feat(agent): strip trailing assistant prefill for proxy provider compat

[xuandung38]

Summary

Some LLM providers and models (via LiteLLM, OpenRouter, etc.) do not support assistant message prefill — requests ending with an assistant-role message get rejected with HTTP 400:

This model does not support assistant message prefill

This happens when the system injects trailing assistant messages to guide model behavior or establish context continuity.

Changes

• Backend: Add per-agent strip_assistant_prefill toggle (stored in other_config)

  • ParseStripAssistantPrefill() on AgentData reads the flag from other_config
  • New stripAssistantPrefill field on Loop / LoopConfig structs
  • Wired through resolver.go → loop initialization
  • Before constructing ChatRequest, if enabled, removes the last message when it has role: "assistant"

• Frontend: Checkbox in LLM Configuration section of the agent General tab

  • i18n support for EN, VI, ZH
  • State managed alongside other LLM config (provider, model, context window)

• Config: Added .gemini/ and .opencode/ to .gitignore (AI tool config dirs)

Files Changed

File	Change
internal/agent/loop_types.go	Add stripAssistantPrefill field + wiring in LoopConfig
internal/agent/resolver.go	Wire StripAssistantPrefill from agent config
ui/web/.../agent-general-tab.tsx	Add state + save logic for the toggle
ui/web/.../llm-config-section.tsx	Add checkbox UI component
.gitignore	Ignore AI tool config directories

When to Use

Enable strip_assistant_prefill on agents that use proxy LLM providers which reject assistant-role prefill messages. This is a per-agent setting — only enable it for agents connected to providers that exhibit this issue.

Test Plan

• Create agent with proxy provider (e.g., LiteLLM → GPT-4o)
• Enable "Strip Assistant Prefill" in LLM Config section
• Verify agent responds without HTTP 400 errors
• Verify agents with native Anthropic provider still work normally (no regression)
• Toggle saves correctly and persists across page reload

[viettranx]

Thank you for identifying this issue with proxy providers rejecting assistant prefill messages! Great catch.

After tracing the full message flow, we found the root cause: team task reminders (lead + member) were injected after the user message, leaving a trailing assistant message that proxy providers (LiteLLM, OpenRouter) reject with HTTP 400.

Instead of a per-agent config toggle, we addressed this at the source:

1. Reordered team reminders — injected before the user message so the user's message is always last (commit 11673da)
2. Added a safety net in the OpenAI provider to strip any trailing assistant message that might slip through in the future

This approach requires zero configuration — all agents using proxy providers automatically benefit without needing to toggle a setting.

Resolved in main via 11673da.

[xuandung38]

The approach makes sense, thanks sir.