Gate, playground and experiments

These tools share the Go client but serve different jobs. They are part of the same repository as semlint, SysOneScript and Studio.

Agent tool-call gate

cmd/gate implements a pre-flight policy gate. Hook adapters in plugin/ support Claude Code and omp. The model supplies narrow judgments; Go code combines them into allow, ask or deny decisions.

make buildclaude --plugin-dir "$PWD/plugin"omp --hook "$PWD/plugin/omp/typesafe-gate.hook.ts"

Run from the repository root, with TYPESAFE_API_KEY set in the launching environment. The build target writes plugin/bin/typesafe-gate. These are separate agent-launch commands, not commands to run inside a SysOneScript program.

The gate considers request intent, whether the target is related, whether the description matches the proposed action, and safety factors such as destructive operations, credentials and network writes. Read-only tool calls bypass model evaluation.

It fails open. Missing credentials, timeouts, bad payloads and panics allow the call. It is an advisory supplement to the agent's normal permissions, not a security boundary.

Configuration comes from .typesafe-gate.json in the project or ~/.config/typesafe-gate.json; partial configuration overrides selected fields:

{  "enabled": true,  "timeout_ms": 3000,  "thresholds": {    "request_wants_changes": 0.40,    "target_unrelated": 0.60,    "ask_network_write": true  }}

make eval uses the fitted corpus. make eval-heldout uses the separate held-out corpus. Both make live judgments; a fitted score does not establish unseen-case reliability.

API playground

cmd/playground contains seven executable scenarios: primitives, structured state, confidence, fan-out, composite questions, selection and errors/metadata.

go run ./cmd/playground 01go run ./cmd/playground 03

The argument is a substring filter. Omitting it runs all scenarios. These are live API demonstrations, not offline tests; they read the key from the environment or working-directory .env.

Experiments

cmd/experiments explores calibration, sensitivity, guardrails, option scaling, throughput, budgets, staged decisions and model-derived features.

go run ./cmd/experiments e3go run ./cmd/experiments e7

Experiments make real requests. Heavy throughput/sustained-load cases are skipped by default, but -heavy or naming a heavy experiment exactly enables them. Do not treat them as normal verification commands. Results depend on the model, input distribution, request grouping and provider behavior.

Repository map