[tang-vu]

Summary

Reload currently selected Library built-in manifests on each resources/restart request. Remove disabled/invalid entries immediately, preserve measured auxiliary containers without granting them authority, and recover on the next request after an unreadable selection.

Closes #7403. Issue: #7403. Reconstructs the independently useful archived #7264 scope on current upstream main c5b5533e9d81714dcdebf5b929765dfbdc998689. Candidate e058a7ac76ee45cd76dc8e3efcd69e1063b12192.

Why this matters

Dashboard resources and restart controls use import-time SERVICES while Library enables/disables built-ins during API uptime. An added service lacks restart authority; a disabled startup entry can retain it.

AI Assistance

The contributor and Codex collaborated. The contributor provided project direction, scope decisions and environment setup; Codex assisted analysis, implementation, tests and review preparation. Listed commands were executed by Codex. Specific human runtime verification is not inferred; independent maintainer/human release review remains required.

Release Lane

• Mainline change targeting main

Changed Surface

• Dashboard API / host agent

Risk And Validation

Risk: High under HIGH_RISK_CHANGE_MAP because this changes live service restart authority after Library selection changes. Native Windows x64, Python 3.14.2; Node 24.12.0/npm 11.6.2/Vitest 4.1.11 where applicable.

18 current-main regressions failed and one passed. Candidate: 47 resources/collection/auxiliary-container tests passed, including all 11 current Library built-ins, authenticated API and real host-agent HTTP restart handling. Docker outcomes are simulated.

• Diff check and touched Python Ruff
• Commit identity gate against exact base/head
• New API tests actually execute in existing dashboard.yml API whole tests job; Dashboard tests execute in its frontend job

Focused API command (from ods/extensions/services/dashboard-api):

python -m pytest -q tests/test_library_resource_restart.py tests/test_resources.py tests/test_resource_collection.py tests/test_auxiliary_container_resources.py

Operational Change Check

• Operational validation intentionally deferred: full fresh install, installed large-model/GPU, Portal, Docker/WSL/macOS/Linux runtime and release fleet qualification. HTTP/fixtures establish only the behavior described above.

Notes For Reviewers

Revert this isolated scope to roll back; preserve model artifacts, persisted configuration and recovery receipts.

Overlap check

Searched all states for Library restart eligibility and checked every live open PR file list: no open PR changes routers/resources.py. #7175 and other lean Library integration branches change installation/lifecycle selection producers; this fixes their live resources/restart consumer after selection changes. Existing #7263/#7390 are separate API template/Windows agent ownership outcomes. Closed #7264 is the archived predecessor, now reconstructed and tested across all eleven current Library built-ins.

Independent maintainer/human review is required before merge; agent review and local test results do not replace the installed/fleet release gate.

Combined integration validation

All three independently useful candidates were composed on base c5b5533e9d81714dcdebf5b929765dfbdc998689; exact combined head b301503f6c80a219a1f5b804f6e8096c50f3d333.

• Native Windows/Python 3.14.2 full Dashboard API suite: 6,567 passed, 60 skipped, 25 failed. All 25 failures reproduced on exact unchanged base. They are existing shell/Windows WSL bridge, symlink/FIFO privilege and filesystem/config fixtures; no new regression failed.
• Isolated Ubuntu 24.04.5 WSL2/Python 3.12.3 full API suite with locked service dependencies: 6,648 passed, 3 skipped, 1 failed. That same failure reproduced on exact unchanged base: an existing Darwin-unavailable fixture mocks Darwin telemetry but leaves real WSL telemetry available, so the handler correctly returns 200 instead of its fixture's expected 503. This is not claimed as a completely green full-suite run.

The 23 new API regressions pass in both full runs. No production change or exclusion was added to hide baseline failures. Individual focused checks remain green; hosted required CI and maintainer/human review remain separate. Linux source was a credentials-free exact Git archive in a dedicated QA directory; the installed product runtime was not changed. These combination receipts establish test interactions only, not installed model/GPU/Portal or fleet qualification.

CI follow-up

Candidate e058a7ac76ee45cd76dc8e3efcd69e1063b12192 adds only explicit fixture types after CI's mypy ratchet identified untyped fixture state. Production files are byte-for-byte equivalent to independently reviewed 898e6069948b658c3789de4c0781ba0c755b9ea3; earlier full-suite/combined receipts apply through that production equivalence. Root independently reviewed the typing-only diff. The affected resource/update checks passed again (49 Linux tests), touched-file Ruff and commit identity gates passed, and the CI dependency-only mypy 2.4.0 preflight targeting Python 3.11 reports 426 established errors, 0 new errors (baseline unchanged). The test-runtime environment includes pytest's typed fixture declarations, so it is kept separate from the workflow's dependency-only type-checking environment.