Agent run review
Review a completed coding-agent run from its prompt, tool calls, patch, checks, and final response; then decide what follow-up it needs.
auto-closehuman reviewfile issuepage on-call
System One,
inside Pi & OMP.
pi-system-one adds one system_one tool to Pi and OMP
for bounded choice, yes/no, and
score decisions. Use it with TypeSafe Jev or any compatible
System One model or endpoint.
01 / INSTALL & CONFIGURE
Install the extension.
Choose a mode.
TypeSafe Jev is the default. Install in Pi or OMP, then choose another mode only when your workflow needs it.
pi install npm:pi-system-one
TypeSafe Jev (default)
Pi uses your saved /login typesafe key or TYPESAFE_API_KEY. No endpoint or model setting is needed.
export TYPESAFE_API_KEY=your-key
Other modes
/so config saves your choice: native handles compatible Choice/Noul, not Score; auto uses native when compatible and TypeSafe otherwise, with no retry on failure; custom uses your endpoint. Mode details ↗
Local / compatible endpoint
SYSTEM_ONE_BASE_URL=http://localhost:8009
SYSTEM_ONE_API_KEY=<optional-if-required>
SYSTEM_ONE_MODEL=<optional-if-required>
EXPLICIT JUDGMENT
/judge
Use /judge when you want to invoke system_one
explicitly instead of relying on the agent to infer that a bounded judgment is needed.
/judge should I ship this? <paste diff plus test summary>
02 / USE CASES
Use cases.
These are examples, not routing rules. The agent can call
system_one when the work narrows to a bounded
choice, yes/no judgment, or
score. Let the normal agent gather evidence and run tools
first; use /judge when you want that judgment path explicitly.
Test failure triage
Classify a failed check from logs, before/after behavior, and environment evidence so the agent knows which path to investigate next.
regressionflaky testenvironmentspec mismatch
Patch acceptance
After deterministic checks have run, choose whether a patch is ready, needs another agent pass, or needs a human decision.
acceptreviseescalate
Review finding severity
Put a concrete code-review finding on an ordered severity scale using reachability, impact, tests, blast radius, and rollback evidence.
lowmoderatehighcritical
Permission boundary review
After hard policy checks, judge whether the trace shows an attempted action outside the scope the coding agent was given.
outside granted scope?0.91 yes
Candidate fix selection
When several fixes already satisfy the hard checks, choose among them using explicit compatibility, performance, complexity, and maintenance constraints.
patch Apatch Bpatch C
03 / FAQ
Common questions.
Coding-agent workflows where the evidence is already available and the
remaining step is bounded. Examples include agent-run review, test-failure
triage, patch acceptance, code-review severity, permission-boundary review,
and choosing among candidate fixes. These are examples, not a fixed routing
table: the agent can choose system_one when the task fits.
Install pi-system-one. The system_one tool uses
TypeSafe Jev by default with TYPESAFE_API_KEY or a saved
/login typesafe key. No endpoint or model setting is needed.
For Pi's built-in classifier, run /so config native.
Native mode supports Choice and Noul, but rejects Score questions.
Use TypeSafe mode for full Score results, or select
/so config auto to route compatible Choice and Noul
requests to a registered Pi classifier and other requests to TypeSafe.
No. Jev is the tested hosted example. Any model, endpoint, or provider implementing the compatible System One contract can sit behind the same tool.
No. pi-system-one exposes the decision primitive.
For automatic model routing in Pi,
Pi-Bifrost
handles that layer and can use System One as a decision engine.