WEBVTT

00:00.000 --> 00:18.000
NIRA v0.5 is a local-first assistant with visible offline status, local storage, and explicit side-effect approval.

00:18.000 --> 00:38.000
The canonical runtime reports healthy deterministic-offline mode. Chat works without pretending that a local model is running.

00:38.000 --> 00:52.000
The Operations Center reads the canonical runtime; its overview counts real roles, tools, workflows, conversations, messages, and loaded models.

00:52.000 --> 01:08.000
The agent board shows sequential routing, specialist, planning, safety, execution, critic, and response activity plus completed handoffs.

01:08.000 --> 01:22.000
Memory evidence is local and count-based. Conversation content is not exposed in this dashboard.

01:22.000 --> 01:36.000
Workflows show registered templates and completed task-plan states from the last real request.

01:36.000 --> 01:48.000
Model routing exposes aliases and cache limits while keeping the unavailable local-model state honest.

01:48.000 --> 02:00.000
Tools and permissions show the real registry. Only read and NIRA-state access are allowed by default.

02:00.000 --> 02:12.000
System Health shows live metrics and a privacy-safe contract without workspace paths, state directories, or session identifiers.

02:12.000 --> 02:44.000
Conversation history is local and user-controlled: search, open, pin, rename, export, and confirmed deletion.

02:44.000 --> 03:15.000
The first real tool inspects this repository while excluding virtual environments, dependencies, caches, and build output.

03:15.000 --> 03:35.000
The second tool reads bounded workspace content and reports the applied size limit.

03:35.000 --> 03:55.000
An attempted path escape is rejected without reading the outside file.

03:55.000 --> 04:20.000
A coding plan reaches process execution and pauses. Denial is the safe default, stops the task, and is never auto-retried.

04:20.000 --> 04:35.000
Permission evidence stores the tool, access class, verdict, timestamp, and reason—never the raw arguments.

04:35.000 --> 05:00.000
A second verification is approved once. The narrow compile command runs through the same registry and permission boundary.

05:00.000 --> 05:12.000
Returning to Workflows shows completed states from the real verification plan, separately from registered templates.

05:12.000 --> 05:25.000
Release evidence: 51 tests, clean dependency and secret scans, and a v0.5 wheel verified outside the source tree.

05:25.000 --> 05:40.000
Honest limitation: no real model profile, optional voice and PyQt outside core, and retrieval/accessibility evaluation still pending.