pressure is reduced before Core opens the next model attempt. Successful native reduction avoids an unsampled local-pressure attempt that would make whole-turn Usage unavailable. First steps, pending selections, new fixed-prefix growth and final hard-budget refusals retain existing admission; this does not repair historical totals or invent missing usage.
249
100
 
250
- The managed route selects Responses, Chat Completions or Anthropic Messages from advertised `supported_endpoints`, not model names or a per-model allowlist. An existing pi protocol is retained only if the server also advertises it. New IDs with complete supported metadata can therefore work without another plugin release. Missing endpoints, disabled policy, unsupported protocols or malformed limits produce explicit diagnostics rather than guesses. Updating pi or seeing a model ID in its catalog does not itself prove the protocol is correct.
101
+ For exact verified `408 / user_request_timeout`, diagnostics describe bounded request composition and observable timing, not a proven payload limit or root cause. Bytes are not tokens; timing is not upload duration. Small, image-free requests can also time out. Follow [request-budget and timeout guidance](./docs/copilot-compaction.md); do not automatically trim history, disable proof, switch models or add retries.
251
102
 
252
- ### Automatic model freshness
103
+ HTTP/SSE liveness separates five-minute byte idle from bounded ten-minute assistant-output silence by default. Consumer work is excluded; WebSocket/explicit `auto` stays native-only. This does not cure supplier HTTP 408s. Images are admitted against actual native-projected MIME evidence; the plugin does not own conversion or infer format support from filenames. [Image compatibility](./docs/image-input-compatibility.md).
253
104
 
254
- Opening Models separately calls non-forcing Remote `githubCopilot.ensureModels()` once for signed-in metadata that is missing, idle, stale, errored or loading. Reopening an error may retry after the shared failure cooldown, without a same-mount retry loop; loading joins the existing Host flight to observe completion, not another network request. Fresh ready metadata does not fetch, and a genuine `unavailable` result with no models does not automatically retry. Normal model use also ensures freshness. The Host shares one in-flight discovery across callers, with no periodic refresh timer. The configurable defaults are **24 hours** maximum metadata reuse (`accountModelTtlMs: 86400000`) and a **5-minute** failure cooldown (`accountModelFailureCooldownMs: 300000`). Explicit successful UI sign-in/account switch forces discovery once; manual **Manage → Refresh models** and visible **Retry** remain available.
105
+ ## Settings and troubleshooting
255
106
 
256
- Last known same-account metadata may remain visible during TTL refresh/loading/error, but that display never authorizes a request. Credential/account/permission invalidation or proof expiry immediately revokes old request evidence; 24 hours is a maximum metadata reuse window, not an OAuth-token extension. Background credential/reset events clear Client state and read status, without forcing a discovery on every token event; next open/use ensures metadata. Status and details toggles themselves stay read-only and network-free.
107
+ Configuration is under `github-copilot`. Credentials, endpoint definitions and a static model catalog are not plugin settings.
257
108
 
258
- Before issuing a new metadata lease, the managed route uses a **5-minute-30-second preflight renewal threshold**: the native SDK's five-minute renewal window plus a thirty-second preparation margin. Metadata preparation consumes this margin; it is not a promise of that much validity remaining at the later lease or wire boundary. A reusable warm snapshot inside that window is retired, and the existing single-flight discovery resolves native OAuth and validates fresh metadata before preparing the model call. Other callers join that flight; it does not force past failure cooldown. Native `Models.getAuth()` receives the same minimum-validity budget, so an insufficiently renewed token fails before model dispatch rather than causing a refresh loop. Status reads remain network-free. This fixes a warm-cache path that could produce `OAuth auth derivation failed ... COPILOT_PREVIEW_METADATA_STALE` during the request's own refresh. The UI's **Updated** timestamp is metadata freshness, not the OAuth refresh timestamp. Credentials, permissions and endpoints are still checked; an already-prepared call that later crosses the renewal boundary or loses its proof still fails closed without automatically replaying the message. Hosted-search continuity/probe gates remain independent and may reject that attempt after a credential notification.
259
-
260
- A definitive `UNKNOWN_MODEL` result triggers one bounded metadata refresh, not a replay of the failed message or an automatic model switch. Generic HTTP/network failures are not guessed to mean an unknown model. Account/token/permission generation checks reject stale results; discovery never copies credentials or rewrites selections/history. A server-enabled new model may precede an older grant's cached ID list without rewriting that grant.
261
-
262
- **Plugin-only development boundary:** fixes in this project must use existing published public APIs and remain in the plugin. Do not patch Core source, installed binaries, `node_modules`, private runtime registries or shared upstream catalogs; do not make a new Core export or upstream Core PR a delivery prerequisite. Read-only inspection and isolated verification against unchanged pinned Core artifacts are allowed. If a stock API cannot support a requested feature, state the limitation and use a tested plugin-local alternative rather than changing Core. The authoritative policy and its machine checks are documented in [AGENTS.md](./AGENTS.md#plugin-only-implementation-boundary).
263
-
264
- Keep `llm-pi-ai` mounted: its OAuth method and the unique Host-only `llm-pi-ai/github-copilot` credential remain the authentication owner even without a canonical model profile. Do not sign out, remove the authorization plugin or copy credentials to hide a route. Model metadata from another dependency copy is not proof that actual Core supports that model or protocol.
265
-
266
- The plugin does not install a global Responses override or recreate a missing canonical profile on login, startup or token refresh. Existing canonical profiles retain their models, protocol, headers and custom fields; normal legacy reconciliation changes only `compat.supportsStrictMode: false`. Removing such a profile requires explicit migration, not an automatic upgrade side effect. The managed route obtains supported protocols from account metadata and does not fall back to a static catalog when discovery fails.
267
-
268
- Previously installed plugin overrides are restored only through a verified ownership journal, to the recorded original `api`/`models` values. Owned public headers are removed only if their values still match. User edits, ambiguous legacy markers and interrupted uncommitted writes remain conflicts rather than being forcibly adopted. Restoring an absent or empty model list returns Core to its default catalog; it does not mean an empty set of models.
269
-
270
- Temporary ownership uses a bounded version-2 journal: raw `api`/`models` preimage and postimage, prepared restoration target, namespace revision and process epoch. Only fixed public Copilot headers can enter the journal; arbitrary header values, credential payloads and custom model extras are never copied into it. Every write uses the namespace revision; current owned fields must still match the recorded values. Conflicts retain the journal and user edits instead of guessing ownership. Deleting a created profile additionally requires its complete raw shape to be plugin-owned, with no base profile, user additions or set secrets. Normal reconciliation merges actual raw user extras only, not schema defaults. These checks are conservative recovery safeguards, not an atomic transaction across settings namespaces and credential storage.
271
-
272
- **Upgrade boundary:** older backups without postimages/epoch are reported as `conflict`, not automatically adopted. Review the existing route and backup before explicitly migrating or removing the marker; do not reconnect or delete it to force ownership. A later invocation never replays a prepared but uncommitted activation/restoration automatically: Core revision counters reset on namespace re-registration as well as restart, and there is no public durable registration identity. Even matching recorded epoch/revision is not enough to prove ownership across that boundary. Steady owned postimages can begin a fresh restoration, and already-restored targets can clear their journal without replaying writes. An existing recorded restoration target that differs from the original preimage remains a conflict for explicit review; a new account model set never authorizes overwriting that evidence. Remove legacy connection fields explicitly during migration. Sign-out itself still deletes only the `llm-pi-ai/github-copilot` credential record and keeps route settings.
273
-
274
- ### Read-only status and explicit repair
275
-
276
- `githubCopilot.status()` and Host `describeGitHubCopilotProviderProfile()` only read stored state and plan legacy canonical changes. They do not write settings, refresh OAuth or test the network. Status separates configured authentication from legacy route states `ready`, `needs-repair`, `not-configured`, `conflict` and `error`. With valid login, `route: not-configured` normally means the optional canonical profile is absent: it is not a login error or a repair request, and it does not prove managed discovery is ready. `ready` only means the inspected canonical configuration needs no repair. Discovery and successful model/search calls remain separate evidence.
277
-
278
- Use **Repair model configuration** (Remote `githubCopilot.reconcile()`) for an explicit, revision-checked repair of an existing legacy profile or verified journal. It does not fetch a model list, create an absent profile, or force conflicts. Successful login, startup and auth-refresh reconciliation also preserve absence. Browser status polling is read-only; the UI separately initiates its one discovery when this user's Start sign-in succeeds. Deploy matching Host and Client bundles.
279
-
280
- Before a grant is persisted or reused, the Host normalizer rebuilds only pi-ai's documented `type`, `refresh`, `access`, finite `expires`, optional `enterpriseUrl`, and optional deduplicated `availableModelIds` fields into a fresh plain JSON object.
281
-
282
- ## Hosted search
283
-
284
- Auto-mode native search identity comes from the captured initiating Session's effective request-header/config or explicit `GenerateOptions`, never the future global chat default. Fixed/fallback Copilot search instead resolves its execution model inside the provider from current account metadata, while preserving a nonempty explicit `github-copilot.searchModel` override; it never borrows the Chat model or another Session's selection. Core `Agent.options` remains the activation seed: model selection overrides request/assembly, with effective config recorded in `Session.requestHeader().config` before tools. Without a proven initiating owner, traditional search is unavailable. After a new model is selected but before its next request, the previous header must not be treated as current prompt guidance. Per-owner plan caches keep different-model Sessions independent. An actual cold managed search first ensures the shared account metadata without force, then derives model facts and applies the existing account/protocol/allowlist/probe gates. Explicit marked `GenerateOptions` can still bind eligible inline requests, uncached when no owner is available, under all existing guards.
285
-
286
- **Credential-change limit:** an OAuth notification during initial lazy metadata discovery cannot distinguish the discovery's own token rotation from an external account change through current public status. That first search deliberately fails closed with `WEB_PROVIDER_UNAVAILABLE`, before probe/wire work. A later user/driver request may retry with refreshed credentials; there is no automatic retry or promise of seamless first-attempt refresh.
287
-
288
- - **Managed-route conversations:** `github-copilot-preview` uses the native adapter, not the custom inline wire; this preserves account evidence, replay and attachment handling.
289
- - **`github-copilot-hosted` through `ctx.web.search()`:** supports account-authorized OpenAI Responses candidates, including the single managed route. Ordinary chat success is not search capability proof.
290
- - **Legacy canonical inline agent-loop path:** eligible `github-copilot` requests support Responses or Anthropic Messages native-search candidates only while that legacy route remains configured.
291
- - **Chat Completions models:** remain usable through normal native transport but do not advertise hosted search.
292
-
293
- ### Session-aware search routing
294
-
295
- The bundle composes a plugin-owned web-service facade while preserving the original official service and its exact configuration in a named realm. It does not edit Core or agent presets. Native `web_search` consumers keep their argument validation, query/source limits, middleware, timeout and presentation; the same routing also covers direct `ctx.web.search` consumers with proven initiating context.
296
-
297
- The **Web search** card under **Plugins → dsh-github-copilot → Details** controls routing across search backends through the public `plugins.bundle.config` slot (keyed by package name). Routing uses revision-checked writes in the existing `github-copilot` settings namespace. When the bundle slot is unavailable, the routing card falls back to **Settings → Models**, or **Settings → Web search** if the Models footer is also unavailable. Only one routing card is registered at a time; leaving the detail page discards unsaved edits. Account sign-in, status, and model refresh remain under **Settings → Models**. Opening the page does not migrate or rewrite settings, and the removed internal hosted-search tuning form is not part of the user-facing configuration. Removing the companion restores the original web service.
298
-
299
- The **Search provider** selector offers **Auto — follow Chat** and the actual search providers registered through the routed facade. **Fallback provider** uses the same catalog, with an additional **None — no fallback** choice. These are search backends, not individual models: Copilot is one backend even when different account models can execute its search requests. Ordinary setup needs only these provider choices, not a separate search model.
300
-
301
- - `github-copilot-search-routing.searchProvider: auto` follows the initiating Chat provider. Copilot's plugin-owned aliases retain their existing ownership and model-capability checks; other Chat provider IDs must exactly match a registered search provider ID. Names, suffixes and model families are never guessed. This convention does not promise the same model or account across independently registered providers.
302
- - A concrete `searchProvider` ID pins the primary backend independently of Chat.
303
- - `defaultSearchProvider` is only the final fallback when no primary matches or the primary search fails. It is attempted at most once, never retried when it is already the primary, and never replaces a successful empty result. `none` disables fallback without disabling the primary.
304
- - Copilot selected explicitly or as the fallback automatically considers at most three eligible Responses candidates from current account route facts, sorted by model ID. The default capability probe must succeed before the final user query is sent once; metadata alone is not proof, and a failed final query is not replayed across models. This internal selection does not change Chat or global defaults. Other backends own their model configuration, if any.
305
- - A saved nonempty `github-copilot.searchModel` remains authoritative; the plugin does not silently replace an invalid override. Its read-only details offer a separate explicit reset to automatic selection (`searchModel: ''`). Saving provider routing never changes this override.
306
-
307
- Existing `searchMode: auto/fixed` settings remain readable without automatic writes. Legacy fixed mode preserves its old default as the primary; fixed plus `none` stays disabled. An explicit save writes both routing keys together with one revision-checked mutation; it does not require the Copilot settings namespace or fetch an account/model list. Stable Remote references keep enclosing page renders from resetting the draft. On a revision conflict, reload saved settings and reapply the intended choices; a rejected or unconfirmed response is not reported as success. Saving proves only configuration persistence, not live search capability. The UI explains that saving adopts the chosen final fallback, including possible API charges; legacy `github-copilot.searchFallback: none` failure-spending restrictions remain until that choice is saved. Unregistered saved IDs remain visible as unavailable instead of being silently substituted.
308
-
309
- `github-copilot.routeWebSearch: false` still delegates to the original configured web service. Otherwise cancellation, unload and captured account-proof invalidation never authorize a fallback. Generic registered backends must honor cancellation; their public interface does not expose an internal pre-network authorization hook. The historical direct Copilot/DeepSeek path retains its stronger owned pre-dispatch guard. Neither path supplies another provider's credentials from Copilot sign-in.
310
-
311
- The catalog contains registrations observed by this facade, not hidden registrations made directly in another realm. Listing it does not call availability checks, model discovery, credentials or search; actual usability is checked for each request. No static list of example providers is presented as installed support. Custom/nonstandard web compositions still require review. Fetch is unchanged. See [implementation and evidence limits](docs/session-search-routing.md).
312
-
313
- Requests go directly to the credential-resolved HTTPS Copilot endpoint after strict host validation: GitHub-hosted `api.*.githubcopilot.com`, or `copilot-api.<signed-in-enterprise-domain>` for an accepted GitHub Enterprise credential. No external gateway receives the credential.
314
-
315
- By default (`probe: true`), search fails closed unless the selected route is canonical `github-copilot` or the plugin-owned `github-copilot-preview`, the account authorizes the model, its verified protocol supports native search, and a bounded capability probe succeeds. Managed-model conversations always use their native adapter; independent Responses `ctx.web` search uses account-bound authorization without the old static-catalog ID restriction. Setting `probe: false` bypasses only capability proof and trusts the selected native protocol; route, account, protocol, endpoint, and authentication checks remain active. The underlying hosted-search provider does not itself fall back. The routed facade may apply the explicitly selected final fallback for eligible failures; aborts and invalidated owner/account proofs remain terminal. Requests containing any Core file block—including files nested in tool-result content—also fail closed to `next()`, preserving Core's file projection instead of letting the hosted-search serializer drop that context.
316
-
317
- Search proof is lazy: attach, settings updates and `credentials/record-updated` for `llm-pi-ai/github-copilot` only invalidate cached plans, without starting network work. The next actual eligible request proves capability again; unrelated credentials are ignored and event bursts do not trigger repeated eager probes. In-flight proofs are cancelled on invalidation/disposal. If credentials change during proof or final auth resolution, the current request fails closed rather than applying account A's proof to account B. Submit a new request after the update; there is no automatic retry loop or implicit `probe: false` fallback.
318
-
319
- ## Reasoning summaries and empty Think disclosures
320
-
321
- On the eligible custom Responses path, an explicit request reasoning effort overrides the provider profile default. The companion validates it against the selected model's declared/native efforts, maps its wire value, and requests `summary: "auto"`, matching the native pi-ai path. No effort selection leaves provider defaults unchanged; Core's `off` omission is preserved rather than represented as a guaranteed server-side disable. Unsupported or missing model metadata fails with a named error before the custom model request; the separate capability probe retains its existing lifecycle.
322
-
323
- The Responses parser preserves public `reasoning_summary_text` and `reasoning_text` events, final-only summaries, and interleaved parts without repeating delta text from completion snapshots. Empty or encrypted-only items do not become fabricated explanations. Some Copilot Responses requests still return encrypted reasoning without public text: [summaries are optional](https://developers.openai.com/api/docs/guides/reasoning), and requesting one does not guarantee it. The companion never decrypts or invents reasoning.
324
-
325
- Requests whose assistant history contains reasoning blocks or opaque `replayState` bypass the custom wire through Core **before probing**. A public summary must not be reconstructed as a raw `reasoning_text` input item, and encrypted replay belongs to Core. This deliberately limits inline search on such histories; the separate Responses-only `ctx.web` provider remains available under its usual route/probe gates.
326
-
327
- On the guarded Chat rendering contract, the Client hides completed, empty or whitespace-only reasoning disclosures for replies whose own recorded provider is `github-copilot` or `github-copilot-preview`, for any valid model ID. It delegates the remaining view to DSH's native renderer. Nonempty summaries, answers, tools, images and actions remain unchanged. Running or interrupted steps, unknown provenance, and non-Copilot replies retain their native rendering; selecting a different model later does not reclassify historical replies. Because this runs on the rendered view rather than the search transport, it also covers native/image request paths and loaded history when the necessary provenance is present.
328
-
329
- Filtering changes only temporary render props. Durable messages, encrypted signatures, replay-state block indexes and token usage are not rewritten, so future requests retain the original reasoning context. There is no DOM polling or whole-page observer. Removing the plugin withdraws the contribution and restores native rendering.
330
-
331
- This optional integration uses the public `conversation.chat.node` keyed slot and `uiConversation` location data. Its guard checks the current plain memo renderer named `AssistantNodeView`; renamed/minified future renderers are left alone. This is a compatibility check, not module-ownership or security proof: a deliberate replacement with identical naming and metadata cannot be distinguished through this registry. It does not block authorization on older Cores. Missing or incompatible extension contracts, or a competing assistant renderer, leave native output unchanged with a named compatibility diagnostic. The new display behavior is not a claim of live GPT-6 transport success; unsupported Core versions may still show empty Think rows. It does not change the `github-copilot.enabled` setting, which controls hosted search only.
332
-
333
- ### Capability warnings and adapter limits
334
-
335
- - `REASONING_EFFORTS_UNSUPPORTED` means some advertised reasoning labels cannot be expressed faithfully by the selected native SDK protocol. They are not guessed, and ordinary requests remain available. Explicit unsupported efforts—including an unadvertised `off`—are rejected rather than silently treated as defaults.
336
- - `INPUT_LIMIT_ESTIMATED_GUARD` means the managed route applies a plugin-owned estimated input guard while retaining the provider's distinct input, context and output capacities. This is not exact provider token counting or a claim that Core exposes a separate input limit; provider overflow remains possible.
337
- - The Core-facing integration uses the published adapter's normal `streamSimple` path. Its advanced `stream` entry explicitly rejects incompatible protocol-specific SDK client objects; this does not disable normal conversation streaming.
338
- - Unknown pricing is represented as unpriced metadata, not a claim that a model is free. Public summaries remain optional provider output.
339
-
340
- ## Managed request budgets and compaction
341
-
342
- On Desktop, automatic recovery and Auto diagnostics use the initiating Agent's current preset engine, including isolated groups, through Core's public preset lookup. A bound preset without an engine never borrows global recovery. Automatic compaction remains the native pressure/error-recovery path, not `/compact` or `/copilot-compact`; segmented recovery remains manual-only. Timeout/transport errors are not reclassified as context overflow.
343
-
344
- The managed `github-copilot-preview` route estimates the complete native request, including current system/tool prefixes, and checks both the independent prompt limit and combined input/output capacity before model transport. Its operational input budget is `min(prompt limit when advertised, context capacity - requested/default output reservation) - safety allowance`. Supplier capacities are not rewritten. Invalid output policy or exhausted headroom fails explicitly; no history is silently truncated and no model is substituted.
345
-
346
- For a marked ordinary agent-loop request with a matching initiating Session and committed request configuration, optional token-meter and enabled stock automatic compaction, the plugin signals local estimated pressure at 90% of that budget. Core owns the existing bounded compaction/rebuild path and commits only a valid smaller checkpoint. It does not compact concurrently inside a frozen request. Missing optional seams disable this early signal, not the final native budget guard; `auto: false` or zero stock overflow retries remain respected. Canonical/user-owned routes and other providers are unchanged.
347
-
348
- Calls explicitly marked `purpose: 'compaction'` bypass the early-pressure signal and use the full hard budget. When no caller or adapter-resolved reasoning effort is present, `prefer-low` selects an advertised supported `minimal` or `low` effort; unsupported controls preserve provider defaults. An explicit or materialized effort wins. The requested summary output cap remains unchanged, preserving Core's auxiliary summary record; the plugin does not blindly raise/lower it or accept truncated output as a checkpoint. `enabled` remains a hosted-search switch, not a switch for this managed-route protection.
349
-
350
- **Existing oversized history is not guaranteed recoverable.** The default manual compact still rejects an oversized summary before model transport. A separately selected [manual recovery engine](./docs/manual-compaction-recovery.md) can summarize compressible, balanced managed-Copilot history through bounded intermediate calls while Core retains the only durable transaction. It is opt-in as a **replacement**, not a second compaction service; merely installing the plugin does not activate it. Indivisible input, oversized fixed prefixes and truncated output fail without replacing history. No automatic model switch or hidden history deletion occurs.
351
-
352
- When that recovery engine and native jobs are available, `/copilot-compact` starts a Session-owned background compaction without holding Desktop's command HTTP request open for the entire summary. Use `/copilot-compact status` or `/copilot-compact cancel` for its latest job. Admission is not summary success; failures remain explicit. Completion does not wake the model or resume a Goal. Native `/compact` stays unchanged, and job status is process-local.
353
-
354
- ## Copilot tool compatibility
355
-
356
- To prevent observed invalid Copilot tool payloads, the package sets the managed route's `compat.supportsStrictMode` leaf to `false` and applies two schema-only fixes when the selected provider is canonical `github-copilot` or the plugin-owned account route `github-copilot-preview`: it removes top-level `sandbox_permissions` and `justification` properties, and rewrites Core's multi-action `update_goal` parameters as a discriminated `oneOf`. Each Goal action then advertises only its legal fields: `complete`, `pause`, and `resume` cannot carry edit or blocker fields; `blocked` requires `blocked_reason`; and `edit` alone exposes replacement fields. Execution still uses Core's original Goal tool and service. Non-Copilot prompt assemblies are unchanged.
357
-
358
- Copilot sessions that need wider file or command access must select sufficient standing permissions before the call. Installation agents must also follow these payload rules:
359
-
360
- - Omit `sandbox_permissions` and `justification` on initial `pwsh` calls.
361
- - Never emit them when approval prompts are disabled or the current mode is already `danger-full-access`.
362
- - Use them only for the single exact-command retry allowed after a real sandbox denial when approval is available and the requested mode is wider.
363
- - Omit the keys entirely rather than sending null, empty, or current-mode values.
364
-
365
- The plugin does not rewrite `$DSH_HOME/AGENTS.md`. Installers may merge these rules into user instructions only with explicit user consent.
366
-
367
- ## Settings
368
-
369
- The plugin's `github-copilot` settings section controls account-metadata freshness, managed request budgets and hosted search. `enabled` still controls hosted search only:
370
-
371
- The plugin details page exposes only **Search provider** and **Fallback provider**. Account-metadata, request-budget, and hosted-search safety values remain schema-backed settings for administrators and compatibility tooling; they are not routine user-facing controls.
372
-
373
- | Key | Default | Scope and meaning |
109
+ | Key | Default | Purpose |
374
110
  |---|---:|---|
375
- | `accountModelTtlMs` | `86400000` | Maximum account-metadata reuse window in milliseconds (24h); does not extend credentials or proof validity. |
376
- | `accountModelFailureCooldownMs` | `300000` | Failure cooldown in milliseconds (5min) for non-forcing discovery; no periodic retries. |
377
- | `chatStreamIdleTimeoutMs` | `300000` | Managed chat byte-idle interval; native assistant-chunk silence is bounded to twice this interval when liveness is enabled (maximum 2147483647 ms). Separate from hosted search. |
378
- | `chatStreamLiveness` | `true` | Observe actual HTTP/SSE bytes without synthetic assistant output; `false` restores native-only timing. Explicit WebSocket/`auto` transport is unchanged. |
379
- | `chatMaxRequestImageBytes` | `20971520` | Native outgoing image budget (20 MiB), not a total JSON/supplier limit. A smaller explicit value lets Core offload older images to mapped read-only paths; original history stays intact, but outgoing image visibility changes. |
380
- | `requestBudgetSafetyTokens` | `4096` | Non-negative safe-integer estimated input allowance; larger values reduce available input. |
381
- | `requestBudgetPressureRatio` | `0.9` | Fraction (0.01–1, step 0.01) for early local pressure, only with supported enabled stock recovery. |
382
- | `compactionReasoning` | `prefer-low` | Choose supported minimal/low only when summary effort is absent; `preserve` keeps native defaults. |
383
- | `enabled` | `true` | Enable both hosted-search surfaces. |
384
- | `providers` | `[]` | Optional route allowlist for both surfaces; empty follows the initiating route. Existing nonempty lists are preserved; Ops must explicitly review any legacy-ID change to `github-copilot-preview`. |
385
- | `includeSources` | `true` | Request provider citations on the inline path. The `ctx.web` bridge always requests and returns sources. |
386
- | `stripServerTools` | `true` | On the inline path, remove local function variants of provider-hosted search tools. |
387
- | `idleTimeoutMs` | `300000` | Inline stream idle timeout and `ctx.web` request deadline, in milliseconds. |
388
- | `probe` | `true` | Require capability proof on both surfaces; `false` explicitly trusts the native protocol. |
389
- | `probeTimeoutMs` | `30000` | Whole capability-probe deadline, in milliseconds. |
390
-
391
- There are no credential-token, API-key, model-catalog, or endpoint settings in this package.
392
-
393
- ## Migration and troubleshooting
394
-
395
- Managed Responses requests omit a narrowly recognized empty reasoning shell from the outgoing payload: a nonempty item ID, absent/completed status, an explicitly empty summary array, absent/empty content, no encrypted content, and no unknown fields. Stored history remains unchanged. Nonempty summaries, opaque reasoning, partial signatures, reference-only items and unknown fields are never discarded. This prevents empty SDK signatures from blocking otherwise complete conversation replay; it does not recover missing historical content or introduce automatic model switching.
396
-
397
- Existing installations must follow the [single-route migration guide](./docs/single-route-migration.md) before removing a native profile. The plugin does not migrate defaults, presets, active sessions or history for you. Old gateway routes and `COPILOT_GITHUB_TOKEN`-style references are not required; review them separately rather than deleting unrelated user configuration.
398
-
399
- - **No sign-in control:** confirm the package is loaded in the active profile and use the baseline-specific UI above. Do not add a native provider just to reveal login.
400
- - **Two Copilot groups after upgrade:** a legacy canonical profile is still configured; it is preserved deliberately. After explicit migration/removal, both composer and `/model` list only the managed group. There is no display-only alias masking a second route.
401
- - **Signed in but a new model is missing:** opening Models ensures missing/stale metadata automatically; inspect accepted/rejected models for `github-copilot-preview`. Use **Retry** after errors or **Manage → Refresh models** to intentionally refresh before the TTL expires. Unsupported endpoints or incomplete metadata produce diagnostics, not static-catalog fallback. Do not repeat login or disable validation.
402
- - **`AUTH` / “API key is invalid” after signing in again:** this label does not prove the conversation cached an old key. On the managed route, an observed model HTTP 401 retires only the exact current token proof, except the narrowly recognized Responses replay-scope rejection described below. The failed request stays failed: no automatic message replay, logout or model switch. Your next request (or model discovery) can perform one native OAuth renewal even before stored expiry, then validate account metadata. Repeated rejection is bounded per account by `accountModelFailureCooldownMs` (five minutes by default, at least one second); even forced discovery cannot bypass this recovery cooldown. If the renewed token is still rejected, wait for the cooldown and inspect account/permission status rather than repeatedly signing out. A new sign-in is not overwritten by a late old response. HTTP 403, network errors and strings mentioning 401 do not trigger this recovery. It observes native HTTP transport only; WebSocket recovery and live endpoint acceptance are not claimed.
403
- - **`input item ID does not belong to this connection` / `input item does not belong to this connection`:** these are the two observed Responses history-reference rejection messages, not proof of an invalid API key. On `github-copilot-preview` only, a bounded clone of the actual HTTP 401 JSON identifies these exact uncoded errors, with matching outer/nested messages; the plugin reports `INVALID_REQUEST` with `COPILOT_RESPONSES_REPLAY_SCOPE_MISMATCH` without retiring shared account proof, aborting neighboring calls, renewing credentials or retrying the request. Unknown, malformed, oversized or ambiguous error bodies retain native authentication handling. Historical error records are not rewritten.
404
- - **Managed Responses replay compatibility:** the public SDK `onPayload` hook normalizes only direct item IDs on complete assistant messages, function calls and reasoning items carrying encrypted content. After an actual HTTP 408, a newly prepared attempt in the same Core step may reuse the exact previously normalized wire payload if it substitutes only matching item-ID references or incomplete ID-bearing items without conflicting fields, and the complete transcript and every other request field agree. The entry is bound to the original turn signal, Session, model and account proof; it expires after 60 seconds, is limited to 2 MiB, and is cleared on the next step/turn end, concurrency, cancellation, credential discontinuity or a different request. It neither creates a retry nor bypasses native policy. Cold-history reference-only, incomplete or unsupported ID-bearing items still fail explicitly with `COPILOT_RESPONSES_REPLAY_UNSUPPORTED`. `call_id` pairing, encrypted bytes, public summaries, `phase`, nested IDs and durable history are unchanged; no hidden history trimming or model switch occurs. This is Copilot-specific, not a universal OpenAI schema change: public OpenAI types require IDs on some replay forms. Native SDK/pinned-Core synthetic tests verify transport shape and isolation, not real Copilot acceptance of an old conversation. Canonical `github-copilot`, other providers and non-Responses protocols are unchanged; install/activation and old-session endpoint acceptance require separate review.
405
- - **Quota TLS verification failure (`COPILOT_USAGE_TLS` / `UNABLE_TO_VERIFY_LEAF_SIGNATURE`):** the quota request uses a reusable, TLS-verified agent combining Node's default and system CA roots, including trusted Windows roots installed by Watt Toolkit / Steam++-style accelerators. `NODE_USE_SYSTEM_CA=1` is no longer required for this plugin-owned request after the updated plugin is loaded; it can still matter for unrelated Desktop traffic. `COPILOT_USAGE_TRUST_UNAVAILABLE` means system roots could not be loaded or the agent could not be created, and no request was sent. Check the Host certificate store and proxy configuration without disabling verification; neither diagnostic proves a specific proxy or live account status.
406
- - **Scope rejection persists after alpha.36:** its original exact-message classifier missed the observed wording without `ID`. Alpha.37 only closes that classification/isolation gap; it does not further rewrite payloads or prove that the rejected history can be replayed. The parallel-request failure is reproduced with the native adapter in a fixture that does not load cron. This establishes a companion error-handling defect, not the upstream reason for rejecting the original item. `Request aborted` alone does not identify its cause; `COPILOT_MODEL_SOURCE_AUTH_FAILED` is a generalized pre-catalog auth-resolution failure, whose cached result may recur without another auth attempt. Compare actual loaded version, request provenance and the same history/model before attributing another machine's success to a plugin downgrade or cron removal. Do not silently discard encrypted reasoning, retry old business work or reset credentials.
407
- - **Canonical route says `not-configured`:** with configured login this is normal managed-only mode; inspect account discovery separately rather than creating a native profile.
408
- - **Delete dialog stays on “Deleting…”:** this update does not prove that hang fixed. Do not repeatedly delete; after an authorized Host stop, inspect persisted settings and follow the migration guide before deciding whether any removal is still needed.
409
- - **No Think text:** the provider may omit public summaries, but nonempty summaries must survive the Responses parser. Check selected effort and named errors. Empty disclosures are hidden only after completion; encrypted replay is never displayed. Reasoning/replay-bearing histories use native Core transport.
410
- - **Hosted search unavailable:** for explicit/fallback Copilot, inspect account discovery/probe diagnostics; ordinary setup no longer requires selecting a search model. A legacy override remains authoritative until you explicitly reset it to automatic selection. Auto following Copilot Chat still requires that initiating model to support search. The custom inline path is legacy-canonical only; a saved provider choice or normal managed chat success does not prove search support.
411
- - **Legacy endpoint/key still present:** reconciliation preserves unowned fields by design. Review explicit migration; never force-remove an ownership marker.
412
-
413
- ## Package entries and source map
414
-
415
- Public exports are `.`, `./client`, `./remote`, `./deployment-baseline.json`, and `./package.json`.
416
-
417
- - `src/index.ts`: authorization bootstrap, dependency-gated Host composition, settings, inline interception, and `ctx.web` registration.
418
- - `src/authorization-controller.ts`: Host authorization and path-level route reconciliation.
419
- - `src/copilot-grant.ts`, `src/copilot-auth.ts`: grant normalization and Host credential lifecycle.
420
- - `src/client.ts`, `src/remote.ts`: Models UI and Client-safe Remote contract.
421
- - `src/account-model-catalog.ts`, `src/account-model-source.ts`, `src/account-model-auth.ts`: bounded account discovery, endpoint/capability normalization and native OAuth binding.
422
- - `src/preview-route.ts`, `src/preview-provider.ts`, `src/pi-provider-bridge.ts`: data-driven account models and the public native adapter/SDK boundary.
423
- - `src/current-provider.ts`, `src/plan.ts`, `src/probe.ts`: owned route facts, candidate planning and capability proof.
424
- - `src/temporary-models.ts`, `src/route-ownership.ts`: recognition and conservative restoration of historical configuration writes; not a new-model routing table.
425
- - `src/model-protocol.ts`: public local facts with explicit ownership limits; no unshipped Core service dependency.
426
- - `src/responses-reasoning.ts`, `src/responses-reasoning-text.ts`: selected-model effort mapping and public summary assembly.
427
- - `src/wire.ts`, `src/wire-anthropic.ts`, `src/traditional-search.ts`: hosted-search transports.
428
- - `deployment-baseline.json`: declared machine-readable compatibility/capability evidence inventory; `scripts/verify-deployment-baseline.mjs` checks its source and test markers for drift.
429
- - `lib/`: generated release output; never edit it directly.
111
+ | `autoSemanticAssessment` | `true` | One auxiliary assessment for locally unknown Auto tasks |
112
+ | `followParentModel` | `false` | Supported child following; explicit child selections win |
113
+ | `excludedModelIds` | `[]` | Exact exclusions through Model preferences |
114
+ | `accountModelTtlMs` | `86400000` | 24h maximum metadata reuse, not token/proof validity |
115
+ | `accountModelFailureCooldownMs` | `300000` | 5min non-forcing discovery failure cooldown |
116
+ | `chatStreamLiveness` | `true` | Managed HTTP/SSE byte observation |
117
+ | `chatStreamIdleTimeoutMs` | `300000` | Byte-idle deadline; separate from search |
118
+ | `chatMaxRequestImageBytes` | `20971520` | Native 20 MiB outgoing-image budget, not a total JSON limit |
119
+ | `requestBudgetSafetyTokens` | `4096` | Estimated input safety allowance |
120
+ | `requestBudgetPressureRatio` | `0.9` | Early pressure with supported enabled recovery |
121
+ | `compactionReasoning` | `prefer-low` | Supported low effort only when none is resolved |
122
+ | `enabled` / `probe` | `true` / `true` | Hosted search / capability proof |
123
+ | `providers` | `[]` | Optional explicit search route allowlist |
124
+ | `includeSources` / `stripServerTools` | `true` / `true` | Inline citations / hosted-tool schema handling |
125
+ | `idleTimeoutMs` / `probeTimeoutMs` | `300000` / `30000` | Search request / probe deadlines |
126
+
127
+ | Symptom | Next safe action |
128
+ |---|---|
129
+ | No sign-in controls | Check active profile and loaded Host/Client; do not add a native provider merely to reveal login |
130
+ | Signed in, missing models | Inspect discovery diagnostics; use visible Retry or intentional Refresh, not repeated sign-in |
131
+ | Two Copilot groups | Review [explicit migration](./docs/single-route-migration.md); no automatic removal |
132
+ | Selection unknown/unavailable | Distinguish lost evidence from read failure; Retry only rereads |
133
+ | Missing/cancelled Turn Usage | Read [usage limitations](./docs/copilot-usage.md); do not infer zero billing |
134
+ | History fails on `github-copilot/auto-model-decision` | Use [detached, check-first recovery](./docs/automatic-model-routing.md#recovering-affected-histories); never replace live history without stopped writers and approval |
135
+ | AUTH, Responses replay-scope or TLS failure | Review [request diagnostics](./docs/model-compatibility-acceptance.md#authentication-replay-and-request-diagnostics) and [usage TLS boundaries](./docs/copilot-usage.md#account-data-boundary); never discard opaque replay, reset credentials or disable TLS as an automatic repair |
136
+ | Hosted search unavailable | Check account/protocol/probe diagnostics; a legacy override stays authoritative until explicitly reset |
137
+
138
+ ## Ownership and further reading
139
+
140
+ DSH Core owns Sessions, tools, sandboxing, attachments, native accounting and other providers. `llm-pi-ai` owns OAuth, token exchange/refresh and normal model transport. This plugin composes the published adapter with authenticated account metadata; new supported account model IDs need no name-based routing patch. Credentials stay Host-only under `llm-pi-ai/github-copilot`.
141
+
142
+ Public APIs cannot replace Core Edit/Delete, restructure its flat picker or make native Add a single-route enforcement mechanism. Retired [Model roles](./docs/dual-model.md) remain compatibility-only. Fixes must stay [plugin-only](./AGENTS.md#plugin-only-implementation-boundary).
143
+
144
+ | Guide | Scope |
145
+ |---|---|
146
+ | [Compatibility acceptance](./docs/model-compatibility-acceptance.md) | Metadata, reasoning, replay and public-interface limits |
147
+ | [Current official-first review](./docs/official-first-020-rc2.md) | Exact rc.2 seams and retained gaps |
148
+ | [Migration](./docs/single-route-migration.md) | Fresh read-only readiness and explicit config-only maintenance |
149
+ | [Distribution](./docs/npm-distribution.md) | Same-byte GitHub/npm publication and qualified installation |
150
+ | [Agent guide](./AGENTS.md), [contributing](./CONTRIBUTING.md) | Owners, task plans, gates and delivery policy |
151
+
152
+ Version history belongs in [CHANGELOG.md](./CHANGELOG.md) and immutable [Releases](https://github.com/cloga/dsh-github-copilot/releases), not the setup instructions. `deployment-baseline.json` retains historical pins without admitting them as current support.
430
153
 
431
154
  ## Build and verify
432
155
 
156
+ Use Node 24 LTS and the exact pnpm version in `package.json`; runtime Node must be >=22.19.0.
157
+
433
158
  ```sh
434
159
  pnpm install --frozen-lockfile
435
160
  pnpm verify
436
161
  pnpm pack --pack-destination artifacts
162
+ pnpm verify:tarball -- artifacts/dsh-github-copilot-<package-version>.tgz
437
163
  ```
438
164
 
439
- Use Node 24 LTS for development and the pinned pnpm version; runtime dependencies require Node >=22.19.0. `pnpm verify` runs the Agent contract check, source and local test typechecking, baseline markers, clean build, Vitest and Node tooling tests, and a real built Host import plus Client/Remote smoke. After packing, run `pnpm verify:tarball -- artifacts/dsh-github-copilot-<package-version>.tgz` to verify archive exports, media, allowed contents and equality to that build. CI gates only exact DSH `0.2.0-rc.2` on Windows and Linux, using its unchanged source runtime and published npm artifacts; the ten earlier pins are retained historical evidence, not release gates. Release publication depends on this exact-target matrix; local tests alone do not establish CI qualification.
440
-
441
- For the optional reasoning UI integration, `pnpm verify:reasoning-ui -- <Core checkout>` runs a synthetic native-renderer, Slot registry and history-assembly fixture against a clean pinned `0.2.0-rc.2` checkout with its Chat dependencies installed. It exclusively creates one temporary test file and removes it only if unchanged. This is local integration/static-render evidence, not a live browser or Copilot API test; the CI definition covers only the current target, without implying this candidate's qualification has run.
442
-
443
- ### Agent-driven development
444
-
445
- From a source checkout, use these read-only entrypoints (no dependencies needed for discovery):
446
-
447
- ```sh
448
- node scripts/agent.mjs describe --json
449
- node scripts/agent.mjs doctor --json
450
- node scripts/agent.mjs plan models --json
451
- node scripts/agent.mjs attribution "DeepSeek Harness (DSH)"
452
- ```
453
-
454
- `agent-contract.json` maps authorization, models, search, client, compatibility, tooling and release tasks to owning files and tests. Plans return unexecuted argument arrays, including the exact package-version archive path. Doctor checks repository prerequisites only: exit 0 means preflight passed, 1 means missing prerequisites, 2 means invalid input/metadata. For clean machine-readable output prefer the direct `node` command rather than parsing pnpm progress logs.
455
-
456
- Attribution follows the actual tool: `Assisted-by: DeepSeek Harness (DSH)` for DSH-assisted changes, not a Copilot App co-author inferred from the model provider. Keep the human Git author and reserve `Co-authored-by` for verified collaborators. User-requested delivery can merge after required review/CI and branch protection without a separate approval prompt. Profile install, sign-out and worktree checkout still require explicit approval. Important updates include post-merge release follow-through under the policy below.
165
+ `pnpm verify` covers contracts, types, baseline, build, tests and built entry smoke. Required CI targets unchanged official rc.2 source and published artifacts on Windows/Linux. Passing synthetic tests does not prove live account/model/search behavior.
457
166
 
458
- Evidence is layered: package presence/import and passing synthetic tests do not prove live DSH activation, account entitlement, model transport or hosted search. Authorization `status()` is read-only; explicit reconciliation/startup may persist configuration, while capability probes run only on actual eligible requests. None of these reports implies live transport success without a real request. Search resolves captured effective initiating Session request config or explicit request options and keeps per-owner plans instead of deriving them from the future global default; unsupported context still fails closed. Release, Ops migration and registry readback remain separate evidence. See the [readiness audit](./docs/agent-readiness.md) for evidence and remaining limitations.
167
+ Agents start with `node scripts/agent.mjs describe --json`, `doctor --json` and `plan <task> --json`; plans are unexecuted arguments, not action approval. Follow [AGENTS.md](./AGENTS.md) for the full change/release policy. Important updates require protected merge and verified dual-channel publication; documentation-only work does not release by default. Profile installation, sign-out and session-interrupting restart need separate authorization.
459
168
 
460
- ## Important-update release delivery
461
-
462
- Important user-requested features, behavior fixes, compatibility fixes, and security or stability fixes include publication after a reviewed merge and green required CI. The agent must continue version preparation, the protected tag/Release workflow and asset verification without waiting for a second request to release. Plan version alignment in the implementation PR where practical; preserve the prerelease channel unless promotion is requested.
463
-
464
- An explicit code-only/review-only/do-not-merge/do-not-release instruction takes precedence. Documentation-only and internal-only changes do not trigger a release by default. A separate version PR still needs required review/CI and branch protection, not another merge-approval prompt; profile installation and session-interrupting restarts remain separate. This is agent delivery policy, not an unconditional publish-on-merge CI trigger.
465
-
466
- Report the published Release URL, version, tag/commit and verified asset SHA-256 before calling release delivery complete. If publication is blocked, report the concrete CI/permission/network blocker and pending step rather than treating a merged PR or local build as a release. Full rules are in [AGENTS.md](./AGENTS.md#important-update-release-delivery).
169
+ Commit attribution uses `Assisted-by` with the actual tool, never the model provider or an unverified collaborator identity.
467
170
 
468
171
  ## Release and checksum verification
469
172
 
470
- `package.json` declares public npm distribution. A release tag must equal `v${package.json.version}`. Versions use standard SemVer prerelease labels (`alpha`, `beta`, or `rc`), each with its matching npm dist-tag; only stable versions use `latest`. The Release workflow performs the frozen install and complete verification gate, packs once (or recovers the original archive on retry), verifies `SHA256SUMS`, publishes the immutable GitHub Release and then publishes those same bytes to npm through OIDC. Either channel failing means delivery is incomplete. First package creation needs an authorized maintainer; staging requires an existing package and is not a first-package bootstrap. Historical releases are not republished.
173
+ 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.
471
174
 
472
175
  ```sh
473
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.0-alpha.88/dsh-github-copilot-0.4.0-alpha.88.tgz
474
- curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.0-alpha.88/SHA256SUMS
176
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.0-alpha.92/dsh-github-copilot-0.4.0-alpha.92.tgz
177
+ curl -LO https://github.com/cloga/dsh-github-copilot/releases/download/v0.4.0-alpha.92/SHA256SUMS
475
178
  sha256sum --check SHA256SUMS
476
179
  ```
477
180
 
478
- PowerShell can verify the same two downloaded files with:
479
-
480
181
  ```powershell
481
182
  $expected = (Get-Content .\SHA256SUMS).Split()[0]
482
- $actual = (Get-FileHash .\dsh-github-copilot-0.4.0-alpha.88.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
183
+ $actual = (Get-FileHash .\dsh-github-copilot-0.4.0-alpha.92.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
483
184
  if ($actual -cne $expected) { throw 'Release checksum mismatch' }
484
185
  ```
485
186
 
486
- The checksum detects download corruption or asset drift; repository controls and the protected Release workflow establish publisher provenance. Never move or reuse a release tag. Increment the package and deployment-baseline versions together for every release. See [CONTRIBUTING.md](./CONTRIBUTING.md) for the change workflow and [SECURITY.md](./SECURITY.md) for private vulnerability reporting.
187
+ Checksums detect corruption/drift; protected workflows and repository controls establish publisher provenance. Report vulnerabilities privately through [SECURITY.md](./SECURITY.md).