llm: routes that name a host, a backup and a window, and a profile that reads them ...

`llm.json` replaces `OLLAMA_BACKENDS_FILE`. It has targets (host and key
environment variable) and routes, where a route is an ordered chain of
`model @ target` entries, each with its own `context_window` and options. The
`default` route now starts on `glm-5.3-flash:cloud` and backs it with
`qwen3.6:35b-128k` on the home host, both at 131072 tokens, so a failover no
longer changes the budget the compressor and `num_ctx` are accounted against.
Embeddings get their own `embedding` route instead of a pinned host.

The engine classifies a failure (network, quota, auth, server, missing model),
cools that entry and moves the call down the chain; a background probe watches
only degraded targets and lets the next turn return to the primary once it is
healthy — it never interrupts a running turn. A profile joins a route with
`llm_route`, and the loader now actually reads it: `llm_route` and
`context_window` were declared on `AgentProfile` and written into every
`config.json` while the loader skipped both, the same silent no-op that
`is_admin_only` used to be, so every profile ran the legacy model list at the
global window. The writer emits them too, or an admin edit would erase the
route from disk.

The turn's resolved entry travels to the UI: the first chunk is stamped with
model and target, `model_info` carries `fallback` when a backup answered, and
the chat header shows a chip only then. `serving_primary` compares the target
as well as the model name, because a route may run one model on two hosts.
`/health` gains an `llm` block with the active entry, target states and missing
models.

`deploy/ollama/` keeps the local model's definition honest: MTP off (slower at
every context on this host), `presence_penalty 0` (the base model's 1.5 breaks
agentic tool calls), and `num_gpu 40` — Ollama's automatic fit claims a 128k
f16 slot fits and then dies mid-prefill with CUDA OOM; 40 of 42 layers leaves
~400 MiB and decodes at ~77 tok/s, 38 layers collapses to 48, and 39 OOMs.
q8_0 KV measured 37 tok/s against 70 on f16 here, so the cache stays f16.

Navi Code: force context compression via typed /compact control message ...

Forced /compact previously sent a chat message, which ran a full agent turn
that only produced summary text instead of running the real context compressor.
Worse, even when wired correctly, forced compact always reported "Nothing to
compact yet — the context is still small" regardless of context size: the typical
navi_code shape is a single long autonomous turn (1 user message + many tool
iterations = one turn), and partition_messages finds nothing to summarize when
turns <= keep_recent. The midturn auto-compress path already bypassed this via
keep_recent_messages (intra-turn split), but compact_stream passed
keep_recent_messages=None, so the fallback was disabled.

Changes:
- WS protocol: {"type":"compact"} control message (distinct from {"type":
  "message"}); rejected while an agent turn is active to avoid racing the agent.
- Agent.compact_stream: forced compression that bypasses the token threshold
  but still runs the real compressor; passes keep_recent_messages=max(12,
  context_keep_recent*2) so a single long turn compresses via intra-turn split
  (mirrors midturn auto-compress). Raises NothingToCompactError when context is
  genuinely too small.
- Orchestrator.run_compact + clear_run: broadcast agent events to subscribers,
  end with done marker, surface NothingToCompactError as an error event.
- Terminal client: ws_client.send accepts str|dict; CompactCommand enqueues
  {"type":"compact"}; TUI distinguishes forced compact (no stream_start) from
  in-turn auto-compress via the _streaming flag.
- Tests: compact_stream (incl. single-long-turn regression), WS handler
  dispatch/rejection, run_compact event/error broadcasting, ws_client send,
  compact command.

Co-Authored-By: Claude <noreply@anthropic.com>

Add Ollama multi-server fallback with in-memory blacklisting ...

- New FallbackOllamaBackend (navi/llm/fallback.py): tries servers and
  models in priority order; on LLMConnectionError blacklists the server
  for the process lifetime, on LLMModelNotFoundError blacklists the
  (server, model) pair — eliminates latency from repeated failed probes
- OllamaBackend now raises typed LLMConnectionError / LLMModelNotFoundError
  instead of bare LLMBackendError; accepts list[str] | str | None for model
- AgentProfile.model changed from str to list[str] (str auto-normalised);
  all profiles updated to ["gemma4:31b-cloud", "gemma4:26b-a4b-it-q4_K_M"]
- New config field OLLAMA_BACKENDS_FILE: path to [{host, api_key?}] JSON;
  when set, registry creates FallbackOllamaBackend instead of OllamaBackend
- ollama_backends.json template added (gitignored — contains API key)
- current_model ContextVar type widened to list[str] | str | None

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Fix LLM hang: stop button during prefill, context guard, timeouts ...

Root cause: during prefill (processing input tokens), Ollama emits no
HTTP chunks. The `async for chunk in stream_complete()` loop body never
executes, so stop_event is never checked — Stop button has no effect.
Same issue with complete() calls (planning, compression): blocking await
with no cancellation path.

Fixes:

_iter_stream_guarded() (agent.py, module-level):
  Wraps any stream_complete() generator. Polls stop_event every 1s while
  waiting for the next chunk using asyncio.wait() — so Stop works even
  during multi-minute prefill. On stop or timeout, calls aclose() on the
  generator which closes the HTTP connection to Ollama → generation halts
  → GPU drops to idle. Applied to both run_stream() and run_ephemeral().

_check_context_size() (Agent method):
  Estimates context tokens (chars/4 + 500 per image) before every LLM
  call. Raises ContextTooLargeError (new NaviError subclass) at 92% of
  ollama_num_ctx — before Ollama ever receives the request.

_run_planning() timeouts:
  Both complete() calls (phase 1 and 2) wrapped with asyncio.wait_for().
  Timeout logged and planning skipped gracefully — execution continues.

New config (config.py):
  llm_complete_timeout = 120s
  llm_stream_first_chunk_timeout = 180s  (prefill budget)
  llm_stream_chunk_timeout = 60s         (inter-token budget)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Initial implementation of the agent system core ...

- FastAPI server with REST API and WebSocket streaming
- Modular LLM backend abstraction (Ollama implemented, OpenAI stub)
- Tool system: web_search (ddgs), filesystem, http_request, code_exec, terminal
- Agent profiles: smart_home, server_admin, secretary
- Tool-calling loop with concurrent tool execution
- In-memory session store with SessionStore ABC for future persistence
- Registry pattern for tools, profiles, and backends
- Orchestrator stub as foundation for multi-agent scenarios