Skip to content

Assess: behavioural-vs-habitat gap lens (privacy-gated, spec-first) [R1 from #340] #428

Description

@russmiles

Follow-up R1 from the microsoft/AI-Engineering-Coach exploration (docs/superpowers/explorations/2026-06-16-ai-engineering-coach-findings.md, issue #340).

Idea

/assess today evaluates habitat evidence — what the repository contains (HARNESS.md, CI, tests, reflections, specs). The MS project's load-bearing difference is that it coaches from behavioural evidence — what the engineer actually did in AI sessions (prompt quality, session hygiene, tool mastery, review cadence).

Triangulating declared habitat against observed behaviour would let /assess detect a new kind of gap — a behaviour-vs-habitat gap, sibling to the existing Habitat Build Gap (cognitive level vs operational mean). A team can have an L4 habitat and L1 session behaviour; only behavioural data reveals it.

Why this is gated

This requires ingesting AI session logs, which is a privacy/consent matter. Out of scope until a privacy/consent design exists. The MS precedent is on-device, read-only, no-phone-home — a posture our hooks/reflections/telemetry already share and could cite.

Scope of a future spec

  • Privacy/consent model: what session data, stored where, retained how long, opt-in.
  • The behavioural axes (see R2 / the MS 5 categories) and how they map to a gap signal.
  • How the gap surfaces in the assessment template alongside the Habitat Build Gap.

Disposition

Deferred / largest of the five. Do not start without the privacy/consent spec. Highest strategic value of the set.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions