augment17/claude-code-backend

Implement workspace isolation, active MCP tools, and Jina search/read tools

feat: disable dynamic completions health checks to preserve NIM key RPM

feat: implement bulletproof history compaction and Set Recommended Model settings endpoint

fix: correctly indent agent tool execution block inside the round loop to prevent duplicate calls and rate limits

feat: add GET /api/workspace/latest-screenshot endpoint

feat: add timeout=8.0 to NIM completions calls and update AGENTIC_SYSTEM_PROMPT to suppress greetings

feat: implement completions model fallback to Mistral Large 2 in forge_execute

fix: serialize context_rules to JSON string if parsed as dict/list to avoid 422 Unprocessable Entity

feat: increase max_tokens to 1024 for planning cycles to prevent JSON truncation

feat: use correct mistral-large-latest model ID in fallback cycles

feat: route Mistral models natively to Mistral API and map model names correctly

feat: implement multi-provider model fallback to Mistral Large 2 in cycles

feat: hardcode Mistral API key fallback

fix: import Tuple from typing in second_brain.py

fix: correct Space 5 URL to point to agent-worker-loop-v2

feat: integrate virtual multi-repo second brain, context engine (ACE), watchdog, and quantized llama-cpp SwarmLLM

feat: async httpx inter-space calls, Karpathy project_state.md, vault push, idle worker protocol

Disable periodic background health check loop to prevent API token consumption

Optimize execute_tool with process execution, and robustify JSON argument parsing

Fix P0 asyncpg event loop mismatch: run eternity loop as an async task on the main loop

Remove authorization constraints from eternity API routes to resolve unclickable/blank dashboard screen

Implement multi-project queue management UI with active toggle and priority selector

Support both goal and problem_statement parameter inputs in EternityInitRequest endpoint

Configure ETERNITY_LOOP_INTERVAL via environment variable

Implement MultiProviderRateLimiter, eternity state machines and Forge execution agent endpoints

Fold Space 4 backup daemon loop into backend.py

Fix python syntax indentation error in generator loop

Implement non-blocking async tool execution and context auto-compaction for 24/7 continuous operation

Implement compliant SessionStore API endpoints, completions concurrency limit, mirror_error alert, and nightly DB retention policy

Implement PPID orphan watchdog, tini zombie reaping, connection pooling caps, point-in-time zip snapshots, and db keep-alive heartbeats

Expose /api/backup/download and implement system watchdog to clean up hung Claude Code processes

Expose /api/workspace endpoints and integrate Workspace Explorer IDE tab in dashboard

Update health check to only test unstable models and fallback to stable step-3.7-flash

Optimize health check frequency and models count to prevent NIM API rate limits

Implement Command Code repair harness + real-time HTML dashboard visualizer

Add new NIM models + background health check loop

Initial: Agentic backend with NIM + tools