[tang-vu]

Summary

Closes #7298. Windows ods.ps1 chat currently bypasses an enabled switchboard and sends neither the native llama-server credential nor UTF-8 prompt bytes. Use ods/current with the saved gateway port/key when the switchboard is enabled; otherwise honor the resolved native endpoint's LLAMA_SERVER_API_KEY and saved GGUF alias. Missing required keys refuse the request. Failed requests remain on the selected route.

This current-main maintenance removes the retired Lemonade branch and replaces its obsolete wire tests. Docker llama-server retains the existing default-model behavior, configured port/API path, and no native-key forwarding.

Why this matters

The shipped Windows CLI is a real client of the configured inference route. On current main, ods.ps1 chat bypasses a selected switchboard; direct native requests omit the required credential, and Windows PowerShell 5.1 string serialization can corrupt Unicode prompts. This restores that public command across the supported current endpoints without bypassing a failed selected route.

Overlap check

Reviewed current main's endpoint resolver and native llama startup contract, plus the existing PR's original scope. This updates #7214 in place; it does not reintroduce the retired Lemonade runtime. The other maintained tang-vu model/UI PRs concern download staging and remote-peer inventory, not Windows CLI request routing. Existing endpoint/readiness contracts remain intact.

Validation

Base: 36b950130c7378c83946213569c3766d2e2b3c8b. Candidate: f6409e57d8283250dce5647694e310b6c7675a98.

The original published head 3273c97e8e8db6d60b662e03a907772242605bf3 is preserved as an ancestor through a normal main merge. Its eight historical wire tests fail five retired-Lemonade assumptions after integration onto this main. The first eight replacement regressions fail seven cases on exact unmodified main, including missing routing/auth and Unicode handling; the unchanged Docker control passes.

Agent-executed native Windows checks on the candidate source (including an independent root replay of all nine HTTP-wire cases at this exact candidate):

• 9 PowerShell 5.1 loopback HTTP-wire cases passed: gateway alias/key, native alias/key/Unicode, missing-key refusals, HTTP failure without route bypass, default Docker behavior, configured Docker port/path, and no native key leakage to Docker.
• Windows local endpoint resolver contracts passed.
• Native llama-server legacy lifecycle contracts: 86 checks passed with process/task/socket fixtures.
• Model-readiness shell contract: 16 checks passed, including the missing-model-file behavior under portable PowerShell 7. The earlier PATH-only run skipped that behavior; the final root replay used the available portable runtime.
• Ruff E/F/W with CI E501 exclusion, whitespace, and all three added author/committer identity checks passed.

Existing Windows workflow explicitly executes test-windows-cli-chat.py; no test-only file is left unexecuted. Root also ran the available portable PSScriptAnalyzer against the CLI with the exact CI settings: zero warnings/errors. Hosted current-head checks are monitored separately.

AI Assistance

Tang Vu (tang-vu) provided project direction, scope decisions and environment setup. Codex assisted source analysis, implementation, isolated tests and review preparation. Commands above were agent-executed; personal human verification or independent maintainer release acceptance is not inferred.

Release Lane

• Mainline change targeting main

Changed Surface

• Installer / bootstrap / lifecycle
• Model routing / Hermes / capabilities

Risk And Validation

Risk: High (Windows CLI model routing/authentication). Loopback fixtures exercised the public CLI without touching an installed service, customer configuration or real model. Full installed model/GPU/Portal, fleet, and independent human/maintainer release approval remain deferred. The command's existing error-reporting/exit convention is unchanged. Rollback is reverting this source scope; no stored data migration is introduced.

Operational Change Check

• Operational validation is intentionally deferred for installed live model, GPU, Portal, hardware/fleet and independent human release lanes. Local HTTP-wire and isolated lifecycle results are recorded above.