Skip to content

ContextLattice v3.23.0 - Agent Fit

Choose a tag to compare

View all tags
@sheawinkler sheawinkler released this 19 Jul 07:46
v3.23.0
befd2e0
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

ContextLattice v3.23.0

Agent Fit

ContextLattice now adapts the briefing around the agent without absorbing the
agent itself. The CLI remains the primary interface. HTTP remains the
integration fallback.

What changed

  • contextlattice agent-fit adds steering publish, replay, acknowledgement,
    watch, runner selection, model selection, profile resolution, profile
    configuration, and explicit context-preparation workflows.
  • Steering supports resumable Server-Sent Events with event IDs,
    acknowledgements, bounded replay, backpressure handling, and an honest pull
    fallback when a harness cannot hold a stream.
  • Runner and model selection combines readiness, capabilities, observed
    quality, context fit, cost, latency, and failures, then abstains when the
    evidence is weak. Public selection is advisory and performs no execution.
  • Effective profiles resolve deterministically from explicit request, stored
    agent profile, then generic fallback, with capability and context-window
    constraints applied before use.
  • Context preparation is opt-in and approval-first. Artifacts are bound to the
    task, profile, source generation, authorization state, expiry, and explicit
    external-worker consumption.
  • Pi and Droid task-worker adapters can consume an entitled selection
    authorization only when it matches the runner or model already named by the
    task. No hidden auto-routing was added.

Product boundary

The public core includes steering delivery, advisory selection, effective
profiles, context preparation, the primary CLI, bounded telemetry, and every
fail-closed execution boundary.

Operator and Enterprise add workspace-isolated governance for steering
delivery and revocation, outcome-calibrated profile rollout, context-prep
budgets and schedules, and explicit selection authorization. These routes
retain bounded owner-only state and hash-linked receipts. They do not run a
model, spawn a process, inject a prompt, write ordinary memory, merge, or push.

Rollback

Deactivate or roll back the affected governance policy by generation. Stop
issuing new selection authorizations or context-prep schedules. Public Agent
Fit remains advisor-only, and external task workers retain process and mutation
ownership throughout rollback.

Assets 2
Loading