TypeSafe AI integration overview
September 20, 2026 ยท View on GitHub
claude-jev-plugin inserts TypeSafe System One judgments into Claude Code's hook lifecycle. TypeSafe is an external semantic classifier. It does not replace Claude, grant permissions, or run an MCP server.
Architecture
flowchart LR
U[User] --> C[Claude Code]
C --> H[claude-jev command hook]
H --> B[Validate and bound state locally]
B --> T[TypeSafe System One API]
T --> V[Typed answers]
V --> L[Local threshold evaluation]
L --> C
H -. no MCP .-> N[No persistent plugin process]
ASCII version
[User]
|
v
[Claude Code] ---> [claude-jev command hook]
|
v
[Validate + bound state]
|
v
[TypeSafe System One]
|
v
[Typed answers]
|
v
[Local threshold logic]
|
+--------> [Claude Code]
No MCP server. No persistent plugin process.
Claude starts a short-lived Node.js process for every matching hook event. Event JSON arrives on stdin. Hook writes either no output or one Claude-compatible JSON object to stdout.
Hook configuration uses exec form:
{
"type": "command",
"command": "node",
"args": ["${CLAUDE_PLUGIN_ROOT}/dist/hooks/pre-tool.js"],
"timeout": 20
}
Tool input remains process data and is never interpolated into shell source.
Activation
Plugin installs disabled by default because activation sends bounded data to an external service and may incur TypeSafe API cost.
/plugin marketplace add dr-dimitru/claude-plugins-marketplace
/plugin install claude-jev@dr-dimitru-claude-tools --scope user
Set key outside project configuration:
export TYPESAFE_API_KEY="..."
Then enable and reload:
/plugin enable claude-jev@dr-dimitru-claude-tools
/reload-plugins
Claude discovers four handlers:
flowchart TD
A[Plugin enabled] --> B[Load hooks/hooks.json]
B --> U[UserPromptSubmit]
B --> P[PreToolUse]
B --> S[PostToolUse]
B --> F[PostToolUseFailure]
U --> U1[Store bounded current request]
P --> P1[Judge proposed Bash, Write, or Edit]
S --> S1[Judge successful Bash output]
F --> F1[Judge failed Bash output]
ASCII version
[Plugin enabled]
|
v
[Load hooks/hooks.json]
|
+----------------+-------+-------+-------------------+
| | | |
v v v v
[UserPromptSubmit] [PreToolUse] [PostToolUse] [PostToolUseFailure]
| | | |
v v v v
[Store bounded [Judge Bash, [Judge successful [Judge failed
current request] Write, Edit] Bash output] Bash output]
Automatic hooks do not require skill invocation.
TypeSafe request
Plugin sends one HTTPS request for all independent questions about one state:
POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY
Content-Type: application/json
{
"model": "jev-latest",
"state": {"tool": "Bash", "tool_input": {"command": "git status"}},
"questions": {
"destructive": {
"type": "noul",
"instructions": "Is this action destructive?"
}
}
}
Response must contain model, usage, and one typed answer per question. Plugin rejects missing answers, unknown options, incomplete distributions, invalid score legends, non-finite numbers, and probabilities that do not sum to one.
Question primitives
Noul
Noul answers a yes/no semantic question as probability from 0 to 1:
{"type": "noul", "noul": 0.96}
Noul has no separate confidence value. Plugin uses it for destructive action, exfiltration, scope, and secret detection.
Choice
Choice selects one declared category and returns full distribution:
{
"type": "choice",
"choice": "environment",
"probabilities": {
"transient": 0.05,
"environment": 0.80,
"code_bug": 0.05,
"permission": 0.05,
"user_error": 0.03,
"no_failure": 0.02
},
"confidence": 0.76
}
Plugin uses Choice to classify Bash failures.
Score
Score rates state across ordered criteria:
{
"type": "score",
"score": 2.8,
"legend": {
"0": "None, it only reads",
"1": "Small, one file or one reversible change",
"2": "Large, many files or shared state",
"3": "Severe, data loss or a forced overwrite of shared history"
},
"probabilities": {"0": 0.0, "1": 0.0, "2": 0.2, "3": 0.8},
"confidence": 0.84
}
TypeSafe supplies typed semantic evidence. Local plugin code decides whether configured thresholds were crossed.