II-Agent Build Variants - Complete Summary

Based on PR #166 and commit history analysis

3 Main Build Modes

1. Full Installation

• Command: pip install -e .
• Size: ~2GB, 95+ packages, 3-5min install
• Dependencies: All core + server + GCP + sandboxing + auth
• What you get: Complete stack for server + REPL
• Docker: docker/backend/Dockerfile, docker/frontend/Dockerfile

Includes:

• Server mode (FastAPI, WebSocket, Postgres, SQLite, Redis)
• REPL mode (interactive CLI)
• GCP (Vertex AI, Cloud Storage)
• Sandboxing (E2B, Playwright)
• Auth (OAuth, JWT, Stripe)

2. Lite Installation (REPL only, no GCP)

• Command: pip install -r requirements-lite.txt && pip install -e . --no-deps
• Size: ~200MB, 45 packages, 30s install
• Dependencies: Core only (agents, MCP/Tools, REPL, file processing, search)
• What you get: REPL mode only
• What's removed:
  • Server stack (FastAPI, DBs, Redis)
  • GCP dependencies
  • Sandboxing (E2B, Playwright)
  • Auth (OAuth, JWT, Stripe)

3. GAIA Benchmark Installation

• Command: pip install -e .[gaia]
• Dependencies: Full + GAIA-specific (datasets, huggingface-hub)
• What you get: Bench tasks and evaluation

---

2 Runtime Modes

REPL Mode (Interactive CLI)

ii-agent --repl [--workspace .]

• Implementation: src/ii_agent/cli/repl.py (+537 lines)
• Features: Tab completion, model/provider selection
• Works with: Lite or Full install

Server Mode (Web interface)

ii-agent [--port 8000] [--host 0.0.0.0]

• Implementation: src/ii_agent/server/app.py
• Features: Full web app with auth, credits, Stripe
• Docker: Full Docker Compose stack included

---

3 Memory/Storage Backends

All three survive and are available:

1. Hashtable: LRU in-memory caching
2. Memvid: QR-encoded video storage
3. DuckDB: Vector search and analytics

---

Adaptive Context Engineering (NEW)

Added in bb33517 on top of d8c3670:

• Checkpoint without eviction (Strategy 5)
• Microkernel generation before checkpoint
• Context modes: SUSPENDED, HIGH_DETAIL, HIGH_CAPACITY, NORMAL
• Dynamic model-specific thresholds (90% for 64K → 15% for 1MB)
• Tools: RecallContext, MicrocontextSubroutine

Does NOT replace lower build variants - it's additive.

---

NVIDIA Provider

From lower commits (f01560f, 9939946, bf3b298):

• NVIDIA_API_KEY support
• Qwen3-coder-480b default
• Bayesian model ranking
• SQLite database integration

---

Python 3.14 Compatibility

From d8c3670:

• Puremagic instead of imghdr
• Unpinned pydantic (was ==2.11.7, now >=2.11.7)
• Build flag: PYO3_USE_ABI3_FORWARD_COMPATIBILITY=1

---

What Tests Were Removed (Orphaned)

From lower commit 51086ca (refactor/tool #144), tools were deleted:

• src/ii_agent/tools/bash_tool.py
• src/ii_agent/tools/str_replace_tool.py
• src/ii_agent/tools/sequential_thinking_tool.py
• 42 other legacy tools

Tests removed in d8c3670 (1529 lines, orphaned tests):

• tests/tools/test_bash_tool.py (565 lines)
• tests/tools/test_str_replace_tool.py (747 lines)
• tests/tools/test_sequential_thinking_tool.py (217 lines)

All deleted code tested tools that no longer exist.

---

Docker Deployment

All survive from lower commits:

• Full Stack: docker/docker-compose.stack.yaml
  • postgres, redis, frontend, backend, sandbox, tool-server, ngrok
• Backend: docker/backend/Dockerfile (uv, playwright)
• Frontend: docker/frontend/Dockerfile (Node 18)
• E2B Sandbox: e2b.Dockerfile

---

Complete Build Matrix

# Install variants
pip install -e .                              # Full (2GB)
pip install -e .[gaia]                       # GAIA benchmark
cat requirements-lite.txt | xargs pip install && pip install -e . --no-deps  # Lite (200MB)

# Runtime modes
ii-agent --repl                              # REPL mode
ii-agent --port 8000                       # Server mode

# Docker
docker-compose -f docker/docker-compose.stack.yaml up

# Python versions
Python 3.10, 3.11, 3.12, 3.13, 3.14
# For 3.14: export PYO3_USE_ABI3_FORWARD_COMPATIBILITY=1

Key Insight

No lower commit variants were removed - only deprecated test files for already-deleted tools were cleaned up. The codebase maintains full backward compatibility while adding new features.