ChatCode

The deciding trait is the mode split. Ask, code, debug and architect are separate lanes rather than one assistant guessing which job you meant, and in practice that means the debug session does not wander into a redesign and the architecture question does not come back as a patch. Once you get used to naming the job first, going back feels careless.

What you accept is a tool built for one market and documented for it. Pick it if that is your market. Pick Zoo Code if you want the same modes without the account.

The risk is recursion with the brakes optional. A task decomposes into subtasks, each subtask runs in isolation and may decompose again, and the tree is described without a depth limit, a spend ceiling or a documented rule for detecting two modes handing work back and forth. Turn on auto-approve, which is a listed feature, and nobody is watching the branch that forgot to stop.

What it does right is granularity of permission. Reading, writing, retrying, mode switching and execution are separate grants, so the gate you leave open is the gate you chose.

1. Applying a diff instead of rewriting a file bounds the blast radius of a bad generation to the hunk, which is the edit format that survives model changes best. 2. Reading files in parallel is a latency decision rather than a quality one, and it is worth noting that the two are described together as though they were the same improvement.

2. Knowledge-graph retrieval is named without a description of what the nodes are or how they are built, so it cannot be compared with anything. No benchmark is published, and none is claimed.

Read the incentive rather than the product. An operator monetises inference capacity, so the editor plugin is distribution for the platform underneath and will be funded as long as that platform matters to the parent. That is a stronger continuity story than most venture-backed tools here can tell, and it is also why nothing is published about what it costs.

Moat: the parent's own customer base. Likely acquirer: none, and no exit is intended. Position: the vendor risk is low and the pricing risk is entirely unknown, which is an unusual pair.

The blocker is identity, not money. Sixty seats cost nothing to install and then require sixty accounts on an operator platform my company has no relationship with, and the published material describes no single sign-on, no directory provisioning, no audit export and no retention statement I could hand to a security reviewer.

There is also nothing that runs unattended, so it never becomes a pipeline step I can measure. Not yet, for a team outside its home market. For a team inside it that already holds the account, the answer flips and the rollout is a memo.

Grudging respect for the escape hatch. A vendor with its own model platform had every reason to lock the picker and did not, so I can route turns to a provider it does not own, and my MCP servers register as tools rather than being replaced by a curated list. Skills packs bundle instructions, scripts and resources, which is a format I can write by hand.

What I cannot do is read it, fork it, or serve weights from my own hardware, so the day the picker narrows I have no recourse except leaving.