The arc closes. Trazum had only ever read prompts somebody else wrote; it writes one now, by asking. Six chapters, all six delivered — the questions, the assembly, the three measured claims, the terminal, the web, and MCP.
What it will not say is the part worth reading. Not that the prompt is good:
that is a judgement about text nobody has run. It says the checklist is complete
with its gaps named, what the prompt costs with the estimate marked as one, and
that trazum optimize recovers nothing from it — which is the only one of
the three this product could ever have staked its name on, because the tool
grading the output is the same tool that would have to find the fault.
One thing the plan named and this release did not build: the optional model-assisted polish. It needs a credential this repository does not have, and inventing what a model would have said is the estimating-and-measuring merge that 1.36–1.40 spent five releases removing. Named, not faked, and still open.
Added#
Chapter six of 1.61: prompt_writer over MCP. An agent asked to write a
prompt for something has the same problem a person has — it does not know what
it has not been told — so this hands it the questions rather than the answers.
Call it with what you know; get back the next question worth asking and whatever
can be assembled so far.
next carries the wording, not just the id. An agent handed a bare id would
either invent the question or skip it, and a question nobody asks is a slot
nobody fills. The two slots that fill no section say so in their own wording —
"this changes the estimate, never the prompt" — because an agent told only
"which model is this for?" would reasonably expect the answer in the text. Any
slot with a null section whose question does not say it fails.
Stateless, and asserted as such: the same answers produce a byte-identical reply, and an earlier call leaves no trace in a later one.
The __proto__ test parses the wire form. In a JS object literal that name
sets the prototype rather than a key, so a test written that way proves nothing;
over MCP the arguments arrive as JSON, where it is an ordinary own property. The
tool writes only catalogue ids into a map with no prototype — the shape CodeQL
named on /api/write, fixed here before it could be written twice.
And a count stopped being a check. rpc.test.js asserted tools.length === 5,
a literal that had to be bumped every time a tool arrived; a number somebody
edits to make a suite pass has stopped asserting anything. It compares against
TOOLS now.
Chapter five of 1.61: the interview on the web. A Write tab and
POST /api/write, the same fourteen questions and the same
prompt-draft document as the terminal.
The route is stateless on purpose. The browser holds the answers and sends all of them every time. A session would mean that endpoint knowing what somebody is halfway through writing, which is the thing the rest of this product refuses to hold — and nothing there calls a model either, so the surface that has a network by definition is held to the same rule as the one that does not.
The browser asks the server what to ask next rather than deriving it. next
and missing look alike and mean different things: missing holds only the
required slots, and the interview carries on through the optional ones.
Deriving one from the other is how a form starts skipping questions, and the
test asserts the two disagree on a real answer set.
Skipping is offered as an answer, not left as an empty box. A question a reader cannot decline is a question they will answer badly to get past — and a bad answer goes into the prompt, where a decline leaves nothing.
An unpriced model comes back with a verdict, not a 400. It is one of the three answers this format gives; refusing it would turn a measurement into a failure the caller has to handle. An unknown slot is refused, because ignoring one would hand back a draft with no way to tell it from an answer the assembly had no use for.
The sentence this mode refuses to say is rendered, and guarded. Copy that exists in a catalogue and appears nowhere is a promise nobody reads.
Fixed#
CodeQL called the route's answer loop a remote property injection, and it was
right about the shape. The first version iterated the request's own keys, so
the property being assigned was a string the caller chose — __proto__ among
them. The slot check would have refused that one, but a write whose target
is caller-controlled is only ever as safe as the guard immediately above it.
The loop runs over the catalogue now, so the writable keys are a fixed literal
set, and the map is created with no prototype to reach in the first place. Sent
__proto__, constructor, prototype and toString anyway, and asserted
Object.prototype is untouched.
Two things an existing guard and a second reading caught in this chapter's own
work. The Write tab shipped without forceMount on the reasoning that
mounting it would cost one request per page load from readers who never open it
— which traded somebody's half-answered interview for a request cheaper than the
page's own assets. The answers live only in the browser, and Radix unmounts
an inactive tab by default. The guard names the property rather than the tabs
and was right. And the mode was put first in the rail while defaultValue still
opened Optimise, so the reader would land with the second row highlighted; the
rail's order and the default tab are one decision, not two.