Skip to content

getting_started.md

Latest commit

 

History

History
56 lines (38 loc) · 1.91 KB

getting_started.md

File metadata and controls

56 lines (38 loc) · 1.91 KB

Getting started

1. Install SAGE (repository clone)

From the SAGE repo root (where startup.sh lives):

  • Linux/macOS: bash startup.sh && source .venv/bin/activate
  • Windows: .\startup.ps1 and activate .venv

Details → INSTALL.md.

2. Prepare a project directory

mkdir -p ~/myproject && cd ~/myproject
sage init

This creates .sage/, memory/, starter rules, and pytest hints. sage init does not install the SAGE package — run startup.sh from the clone for that.

3. Run

export SAGE_MODEL_PROFILE=test   # optional — one small model per role (laptop/CI)
sage doctor
sage run "Your goal here" --auto
  • --research (default) — interactive plan checkpoints (a / r / e).
  • --auto / --silent — less interaction; see CLI.md.

Overview and first-run tips → README.md. CLI reference → CLI.md. Spec vs implementation: ARCHITECTURE_STATUS.md.

Benchmarks & RL (reference)

  • sage bench
  • sage bench --compare-policy — static vs learned routing (needs checkpoints)

Phase 5 — offline RL

  • sage rl collect-synth --rows 650
  • sage rl export --output datasets/routing_v1.jsonl
  • sage rl analyze-rewards --data datasets/routing_v1.jsonl
  • sage rl train-bc --data datasets/routing_v1.jsonl
  • sage rl train-cql --data datasets/routing_v1.jsonl
  • sage rl eval-offline --data datasets/routing_v1.jsonl --checkpoint memory/rl/policy_cql.joblib
  • scripts/train_routing_policy.py — BC then CQL in one script (from repo root with venv activated)

Phase 6 — simulator

  • sage sim generate --count 1000 --out datasets/sim_tasks.jsonl
  • sage sim run --tasks datasets/sim_tasks.jsonl --workers 4
  • docker build -f sim/Dockerfile -t sage-sim:latest .
  • sage sim run --tasks datasets/sim_tasks.jsonl --workers 4 --docker