Summary

Addresses #6318 (the "manually manage the base model" half). Lets you set a model id that isn't in the catalog when editing an existing agent, so an agent-type provider (codex-cli / claude-code / gemini-cli) pointed at a third-party base can be configured from the dashboard.

Background

• The driver layer already passes any model through: each agent driver's model_flag (crates/librefang-llm-drivers/src/drivers/{codex_cli,claude_code,gemini_cli}.rs) falls through _ => Some(stripped.to_string()), so an unknown model id reaches the CLI verbatim. There is no hard rejection — this is not a backend bug.
• The third-party base itself is selected via the CLI's own env (OPENAI_BASE_URL, ANTHROPIC_BASE_URL, …), which LibreFang doesn't manage.
• The create form (AgentManifestForm) already drops to a free-text <input> when the provider has no catalog models, so custom ids were enterable there.
• The gap was the agent detail drawer's model editor: a plain <select> populated only from catalog models, showing a disabled "No models" option for agent-type providers (whose catalog is empty) — so you could not change an existing agent to a custom / third-party model from the UI.

Change

crates/librefang-api/dashboard/src/pages/AgentsPage.tsx — the detail-drawer model field becomes an <input list=…> + <datalist> combobox. It still suggests the provider's catalog models, but now accepts a typed-in id. Provider-not-selected / loading states move to the input's placeholder; the disabled "No models" dead-end is gone. The existing auto-select-single-model effect (#5917) and the per-id save flow are unchanged.

Removed the now-unused agents.no_models i18n key from en/zh/uk (locale parity preserved; the dead-key coverage test enforces this).

Out of scope

The other half of #6318 — auto-detecting the model an agent CLI actually ran — is a separate enhancement. codex-cli already does this via parse_model_from_banner; claude-code's JSON output carries a model field that isn't parsed yet. Left for a follow-up so this UI fix can land independently.

Verification

In crates/librefang-api/dashboard/:

• pnpm typecheck — clean
• pnpm lint — 0 errors (warnings are pre-existing baseline)
• pnpm test --run — 838 passed (85 files), including the locale dead-key + parity gates
• pnpm build — succeeds

No render-test was added for the model editor: AgentsPage.test.tsx documents that the page has no render harness (~20 hooks) and tests the extracted SystemPromptSection directly. The change is a <select>→<input list> swap with identical value/onChange/disabled wiring, covered by typecheck + build.