Choose the agent that writes the plan

After the film factory produced its four completed movies, I asked whether I could return to Codex and request a film of SC-000001. The answer exposed a dependency I had never requested: its sole planning call was wired to a signed-in Claude command-line tool.

I explicitly required the choice to remain mine, whether Codex planned, Claude planned, Copilot planned or another interface supplied the book, while the same theatre and build route continued to execute it. The permanent theatre and the build route should not require an unrelated model subscription simply because that model had helped construct them.

The amendment now gives the planning stage a selectable command and an option for the calling agent to write the book itself. A live Copilot trial reached a checker-clean plan; the other routes retain their own, narrower evidence.

One document is the connection

The planner’s output is a prompt book containing the film’s movements, gestures, scenery cues, lights, views and music on the script’s clock. The following stages compile that document, build its media, simulate and measure the scene, record seven cameras and assemble the movie.

Keeping that document explicit made the dependency removable. The planner does not need a private connection to Blender. It needs to read the script, timing, exact dialogue media, standing directions and equipment catalogue, then write a book that the same checker accepts.

Claude Code replaced the hard-wired call with presets for Claude, Codex and Copilot, plus a custom command option. A custom command can receive the prompt, production folder, workspace and selected model through placeholders; a command without a prompt placeholder receives the prompt directly as its input. The factory retains each round’s command, output, errors and exit status.

These are implemented invocation paths. Their existence does not establish that every possible model can plan the show well, or that every preset has produced a complete new film. The contract stays the same while the caller chooses who attempts it.

Let the current agent write the book

The most direct option is --planner none. The name means that the runner launches no separate planner process. Instead, it writes the exact task into plan/NEEDS-PLANNER.md and stops with a distinctive exit code:

PLANNER_EXIT = 3   # the runner stopped because the caller must write or repair prompt-book.json

Exact implementation and retained trial evidence. Retained source file.

The calling agent reads the note, writes or repairs prompt-book.json, checks it, and invokes the factory again, using the same handoff to receive checker problems or actual physical failures if the performance needs another correction. The model receives a specific piece of work in the existing performance folder, rather than a request for me to coordinate another application.

The plan stage also records the book’s own bytes among its inputs. Removing or editing the book makes the stage check it again. A previously accepted plan therefore cannot supply a passed state to a changed file merely because the surrounding script remained the same.

Keep the trial separate from the delivered films

The portability trial used SC-000002 while preserving and later restoring its production book. Caller-planned mode wrote the handoff note and returned code 3. A small custom test command copied the retained book back into place and passed the planning stage. Those operations tested the connection without making another film.

The live Copilot command-line tool, configured with model selection auto, then wrote a checker-clean book in two rounds of about 34 and 30 seconds. Its candidate remains saved separately. The trial establishes real planning through that tool; it does not identify a particular served model from the word auto or prove another complete production through Copilot.

The Codex command-line preset was written but could not execute in that trial because a required executable was missing from its installed package. No Grok command was tested. These are dated limits of the recorded trial. Another Codex thread subsequently began a cached factory invocation using caller-planned mode, which has to be judged from its own run and release records.

The four-film repeat belongs to the preceding frozen code. This amendment changes the runner afterward, preserving that successful history while adding the caller’s choice. The next agent still has to meet the same script, theatre, physical and delivery requirements, and can now receive the exact unfinished planning task without signing into Claude.