fix(advisor): five real-world bugs blocking client-side mode

[agentforce314]

Summary

Real-world bug fixes for the client-side advisor mode shipped in #182. The original PR passed all unit tests but failed when actually exercised against a live provider (haiku-4-5 main + opus-4-7 advisor, both routed through litellm.singula.ai). Each bug below was caught during that live verification.

The five bugs

1. `agent_loop.py` had zero advisor wiring
Phase-03 (#179, landed via #181) wired the advisor into `_call_model_sync`, but the REPL, TUI, and headless paths all route through `run_agent_loop` — not `query.query`. The advisor was effectively dead code outside the subagent path. Fix: replicate the activation predicate + schema/instructions injection + ctx snapshot inside the `run_agent_loop` per-turn loop.

2. Duplicate `advisor` tool name in tools[]
`agent_loop.py` iterated `tool_registry.list_tools()` which doesn't filter `is_enabled()` — so the hidden `AdvisorTool` leaked into the base list. When client-side mode then appended the advisor schema, the API got two entries with name `advisor` → `BadRequestError: Tool names must be unique`. Fix: filter by `is_enabled()` at both the agent_loop build site and (belt-and-suspenders) in `_call_model_sync`.

3. Provider constructor rejected `default_model` kwarg
`execute_client_advisor` passed the raw config dict from `get_provider_config()` straight to the provider constructor. That dict uses `default_model`; the constructor expects `model`. Crashed `OpenAIProvider.init()`. Fix: explicit key translation (`api_key`, `base_url`, `model` only).

4. Forwarded conversation ended with assistant message
The advisor is invoked from inside a `tool_use` block, so the natural tail is the assistant message that contains it. Vertex-fronted Anthropic (via litellm) rejects this: "This model does not support assistant message prefill. The conversation must end with a user message." Fix: append a synthetic user "please advise" turn when needed.

5. `tool_use`/`tool_result` blocks in forwarded messages
The advisor is called with `tools=[]` (it just emits text). But proxies reject `tool_use`/`tool_result` blocks when no `tools=` array is sent: "Anthropic doesn't support tool calling without tools= param specified". Fix: flatten all tool blocks to `[Tool call: ...]` / `[Tool result: ...]` single-line text summaries. Preserves substance without the typed schema. Drops `thinking` blocks too.

Verified end-to-end

Live run config:

• Main: `claude-haiku-4-5` via `openai` provider pointing at `litellm.singula.ai`
• Advisor: `claude-opus-4-7` via the same proxy (auto-detected by the "main is proxy → use main provider for advisor" routing from feat(advisor): client-side mode for 3P providers (any provider, any model) #182)
• Task: scaffold a Vite + TypeScript + React + react-router-dom blog

Observed flow:

1. Worker called advisor BEFORE writing code
2. Advisor replied with concrete planning guidance (package.json deps, tsconfig settings, router structure)
3. Worker wrote 9 React blog files
4. Worker called advisor again at the end
5. Advisor caught that the first Write batch had empty content placeholders and instructed the worker to redo with actual content
6. Worker redid; final state: 9 working files on disk at `/tmp/blog-test/`

Test plan

• tests/test_advisor_helpers.py — 32 passed
• tests/test_advisor_orphan_pairing.py — 6 passed
• tests/test_advisor_chat_response_roundtrip.py — 6 passed
• tests/test_advisor_command.py — 14 passed
• tests/test_advisor_request_wiring.py — 13 passed
• tests/test_advisor_client_side.py — 34 passed (2 new tests for proxy routing + fallback)
• tests/test_app_state.py — 16 passed
• tests/integration/test_advisor_smoke.py — 3 passed
• Total: 124 advisor tests + tool_system/settings regression all green

🤖 Generated with Claude Code