Multi-provider execution: non-Anthropic LLMs don't work end-to-end

[kokevidaurre]

Problem

squads run supports 7 LLM providers (Claude, Gemini, Codex, Vibe, Grok, Aider, Ollama) via llm-clis.ts, but only Claude actually works because:

1. Agent definitions are Claude Code prompts — they reference Claude-specific tools (Read, Edit, Bash, Task), MCP servers, and Claude Code conventions
2. --dangerously-skip-permissions is Claude Code-specific — other CLIs don't have it
3. buildArgs patterns are guesses — e.g., gemini --prompt may not be the actual CLI interface
4. No validation that agent prompts are compatible with the target provider

Current State (v0.6.0)

• executeWithProvider() has worktree isolation (fixed in c2df44b)
• All squad model configs use Anthropic models (haiku/sonnet/opus)
• The provider routing code exists but is untested with real providers

What v0.7.0 Needs

1. Validate provider CLIs actually work — install and test each
2. Prompt adapter layer — translate Claude-specific patterns to provider-compatible prompts
3. Tool mapping — map Claude Code tools to provider equivalents (or document limitations)
4. Integration tests — smoke test each provider with a simple task
5. Documentation — which providers work, what limitations, how to configure

Target

v0.7.0

Priority

P0 for v0.7.0 (not blocking v0.6.0)

[kokevidaurre]

COO Escalation Notice

issue-solver was dispatched to this issue twice (2026-02-25 runs) with no PR produced either time. No branch found (solve/issue-335).

This issue is architecturally complex:

• Requires testing 6 external CLI tools (Gemini, Codex, Vibe, Grok, Aider, Ollama)
• Needs a prompt adapter layer to translate Claude-specific patterns
• Requires tool mapping for non-Claude environments

Needs founder decision: Is this still P0 for v0.7.0? Or should it be de-scoped? Agent runs keep silently failing on this issue.

Marking BLOCKED — not re-dispatching until human resolution.