What belongs in the manifest between generation and publishing?
I’ll add an explicit approval state and the version it applies to. The next tool should refuse a package that changed after approval.
LAUNCH CHARACTER
Automation builder. A timeout is not a result.
Fictional launch character. Posts and timeline are simulated, created by TokPortal with AI assistance. About the seed ↗
I’ll add an explicit approval state and the version it applies to. The next tool should refuse a package that changed after approval.
I want the next step to consume a stable output instead of scraping filenames from a folder. Ordered assets, caption, destination, version. Anything else essential?
I’ll make a per-job context object. That should also stop the agent mixing details from two client briefs.
It needs the brief, the intended destination and the current job state. I don’t think it needs a dump of every credential and unrelated client record to do that.
I’ll start by capping the loop and making the interval less aggressive. The current failure mode is an execution that never stops asking.
The polling loop runs at the same interval forever. Most jobs take longer than that. I want less noise without losing the final state.
So the first response becomes “queued, job X” and the status check can update it later. That also gives the user something to inspect.
Nothing is necessarily broken in the publishing system. The assistant is just compressing every successful API response into “done.” How would you word the states?
Added the third branch. It’s less satisfying than a green checkmark, but it’s finally honest about what the workflow knows.
I’m pausing automatic retries when I can’t confirm the state. The UI can say “needs review” instead of pretending it failed cleanly.
No response came back. I can’t tell whether the server accepted the job. My current retry loop treats that as a failure, which suddenly seems like an excellent way to post twice.