Glossary

The words this site and the stack use in a sense of their own. Each has one definition here, and where the stack defines the word or uses it clearly, the place it does. Where the sources use a word in more than one sense, the entry says so.

Claude Fable 5.2 wrote this page and its data file. Its definitions were checked against their sources on . The maintainer has not yet reviewed it.

The ladder

coherence
Software that still reads as one design: one way to do each thing, one name for each concept, documents that are still true, and checks that mean what they say.
This site's coherence pages
failure
One way agent-written work goes wrong. A row of the coherence grid.
This site's coherence pages
hold
Whatever stops a failure or catches it. A cell of the coherence grid.
The repos tool also holds a push or a sync that it will not run yet, which is a different sense.
This site's coherence pages
ladder
The levels a hold can sit on, from a person's judgment down to the shape of the system, each depending less on anyone remembering than the one before.
The fuz-stack skill has ladders of its own, for styling and for dependency injection, which are unrelated to this one.
This site's coherence pages
rung
One level of the ladder. From the top: a person decides each time, a session carries out a procedure, a written rule is loaded, a tool gives a verdict, and the shape of the system rules the failure out.
This site's coherence pages
coverage
How much of the stack has a hold: throughout, or in places. It says nothing of how strong the hold is, which is its rung.
Test coverage and conformance coverage are the usual senses, and the pages use those too.
This site's coherence pages

Checks

gate
A check that gives a pass-or-fail verdict on a change.
On the values page a gate is a kind of value, one that has to hold before anything else is weighed. The skills also use the word for a check on a request at runtime.
This site's table of gates
probe
A string that a named file, check script or CI step of a repo must have, or must not have, for a row of a table on this site to still stand. A probe that passes says the text is still there, not that the check it names is sound.
This site's table of gates
local gate
The command a session runs in a repo before work counts as done, with its exit code as the pass condition. Often just "the gate".
The fuz-stack skill on the gate
release audit
A check that walks the dependency graph of every shipped binary and refuses a forbidden crate, a test-only feature, a broken layering rule, or a lint table that has drifted.
The skill names it by its command, check-release.
The spine reference on check-release
drift
Two things that are meant to match no longer matching: a generated file and its source, a column list and the live schema, two twins, a doc comment and the code.
The coherence pages also say that a design drifts, meaning it stops reading as one design.
This site's own usage.
fixture
A test case kept as files: an input, and an expected output that a task generates and nobody edits by hand.
A shared JSON table of cases that two implementations both read is called a case table.
The testing reference on fixture tests
pinned count
A committed number of cases a check is expected to grade or pass, so that the check fails when the number moves.
ledger
A committed list of known, accepted differences from an oracle, each recorded with a tracking id or a reason. A difference that is not on it fails, and so does an entry that no longer matches anything.
The sources also say catalog, recorded difference and declared difference for the same idea, and tsv has a ledger of comments that is something else.
The twin reference on enforcement
mutant
A small break made to the code on purpose, to check that some test fails because of it.
fuz_schema's mutants

Twins and comparisons

oracle
What a project's output is compared against and expected to agree with: an independent implementation, an outside suite, a twin, another build, a property the output must satisfy, or an earlier version.
This site's table of oracles in use
twin
One of two separately written implementations of the same thing, held to the same observable behavior by shared tests. Most often a TypeScript one and a Rust one.
reference implementation
An independent implementation outside the project that already has the right behavior, as Prettier is for the tsv formatter and Postgres is for pglet.
The twin reference also calls the TypeScript server of a pair the reference server, though it is inside the project. A skill's reference docs are a different sense of the word.
This site's own usage.
differential
A test that puts the same case to two implementations, normalizes the answers, and fails on a difference that is not declared.
The twin reference on enforcement
cross-backend suite
A test suite whose same test bodies run against each backend of a server over real HTTP, WebSocket and SSE.
spine
The shared backend layers an app server is built on: database, auth, HTTP, realtime and actions. They exist as fuz_app in TypeScript and as the spine crates of the fuz repo in Rust.
Some of the Rust docs mean only the Rust crates by it.

Working with agents

session
One run of a coding agent, which starts with no memory of earlier runs beyond the files it loads.
Session is also the name of a rung, and in the backend docs a login session.
The grimoire skill, on stateless agent sessions
agent harness
The program that runs a coding agent, carries out its tool calls and applies its permission settings.
A test harness or a measurement harness is a different thing: the code that runs a comparison.
This site's own usage.
skill
A folder of instructions in the Agent Skills format, a main file and optional reference docs, that an agent loads for a kind of task.
This site's skills
maintainer
The person who sets a project's scope and design, and to whom versions, publishing, deploys and new dependencies are reserved.
The sources also say owner, operator and the human for this person.
The fuz-stack skill's shared rules
kept action
An action the coherence pages list as kept with a person, such as publishing, deploying or changing a version, with whatever stops an agent from taking it.
This site's table of actions kept with a person
deny rule
A permission rule in the agent harness's settings that refuses a command whose text starts with a given prefix.
The repos tool's docs on what agents push through
dry run and wetrun
A dry run shows what a command would change and changes nothing, and wetrun is the flag that makes it do the work. zap, and a new command in the stack's Rust tools whose effect is hard to undo, are dry runs by default. Not every release or deploy command in the stack is.
zap's concepts

Planning

planning layer
Where plans and notes that span sessions are kept, outside the code repos. The grimoire skill describes one form of it, a separate repo of Markdown.
The fuz-stack skill's shared rules
membrane
The one-way rule that a planning layer may read every repo while nothing public refers to its contents.
The grimoire skill's key concepts
lore
A planning layer's notes on one subject, usually a repo: its role among the others, decisions and their reasons, and what is next. It holds what the repo does not say and does not restate what the repo does.
The grimoire skill on lore
quest
A goal that spans two or more repos or has several phases, kept as one file with its tasks and deleted when the goal is met. Work within one repo stays in that repo's lore.
The grimoire skill on quests
scry
A check, run as code with no model in it, that blocks when it fails: of a planning layer's claims against the repos, or of a policy across repos.
The skill names the practice scrying.
The grimoire skill's supporting layers
survey
A read-only observation across repos, run as code with no model in it and written out as findings to triage. It never blocks, and each run replaces the last.
The skill names the practice surveying.
The grimoire skill's supporting layers
arc
A long-running implementation effort across many sessions, kept to one checkout with at most one live session.
The grimoire skill on arcs
writer and reviewer loop
Within an arc, one writer edits and a read-only reviewer reports findings, in turns, until a round of review leaves nothing worth fixing.
The skill calls it the review loop.
The grimoire skill on arcs