Skip to content

Add OpenCode GO & vLLM Routing#6

Open
giraudremi92 wants to merge 1 commit intostevibe:mainfrom
giraudremi92:main
Open

Add OpenCode GO & vLLM Routing#6
giraudremi92 wants to merge 1 commit intostevibe:mainfrom
giraudremi92:main

Conversation

@giraudremi92
Copy link
Copy Markdown

Add OpenCode GO & vLLM Routing

Summary

Added support for two new providers in ToolCall-15:

  • OpenCode GO (opencode) - Connect to OpenCode GO agents
  • vLLM (vllm) - Connect to local vLLM instances for benchmarking local models

Changes

  • lib/models.ts - Added opencode and vllm to ProviderName type and routing logic
  • .env.example - Added OPENCODE_GO_URL, OPENCODE_API_KEY, and VLLM_HOST environment variables

Usage

OpenCode GO

OPENCODE_GO_URL=https://opencode.ai/zen/go/v1
OPENCODE_API_KEY=your_key
LLM_MODELS=opencode:agent-model

vLLM

VLLM_HOST=http://localhost:8000
LLM_MODELS=vllm:Intel/Qwen3.5-122B-A10B-int4-AutoRound

Testing

npm run typecheck
npm run lint

juliusalba added a commit to juliusalba/agentic-toolcall that referenced this pull request Apr 7, 2026
CRITICAL fixes:
- scoreModelResults() now accepts scenarioPool parameter — enterprise
  and memory suites were silently scoring 0% across all categories
  because it was hardcoded to base SCENARIOS (audit finding stevibe#1)
- TC-06: Added partial credit for translating 1 of 2 languages;
  relaxed text matching from exact-string to includesText (finding stevibe#2)
- TC-09: Datetime matching now accepts multiple ISO formats —
  "2026-03-21T08:00", "08:00:00Z", etc. (finding stevibe#3)
- EC-12: Cron validation now parses 5 fields properly instead of
  loose string.includes() that matched false positives (finding stevibe#5)

IMPORTANT fairness fixes:
- TC-03: Models that verify common knowledge with tools now get
  partial credit (1pt) instead of fail (0pts) — removes bias against
  safety-trained models (finding #9)
- MR-10: Expanded admission phrases from 5 to 13 patterns for
  "I don't know" detection (finding stevibe#6)
- MR-11: Contradiction detection no longer uses fragile exact string
  match; uses regex pattern matching (finding stevibe#7)
- EC-02: Cursor detection now searches all arguments, not just
  URL+body+headers concatenation (finding #10)
- EC-05: Turn ordering now allows same-turn parallel calls with <=
  instead of strict < (finding #11)
- MR-04: Requires 2 of 3 key facts (Acme, $50M, Series C) instead
  of just 1 (finding #15)
- EC-03: MCP tool name matching tightened from includes("issue") to
  specific create-issue variants (finding #19)
- MR-07: Generic queries now return partial results to reward
  targeted queries (finding #20)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant