Options, inputs and output

This page lists the plugin options, the inputs of the pr_risk tool and the fields of its result. For what the values mean, see How a risk level is computed.

Plugin options

Open the options with /plugin configure dev-kit@cc-toolkit. A change takes effect after /reload-plugins.

Option	Default	What it does
typesafe_api_key	empty, reads TYPESAFE_API_KEY	The TypeSafe API key. Without one, the tool doesn’t exist and nothing is sent.
typesafe_endpoint	https://api.typesafe.ai	Base URL of the API. /v1/systemone is appended.
typesafe_model	jev-1.14.0	Model name or alias. Pin a version to keep levels stable across Jev releases.
risk_label	true	Whether pr-open labels the pull request.
risk_label_name	risk: {level}	The label name. {level} becomes the level.
risk_levels	low, medium, high, critical	Names of the four levels, from lowest to highest. Any count other than four falls back to the default.

Tool inputs

Claude calls the tool as mcp__dev-kit__pr_risk. pr-open calls it for you, and Claude can also call it when you ask how risky a branch is. There’s no slash command.

Input	Required	What it does
base	yes	The branch the change merges into, such as main. The tool diffs origin/<base>...HEAD, or <base>...HEAD without a remote branch.
slices	no	The changed files grouped by purpose, each {title, files}. Without it, Claude Haiku groups the files. See Group the files into slices.
context	no	The pull request’s and ticket’s titles and descriptions. It tells Jev what each slice is for, and it’s where the tool looks for manual evidence. A Risk section that pr-open wrote earlier is removed before it’s sent.
testing_policy	no	The repository’s testing rules, in a few sentences. It decides whether a slice without tests breaks those rules.
execution_context	no	Up to 8 entries of short, exact source excerpts for changed files, each checked against a full base or HEAD commit ID. Only the matching slice’s diff request receives them.
risk_policy	no	calibrated-v1 (default) or the stricter, unvalidated evidence-v1. It affects only the two-axis fallback.
risk_experiment	no	atomic-v1 adds experimental questions whose answers appear under experimental. They don’t change the level.

Everything in context, testing_policy and execution_context is sent to TypeSafe, along with the diff and the commit messages.

Result

The tool returns JSON. A result that starts with Risk assessment unavailable: means no assessment was made, and the rest of the text gives the reason.

The whole change

Field	What it holds
level	The risk level, one of the four risk_levels.
scorer	factors, or two-axis fallback when no factor request answered.
impact.level, likelihood.level	The worst impact and the worst likelihood across slices.
complexity	The highest complexity across slices.
evidence	needed: the manual evidence types Jev finds plausible. provided: those found in context. credited: whether any slice’s likelihood was lowered for them.
adjustments	The rules that moved the level, in plain sentences.
confidence, lowConfidence	See Confidence and lowConfidence.
intents	What the change does, from the intent requests.
types	The change types found in any slice.
flags	Risk flags, each {key, label}.
slices	One entry per slice, described below.
label	enabled, the label name for this level, and all label names, so pr-open can remove the old one.
coverage	How much of the change was assessed, described below.
model, base, riskPolicy, slicing	The Jev model that answered, the resolved base, the policy used, and where the slices came from (caller, model or size).

Each slice

Each entry of slices holds title, files, level, impact, likelihood, complexity, evidence, testGap, adjustments, intents, types and flags. A slice whose factor requests all failed holds factors: "missing" and keeps its diff request’s level, impact and likelihood.

Coverage

Field	What it holds
requests	The number of diff requests sent.
filesAssessed, filesNotAssessed	Files read and files beyond the request cap.
filesNotAssessedSample	Up to 20 of the files not assessed.
filesExcluded	Lockfiles, binaries and generated files left out.
truncatedHunks	Hunks cut to fit a request.
factorsMissing	Slices, or parts of a split slice, without factor answers.
partial	Whether code went unread or a slice lacks factor answers.

Change types and flags

The diff request asks about these change types. The ones in the table raise a flag:

Flag	Raised by
Touches security	authorization, authentication, secret_handling
Changes persisted data	schema_or_data_migration, data_removal_or_rewrite
Changes a public contract	api_contract
Changes CI, infrastructure or deployment	ci_cd_pipeline, cloud_infrastructure, runtime_environment
Logic without tests	a slice with code that adds, changes or fixes behavior and has no test

The other change types describe the slice without raising a flag:

• Risk signals: irreversible_external_effect, infrastructure_state_dependent, team_only_scope, data_access, dependency_added_or_removed, feature_flag_gated, ui_components, ui_interactions, build_and_dev_tooling, dependency_version_update, ai_assistant_instructions.
• Tests: unit_or_integration_tests, e2e_tests, performance_tests, infrastructure_tests.
• What much of the slice is about: observability, error_handling_and_validation, visual_styling, configuration, documentation, generated_artifacts, release_metadata.

The intents are new_functionality, behavior_change, behavior_condition_change, bug_fix, refactor, rename_or_move and code_removal.

Labels

When risk_label is on, pr-open adds the label for the level and removes every other risk label the pull request carries. It creates a missing label once, and never changes an existing label.