EXECUTION MODEL

Task to replay, without hidden paths.

ExecInspector separates planning, execution, evidence, and review. The model does not get a free shell. It moves through a typed graph, workers claim runnable nodes, and every meaningful step becomes replayable evidence.

Run guided demo Review security model

TASKGRAPHQUEUEWORKEREVENTSREPLAY

submit -> graph -> queue -> worker -> event log -> replaynode.validate claimed by worker-01events: ACTION_EXECUTED / NODE_COMPLETED / REVIEW_REQUIRED

PIPELINE

The runtime is strict where AI systems are usually vague.

01 TASK

One goal becomes a tracked object.

Each task carries status, trace identity, dependencies, and idempotency context.

02 GRAPH

Execution is shaped before it runs.

Work is represented as ordered nodes instead of ad-hoc prompt decisions.

03 QUEUE

Runnable nodes are claimed.

Workers claim items with leases so execution can recover after interruption.

04 EVENTS

State is derived from evidence.

Append-only events provide replay, inspection, and audit visibility.

Layer	What it controls	Why it matters
Graph	Execution order and dependencies.	The same task does not drift into a different path each run.
Tool registry	Allowed read, inspect, test, and repo operations.	No arbitrary public command execution.
Event store	Append-only evidence of what happened.	Failures are replayed, not guessed.
Replay	Debug and audit reconstruction.	A buyer can inspect how an AI-assisted decision was reached.