@standardagents/skill 0.41.26 → 0.42.1

package/README.md

@@ -14,6 +14,13 @@ The first bundled skill is:
 - `agentbuilder` - guidance for composing models, prompts, tools, hooks, APIs,
   and subagents in Standard Agents projects
 
+The skill also covers TypeSafe AI decision inference through `state.decide()`:
+explicit context, named choice/options, score/levels, and boolean questions;
+native Jev conversion; preservation of confidence and probabilities; and
+caller-owned 0–1 thresholds. Decision calls automatically log operations and
+cost without adding conversation messages. Full usage lives in the
+`@standardagents/typesafe` package README.
+
 ## Installation
 
 Recommended:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@standardagents/skill",
-  "version": "0.41.26",
+  "version": "0.42.1",
   "private": false,
   "publishConfig": {
     "access": "public",

package/skills/agentbuilder/SKILL.md

@@ -542,6 +542,28 @@ Register the hook id in the side A prompt's `hooks` array. Why this split and no
 
 ---
 
+## Decision inference
+
+For classification, scoring, or yes/no probabilities, use
+`state.decide({ provider: typesafe, model: 'jev-latest', context, questions })`
+with `typesafe` from `@standardagents/typesafe` and BYOK `TYPESAFE_AI_API_KEY`.
+Keep conversational `defineModel` choices separate. Scope credentials as needed,
+but set destination URLs only through trusted instance/deployment configuration.
+Supply explicit context and
+named questions: `choice` uses described `options` (or null), `score` uses ordered
+`levels`, and `boolean` uses optional true/false criteria. Write the actual
+question in `instructions`; its map key only identifies the answer. TypeSafe
+maps context→state, options/levels→criteria, and boolean→noul. Preserve full
+probabilities, confidence, and score legends; boolean answers return a 0–1
+`probability`, so choose a threshold rather than coercing it. Thread calls log
+request/response/duration/status/cost automatically, persist before the first
+message, and never append messages;
+explicitly use `state.injectMessage()` when needed.
+Automatic logging and optional `name`/`trigger` attribution are AgentBuilder
+extensions, not spec requirements. Standalone `typesafe({ apiKey }).decide(...)`
+creates no thread logs. Read the package README for limits and the complete
+conversion contract.
+
 ## Model selection
 
 **Rule: never write a model string the user did not request and `current-models` did not produce.**