feat(advisor): client-side mode for 3P providers (any provider, any model)

[agentforce314]

Summary

• Extends /advisor to work on any provider (GLM, Gemini, OpenAI, OpenRouter, Bedrock-shimmed Anthropic, etc.) by adding a client-side execution path. The server-side Anthropic beta stays the optimized default when applicable; client-side activates for 3P providers, non-server-side advisor models, or when forced via --client.
• New flags: /advisor <model> --client (force client-side), /advisor --no-client (clear), /advisor unset (clear all).
• Now also supports cross-provider pairings: e.g. GLM main loop with Anthropic Opus advisor, or Anthropic main with Gemini advisor.

Mode-decision table

Provider	Force --client	Advisor model	Mode
1P Anthropic	off	opus-4-6 / sonnet-4-6	Server-side (unchanged, optimal)
1P Anthropic	on	any routable	Client-side
1P Anthropic	off	non-server-side (e.g. haiku, gemini)	Client-side
3P	n/a	any routable	Client-side
any	any	unroutable model	Inactive

What changes

Activation

• decide_advisor_mode(provider, main_model, advisor_model, force_client_mode) returns server_side / client_side / inactive.
• infer_provider_for_model(model) does exact-match in PROVIDER_INFO, then well-known prefix fallback (claude-* → anthropic, gemini-* → gemini, zai/* → glm, <vendor>/* → openrouter, etc.). Rejects ambiguous strings rather than guessing.

Client-side path

• AdvisorTool registered in src/tool_system/tools/advisor.py with is_enabled=lambda: False (hidden from /tools listings; registry still finds it for dispatch).
• _call_model_sync injects the regular-tool-shape schema only when client-side mode is active for the request. Cache-friendly: appended after the regular tools and after the cached system blocks.
• AdvisorTool.call(input, ctx) reads ctx.messages, strips prior advisor blocks via build_advisor_forwarded_messages, calls execute_client_advisor.
• execute_client_advisor resolves the advisor's provider via the existing get_provider_class + get_provider_config (reuses user-configured API keys), makes one non-streaming chat() call with the advisor's own system prompt and empty tools, returns (ok, text_or_error). Never raises — degrades to a tool_result error on failure.

/advisor command

• Drops strict is_valid_advisor_model rejection (the opus-4-6/sonnet-4-6 server-side gate). Now only requires the model to route via infer_provider_for_model.
• Parses --client / --no-client flags (order-insensitive).
• Status output shows the actual mode picked: Advisor: <model> — active (server-side|client-side|inactive) [--client forced?].

Settings + state

• SettingsSchema.advisor_client_mode: bool = False
• AppState.advisor_client_mode with _on_advisor_client_mode_change (persists to global settings + invalidates cache, matches the existing advisor_model handler pattern).

TUI

• AgentBridge._emit_advisor_events extended: also detects client-side tool_use(name=\"advisor\") on assistant messages plus the matching tool_result on subsequent user messages. Emits the same AdvisorEventMessage types so the existing widget renders both flavors.

Loosened gates

• can_user_configure_advisor: was 1P-only, now env-only. /advisor works on any provider.

Test plan

Verified locally (/Users/ericlee2/workspace/clawcodex/.venv/bin/pytest):

• tests/test_advisor_helpers.py — 32 passed (1 updated: 3P provider now allowed)
• tests/test_advisor_orphan_pairing.py — 6 passed
• tests/test_advisor_chat_response_roundtrip.py — 6 passed
• tests/test_advisor_command.py — 14 passed (4 updated for new semantics)
• tests/test_advisor_request_wiring.py — 13 passed (3 inactive cases now correctly activate client-side)
• tests/test_advisor_client_side.py — 30 new tests (mode decision table, model routing, forwarded message stripping, execute_client_advisor success/failure/empty, AdvisorTool call paths, schema shape)
• tests/test_app_state.py — 16 passed
• tests/integration/test_advisor_smoke.py — 3 passed
• tests/test_tool_system_tools.py + test_tool_registry_pipeline.py — 99 passed (regression)
• tests/test_settings*.py + tests/test_query*.py — 70 passed (regression)
• Total: 120 advisor tests + 169 collateral tests passed

Design choices

1. Server-side stays the optimized path — when 1P + valid models, no behavior change. One roundtrip, prompt-cache friendly, beta-header gated.
2. Cache stability preserved — both modes append schema + instructions at the end. Toggling /advisor doesn't churn the cached prefix.
3. Reviewer doesn't see its own prior consultations — build_advisor_forwarded_messages strips advisor blocks before forwarding. Each consultation is independent.
4. Failure mode is degradation — execute_client_advisor catches every exception and surfaces as is_error=True tool_result. The turn doesn't die.
5. Reuses existing provider config — no new API key surface. The advisor's provider goes through the same registry as the main loop.

🤖 Generated with Claude Code