dsh-jev-plugin

Configuration

Field	Default	Description
apiKey	—	TypeSafe API key; takes precedence over apiKeyEnv
apiKeyEnv	TYPESAFE_API_KEY	Environment variable used when apiKey is unset
baseURL	https://api.typesafe.ai/v1	API root; trailing slashes are removed
model	jev-latest	Default model; tools can override it per call
timeoutMs	10000	Call timeout in milliseconds
retry.maxAttempts	3	Maximum attempts, including the first
retry.baseDelayMs	500	Initial backoff in milliseconds
retry.maxDelayMs	5000	Maximum delay between attempts
maxStateChars	64000	Maximum serialized state length; 0 disables the limit
confidence.approveAt	0.8	Confidence threshold for automatic acceptance via verdictFor
confidence.escalateBelow	0.5	Escalate below this confidence; must not exceed approveAt
policy.enabled	false	Enable the circuit breaker and minimum call interval
policy.failureThreshold	5	Consecutive failures before opening the circuit
policy.openMs	30000	Time before allowing a probe call
policy.minIntervalMs	200	Minimum interval between calls
cache.enabled	false	Cache responses for identical model, state, and questions
cache.maxEntries	100	Maximum cache entries; LRU eviction
cache.ttlMs	60000	Cache entry lifetime
guard.enabled	false	Enable the tools/pre-execute risk gate
guard.tools	[]	Tools to check; empty means all tools
guard.question	Built-in risk question	Question submitted to Jev
guard.levels	["low","medium","high","critical"]	Ordered risk labels; 2–10 entries
guard.denyAt	3	Deny at or above this score
guard.askAt	1	Request approval at or above this score
guard.escalateOnLowConfidence	true	Request approval below confidence.escalateBelow
guard.reviseAt	Highest level index	Reject with revision guidance; equal to denyAt by default, leaving no revision band
guard.onError	"allow"	Gate failure behavior: fail-open (allow) or fail-closed (deny)
rules.enabled	false	Enable synchronous offline denial rules; no API key needed
rules.tools	[]	Tools to check; empty means all tools
rules.deny	[]	Additional case-insensitive denial regexes; invalid patterns fail at load time
review.enabled	false	Enable tools/post-execute review; may rewrite completed tool results
review.blockAt / review.onError	0.8 / "accept"	Correction probability threshold and error behavior
routing.enabled	false	Route models on agent/request using request complexity and routing.models
context.enabled	false	Enable context pruning on agent/pre-step
context.maxDrops / context.shadow	0 / false	Drop limit (0 means unlimited); shadow mode reports without changing context
intent.enabled	false	Decision logic exists, but instruction injection is not connected; enabling this does not inject messages
quotaCooldownMs	900000	Cooldown after 402, independent of policy.enabled; 0 disables it
enableDecide / enableEvaluate	true	Register the corresponding tool
telemetry.log	false	Emit a sanitized structured record for each call
telemetry.sampleRate	1	Sampling fraction from 0 to 1
telemetry.errorRateAlert	0.5	Error rate threshold for degraded health
telemetry.alertMinCalls	10	Minimum sample size before reporting health

Validate gate thresholds for your workload: scores and confidence can vary. Changing guard.levels clamps default thresholds into range; explicitly configured out-of-range values fail validation. Start context pruning in shadow mode before allowing it to remove messages.

Oversized state is rejected, never truncated, so a decision cannot silently use incomplete input. Policy, caching, and the risk gate are disabled by default. Intent injection remains unavailable because the host lacks a pre-step message channel that records plugin authorship.

Prefer an environment variable over storing the key in cordis.patch.yml:

export TYPESAFE_API_KEY="..."

PowerShell:

$env:TYPESAFE_API_KEY = "..."

If storing a key in configuration, use the secret role and restrict file permissions. The plugin redacts API keys from tool output, error messages, and logs.