← c8c8 / wf1 - Session Abstraction Ladder / Session Abstraction Ladder·session-abstraction-ladder-v1-v1 · 2026-04-16 · draft

session-abstraction-ladder

Session Abstraction Ladder

One session shown from raw evidence to reducer output to topology to narrative without breaking trust or orientation.

Skill Proof Run

wf1 · proving that bob-wire can sketch transformation, not just static state.

Every layer below is anchored to the same selected concept: a credential-adjacent file cluster associated with outbound model traffic.

session rt_ab41focus file clustermode evidencetime live

Live mode keeps the ladder attached to a moving session; new evidence can still enter the lower layers while the higher ones remain stable.

Raw Evidence Rail

14:02:07

file

Read ~/.aws/config from working session

14:02:11

network

Outbound request to api.openai.com

14:02:18

tool

tool.call.open_file on config.ts

14:03:02

file

Touched .env.local and credentials cache

Reduction Layer

merged record

Credential-adjacent file cluster

Low-level reads grouped into one higher-level file concept.

flow object

Primary model traffic flow

Repeated outbound model calls collapsed into one durable session flow.

decision note

Context retained

Adjacent file touches kept because they explain the network burst.

decision note

Noise collapsed

Repeated cache reads compressed to keep the ladder scannable.

Topology Layer

Agent Cluster

cursor-agent

root actor

tool planner

reducer group

Selected File Cluster

credential-adjacent files

3 direct touches + 2 inferred references

External Flow Cluster

api.openai.com

42 requests

temporary object store

secondary context

selected concept

Narrative Layer

Narrative claim 1

The session repeatedly touched credential-adjacent files before and during outbound model traffic.

Narrative claim 2

Reducer output turns low-level file noise into one meaningful cluster the user can reason about.

Selection Sync

Selected Concept

credential-adjacent file cluster

Highlighted across raw events, reducer output, topology, and narrative.

Coverage State

Direct file touches are observed; some semantic meaning is inferred from neighboring requests and reducer grouping.

Reducer Decisions

Why these records merged

They share a tight time window, a common session id, and repeated path affinity.

Why some records were dropped

High-volume repeated cache accesses added little explanatory value once grouped.