Complete local setup assistant acceptance evidence

.github/workflows/ci_check.yml

@@ -13,6 +13,10 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
   cancel-in-progress: true
 
+defaults:
+  run:
+    shell: bash
+
 jobs:
   ci-check:
     name: CI Check
@@ -100,5 +104,8 @@ jobs:
       - name: Validate product specification catalog
         run: pnpm run validate:specifications
 
+      - name: Validate setup acceptance ledger
+        run: pnpm run validate:setup-acceptance
+
       - name: Validate Git diff
         run: git diff --check

.github/workflows/copilot_commit.yml

@@ -7,6 +7,10 @@ on:
       - '!master'
       - '!develop'
 
+defaults:
+  run:
+    shell: bash
+
 jobs:
   copilot-commits:
     if: ${{ vars.COPILOT_BOT_LOGIN == '' || github.actor != vars.COPILOT_BOT_LOGIN }}

.github/workflows/copilot_pull_request.yml

@@ -5,6 +5,10 @@ on:
   pull_request:
     types: [opened, reopened, closed, synchronize]
 
+defaults:
+  run:
+    shell: bash
+
 jobs:
   copilot-pull-requests:
     if: ${{ (vars.COPILOT_BOT_LOGIN == '' || github.actor != vars.COPILOT_BOT_LOGIN) && github.event.pull_request.head.repo.full_name == github.repository }}

.github/workflows/copilot_pull_request_merge_queue.yml

@@ -1,4 +1,8 @@
 name: Copilot - Pull Request Merge Queue
+
+defaults:
+  run:
+    shell: bash
 run-name: Copilot PR · ${{ github.event_name }}:${{ github.event.action }}
 
 on:

.github/workflows/hotfix_workflow.yml

@@ -1,5 +1,9 @@
 name: Task - Hotfix
 
+defaults:
+  run:
+    shell: bash
+
 on:
   workflow_dispatch:
     inputs:

.github/workflows/release_workflow.yml

@@ -1,5 +1,9 @@
 name: Task - Release
 
+defaults:
+  run:
+    shell: bash
+
 on:
   workflow_dispatch:
     inputs:

.github/workflows/setup_platform_smoke.yml

@@ -0,0 +1,43 @@
+name: Setup Platform Smoke
+
+on:
+  pull_request:
+    types: [opened, synchronize]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+defaults:
+  run:
+    shell: bash
+
+jobs:
+  setup-platform-smoke:
+    name: Setup package and local session (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 20
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest, macos-latest]
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          persist-credentials: false
+      - uses: pnpm/action-setup@v5
+        with:
+          version: 10.12.4
+          standalone: false
+      - uses: actions/setup-node@v7
+        with:
+          node-version: '24.x'
+          cache: pnpm
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm run build
+      - run: pnpm run validate:build
+      - run: pnpm run typecheck
+      - name: Local setup fixtures only
+        run: pnpm exec jest src/application/usecases/setup/__tests__/setup_session_coordinator.test.ts src/cli/__tests__/web_setup_server.test.ts src/cli/__tests__/setup_session_guard.test.ts src/cli/__tests__/setup_apply_snapshot.test.ts src/tooling/__tests__/setup_acceptance_path.test.ts --runInBand --silent
+      - name: Isolated npm package checks
+        run: pnpm run validate:npm-package && pnpm run smoke:npm-package

build/web/index.html

@@ -5,7 +5,7 @@
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <meta name="color-scheme" content="light dark" />
     <title>Copilot · Setup studio</title>
-    <script type="module" crossorigin src="./assets/index-CcOfHgj1.js"></script>
+    <script type="module" crossorigin src="./assets/index-CaPMoT2_.js"></script>
     <link rel="stylesheet" crossorigin href="./assets/index-7TY0kYbf.css">
   </head>
   <body>

docs.json

@@ -287,6 +287,11 @@
           "href": "/development/testing",
           "icon": "vial"
         },
+        {
+          "title": "Local setup assistant review",
+          "href": "/development/setup-assistant-review",
+          "icon": "clipboard"
+        },
         {
           "title": "Product specifications",
           "href": "/development/specifications",

docs/authentication.mdx

@@ -54,6 +54,10 @@ For Projects, this early choice is only **whether you want the integration**.
 Listing private organization Projects requires authorization, so the assistant
 cannot reliably ask you to select Project numbers before the setup PAT exists.
 If you opt in, the guided setup PAT requests organization **Projects: read**.
+The read-only check lists at most one accessible organization Project. A
+non-public Project confirms the read grant. If the result contains only public
+Projects or is empty, discovery can continue but the PAT grant remains
+`Unverifiable`; access to each selected Project is checked again later.
 After you enter that PAT, the assistant lists the accessible Projects and lets
 you choose their numeric Project numbers and, when available, their Status
 options. You can enter a Project URL or number manually if discovery is
@@ -108,7 +112,7 @@ these states:
 |---|---|---|
 | `✅ Verified` | A safe authentication-bound GitHub operation proved the requested read capability. | Continue. |
 | `❌ Missing` | GitHub deterministically rejected a required capability after identity and repository access were established. | Stop before the dependent mutation and name the permission to grant. |
-| `? Unverifiable` | GitHub does not expose a safe non-mutating proof of the PAT grant, or the response was ambiguous/transient. | Required reads block unless the exact successful public-repository read has separate operational evidence. Required writes pause for a separate explicit acknowledgement and remain non-verified. |
+| `? Unverifiable` | GitHub does not expose a safe non-mutating proof of the PAT grant, or the response was ambiguous/transient. | Required reads block unless an exact successful public-repository read or organization Projects list has separate operational evidence. Required writes pause for a separate explicit acknowledgement and remain non-verified. |
 
 A `403` is not automatically a missing-permission result. Copilot reports it as
 `Missing` only when bounded GitHub metadata explicitly identifies a permission
@@ -130,6 +134,9 @@ access, so setup requires explicit acknowledgement for those rows. Public
 organization Issue Types reads remain `Unverifiable` as PAT evidence.
 After valid token identity, a successful public repository read can be used
 for that exact operation while its PAT permission row stays `Unverifiable`.
+An organization Projects list containing only public Projects, or no Projects,
+is similarly usable for discovery without proving the PAT grant. A malformed
+or denied Projects response is not usable.
 The public organization members list can omit concealed members, so it cannot
 establish Members-read capability. Copilot instead checks the authenticated
 user's active organization membership through GitHub's permission-bound
@@ -271,8 +278,8 @@ For comment-driven assistance, read-only commands are available to anyone who ca
     Read the required-permissions table before creating the token, then review
     the permission-check table after entry. A `❌ Missing` required row must be
     corrected before setup can continue. An ambiguous required
-    `? Unverifiable` read blocks; only a successful public-repository read can
-    remain unverified as PAT evidence but operationally usable. A required
+    `? Unverifiable` read blocks unless a successful public-repository read or
+    organization Projects list has separate operational evidence. A required
     write row means to compare the PAT settings with the
     requested access level and explicitly acknowledge it; it is never a pass.
   </Step>

docs/development/architecture.mdx

@@ -203,6 +203,18 @@ graph tests reject transitive dependencies from these use cases into CLI,
 infrastructure, Action, or browser code, and browser components can import
 only redacted application contracts as types.
 
+`SetupSessionCoordinator` now owns the common repository → choices → setup PAT
+→ plan → credentials → Apply sequence for terminal and browser setup. Its
+semantic phase ports compose the existing use cases; it checks live-session
+state after asynchronous work and classifies a possible write as partial if
+the result is unknown. The initial setup workflow emits value-free resource
+transitions through an optional progress port. The local Action path passes
+those facts to the browser bridge, while the final structured receipt remains
+authoritative for completed, skipped, and uncertain resources. The browser
+never sees provider request bodies or credential values.
+The [fixture-only acceptance review](/development/setup-assistant-review)
+records the human browser, language, accessibility, and platform gates.
+
 Inside `web/src`, `App.svelte` is only the page shell. The session client owns
 same-origin bootstrap, polling and revision-bound commands; it holds no pasted
 PAT in a store. Presenters in `components/` render progress, prompts, context,

docs/development/index.mdx

@@ -34,5 +34,6 @@ Documentation changes must also update `docs.json`, links, component usage, icon
 - [Build artifacts](/development/build-artifacts)
 - [Product specifications](/development/specifications)
 - [Testing](/development/testing)
+- [Local setup assistant acceptance review](/development/setup-assistant-review)
 - [Agents](/agents)
 - [Security & Operations](/security-operations)

docs/development/setup-assistant-review.mdx

@@ -0,0 +1,74 @@
+# Local setup assistant acceptance review
+
+The [350-case acceptance ledger](../../specs/local-web-setup-assistant-acceptance.json)
+is the source of truth for case status. Automated tests establish structural
+and behavioral facts. The 40 human rows remain **open** until a reviewer
+records the environment and observation for each row. Do not infer a visual,
+linguistic, assistive-technology, or operating-system pass from Jest, SSR HTML,
+or a screenshot alone.
+
+## Safe fixture setup
+
+1. Build the package with `pnpm run build` in a disposable checkout. Do not
+   run `copilot setup` against this or any other real repository for review.
+2. Run `node scripts/serve-web-setup-review.cjs pending`. The script serves
+   the compiled browser UI and canned redacted views on `127.0.0.1`. It has
+   no GitHub, filesystem mutation, PAT, Issue, or Action adapter. Open the
+   printed local URL and enter the printed **fixture** pairing code.
+3. Repeat with `action-required`, `blocked`, `partial`, `completed`,
+   `cancelled`, and `expired`. Stop the previous fixture before switching.
+   The fixture rejects answer and close commands; use it to review visible
+   content, navigation, focus and contrast, not to claim a successful setup.
+4. Keep the browser console and network panel available. Record any missing
+   static asset or unintended external request. The GitHub PAT settings link
+   is informational; do not follow it to create a token for this review.
+
+## Required observations
+
+For each of the seven states above, review light and dark modes (`U041–U054`).
+At 320 px, tablet width, desktop width, and 200% browser zoom, check that the
+status, completed facts, next action, affected resource and inspection action
+remain visible without horizontal page scrolling. Verify the focus indicator,
+contrast, reduced-motion behavior, headings and landmarks. A screenshot may
+supplement the recorded observation but cannot replace keyboard or screen
+reader inspection.
+
+For each of English, Spanish, French and Portuguese, review keyboard order,
+screen reader announcements, 200% zoom, 320 px and tablet layouts (`U075–U094`).
+Check the pairing field, action-required prompt, progress changes, error text,
+resource receipt and PAT cleanup link. Use a named screen reader and browser;
+record actual announcements and any field whose label or correction is unclear.
+English is the CLI language. Review the CLI's English prompts and result
+wording separately. Review the meaning and consistency of the four web
+languages with a fluent reviewer; automated key parity is not linguistic
+approval. No other web language is in scope.
+
+On macOS, Linux and Windows, inspect packaged `--web` launch and browser-open
+fallback (`I037–I042`) in an isolated fixture repository with fake ports and
+no network credentials. Record the package version, OS, Node version,
+browser, launcher output, URL, pairing behavior and whether fallback
+instructions are usable. The CI fixture jobs check build and local server
+behavior, but they do not replace these desktop observations. Do not use this
+project or a live GitHub repository as a setup target.
+
+Controlled GitHub account switching, 2FA on GitHub, wrong-account handling,
+and PAT deletion require a separate authorized human review with test accounts
+outside this repository. They remain open here. Never create a real PAT, test
+Issue, or Action run to fill a ledger row in this task.
+
+## Evidence record
+
+For each human row, record:
+
+| Field | Required content |
+|---|---|
+| Case ID | Exact `U` or `I` ID from the ledger |
+| Reviewer and date | Person who performed the observation and local date |
+| Environment | OS/version, browser/version, viewport, zoom, palette, locale, screen reader/version when relevant |
+| Fixture | One of the seven states and the local build/commit ID |
+| Observation | What was read, announced, focused or blocked; include the expected next action |
+| Result | Pass or fail with a link to a redacted screenshot/recording if useful |
+
+Update the ledger row only after direct observation. A pass needs all fields;
+leave it open when a platform, assistive technology or language reviewer is
+unavailable. Run `pnpm run validate:setup-acceptance` after editing it.

docs/how-to-use.mdx

@@ -68,6 +68,11 @@ already deleted the setup PAT in GitHub, run `copilot doctor --read-only`
 with a valid setup PAT instead. A failed or partial Apply does
 not offer this button: inspect the itemized result and use the CLI command
 after resolving the uncertain resources.
+While Apply runs, the page reports each resource as in progress and then as
+completed, skipped, or needing inspection. If the process stops mid-write,
+the operation may already have reached GitHub; inspect that resource before
+starting a new setup session. A new process always requires a fresh review
+and approval.
 The language selector starts in English and offers Spanish, French, and
 Portuguese. Switching language does not change the repository locale or reset
 your answers. Question headings, explanations, options, permission details,

docs/security-operations/operations/troubleshooting.mdx

@@ -28,14 +28,21 @@ not imply a setup failure: open the exact `127.0.0.1` URL printed by the CLI.
 Enter the terminal's 16-character pairing code in the page; after a refresh,
 enter it again. Do not share the code. TCP loopback can be reached by other
 local OS users; the code, not the public URL or OS-user identification, grants
-access to this setup session. After five wrong codes, restart setup.
+access to this setup session. After five wrong codes, pairing and takeover
+pause for 30 seconds; wait, then retry the code printed by the launcher.
 If the local bind or bundled assets fail, stop and use `copilot setup` in the
 terminal. A second browser tab is read-only until you select **Take control
 in this tab**; the former tab then cannot submit decisions. A rejected stale
 answer means the page should refresh to the current decision, not replay it.
-If the page closes during Apply, inspect the terminal result and run
-`copilot doctor` before retrying; do not assume that completed local or remote
-writes were rolled back. No local session shutdown revokes a PAT in GitHub.
+If the page closes during Apply, reopen it and pair with the launching
+terminal's code while that process is alive. The per-resource progress view
+shows completed, skipped, in-progress, and uncertain operations. If the
+process exits during a write, treat its result as needing inspection: check
+the local checkout and corresponding GitHub resource names/scopes, then run
+`copilot doctor --read-only` with a valid setup PAT before retrying. The
+previous approval and secrets are not restored by a new process. Completed
+local or remote writes are not rolled back. No local session shutdown revokes
+a PAT in GitHub.
 
 If setup reports a lock belonging to a stopped process, it will not delete
 that lock automatically: another setup could have acquired the same path in
@@ -59,6 +66,13 @@ directory. The lock contains no PAT or other credential.
   </Accordion>
 
   <Accordion title="PAT permission table shows Missing or Unverifiable" icon="key">
+    In the web assistant, a PAT-related blocked result lists the required
+    grants that could not be confirmed, including scope, access level, and
+    status. It does not display the PAT. The terminal's aggregate error alone
+    may not name the grant; use the browser result before closing the local
+    session. If GitHub identity itself was rejected, fix account, repository
+    selection, token expiry, or organization approval before reviewing grants.
+
     **`❌ Missing`:** GitHub rejected a safe read-only capability probe after
     setup confirmed the token identity and repository selection. Grant the named
     repository or organization permission to the correct resource owner, ensure
@@ -89,6 +103,10 @@ directory. The lock contains no PAT or other credential.
     workflow files remain `Unverifiable` unless repository metadata proves the
     repository is private. Public organization member and issue-type responses
     are likewise inconclusive as PAT evidence or full-membership readiness.
+    An organization Projects list with only public Projects, or no Projects,
+    can support discovery while the `Projects: read` grant remains
+    `Unverifiable`; a non-public Project proves that read access. A denied or
+    malformed list still blocks and must be corrected or retried.
     Members read is verified only by a valid active self-membership response
     from the permission-bound endpoint; Issue Types, ambiguous responses, and
     writes remain blocking.

docs/security-operations/security/self-hosted-runners.mdx

@@ -16,6 +16,16 @@ A self-hosted runner that can access agent credentials MUST be dedicated or isol
 
 A shared persistent runner is not equivalent to an ephemeral trusted runner.
 
+The `codex` label can route a job to macOS, Windows or Linux. Install Git and
+Bash on every runner; Windows needs Git for Windows with its Bash available to
+the Actions runner service. Installed workflows select `shell: bash` for
+`run` steps so Windows does not create PowerShell scripts subject to local
+execution policy. Keep Node.js 24 and pnpm provisioning available to jobs that
+build the Action. Verify each operating system with an actual runner job before
+claiming Action-wide compatibility; the setup fixture matrix exercises only
+its isolated setup paths. In particular, inspect timeout and cancellation of
+agent child processes on Windows separately from POSIX process-group behavior.
+
 The Action passes agent children an allowlisted runtime environment plus only the credential selected for that provider; unrelated GitHub, cloud, database, and PAT variables are not inherited. Codex review capabilities are forced to `--sandbox read-only --ignore-user-config --ignore-rules`; fixer capabilities use `workspace-write`, with network and extra writable roots disabled. Cursor receives a per-invocation configuration/home with sandboxing and default-deny network policy. OpenCode receives a pure, inline default-deny permission profile with shell, task, skill, web, and external-directory tools disabled. Dangerous bypass flags are rejected for every runtime. These controls complement rather than replace runner/container isolation.
 
 Configured verification commands also receive a temporary `HOME` and credential-free allowlisted environment. They are repository processes, not model tools, so the runner must still prevent them from reading unrelated host paths through OS-level isolation.