Skill Forge rewriter/gate ignore agents.defaults.model, fall back to region-blocked provider default

[arelchan]

Summary

Skill Forge's auxiliary LLM calls (the query rewriter and the LLM gate)
do not use the configured agent model (agents.defaults.model). When no
Skill-Forge-specific model is set, they fall through to the provider's
hardcoded default_model, which for the openrouter provider is
openrouter/anthropic/claude-sonnet-4-5. If that default model is unavailable
(e.g. region-blocked), the rewrite/gate step hard-fails even though the agent is
configured with a working model -- so the main turn succeeds while
skill.rewrite / skill.gate error out.

Evidence (from tracing)

Config: agents.defaults.model = "minimax/minimax-m3", provider openrouter.

In one turn, the trace shows:

• llm.call with invocation_source = session.turn (main reasoning) ->
  llm.model = minimax/minimax-m3, status OK.
• llm.call with invocation_source = skill.rewrite ->
  llm.model = openrouter/anthropic/claude-sonnet-4-5, status ERROR:
  OpenrouterException - {"error":{"message":"This model is not available in your region.","code":403}}.

So the rewriter used a different model than the configured minimax/minimax-m3,
and that model is region-blocked. The rewriter then falls back to
need_retrieval = True with an empty rewritten query.

Root cause

• raven/context_engine/factory.py: QueryRewriter(provider, max_tokens=...)
  is constructed with no model; LLMGateFilter(..., model=skill_forge.llm_gate_model or None, ...)
  is None unless llm_gate_model is set.
• QueryRewriter.analyze / LLMGateFilter.filter call
  provider.chat_with_retry(...) without a model, so they use
  provider.default_model.
• raven/providers/registry.py: the openrouter provider spec hardcodes
  default_model = "openrouter/anthropic/claude-sonnet-4-5".

The main agent loop passes model=agents.defaults.model explicitly on every
call, so it is unaffected; only the auxiliary Skill-Forge calls hit the
hardcoded default.

Impact

• Query rewrite and skill gate fail whenever the provider's hardcoded default
  model is unavailable/region-blocked, regardless of the configured agent model.
• The rewriter has no config override (unlike the gate's llm_gate_model),
  so this cannot be worked around by configuration alone.

Proposed fix

Make the Skill-Forge auxiliary calls inherit agents.defaults.model when no
Skill-Forge-specific model is configured:

• Thread the configured agent model into QueryRewriter / LLMGateFilter in
  factory.py (add a model parameter to QueryRewriter).
• Pass that model on the chat_with_retry calls, so rewrite/gate use the same
  working model as the main loop by default; a Skill-Forge override (e.g.
  llm_gate_model) still wins when set.

Workarounds (until fixed)

• Gate: set skill_forge.llm_gate_model to a model available in your region
  (e.g. minimax/minimax-m3).
• Rewriter: no config knob; either disable it (skill_forge.rewrite_enabled = false)
  or apply the code fix above.