English · Italiano
Vascend Official Plugins
Official marketplace of the Vascend
plugins for Claude Code: a method that makes an
AI agent think, execute and remember in a structured, traceable and
verifiable way, with zero dependencies.
Installation
/plugin marketplace add VascendTechnologies/Vascend-Official-Plugins
/plugin install vascend@vascend-official-plugins
After pushing to this repo, to update:
/plugin marketplace update vascend-official-plugins
/reload-plugins
Available plugins
| Plugin | Description | Install |
|---|
vascend | Danilov method for Claude Code: structured INDEX/DEFINITIONS/RELATIONS prompts and DanilovGoal execution tracked with one-hot bits, with a chained HMAC signed Trace and a deterministic verdict. | /plugin install vascend@vascend-official-plugins |
The idea: the memory palace
The method borrows a real technique, the memory palace (method of loci):
to remember something, you picture yourself walking through a familiar place and
"place" each item in a fixed room. It is the trick Patrick Jane uses in The
Mentalist, the mental castle where every piece of information has its exact
spot and you get back to it by walking there.
Patrick Jane in The Mentalist (CBS / Warner Bros.)
Vascend makes it literal for an AI agent:
- the plan is the castle;
- each task is a fixed room: bit K is always room K, it never moves;
- completing a task means turning on the light in that room (only via
mark.js);
- the goal is reached when the whole castle is lit (
state == MASK_TARGET),
and validate() returns TRUE.
And the castle does not stay behind the agent's eyes: it builds it by
answering. The chat is the castle and every line written is a room made
visible, so the reader walks through the reasoning instead of reading a summary
of it. state is the map of the lit rooms, missing the ones still dark: you
always know which one to go back to.
The three concepts
1. Structured prompts: INDEX / DEFINITIONS / RELATIONS
A new format, designed for machines before humans. Instead of describing
everything in prose, where "what it is", "what properties it has" and "how it
connects" get mixed together, it separates the three layers into a compact
numeric notation:
INDEX # the bare concepts, one word each, numbered
1 = node
2 = edge
DEFINITIONS # the instances with the properties that matter
@1[start]: color=#3B82F6, shape=stadium
@1[end]: color=#10B981, shape=circle
RELATIONS # the links: → directed, ↔ symmetric
@R1: @1[start] → @1[end] [ label=ok ]
OUTPUT: Mermaid diagram
Separating the layers triggers the model's structured parsing and yields output
that is more faithful where the values matter: diagrams, image generation
prompts, configurations, executable plans.
2. DanilovGoal: one-hot bit execution, verifiable
Each task is a bit. It turns on only by running mark.js, which appends
a line to the Trace, signed with a chained HMAC. The verdict
validate() == (state == MASK_TARGET) is computed by the script from the
signed data, never by the agent: it is math, not an assertion. Hand written
lines have no valid signature and are rejected, and the Trace stays tamper
evident.
3. Compact notation: how the agent answers and remembers
During a DanilovGoal the agent does not answer in prose. Everything it does
becomes one relation per line:
read LotTable.tsx>multi_selection | where to hook the props
edit LotTable.tsx>multi_selection | props + checkbox column
run pytest>backend_green | 6 passed
One line instead of a whole sentence. The format is
<action> <entity>>goal | note and uses about 2 syntax tokens per line, versus
about 5 for a notation loaded with @, :, unicode arrows and brackets. Over a
full turn the effect is large: far fewer tokens generated, and since the
agent already thinks in structured form instead of prose, the reasoning is
shorter and faster, both in tokens and in time.
The same lines are stored in .vascend files, a relation text format designed
to be reinjected into the context cheaply, grouped by @plan (the
session/goal):
