Trellis threads

Think about how email threading works. You reply to a message, and your reply folds into the conversation it belongs to — you get one continuous line of exchange, instead of a hundred disconnected messages that all happen to be about the same thing. Trellis does exactly that for the work your agents do. Every piece of work — a fix, a decision, a half-finished idea — gets folded into the ongoing thread it belongs to, across sessions and across agents, instead of scattering into file after file nobody can piece back together later.

A Trellis thread outlives any single session. It survives past the agent that started it, past the day it was opened, past whichever tool you happened to be using at the time — different from what a "thread" means in most AI chat tools, where it's just one session's temporary state, gone the moment the session ends.

The diary underneath

Underneath the surface, a Trellis thread is really one long diary — the kind that can only be added to, never edited or erased. Each sphere keeps exactly one of these, and it's the sole source of truth for everything that happened. Every other view you ever see — a list of open work, a summary, "what's blocked right now" — is generated fresh from that diary on request, not stored separately. Delete one of those generated views by accident, and it comes back identical, because it was never anything more than a readout of the diary in the first place. Nothing drifts out of sync, because nothing else is allowed to be authoritative.

The four questions every new piece of work answers

Before an agent writes anything at all, it works through a short sequence of questions — a judgment call the agent makes in the moment, not a formula a computer calculates for it. No similarity scores, no ranking algorithm guessing which thread something belongs to. The reasoning behind that choice: the agent sitting in the actual conversation, watching the work happen, is a better judge of where it belongs than a system trying to reconstruct that context afterward from bare text.

1. Has this been touched before? If not, it opens a brand-new thread.
2. If it has — how does it relate? Does it continue something already underway (connect it), take a genuinely new angle on work that's already mature (open a side-thread), or tie two existing threads together (merge them)?
3. Has it matured into something real? If so, it graduates into a full project.
4. Always, regardless of the above: leave a clear note about what should happen next, and keep anything produced in the thread's own home rather than scattered elsewhere.

The refusal has teeth

The tool will simply refuse to save any work that hasn't first declared where it belongs — no exceptions, and no way for an agent to quietly skip it under time pressure. In stricter spheres, it goes one step further: nothing at all can be written to the sphere in a session until that declaration has been made, enforced automatically before the very first file touches disk.

Threads can only ever connect one way — like a family tree, never a loop back on itself — and a quick check runs before every new connection to make sure that stays true. And when an agent genuinely isn't sure where something belongs, there's a specific way to flag that: it forces a real one-on-one conversation with a human before anything gets classified, so an uncertain guess can never quietly pass as "figured out."

Graduation is permanent

When a thread matures into a real project, that moment gets marked permanently in its history — the same idea as a milestone marker in version-history tools, something that can't later be moved or erased. The thread keeps its same identity and its entire past; only its status changes going forward. From graduation onward, it follows a defined path: explored, then defined, then built, then waiting for a go-ahead, then actually running, and eventually archived.

The radiator, not the fridge

There's a view built specifically to show you what needs your attention, and it's worth knowing by name — the standard calls it "the bird" (fugl in the original Norwegian), because it's meant to be a bird's-eye view. Anything blocked, and anything specifically waiting on you, sits at the top, always. As the standard puts it: a list sorted by what was touched most recently is a fridge you have to open to see what's inside. This one radiates heat toward you without being asked.

What's proven, and what's still ahead

Trellis shipped only after its own internal test suite passed all 107 of its checks, and after five separate attempts to deliberately break it, each one hunting specifically for ways it could fail. Those attempts turned up more than 60 real issues, nineteen of them serious, and every one was fixed before the system was considered done.

One thing worth being honest about: threads that span more than one sphere at once aren't built yet. Today, each sphere keeps its own separate diary. Letting a single thread follow work across multiple spheres is a real direction the standard is heading, not something you can rely on today.