Claude Code plugin marketplace
Two skills for working in throughline IDD requirements repositories. Install once and they apply in your own repos. Your graph stays committed there, as it must be — throughline is git-native and the items are the product. The skill is what stays out: nothing is vendored into your repo, and nothing is installed on your laptop.
/plugin marketplace add https://skills.iddn.uk/marketplace.json
/plugin install tl@throughline-skills/plugin marketplace add rhodium-org/throughline-skills
/plugin install tl@throughline-skillsWorks in the terminal and at claude.ai/code.
In the browser you can also use Manage plugins → Marketplaces → Add
marketplace and paste either source. To pick up a new version:
/plugin marketplace update throughline-skills.
Initialises a graph by a structured discovery conversation instead of
the fixed register set tl init lays down: the sources, then
the domain, then what the work needs from its graph, and only then the
registers — each with a rationale, each rejection recorded. The reasoning
lands in idd/shape/; the graph for the work itself in
idd/<name>/, named during discovery. Everything it
writes is AI-origin and proposed; a human ratifies.
A challenge pass over a graph before a human is asked to ratify it: the
twelve questions tl check cannot ask, with a script for the
deterministic ones — link strength, prose that cites an item the links
never record, siblings naming one thing two ways, branches no machine
verifies, universal claims one child realises, and more.
Also re-checks an implementation graph against a specification it composes once that specification moved: stamped links, suspect-link triage, and spec items nothing implements.
Repositories holding several throughline graphs composed with
tl-compose: picking the right binary per graph, wiring sibling
graphs by path, gating them all, and ratifying items whose
grounding chain crosses graphs.
An objective assessment of one piece of work on a graph: items and acceptance, sources and citations, commits and editions, items amended after ratification, generated documents, tests, and from the session transcripts the wall-clock and active time, tokens and tool calls. A script measures; a person writes the narrative beside the numbers; a copy joins a collection whose index compares cases, for or against working this way.
Each skill activates on its own when Claude notices the situation it covers, or you invoke one by name.
This is designed for it. Everything runs in the Claude sandbox: the skill
installs the CLI it needs itself with pip install throughline-compose,
which needs Python 3.11 or later — cloud sessions have it. Graphs wired to
each other by path need no network at all. Graphs adopting an
external standard by url need the sandbox to reach that host, so
use https:// URLs and never an SSH host alias, which will not
resolve there.