FLAB: Consolidate session recovery and Copilot model selection

.github/agents/fixlab-autofix.agent.md

@@ -37,6 +37,15 @@ only for information that cannot be established safely from the repository.
 
 ## Bugfix loop
 
+Before validation in a new isolated worktree, plan `fixlab onboard WORKTREE
+--from SOURCE --runtime RUNTIME`. After scoped setup approval, add `--yes`.
+The source profile declares non-sensitive configuration in `worktreeSetup.files`.
+Never copy credentials or browser state. Missing configuration, authentication
+or denied runtime access is a blocker; do not retry unchanged approvals or
+grant blanket path access. Sign-in and MFA remain human-controlled.
+In read-only modes, do not provision or alter configuration; report missing
+prerequisites as blocked.
+
 1. **Classify scope.** Identify the smallest component in the profile that owns
    the behavior. Treat its source paths as the initial edit allowlist.
 2. **Reproduce or measure.** Locate the affected code and establish a failing
@@ -92,9 +101,24 @@ pass, remove transient validation artifacts, review the effective diff, commit
 and push that checkpoint, and create or update a draft pull request. Run
 remaining independent builds and browser journeys while the draft is open,
 push later reviewed checkpoints to the same branch, and never mark the PR ready
-until every required gate passes or the developer explicitly accepts the
+until every required local gate passes or the developer explicitly accepts the
 recorded outcome.
 
+After required local validation and approval gates pass or have an explicitly
+accepted outcome, publish the authorized PR by marking it ready for review.
+Do not wait for remote PR CI or build completion before publication; publication
+can trigger those checks. Verify the PR is no longer draft before reporting the
+PR stage passed. Report remote checks as pending, not passed, and leave merge
+policies intact. Use `gh pr ready <number>` for GitHub or
+`az repos pr update --id <id> --draft false` for Azure DevOps, then read back
+`isDraft` from the provider to confirm it is false. If publication or state
+verification fails, report the PR stage blocked or failed, not passed.
+
+Prefix FixLab-created pull-request titles with `FLAB:` followed by a concise
+description of the coding change. Use the repository's PR template and clear
+Markdown sections for the problem, changes, validation and risks. Keep
+validation checkboxes accurate; do not describe a pending gate as passed.
+
 ## Guardrails
 
 - Handle one bug at a time; do not bundle unrelated fixes or refactors.

.github/agents/fixlab.agent.md

@@ -13,5 +13,8 @@ review, local validation, live validation when configured, evidence, and pull
 request outcome. Never report a skipped or timed-out gate as passed. Never stop
 unrelated processes or expose credentials and private test data.
 
+Prefix FixLab-created PR titles with `FLAB:` and format their descriptions with
+the repository's PR template, clear validation evidence and remaining risks.
+
 Run `npm run validate` for changes to FixLab itself and summarize exact evidence
 and remaining risks.

.github/workflows/ci.yml

@@ -11,6 +11,16 @@ permissions:
   contents: read
 
 jobs:
+  windows-containment:
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version-file: .nvmrc
+      - run: npm ci
+      - run: node --test test\windows-executor.test.js test\dashboard-executor-lease.test.js
+
   test:
     runs-on: ubuntu-latest
     outputs:

CHANGELOG.md

@@ -4,6 +4,21 @@ All notable changes to FixLab are documented in this file.
 
 ## Unreleased
 
+- Keep managed frontend input open so React-style development servers do not
+  stop on EOF; preserve one-shot helper behavior and owned process cleanup.
+
+- Add approved, configuration-only worktree onboarding with `--from`, preserved
+  custom files, missing-browser-config preflight, and reviewed exact-directory
+  grants for linked dependencies; keep credentials and MFA human-controlled.
+- Make chat detect incomplete profiles, open an explicit interactive setup
+  request, and verify Doctor before launching a new dashboard.
+
+- Give Windows tooling descendants one inherited hidden console while retaining
+  captured input/output, Job Object containment, and termination proofs.
+- Publish authorized pull requests after required local validation and approval,
+  verify draft status is cleared, and leave publication-triggered remote builds
+  pending rather than waiting for them before publication.
+
 ## 0.7.3 - 2026-10-05
 
 - Wait for the native keep-awake inhibitor to become ready before reporting

README.md

@@ -45,6 +45,19 @@ agency copilot
 For direct mode, install and authenticate GitHub Copilot CLI, then use
 `--runtime copilot` or set `FIXLAB_RUNTIME=copilot`.
 
+Select a direct Copilot model with `--model <model>` or `FIXLAB_MODEL`.
+For the account-dependent HydraFusion research preview:
+
+```powershell
+fixlab dashboard C:\path\to\application --runtime copilot --model hydrafusion --open
+```
+
+The Copilot dashboard also offers a per-job model selector. Leave it at
+Copilot default to avoid an override. Model selection is not supported by the
+Agency runtime; unavailable models produce runtime errors rather than silently
+switching models. Preview models may buffer stage updates, and FixLab does not
+guarantee lower token usage, charges, or elapsed time.
+
 ### 2. Install and initialize FixLab
 
 ```shell
@@ -81,6 +94,15 @@ For explicit plan-first onboarding without entering the interactive shell:
 fixlab onboard C:\path\to\application
 ```
 
+For a new worktree of an already configured repository, use
+`fixlab onboard C:\path\to\worktree --from C:\path\to\application`.
+Review the source profile and its `worktreeSetup.files` allowlist, then rerun
+with `--yes` to provision only those non-sensitive files and run setup.
+Onboarding also discovers external linked `node_modules` directories and
+requires explicit approval before recording their exact runtime path grants.
+Authentication and MFA remain interactive; credentials and browser state are
+never copied. See [Repository onboarding](docs/repository-onboarding.md).
+
 The first onboarding run creates the repository-owned profile and stops for
 review. After configuring that profile, rerun to inspect the setup plan, then
 approve dependency restoration, Playwright setup, browser authentication, and
@@ -187,7 +209,8 @@ visible risk. FixLab never reports a skipped or timed-out gate as successful.
   authentication, application startup, and configurable doctor timeouts before
   expensive work begins.
 - **Early draft pull requests** that become visible after focused regression
-  checks while independent builds and browser validation continue.
+  checks while independent local builds and browser validation continue, then
+  publish after local validation and approval without waiting for remote CI.
 - **Playwright evidence** including measurable assertions, screenshots,
   videos, traces, failures, explicit data-source behavior, connect-and-retry
   authentication, and reuse of an already started React frontend.

agents/fixlab.md

@@ -38,6 +38,15 @@ instructions provide authoritative commands.
 
 ## Workflow
 
+Before validation in a new isolated worktree, plan `fixlab onboard WORKTREE
+--from SOURCE --runtime RUNTIME`. After scoped setup approval, add `--yes`.
+The source profile declares non-sensitive configuration in `worktreeSetup.files`.
+Never copy credentials or browser state. Missing configuration, authentication
+or denied runtime access is a blocker; do not retry unchanged approvals or
+grant blanket path access. Sign-in and MFA remain human-controlled.
+In read-only modes, do not provision or alter configuration; report missing
+prerequisites as blocked.
+
 1. Read repository instructions and inspect the effective source state.
 2. Identify the smallest affected component and relevant existing tests.
 3. For code changes, use an isolated worktree when the current workspace is not
@@ -70,9 +79,24 @@ commit and push that checkpoint, and create or update a draft pull request.
 Run remaining independent builds, application checks, and browser journeys
 while the draft is open. Push later reviewed checkpoints to the same branch,
 keep validation status and risks current in the PR, and never mark it ready for
-review until every required gate passes or repository policy records an
+review until every required local gate passes or repository policy records an
 explicitly accepted outcome.
 
+After required local validation and approval gates pass or have an explicitly
+accepted outcome, publish the authorized PR by marking it ready for review.
+Do not wait for remote PR CI or build completion before publication; publication
+can trigger those checks. Verify the PR is no longer draft before reporting the
+PR stage passed. Report remote checks as pending, not passed, and leave merge
+policies intact. Use `gh pr ready <number>` for GitHub or
+`az repos pr update --id <id> --draft false` for Azure DevOps, then read back
+`isDraft` from the provider to confirm it is false. If publication or state
+verification fails, report the PR stage blocked or failed, not passed.
+
+Prefix FixLab-created pull-request titles with `FLAB:` followed by a concise
+description of the coding change. Use the repository's PR template and clear
+Markdown sections for the problem, changes, validation and risks. Keep
+validation checkboxes accurate; do not describe a pending gate as passed.
+
 Runtime-synthesized Playwright scenarios are transient validation artifacts by
 default. Remove their source files and validation-only configuration edits
 before review, commit, push, or pull-request creation. Do not add newly
@@ -92,6 +116,14 @@ behavior only rather than validation of the selected backend or its data.
 
 ## Safety boundaries
 
+Keep execution economical without weakening evidence: reuse verified session
+context instead of rediscovering unchanged repository facts, batch independent
+reads, and keep command output compact. Handle bounded work directly; delegate
+only independent tasks that justify separate context, not routine searches or
+one continuous investigation. On retry, rerun the failed gate and any affected
+prerequisites, not unrelated completed stages. Required checks still apply;
+never omit them to claim token savings.
+
 - Never select production automatically.
 - Never expose, copy, or commit credentials, tokens, customer data, or private
   test records.