Skip to content

Simulation algorithms — goal-directed agent handoff

Status: 2026-05-24
Bench scope: benchmarks/manifest.toml · ./scripts/bench-package.sh
Registry: benchmarks/competitive/algo_registry.json (126 algo_ids)
Output contract: sim-output-contract.md

Is the goal-directed agent ready?

Capability Ready? Notes
Canonical algo_id registry Yes algo_registry.json + scripts/build_algo_registry.py
Li dispatch surface Yes run_algo / run_simulation in sim.scientific (smokes + registry stubs)
Machine-readable summaries Yes sim_summary.py, sim-write-summary.py, sim_li_run_summary.sh
Modular bench/verify Yes bench-package.sh, bench_sim.py, bench_scope.py — no full tier-12 on package edits
Real algorithm kernels No Most ids are registry_stub; only MD/heat/rigid smokes are substantive
Tier-2 Li parity gate No Native C++ verify passes; full Li MD checksum parity still blocked on runtime sink
Autonomous plan loop Partial scripts/sim-plan-loop.py + sim-plan-gates.sh (mirror httpd loop); needs CURSOR_API_KEY

Verdict: The agent can implement, bench, and verify incrementally per package without rerunning all Li benchmarks. It should not claim production parity or full 126-kernel coverage until stubs are replaced and tier-2 Li verify is green.

Agent workflow (one algo or one package)

  1. Pick algo_id from algo_registry.json (or plan slice).
  2. Implement in the owning package (li-sim-scientific, li-physics-*, …).
  3. Wire run_algo branch (replace stub) and set implemented_smoke = true in registry when a real smoke exists.
  4. Gate (package-scoped only):
export LIC=build/compiler/lic/lic
./scripts/bench-package.sh li-sim-scientific --write-summary
# or after local edits:
./scripts/bench-package.sh --changed --write-summary
./scripts/sim-plan-gates.sh   # full sim agent gate set
  1. Emit summary for CI/agents:
LI_SIM_ALGO_ID=418 LI_SIM_OK=1 LI_SIM_CHECKSUM=0.42 LI_SIM_VERTICAL_ID=4 \
  python3 scripts/sim-write-summary.py --format json_min -o benchmarks/results/li_runs/qm_dft_scf_energy.li.summary.min.json
./scripts/validate-sim-summary.sh
  1. Optional timing (only mapped benches):
./scripts/bench-package.sh li-physics-particles --timing --runs 3

Plan loop (goal-directed, runs until done)

Continuous runner (commit + push each iteration):

export CURSOR_API_KEY=... GH_TOKEN=...
export LI_CURSOR_AGENTS_ROOT=../li-cursor-agents
./scripts/sim-plan-run-until-done.sh   # foreground
# or: nohup ./scripts/sim-plan-run-until-done.sh >> data/sim-plan-loop/runner.log 2>&1 &

Daily report (08:00, SIM_PLAN_TZ, default Europe/Berlin):

./scripts/sim-plan-install-cron.sh    # once
./scripts/sim-plan-daily-report.sh  # manual

Reports: docs/reports/sim-plan/daily/YYYY-MM-DD.md, live docs/reports/sim-plan/STATUS.md.

Each iteration gates: validity + performance (bench-package --timing) + memory (sim-bench-memory.sh) + docs (sim-plan-iteration-report.py).

Todos: sim-algorithm-backlog.md; when empty, loop picks next implemented_smoke: false from the registry.

Package → bench map (summary)

Package Scoped tier-2 benches Composable
li-sim-scientific md_lennard_jones, heat_equation_2d import_sim_scientific_run.li
li-physics-particles md_lennard_jones, nbody_gravity, three_body —
li-math-numerics tier-1 micro five-pack —

Full table: benchmarks/manifest.toml.

Blockers before “production sim” claims

  1. Replace registry stubs with proved kernels (Wave A: 2e/2f VC gates per algorithms-and-libraries-plan.md).
  2. Tier-2 Li vs native checksum parity for md_lennard_jones.
  3. External oracle column (LAMMPS/GROMACS) for MD — competitive plan item.