v0.5.0

What's New

ollamaMQ now supports decision-model backends ("System One" models) alongside Ollama and
LM Studio. These are non-generative servers — clm-serve (CLM),
laya-serve (Laya), kev.serve (Kev)

• that answer POST /v1/systemone with typed decisions (choice / score / noul) and calibrated
  probabilities instead of text.

They share no /api/* route with Ollama, so this required a real third API family rather than
another backend URL. Requests are still forwarded byte-for-byte; nothing is translated.

Added

• POST /v1/systemone is proxied, and routed only to backends of this family. /api/chat,
  /api/tags and /v1/chat/completions are never sent to them, and they never receive a route
  their family cannot serve - such a request is refused, not misrouted.
• Automatic detection of CLM and Kev from GET /v1/models answering {"models":[{"name":…}]}
  (OpenAI-compatible servers answer {"data":[{"id":…}]}, so the two are unambiguous).
• kind: systemone per-backend config key to pin the family, needed for Laya, which exposes no
  listing endpoint and cannot be positively detected.
• GET /health probe fallback - a 2xx marks a backend online when no other route identifies it.
  /health stays probe-internal and is never a proxied route.
• models: entries now make a name routable, not merely pinned. This is the only way to reach a
  backend that cannot list its models.
• Model listing falls back to proxy state. A backend that does not implement the listing route
  now contributes its known model names, and a listing request returns 200 instead of being passed
  through to a backend's own 404.
• Per-backend token: works for this family out of the box - all three servers use
  Authorization: Bearer (CLM_API_KEY / LAYA_API_KEY / KEV_API_KEY).

Changed

• GET /v1/models now always answers in OpenAI's data/id envelope, regardless of which
  backend replied first; GET /api/tags always uses models/name. Previously the first answering
  backend decided, so adding a decision-model backend could strip data from the reply and break
  OpenAI SDK clients.
• Decision-model names appear in the aggregated /v1/models and /api/tags listings even though
  only /v1/systemone can serve them. ollamaMQ aggregates one namespace and routes by capability.
• Model load/unload is explicitly not available for this family - these servers expose no model
  control API, no keep_alive and no /api/ps; residency is configured on the server itself
  (e.g. LAYA_PRELOAD, LAYA_MAX_LOADED).

Fixed

• Startup no longer waits the full readiness window before applying model config when every backend
  is a decision-model server.
• A pinned backend family and declared model names no longer wake the scheduler on every health round.

Known limitations

• Only /v1/systemone is proxied. /v1/rank (CLM) and /v1/systemone/permute, /separate (Kev)
  are not.
• A 404 from a backend is not yet treated as a family rejection, so an unclassified backend remains
  a candidate for routes it cannot serve.
• The end-to-end suite does not yet cover this family (it is covered by unit-level mocks).

Upgrade notes

No config changes are required for existing Ollama / LM Studio / OpenAI deployments — their behaviour
is unchanged and pinned by tests. To add a decision-model backend:

backends:
  - url: "http://127.0.0.1:8700"          # CLM or Kev: auto-detected
  - url: "http://127.0.0.1:8000"          # Laya: must be pinned
    kind: systemone
    token: "laya-key"
models:
  - name: multilingual
    backends: ["http://127.0.0.1:8000"]   # selector required with more than one backend

v0.4.2

• rework of the dispatcher, kudos to @miroslavbaudys

v0.4.0

• New features: appconf.yaml config file, model load/unload control (TUI + admin API), API-family-aware scheduling with remembered bad endpoints, fail-fast 503, enforced per-model concurrency limits
• Changed: deterministic bounded model matching (with a heads-up for clients that relied on substring matching), Docker pinning + BACKEND_URLS docs
• Fixed: hanging queue, Ollama -> LM Studio misrouting
• New config keys table with defaults

v0.3.0

• added modal controls via Ollama & LM Studio API
• U for Unload model
• L for Load model

v0.2.8

Added

• --host / -H flag - choose the interface to bind to. Default is now 127.0.0.1 (loopback only).
• Optional API-key authentication - set OLLAMA_MQ_API_KEY to enable. When set, all proxy routes require the key via Authorization: Bearer or X-API-Key: ; missing or wrong keys get 401 Unauthorized. /health stays unauthenticated for healthchecks. Key comparison is constant-time.
• Docker: new HOST environment variable (defaults to 0.0.0.0 in containers, preserving external reachability) and OLLAMA_MQ_API_KEY support via -e.

Changed

• Default bind address changed from 0.0.0.0 to 127.0.0.1 - the binary no longer accepts connections from other machines unless you explicitly pass --host 0.0.0.0. If you relied on LAN access (e.g. via the systemd unit), add --host 0.0.0.0 to ExecStart and consider setting OLLAMA_MQ_API_KEY.

Security

• The proxy is no longer exposed to the local network by default, closing the unauthenticated-access gap where anyone on the LAN could submit inference requests with a plain curl.
• RFC 7235-compliant 401 responses (WWW-Authenticate: Bearer, Content-Type: application/json); Bearer scheme matching is case-insensitive.

v0.2.7

• add lmstudio support
• add detection of AI backend instances
• add list of backend's models
• add active / last used / loaded model(s)

v0.2.6

• fixed docker composer to support multiple ollama backends

v0.2.5

• add all endpoints for ollama
• add flag to allow all request into a proxy --allow-all-routes for example to claude code local use
• add support multiple ollama urls backends

v0.2.4

• update TUI
• add option to BOOST user requests 1:5
• add option to have WIP user (every request is priority)

v0.2.3

• add extra ollama api to support various ollama/openai client @wuhongzhi
• fix queued requests counter
• blocking / unblocking users in main list
• fix always active list selected index
• fix gracefull quit