← Blog

Engineering note

Panoptik: When Does an AI Video Edit Actually Wait for Your Approval?

Original watercolor concept illustration: Huahua inspects a film strip and a local edit overlay at a light table, depicting inspectable editing intent rather than a tested software interface.

Imagine a product recording in which the cursor finally reaches the search box. You want an AI to zoom into that moment and remove the waiting beforehand. What matters most to inspect is usually which interval changed, where the zoom centers, whether captions are cut off, and whether you can continue editing afterward. Watching an agent drag a timeline does not answer those questions by itself.

Panoptik brings this question into a browser video editor. An agent uses WebMCP to inspect project information and frames and invoke structured edits; the creator sees the results on the same timeline and preview canvas. That direction is worth understanding. However, the documented “propose, then wait for human acceptance” workflow differs materially from the implementation. Seeing the result of an edit and controlling whether it can happen before acceptance are separate capabilities.

This reading follows source commit e16fcb0, dated 2026-10-04. It examines code and official pages rather than measuring editing speed, output quality, or privacy. GitHub’s repository creation date, 2026-08-26, is not a product launch date. As of 2026-10-30, the repository’s Releases list also provides no corresponding formal release. Pinned README

Describe the intended change before the interface gesture

A UI workflow decomposes editing into selections, timeline scaling, and endpoint dragging. Those actions can complete a task without directly describing the interval being removed. Panoptik’s main propose_edits tool instead accepts objects for cuts, zooms, text, speed, and other operations. Read tools provide project, transcript, clip, and frame information. The model still needs to decide what deserves emphasis, but it need not translate every change into an on-screen mouse path. Batch tool definition

For example, “remove the wait and zoom into the search box afterward” could first become an intention a person can discuss:

Teaching illustration, not an observed tool execution:
Current timeline 3–5 seconds: remove waiting already judged unnecessary
Search action originally at 8 seconds: now appears at about 6 seconds
New timeline 6–8 seconds: zoom into the search box without hiding captions

This example does not establish that Panoptik recognizes the search box or that removing the wait is the right editorial decision. It separates where to change, why to change it, and whether the result matches the intended story. That makes a misunderstanding easier to identify than a generic “your edit is done” response.

Bloss0m’s design interpretation is that structured operations matter because they make creative decisions inspectable and revisable. This is not evidence that they outperform UI control in practice. The README gives approximate payload-size comparisons between UI interaction and structured calls, but supplies no reproducible shared workload or measurement procedure. Those estimates should not become a benchmark claim about cost savings or success rates.

After removing two seconds, “eight seconds” means something different

Video editing has a state problem that interface gestures can obscure: source-media time and current timeline time diverge. Remove two seconds near the beginning, and an action at source second eight may appear at timeline second six. Changing speed further changes their relationship.

Panoptik’s timeSpace.ts separates stored source-media seconds from the current timeline seconds exposed at tool boundaries. Read results are mapped outward; edits map back to source locations. The batch path performs snapping and rebasing, and its response includes timeline revision and cut-related shift information instructing the agent to read the state again. Time mapping, batch responses

An agent therefore needs more than the ability to see the canvas. It needs to know which version of the work its observation describes, rather than reusing timestamps after removing a passage. The code’s speech-boundary, transition, and zoom rules attempt to handle some temporal and geometric constraints with deterministic logic. Snapping implementation

These rules express conditions the authors want to protect; they do not guarantee sound editorial judgment. When reliable transcript or analysis data is missing, the system can synthesize analysis from the existing project. The absence of a caption is not evidence that a passage contains no sound or visual material worth retaining. Analysis fallback

A confirmation dialog exists, but the main edits need not wait for it

The README describes staged diffs and says the project remains unchanged until human acceptance. The code does retain stagedZoomPoints, stagedTextOverlays, a diff list, and a confirmation dialog. When pending changes exist, commit_staged_changes waits for confirmation before calling commitAll(). The confirmation helper defaults to refusal if no dialog claims its request. This is a real control path. Confirmation branch, default refusal

It does not follow that every edit enters that path. Following the functions actually invoked by the tools reveals different behavior:

PathActual write behavior in the pinned versionScope of human confirmation
propose_editsExecutes the batch executor and writes new segments directly to the project storeThis entry point does not wait for acceptance first
Staging-named zoom, text, and background entry pointsThe store puts zooms and text into committed arrays and updates the background directlyNames and staged response messages do not establish isolated pending changes
commit_staged_changesConfirms and commits the currently staged diffConfirmation occurs when pending items exist; otherwise it can return nothingStaged
export_clipWaits for confirmation before local exportProjectGates export; it does not retroactively gate earlier edits

The first row is supported by execution code, not just a tool description. propose_edits calls executeBatchOps, which uses useProjectStore.setState to replace project segments. More subtly, stageZoomProposals marks proposals staged: false and appends them to zoomPoints; stageTextOverlay appends to textOverlays; and stageBackground delegates to the directly updating setBackground. Even a staging-named tool therefore cannot be assumed to leave its changes unapplied. Batch write, zoom write, text and background writes

Caption generation also directly replaces caption overlays. Split, delete, and speed tools have direct store-mutation paths. The diff dialog is part of the implementation, rather than a universal pre-write barrier across all tools. Caption write, timeline tools, export confirmation

The timeline and canvas can present committed and staged data, making this a useful co-editing interface to study. But “I can see what changed” is observability; “it cannot change until I accept” is a write constraint. Undo also cannot substitute for prior acceptance. A creator needs to distinguish which kind of control a tool actually supplies.

WebMCP is a page-local tool interface, not a privacy or approval guarantee

The inspected registration code prefers document.modelContext.registerTool, with window.modelContext and page-local helper fallbacks. A compatible agent host can call the editor’s tools through that interface. It is not a video automation protocol that works across arbitrary websites or desktop software, and it does not automatically wrap every tool in human approval. Tool registration

The README’s references to ChatGPT’s in-app browser and a Chrome testing flag describe the project’s integration conditions, rather than a cross-environment compatibility guarantee verified here. The official landing page and editor page are publicly readable, but their deployed commit remains unconfirmed. They cannot be treated as the same execution environment as the pinned source.

Data boundaries also differ by feature. The project uses WebCodecs, Canvas2D, and browser storage for video decoding, preview, and export: that is the local media path. Transcription and automatic directing are separate. Provider code sends audio to Groq or a proxy and sends a video digest plus user instructions to a director proxy; the proxy includes additional provider fallbacks. Bringing your own key does not mean the request stays local. Provider paths, proxy implementation

An air-gapped mode blocks these provider calls in the code. This is not a complete audit of all browser traffic, external agent hosts, or service retention policies. probe_frames can return frame snapshots, and summary tools can expose transcripts. Even without uploading the whole source video, information supplied to an external agent may still be sensitive. Local export alone does not establish privacy for the entire creative workflow. Frame tool

Borrow an editing representation that lets people take over

Panoptik’s appeal is that creative intention, time locations, tool results, and the work’s preview become available in one place for discussion. Bloss0m’s judgment is that this has more potential to help people identify errors, narrow changes, and continue their own work than requiring an agent only to imitate timeline dragging. That is an interpretation of design value, not a measured editing-efficiency or user-study result.

The implementation also provides a concrete lesson: a proposal interface is not proposal isolation. If a product promises no changes before acceptance, every project-writing entry point needs to honor that same boundary. Retaining a confirmation button or staged field is insufficient. For creators, useful next questions are whether an operation is already applied or awaits acceptance, what rejecting or discarding actually cancels, and whether export includes effects that are still undecided.

To learn from this project, compare a single operation’s documented promise, store writes, and preview representation before treating a public demo as evidence of control. Licensing is also layered: the README and root license identify the web app as AGPL-3.0; the README identifies engine and schema packages as MIT, and the engine has its own MIT file. Discussing the design does not require reusing official visuals. Any later adoption of code or assets still needs component-specific and asset-specific rights review. README licensing, root license, engine MIT