CLI model catalog and manual metadata

Эта тема ещё не переведена на русский; отображается английский оригинал. Читать оригинал на английском

omniroute models reads the selected server’s public model catalog first, including its custom and synchronized entries. --output json and --output jsonl return all matching rows, without a table truncation message. Interactive tables show at most 50 rows. JSON preserves public context limits, modalities, capabilities, model source when supplied by the server, and API-format metadata; credentials and compatibility headers are not output. A legacy /api/models fallback is used only when the public endpoint returns HTTP 404, 405 or 501, never after an authentication failure. omniroute completion refresh uses that same catalog. A failed model-catalog request returns a nonzero exit code and preserves the previous completion cache. Bash, Zsh and Fish completion scripts advertise the manual model subcommands.

Manual model lifecycle

These operations use the existing authenticated /api/provider-models management API. They honor the same global --context and --base-url selection as other CLI commands. Use the canonical provider ID and the provider’s model ID separately; a model ID containing slashes is valid. Use a configured management context rather than putting credentials into command arguments.

omniroute models openai --output json
omniroute models manual openai
omniroute models add openai my-model --name "My model" --context-window 32768 --dry-run
omniroute models add openai my-model --name "My model" --context-window 32768
omniroute models edit openai my-model --context-window 65536
omniroute models edit openai my-model --clear-context-window
omniroute models remove openai my-model --dry-run
omniroute models remove openai my-model --yes

Examples describe syntax; replace provider/model IDs with entries belonging to your server. The add operation does not create a provider connection.

• add refuses an existing custom entry. Optional --max-output-tokens sets the output limit at creation; --api-format sets the model’s API format and its matching endpoint metadata, such as embeddings rather than chat.
• edit requires an existing manual entry. Its context limit uses the backend’s independent manual context override; clearing it removes that override, not the original input limit or synchronized metadata.
• remove requires --yes, except with --dry-run. It requests resetOverride=true, preserving a synchronized entry with the same ID. Bulk deletion is not exposed.
• --dry-run performs the authenticated preflight read but no write or inference.
• Every actual mutation reads the manual catalog back. persistenceVerified: true means the requested metadata was observed after the write; it does not mean the model can generate text or call tools. inferenceValidation is always not-run.
• A readback failure returns a nonzero exit code. The write may already have taken effect; inspect the server before retrying. Concurrent writers are not serialized by this CLI, and the existing API does not provide a compare-and-swap transaction.

Test & Add: not implemented

Manual entries are unverified. There is no --validate flag and these commands do not issue inference requests or execute tools. A future opt-in validation flow must separate inference credentials from management authentication and establish which exact provider connection produced all three proofs: generation, a nonce-bound synthetic tool call, and continuation after a synthetic tool result. No real tool execution is needed for those proofs.

The existing x-omniroute-connection pin alone is insufficient as a proof receipt: credential selection can intentionally release a pin for cooldown, quota or excluded-account cases, and session affinity participates in selection. See src/sse/services/auth.ts and the forcedConnectionId handling in src/sse/handlers/chat.ts. Before implementing Test & Add, decide a strict-pin or verifiable connection-receipt contract. Tests must then prove no persistence after any failed, truncated, timed-out or wrong-connection response. This is a separate backend contract decision, not a promise made by manual CRUD.

---

Исходный код OmniRoute (a58000c7685f)