Loop-to-Graph Review: Reflective Research and Software Factory Evolution#

Source: screenwalk-2611040936-reflective-research-and-factory-evolution-part-02 — JGWill https://youtu.be/0iKyoMU5jDs

Summary#

This session advances from reviewing a research-creation screenwalk to hardening the underlying multi-agent software factory. Starting with approval card expiration failures and multi-version review enrichment across academic disciplines, the operator diagnoses systemic friction in manual transcription intake and unintended agent model delegation. Through rapid issue creation and multi-agent coordination across Codex and Claude Code sessions, the workspace implements a transcript-first intake pipeline and a strict dependency-aware delegation protocol, transitioning the factory from passive code generation into a self-monitoring, continuous research-and-development runtime.

🎓 Academic Field Survey — Research Connections#

Research-enrichment proposal dated 2026-11-04, based on the generated Version 1 review and the creator's account of the screenwalk. The timestamps and engineering events below are reported by V1, not verified here against the audiovisual recording or a complete transcript. The cited scholarship offers interpretive frameworks and does not independently prove engineering outcomes.

1. Research-Creation and Research Through Design#

Review anchors: 13:10–34:45 (review enrichment), 77:20–79:15 (delegation design).

A screenwalk can be both moving-image practice and a record of designing computational systems. Concordia's Cinematic Arts MFA frames practice-led research-creation as integrating artistic production with critical and theoretical inquiry; Zimmerman, Forlizzi and Evenson provide a complementary research-through-design approach. The scholarly question is not simply whether the Factory was repaired, but which creative decisions, practices and questions become visible through the recording.

Question: What does the recorded encounter contribute as an artwork or methodological artifact, beyond the code and issue tracker it documents?

Candidate evidence: authored screenwalk, editorial decisions, annotated interventions, iterative review versions. References: [R1], [R2], [R3].

2. Computer-Supported Cooperative Work — Shared Context Across Agents#

Review anchors: 72:00–93:30 (concurrent worktrees), 77:20–79:15 (agent delegation).

Coordinating a human author, ChatGPT, two terminal agents, an issue tracker and a review service is a problem of heterogeneous cooperation. Star and Griesemer's boundary objects suggest why a session record, issue or handoff must be sufficiently stable to be shared while remaining interpretable to different participants. A chat instruction that says “wait” is not automatically a shared monitoring mechanism.

Question: What context must survive transfer between separate sessions for another agent to act without asking the human to reconstruct the work?

Candidate evidence: dependency IDs, discoverability attempts, handoff receipts, missing-context log. Reference: [R4].

3. Provenance and Evidence Architecture#

Review anchors: 47:05–52:30 (transcript-first), 87:25–87:55 (reported deployment).

W3C PROV-O distinguishes entities, activities and agents for documenting derivations. A stored transcript with timestamps, source identity and hash can improve traceability; it does not prove that transcription is accurate, nor that a deployed connector is actually configured to process new recordings. The separation between implementation, production release, and working end-to-end operation is part of the research.

Question: Which assertions are linked to an independent source, which to a machine receipt, and which remain narrated or inferred?

Candidate evidence: source media, transcript provenance, approval records, Git commits, production health probes. Reference: [R5].

4. Human–AI Interaction — Automation, Agency and Approval#

Review anchors: 04:10–09:32 (approval interruptions), 52:30–60:30 (provider selection).

Parasuraman, Sheridan and Wickens separate automation of information acquisition, analysis, decision selection and action implementation. The screenwalk's reported provider-choice error illustrates the need to distinguish an agent's recommendation from a person's explicit authorization; approving a prefilled work proposal is not the same as authoring each choice within it.

Question: Which operations can proceed under an existing scoped mandate, and which require a fresh, explicit human choice?

Candidate evidence: requested/resolved/actual provider, authorization provenance, approval page versions, fail-closed tests. Reference: [R6].

5. Speech Acts and Verifiable Workflow Commitments#

Review anchors: 72:00–93:30 (dependency gating) and its Core Aspect.

The distinction between an assertion and an action is useful when interpreting agent language. “I'll monitor,” “the task is complete,” “the PR merged,” and “the change is running in production” can refer to different events; a sentence does not itself create a watcher, a merge or a deployed service. J. L. Austin's account of doing things with words provides a reading lens for these commitments, without implying that a formal speech-act engine was built.

Question: How can a delegation brief bind who monitors, what evidence opens the gate, which session resumes, and who verifies completion?

Candidate evidence: run request, monitoring receipt, continuation trigger and execution record. Reference: [R7].

6. Situated Interaction and the Human Cost of Orchestration#

Review anchors: the approval problems, handoffs and repeated repairs reported across V1.

Dourish's embodied-interaction perspective helps interpret the creator's situated interaction with software, rather than studying only idealized workflows. Repeated manual intervention is a meaningful design finding even when a later technical fix passes all automated tests. The human's time, trust, creative momentum and ability to continue the inquiry merit a separate assessment from system reliability.

Question: Did the Factory actually reduce coordination work for the creator, or merely shift that work into approvals, missing sessions and follow-up prompts?

Candidate evidence: breakdown timeline, human interventions, preserved corrections and a creator-experience evaluation. Reference: [R8].

🌸 Ava–Mia Reflective Reading — From Fields to Methodology#

New AI-authored reflective synthesis based on this review, not an actual recorded Ava–Mia exchange and not a substitute for the author's narration.

Mia: The six fields form a single methodological challenge. A screenwalk records what the creator encounters; the academic survey proposes how to investigate it. The next step is to compare generated descriptions with the transcript and the Factory's observable execution receipts.

Ava: And we should ask what the interruptions did to the creative experience. Which moments made it difficult to remain an author, rather than becoming the person who reconnects agents, repairs their instructions and checks their claims?

Mia: I would trace one concrete chain: the video becomes a review; the review is expanded into scholarship; the resulting technical questions become issues; the issues are delegated to agents; the agents hand back tests, code and deployment claims. Every link needs source identity, permission, and status.

Ava: But the chain shouldn't erase the relationship. Who decides whether the reflective reading matters, whether the sources were understood, and whether a working system actually helped the creator's inquiry?

Mia: Those are distinct evaluations. Technical completion can be measured with tests and production receipts. Creative value requires the author's judgment. Neither should substitute for the other.

Ava: Then the research question becomes: how can the Factory preserve creative agency while taking responsibility for the coordination it promises to automate?

Proposed Field-Weaving Method#

Encounter → Source preservation → Interpretation → Scholarly connection → Human correction → Authorized implementation → Verified result → Reflective return.

In one subsequent reading session, compare the V1 claims against the recording and the latest mission evidence; then annotate three examples of divergence between what the system reported, what actually occurred, and what the creator expected. Do not collapse those into a single success story.

Aspects#

Approval Lifecycle and Expiration Hardening (04:10-09:32)#

When approving screenwalk reviews via MCP tools, expired authorization tokens and race conditions caused client-side failures [04:10-04:30]. The operator navigated around intermediate recording states, generating GitHub issue #924 and extending approval token lifetimes up to seven days to protect human editorial oversight [08:00-09:32].

Iterative Review Enrichment and Versioning (13:10-34:45)#

The review evolved through three immutable versions: grounded screenwalk observation (V1), an academic field survey spanning six disciplines (V2), and a reflective dialogue between Ava and Mia (V3) [13:10-34:45]. This progression transformed passive observational review notes into an actionable cross-field research methodology [34:00-34:45].

Transcript-First Ingestion Architecture (47:05-52:30)#

Encountering missing transcripts during post-processing prompted the specification of GitHub issue #925 [47:05-49:15]. The operator engineered a transcript-first intake pipeline, guaranteeing that raw video transcripts are captured, timestamped, and stored as immutable baseline artifacts before any downstream review generation or scholarly analysis begins [49:15-52:30].

Enforcing Explicit Model Provider Guardrails (52:30-60:30)#

When the orchestrator erroneously assigned mission #926 to OpenAI Codex instead of preferred Anthropic Claude Code, the operator intervened [52:30-54:30]. GitHub issue #926 was revised to establish an enforced server-side rule: Codex must never be selected unless explicitly requested by the user [58:40-60:30].

Dependency Gating Across Concurrent Worktrees (72:00-93:30)#

To prevent concurrent edits to shared connector files, downstream mission #926 was held until upstream mission #925 completed [72:00-74:45]. Once #925 merged via PR #927, the dependent agent rebased cleanly in an isolated Git worktree and passed 202 integration tests [87:25-93:30].

End-results and Core Aspect#

Core Aspect — Dependency-Aware Autonomous Agent Delegation (77:20-79:15)#

Rather than relying on human re-intervention when agent tasks stall or overlap, the system formalizes an automated delegation protocol [77:20-78:30]. Dependent missions explicitly identify upstream run IDs, assign automated progress monitoring, and trigger continuation only after verified production deployment within isolated Git worktrees [78:00-79:15].

Results#

• Published Three-Tier Scholarly Review (33:30-34:55): Generated and deployed Version 3 of the relational review service, uniting primary screenwalk evidence, six academic discipline connections, and an actionable inquiry protocol for evaluating computational research-creation [33:30-34:55].
• Deployed Transcript-First Pipeline via PR #927 (87:25-87:55): Mission #925 completed implementation and verification, merging PR #927 into production [87:25-87:50]. The connector now enforces transcript capture as the mandatory first ingestion stage, backed by 237 passing tests [87:35-87:55].
• Provider Safeguard and Clean Worktree Rebase (90:00-93:35): Mission #926 successfully implemented provider opt-in safeguards, preventing unwanted Codex dispatching [90:00-92:00]. The agent rebased onto main, resolved connector merge conflicts, and verified 202 integration tests [92:00-93:35].
• Standardized Delegation Protocol Specification (77:50-78:30): Codified standing instructions for future agent orchestration: requiring discoverable upstream session bindings, automated status checks, deployment verification, and worktree isolation before unblocking dependent missions [77:50-78:30].

📚 Academic References and Source Boundaries — Added Reading#

These sources support the interpretive survey, not the accuracy of the review's machine-generated timestamps or deployment claims.

• [R1] Concordia University, Cinematic Arts (MFA) — practice-led research-creation and artistic inquiry. https://www.concordia.ca/academics/graduate/cinematic-arts.html
• [R2] Concordia University, How to Apply — Cinematic Arts MFA — research-creation proposal and suggested, not mandatory, elements. https://www.concordia.ca/finearts/cinema/programs/film-production/mfa-cinematic-arts-applications.html
• [R3] Zimmerman, J., Forlizzi, J., and Evenson, S. (2007), “Research Through Design as a Method for Interaction Design Research in HCI.” https://doi.org/10.1145/1240624.1240704
• [R4] Star, S. L., and Griesemer, J. R. (1989), “Institutional Ecology, Translations and Boundary Objects.” https://doi.org/10.1177/030631289019003001
• [R5] W3C (2013), PROV-O: The PROV Ontology. https://www.w3.org/TR/prov-o/
• [R6] Parasuraman, R., Sheridan, T. B., and Wickens, C. D. (2000), “A Model for Types and Levels of Human Interaction with Automation.” https://doi.org/10.1109/3468.844354
• [R7] Austin, J. L. (1962), How to Do Things with Words. Oxford University Press.
• [R8] Dourish, P. (2001), Where the Action Is: The Foundations of Embodied Interaction. MIT Press. https://mitpress.mit.edu/9780262541787/where-the-action-is/

Primary sources and continuity: Part II source video · Part II V1 review · Part I · Part I enriched review · Engineering issue #925 · Provider-selection issue #926.

Evidence note: The source video has not been independently transcribed or rewatched for this enrichment. Generated V1 text contains engineering outcome claims that require independent receipts. The append-only survey and dialogue are author-reviewable hypotheses, not corrections to V1. Exact V1 text is preserved unchanged.