Spaces:
Sleeping
Sleeping
Commit History
-
fix: strip ANSI codes in _run_tests() so ✓/✗ count correctly
natnael kahssay6b28995 -
feat: replace handcrafted user_messages with real MOA session traces
natnael kahssay +1: Claude Sonnet 4.6- tasks.py now uses actual messages from real dev sessions: task_001: provider:list error, vertex auth confusion task_002: p-retry AbortError boot failure, exponential backoff request task_003: append-only structure request, event-store schema questions - model now trains on the real gap: terse/imprecise human intent → correct TS Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>bb5a5ec -
feat: multi-turn tool-using RL environment (RFC 005 pattern)
natnael kahssay +1: Claude Sonnet 4.6Actions are now tool calls (read/edit/bash/submit) instead of one-shot file submissions. The agent navigates the sandbox across multiple steps, exactly like Claude Code, then calls submit to run vitest and get reward. - MOAAction: {tool, params} replaces {file_path, content} - MOAObservation: adds tool_result, user_messages, step_count - moa_env.py: dispatches read/edit/bash/submit with sandbox path safety - tasks.py: adds real user_messages per task (Step 1 intent alignment) - Max 20 steps per episode; auto-submits on step 20 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>5d3d3ff -
feat: use real moav2 source as RL task suite — symlinked sandbox, 3 real service tasks
natnael kahssayce25387 -
fix: embed task content directly, self-contained vitest sandbox
natnael kahssay38cd72d -
add .gitignore, remove build artifacts
natnael kahssay367d69f -
initial moa rl environment
natnael kahssay50a0d81