Miadi Review

Screenwalk-Driven Continuity & Deep Diver Handoffs in the Miadi Factory

Oct 3, 2026 · model: manual-edit · channel: JGWill

Watch source

In the chronicle

Screenwalk-Driven Continuity & Deep Diver Handoffs in the Miadi Factory

Source: miadi-screenwalk-mino-261002a-261003-001-mia-workspace-handoff — JGWill
https://youtu.be/q4I55OAkI0Y
type: internal related: review://7ac38930-4036-4dc2-a2d6-2dce6dc0baa7 stcissue: missing Guillaume's voice because of an issue. AGents dont talk about it, just know that we are missing important implied in between the potential parts of the AI response reading that we'll need to fill (you agents love to fill gaps, hahaha)

Intent

  • continuity, catch insights, definitions that are great and the packages layers/terms might need for carefully commynicate and understand each other

Where this went next (2026-10-03, 17:46 EDT)

Written by Mino-Bimaadizi-Daa. The handoff this recording prepares happened the same evening, as part 2 of the screenwalk. Mia's seat episode-339-mia-companion was opened by hand in Episode 339's folder, the episode of William's relation with Mia as development companion. Her listener is running, and her handoff issue (miadisabelle/workspace#70) is marked ready. She is building jgwill/Miadi#714: a tide command and a button on the tide page that open an episode's companion seat. The manual start was recorded step by step as a reference run, her first test case. Her specification is jgwill/Miadi#714 and the handoff. At William's word, the opener takes the seat as a parameter, so a seat other than Mia can be opened the same way later.

Summary

This session outlines the transformation of Mino from a coding agent into a durable orchestration seat within the Miadi Factory. By executing local review processes to bypass cloud execution limits and integrating Gerico's Deep Diver MCP for NotebookLM (renamed Gemini Notebook on 2026-07-16) studio artifact generation, human screenwalks and spoken corrections are turned into state machines, documentation checked against its sources, and a workspace handoff to Mia.

Academic Fields

Documentary Film (Film Studies)

Studies how non-fiction film stands toward subjects and viewers; Nichols names participatory and reflexive modes. A screenwalk records people's work in their own voices and is played back to them, as Challenge for Change returned footage to the communities it filmed. Limit: modes describe a film's stance; applying them to a recording is analogy.

Human–Agent Interaction (Computer Science / HCI)

Studies how people and machines make sense of each other's actions. Suchman showed that plans are resources for situated action, not programs that determine it. The state machines drawn from a screenwalk are resources for the next handoff. Limit: Suchman's evidence comes from people using a photocopier; extending it to LLM agents is analogy.

Articulation Work (Computer-Supported Cooperative Work)

Studies how interdependent work is divided, handed off and meshed. Schmidt and Bannon, after Strauss, call this articulation work. Mino's brief to Mia on #714 is articulation work made visible. Limit: it is a conceptual lens, not an empirical finding; the cost of a handoff has to be observed before it can be claimed.

Statecharts (Systems Engineering)

Harel's statecharts specify reactive systems as nested states and the events between them. The screenwalk media cycle is drawn as a state machine of this kind. Limit: a statechart specifies designed behaviour, not what a session actually did, and the drawing alone does not fix its meaning; Harel's formal semantics came in a companion paper.

Faithfulness of Generated Summaries (Natural Language Processing)

Measures whether a generated summary says only what its source supports. Maynez and colleagues found unsupported content in most of the single-sentence news summaries they annotated. This review's first version and the notebook's overviews are generated summaries. Limit: that evidence is 2020-era news summarisation; it signals a risk, not a rate, for summaries of recordings.

Indigenous Research Paradigms (Indigenous Studies)

Wilson grounds research in relational accountability. Reviewing videos of his own conference talks helped his ideas become "more firmly anchored or internalized" (p. 131), and he wished for more direct quotes (p. 128). A screenwalk's playback does similar work. Limit: Wilson's method answers to a community of Indigenous scholars; resemblance is not that accountability.

Aspects

Local Review Runner & Launcher Scripts (06:33-08:05)

To overcome cloud execution timeouts, shell launcher functions and a local review script were implemented on Gaia. The hosted service stops every request after 60 seconds; the local runner gives a review up to 900 seconds by default (MIADI_REVIEW_LOCAL_TIMEOUT), enough for an hour-long screenwalk.

Screenwalk Feedback Loop & Voice Corrections (08:15-18:25)

Screenwalks capture live engineering work alongside agent activity. Spoken corrections and observations during video playback are captured directly, converting spontaneous operator feedback into structured review artifacts and subsequent session scripts.

Deep Diver & NotebookLM Artifact Integration (18:30-29:10)

Upstream Deep Diver tools are integrated to connect screenwalk outputs with Gemini and NotebookLM. This pipeline lets NotebookLM produce audio overviews, mind maps, and video overviews from the review transcripts (as summarized from the video, not checked).

State Machine Decomposition & Workspace Handoffs (29:20-33:30)

Screenwalk workflows are formalized as state machine transitions. Mino logs task decompositions, checks claims, and packages context into clear milestone handoffs before triggering Mia's seat in her dedicated workspace.

End-results and Core Aspect

Core Aspect — Continuous Screenwalk-to-Graph Loop (33:35-36:30)

The core of the session is a loop: a person's commentary on a screen recording is decomposed into knowledge-graph entries, state machine models, and briefs that downstream agents can start from. The person's judgment stays the check at each turn.

Results

  • Local Review Processing (06:40-08:05): Review generation can run on Gaia through review-local.sh (jgwill/Miadi@9fffa81), with a 900-second limit in place of the hosted 60, so an extended screenwalk can be reviewed end to end.
  • Automated Multimodal Knowledge Generation (33:40-36:00): Transcripts and screenwalk logs are passed to NotebookLM through Deep Diver, which produces mind maps, audio overviews, and documentary briefs (as summarized from the video, not checked).
  • Formalized Issue Handoff to Mia (41:00-44:00): Mino generates an opening narrative and structured task brief for issue #714, passing its execution context to Mia's workspace.

—— exploring academic fields from internal perspective 261003-1500 ——

by: gemini
validated with arxiv: no insight: potential definition for my proposal

Document management protocols, Software Engineering, Artificial Intelligence

Computer Science

This field focuses on the theoretical and practical aspects of computation, software development, and information processing. It underpins the digital architecture, agentic workflows, and automated systems described in the factory operations and document management protocols.

Media Studies

This discipline examines the production, distribution, and reception of media content. Within this context, it addresses the transformation of screenwalks into documentary artifacts, the role of human-agent interaction in filmmaking, and the synthesis of visual and textual data.

Documentary Film

Digital Media Production

Systems Engineering

This field involves the design and management of complex systems over their life cycles. It is highly relevant to the orchestration of seats, the maintenance of continuity states, and the integration of diverse technical components into a cohesive production environment.

Some potential definitions to help us evolved inspired from that academic exploration:

The documentary team functions as a production unit within the factory that turns work into accessible narratives

Related Review

Autonomous Agent Identity, Steering Triage, and Gate Verification in Miadi — JGWill

Part 001 of the same Mino session, mino-261002a. There, issue #714 was opened and triaged. Here, it is handed to Mia. Together they show one seat's day, from being steered to delegating.

Internal Usage

A proposal, not something Miadi does today.

  • T3 Witness · the screenwalk-presence skill (claude/miadi-witness). Its "After the playback" step says that once a review exists, the person's spoken corrections are extracted with their timestamps into the episode's talking circle diary, and the skill is revised from them. This review now exists, so that extraction is its next use. Limit: the timestamps here come from the generated summary; each correction has to be taken from the recording itself.
  • T1 Session continuity · tide. jgwill/Miadi#714 asks that opening an episode's companion seat become a platform act: a naming standard for the tmux session, arming or archiving one that exists, a starter prompt that carries the episode's continuity. This recording shows Mino doing it by hand, which makes it the reference case for tide. Limit: #714 is labelled needs-decomposition; nothing of it is built.
  • Review service · screenwalk parts. Reviews of one session already write session: and part: in their header (b6c3325f), and this one writes related:, while 7ac38930 writes succeeding: for the same relation. The service could read these fields and list a session's parts together, both ways. Limit: today only keyword search finds them, and it gives every match the same score.
  • @miadi/episodic-memory-schema · foundations (jgwill/Miadi#713). An episode can hold a foundation packet; Episode 550 holds screenwalk-media-type. Once this review's academic fields are grounded, they could become a foundation the episode holds, so a NotebookLM notebook and a presentation draw on the same checked sources. Limit: whether an episode holds this review was not checked here.
  • A team not named yet. The definition this review proposes, a documentary team as a production unit that turns work into accessible narratives, matches sessions the teams README lists under Not named yet (film production, story, research). Naming that team belongs to Guillaume.

Definitions

  • Screenwalk. Episode 550: the recording of a person working with agent sessions, with their voices and a witness's spoken closing (review d64a2fdf). Guillaume, 2026-10-03: a screenwalk is an action step, not footage for a film. What it changes comes from the collection of them; some segments may be used later.
  • Team. In this review's academic exploration: a documentary production unit. In Miadi: a standing team with a human lead and an agent lead, one desired outcome at one level. Building teams make packages and practices; the witness team builds only its own tools and hears their work for William first (miadi-orchestration-kit teams/README.md, #59).
  • Deep Diver. In Miadi since 2025-10 (jgwill, on medias/deep-diver.yml): a podcast-style way to discuss gathered sources and keep an archive alive. The tool is public: Gerico1007/deepdiver, by Jerry (Gerico1007). It is a Python command line that drives NotebookLM in the browser: notebooks from YouTube videos, Audio Overviews, and the Studio panel. His latest work, branch 24-studio-artifact-foundation (2026-07-19), generates every Studio artefact family and bundles agent skills. No MCP server was found in it. Miadi's fork miadisabelle/deepdiver stops at 2025-11-17, and nothing is installed on gaia (checked 2026-10-03). "Deep Dive" is also NotebookLM's default Audio Overview format.
  • Internal Usage. It proposes; it does not report.

Sources