Developer tools¶
Run tools from the repository root with the project Python environment. Build native
components first where required: make all. Use --help to inspect options.
Benchmarks are opt-in experiments; they are not ordinary regression tests or latency guarantees.
Current tools¶
| Tool | Purpose / inputs |
|---|---|
| build_source_release.py | Export a specified Git commit; explicit --working-tree creates a labeled local candidate |
| sync_public_source.py | Preview/apply an archive to a clean public checkout; never commits or pushes |
| build_agent_sdk.py | Export the local SDK wheel for external Agent development |
| check_documentation.py | Check English guides/docstrings, local links and headings; add --include-internal for current internal documents |
| capture_desktop.py | Native desktop captures using controlled display fixtures; requires Desktop dependencies |
| benchmark_desktop_controls.py | Measure desktop control behavior with a prepared fixture directory |
| benchmark_prepared_import.py | Compare complete source preparation and optional CSV; supply permitted historical input |
| benchmark_prepared_runtime.py | Fixed synthetic execution/Agent/Tools workload from the checked-in profile |
| benchmark_prepared_real_window.py | Actual source peak-window measurement; supply permitted historical input |
| benchmark_prepared_checkpoint.py | 35-second run spanning the real periodic financial checkpoint |
| benchmark_experiment_runtime.py | Runtime scheduling and communication experiments |
| benchmark_live_experiments.py | Live multi-participant load/failure experiments; writes under build/benchmarks/ |
| benchmark_platform.py | Standalone native/market/execution/policy/Desktop component comparisons; not the whole prepared Agent path |
| probe_agent_runtimes.py | Opt-in interpreter provisioning probe; downloads require explicit flags |
| replay_platform_run.py | Compatible raw-source recorded-run Replay into an empty destination; not the prepared-operation Replay API |
| network_demo.py, network_sender.py | Finite local UDP feed/recovery simulation and offline comparison |
| validate_contract_fixtures.py | Bounded compatibility-fixture checks, not current-runtime certification |
| agent_probe_records.py | Compact measurement helper, not a standalone product command |
Lightweight entrypoints¶
python tools/build_agent_sdk.py --help
python tools/benchmark_prepared_import.py --help
python tools/benchmark_prepared_runtime.py --help
python tools/replay_platform_run.py --help
python tools/validate_contract_fixtures.py
Historical benchmark names no longer present in this folder are not aliases for the current tools. Do not reuse an old command with a renamed executable or compare results across changed workloads as though they were the same experiment.
Running measurements¶
python tools/benchmark_prepared_runtime.py --include-agent --output build/runtime-study.json
python tools/benchmark_prepared_import.py --source /absolute/path/feed.bin --output build/import-study
python tools/benchmark_prepared_checkpoint.py --output build/checkpoint-study.json
The runtime command measures policy cases by default. Add --include-agent to also
measure model IO, Agent CPU parsing and complete Tools IPC. It does not switch architectures.
The import command defaults to a limited prefix; add --full-day only for an intentional
full-source experiment. Inputs and output-path requirements are tool-specific. Use a new,
dedicated destination and inspect --help before running a costly measurement.
The fixed profile is consumed directly
by the runtime/checkpoint/real-window scripts. Preserve workload and thresholds with results.
Fresh outputs should normally go under ignored build/; the current Live probe's default
docs output is a legacy location, not approval to publish a new internal report.
For a local network demo use the network guide. For test commands, CI scope, optional Desktop setup and measurement limitations see verification. Source publication is a separate reviewed action; generating a source archive does not push, tag or publish anything.
Source export and sync: release procedure.
Website and tutorial¶
build_pages.py renders the existing public Markdown into the GitHub Pages site,
builds search and checks local URLs/anchors. Install requirements-site.txt first.
record_desktop_demo.py captures a real synthetic-data tutorial through Desktop controls;
its optional encoder is pinned in requirements-demo.txt. It verifies a fill and a
recorded Live run, uses a temporary workspace and makes no model calls.
See website setup and recording for local preview and deployment.