ENT-1149 · V1 implementation review

Scoped at every hop. Bounded at every edge.

Many containers per project, each with a developer-set stable ID. Developers assign each Thread to one Learning Container. Persisted AG-UI runs become evidence-backed Insights, and eligible Insights become versioned Skills. API, UI, and CLI share the same Project scope and release flag.

Web + Channels No transcript upload One cross-surface flag

01AssignThe Runtime picks one stable Container ID

02PersistIntelligenceAgentRunner stores AG-UI events

03LearnThe harvester derives one exact run snapshot

04PublishCited Insights become reviewable Skills

1 : NProject to Containers

0 : 1Thread to Container

512 KiBDerived snapshot limit

500Maximum items per manifest list

What ships

A narrow V1 with clear owners

CONTROL

Project-owned setup

Create many Learning Containers with developer-set IDs, inspect evidence, review Skills, and start runs in the Intelligence shell.

RUNTIME

One assignment hook

The same learning.containerId hook assigns web and Channel Threads through their normal create and lock calls.

SOURCE

Persisted events

IntelligenceAgentRunner AG-UI events remain the source of truth. The Runtime never uploads a second transcript.

Review path

Read the system in eight short pages

ArchitectureThread assignment, source-to-sink flow, and failure boundaries Merge readinessThe review gaps, their fixes, and the local proof Data modelProject scope, stable IDs, immutable inputs, and provenance Product UIMany Containers, Insights, Skills, and runs CLIThe one-command Skill bundle contract Worked exampleA Runtime assignment becomes a reviewed Skill Scale proofExact workload, checks, and results Social previewThe review map at share-card size