~/project/.dsh/skills/ProjectLoads for this repository only. The right place to read a bundle before you trust it with the whole machine.
Open-source agent harness (developer preview)
dsh reads skills straight out of folders you point it at, so a Felo bundle is one copy away from being callable in a session — and one delete away from being gone.
Transcripts are the slowest skill to run and the simplest to install. Copy the folder once and the session can reach for it whenever a task cites a video.
~/project/.dsh/skills/ProjectLoads for this repository only. The right place to read a bundle before you trust it with the whole machine.
~/.agents/skills/UserLoads in every project, and is the same folder other harnesses on this machine look at.
mkdir -p .dsh/skills
cp -r felo-skills/felo-youtube-subtitling .dsh/skills/List the skills you can call from this project.Timestamps are the difference between a summary you have to trust and a document you can check. Every line points back at the moment it was said.
We are moving the default to the async client in this release, and the sync path becomes a wrapper around it.
If you were passing a custom retry policy, the signature changed — that is the one that will break you.
So the upgrade is safe if we are not overriding retries?
Correct, and the old endpoint keeps working until the next major.
We also fixed the connection reuse bug people reported on long-lived workers.
If you were passing a custom retry policy, the signature changed — that is the one that will break you.
The useful version of this skill is retroactive: the talk you saved three weeks ago and never got to.
Pull the transcript from this release stream and list every API the presenter says changed, with the timestamp for each.What comes backA change list you can act on, each entry traceable to a moment you can replay before you commit to a refactor.
Compare what this talk claims against our current implementation and flag anything we are not doing yet.What comes backA gap list rather than a summary, which is the version that actually changes what you work on next.
Video tooling fails quietly. These checks catch the two failure modes that matter.
A transcript with no timestamps cannot be checked against the video, which removes most of its value for engineering work.
Ask the session which URL it read. Auto-caption paths sometimes resolve to a different upload or a re-upload with different timing.
Search the transcript for the phrase in your pull quote. If it is not there, the summary drifted from the source.
Video is a weak source in specific, predictable ways. Knowing them keeps you out of trouble.
A presenter can be wrong, out of date, or simplifying for the room. Treat the transcript as evidence of a claim, not proof of it.
Captions carry speech. If the important part was a diff on screen, the transcript will not have it — ask for the accompanying repo or deck.
Expect mangled product names. Grep for the concept rather than the exact string when you search the result.
Five skills written for DeepSeek Harness. They follow the same pattern on this platform, so you can combine them in one session without re-explaining your setup.
Start with a skill, then compose the exact models, live data, knowledge, and output tools your dsh workflow needs — the same key every Felo skill already uses.
Use cost-efficient LLM access when your agent needs custom reasoning or generation.
Give agents current web and social signals instead of relying on stale context.
Ground work in reusable knowledge, then turn it into shareable deliverables.
Extend the workflow from research to visual and office-ready results.