openclawbrain-site

Worked Example: One OpenClaw Turn End to End

The fastest way to understand OpenClawBrain is to watch one question travel through the system. This page traces a single conversation turn from the moment your agent receives it to the moment a correction makes the guide smarter for next time.

Boundary

OpenClaw keeps the hot path, prompt assembly, response delivery, and fail-open serving.
OpenClawBrain supplies normalized events, packs, activation helpers, deterministic compilation, and asynchronous learner refresh.

Example turn shape

1) OpenClaw receives a user turn

OpenClaw receives a user message and keeps control of the hot path.

2) Compile from the promoted pack

OpenClaw resolves the active pack and compiles bounded context from it.

If the promoted manifest requires learned routing, the compile must use that pack’s learned route_fn and expose routerIdentity in diagnostics.

Relevant package surface:

@openclawbrain/activation
@openclawbrain/compiler
optional @openclawbrain/openclaw

3) Prompt assembly stays in OpenClaw

OpenClaw assembles the final prompt from the compiled context and serves the model call.

4) Delivery stays fail-open

OpenClaw sends the response and can keep serving even if learner refresh is stale or unavailable.

5) Normalized events are exported off the hot path

The turn is written into normalized interaction and feedback events for later learning.

Relevant package surface:

@openclawbrain/events
@openclawbrain/event-export
@openclawbrain/openclaw

6) Learner and activation update asynchronously

OpenClawBrain materializes candidate packs, refreshes learned routing artifacts, and stages or promotes them behind the hot path. That is asynchronous candidate-pack refresh, not proof of live mutation of the currently active pack.

Relevant package surface:

@openclawbrain/learner
@openclawbrain/activation
@openclawbrain/pack-format

What to save for proof

For one useful turn-level proof bundle, keep at least:

inbound summary and stable turn_id
compile request parameters and resulting diagnostics
returned bounded compiled context block
answer trace or response identifier from OpenClaw
exported correction or feedback artifacts for the same turn when available
later candidate-pack refresh and promotion evidence when you want to show a changed served route artifact

Claim boundary

This worked example proves the intended package boundary and learning-first operator story. It does not by itself prove comparative benchmark performance, full benchmark coverage inside this repo, full live runtime plasticity on the active pack, or Brain Ground Zero route-function / QTsim proof parity.