Ubiquitous Language: Project Thesaurus Manager

You enforce naming consistency across the codebase by maintaining a living thesaurus of domain terms and consulting it every time something needs a name.

Four modes:

• Naming consultation (frequent) — everything in this file
• Thesaurus generation (rare) — read references/generating-thesaurus.md
• Naming audit (periodic) — read references/naming-audit.md
• History mining (optional) — read references/git-history-mining.md

when ## Unresolved items need evidence: which name came first, which replaced which, which is dying. Offered after generation, or on demand against an existing thesaurus.

Foundations

This skill combines two bodies of knowledge:

• Domain-Driven Design (DDD) by Eric Evans — ubiquitous language, bounded contexts, aggregate naming
• [First Principles Framework (FPF)](https://github.com/ailev/FPF/blob/main/FPF-Spec.md) — a transdisciplinary "operating system for thought" that provides formal tools for semantic precision: bounded contexts as declared semantic frames, polysemy unpacking, lexical firewalls, cross-context bridges with loss notes, term continuity relations, and anti-explosion naming control. References like "FPF F.5" or "FPF A.1.1" point to specific sections of the FPF specification.

Core Principle

"A project should use a single, shared vocabulary. Every name in code, docs, APIs, and conversations must map to a term in the thesaurus. If a concept isn't in the thesaurus — add it before naming anything."

>

— Domain-Driven Design, Eric Evans

The codebase is primary evidence, not automatic authority. Use code to discover which terms are currently in circulation. Use the thesaurus and user input to decide which terms SHOULD be canonical.

• For what exists today — derive from code (classes, DB schemas, API routes, events)
• For what should become the standard — ask the user/domain expert
• If code contradicts the approved thesaurus — the thesaurus wins for new code
• If the user says "fix legacy naming" — the user's directive overrides the codebase;

map existing code names to ## Legacy lines and use the user's terms as canonical

Tacit knowledge: For areas not yet implemented, the most important domain knowledge exists only in experts' heads, not in any artifact.

Thesaurus File

Locating the thesaurus:

1. If the user specified a path — use it
2. If THESAURUS.md already exists somewhere in the repo — use that location
3. Default: docs/THESAURUS.md

Single source of truth for domain vocabulary.

Versioning

The thesaurus declares its format version and the skill that maintains it in YAML frontmatter — machine-readable, outside the body, still one grep away:

---
thesaurus-format: "2.0"
skill: ubiquitous-language
---

# Project Thesaurus

Quote the version — unquoted 2.10 is the YAML float 2.1.

• rg -n '^thesaurus-format:' THESAURUS.md → the version in one hit. No key = 1.0

(the pre-index prose layout shipped before skill 2.0) — unless the file already has `- **Term** Id kind: ` Index lines: that is an unstamped 2.0 file from plugin 9.2.0; just add the stamp.

• Format major = skill major (metadata.version in this file's frontmatter). Skill

minor/patch releases never change the format.

• Read any format ≤ your own; write only the current one. A major gap means

migration first (see generating-thesaurus.md); a minor gap means new optional tokens — older readers keep working, you may add the tokens as you touch lines.

• A thesaurus with a format newer than yours: read it, don't rewrite it — tell the user

to update the skill.

• Only the format is versioned in the file. The skill's own version is not recorded there —

it would go stale on every edit and git log already answers "who wrote this". skill: is a pointer, so an agent without the skill knows what to install.

Layout: grep-first

The file is designed so that one `rg`/`grep` for any name answers "what do I do with this name?" without reading the surrounding text. Five sections, fixed order:

Registry invariant: every name known to the project appears in exactly one registry line — an Index line (as Term, Identifier, or avoid), a Forbidden line, a Legacy line, or an Unresolved header. Each kind of registry line has its own shape, so the shape of the hit tells you its status and the line itself tells you the canonical name. No -B/-A context needed.

Registry lines are bullet lines with labelled tokens, not Markdown tables: tables need | (an alternation in rg), match by column position, and get re-padded by formatters. Tokens (kind:, ctx:, avoid:, use:, in:, →) are position-free, formatter-proof, and need no escaping.

# Project Thesaurus

## Index

- **Order** `Order` kind:aggregate avoid: `Purchase`, `Transaction`, `Buy`
- **Order Line Item** `OrderLineItem` kind:entity avoid: `LineItem`, `OrderItem`, `Item`
- **Order Placed** `OrderPlaced` kind:event avoid: `OrderCreated`, `NewOrder`

## Terms

### Order
- **Definition**: A customer's confirmed request to buy one or more products at agreed prices.
- **NOT**: A payment (that's `Payment`), a shipment, or a draft cart (that's `Cart`).
- **Related**: Order Line Item, Order Placed, Cart

## Forbidden

- `Manager` use: `OrderFulfillment` — hides responsibility; name the activity

## Legacy

- `UserManager` → `Customer` + `CustomerRegistration` in: `src/legacy/` — split in v3

## Unresolved

### Account — one word, two concepts (billing vs auth)
- **Found in**: `billing/Account.ts` (balance), `auth/Account.ts` (login)
- **Question**: Two bounded contexts, or one of them a naming mistake?
- **Impact**: 18 files
- **Options**: `BillingAccount` + `UserAccount`; or contexts Billing / Identity

Index line

- **<Term>** `<Identifier>` kind:<kind> [ctx:<Context>] [avoid: `<name>`, `<name>`]

• One line per concept. All names for that concept live on that line — this is what

makes reverse lookup (rg Purchase → "use Order") a single hit.

• Backticks around every identifier-like name. ` rg 'Order' ` is an exact match;

rg Order would also hit OrderLineItem and Reorder. rg -F '**Order**' is the exact Term.

• Sorted alphabetically by Term. Entries under ## Terms follow the same order.
• avoid lists live only here. Entries do not repeat them — one place, no drift.

Forbidden and Legacy lines

- `<Word>` use: `<Identifier>`[, `<Identifier>`] — <why>
- `<OldName>` → `<Identifier>`[ + `<Identifier>`] in: <files/modules> — <note>
- `<OldName>` → — (see `<X>`, `<Y>`) in: <files> — retired

use: always points at an Index Identifier. → is the legacy marker: A + B means the old name was split; → — means retired with no single successor.

Entry

### [Term]
- **Definition**: What this concept means in the business domain — one sentence
- **NOT**: What this term does NOT mean; name the neighbouring term it is confused with
- **Related**: Other Index Terms this connects to, written exactly as in the Term field

Optional lines when they carry real information: **Broader**, **Narrower**, **Part of**, **Has parts**, **Example**. Write them on one side only — rg gives the inverse for free, mirrored copies only drift. Minimal viable entry is one line: - **Definition**: … — the Index line already holds the rest.

Anchor grammar (so lookups are one regex): the header is exactly ### <Term> or ### <Term> (<Context>) — nothing else. Tags, status, and context prefixes belong in the Index line, not in the header. Find an entry with rg -n '^### Order( \(|$)' — \b alone is not enough, it would also match ### Order Line Item.

The thesaurus captures concepts, not behavior. It's strong at nouns (entity names, roles, process names) but won't replace behavioral specs for business rules. Don't try to turn the thesaurus into a specification — keep entries short. If a concept has a critical invariant, note it briefly in the definition, not as a separate section.

Non-English domains: If the business domain operates in a non-English language, the Term uses the original language — the thesaurus should reflect how domain experts actually speak. The Identifier carries the code form: `- **Счёт-фактура** Invoice kind:entity`. This is exactly why both fields exist.

Lookup Protocol

Look up before inventing. This is the single most important step. Most naming tasks don't need a new term — the right name is already there.

1. Locate the thesaurus (see above). If absent, tell the user and offer generation.

Check rg -n '^thesaurus-format:' — no key or 1.x means the old layout: the protocol below still works by plain text search, but offer migration once, up front.

1. Read the Index if it has ≤ ~60 lines — it is the entire vocabulary at one line

per concept, cheaper than any search. For larger files, search instead.

1. Search every candidate name you are considering, plus whatever the surrounding

code already calls the thing. Use rg (the Grep tool) or grep -E — same patterns:

   rg -n -i 'invoice' docs/THESAURUS.md             # any role, any section
   rg -n '`OrderLineItem`' docs/THESAURUS.md        # exact identifier as seen in code
   rg -n -F '**Order**' docs/THESAURUS.md           # exact Term (not "Order Line Item")
   rg -n 'avoid:.*`Purchase`' docs/THESAURUS.md     # is this word a banned synonym?
   rg -n 'kind:event' docs/THESAURUS.md             # all terms of one kind
   rg -n 'ctx:Billing' docs/THESAURUS.md            # everything one context owns
   rg -n '`Basket` →' docs/THESAURUS.md             # legacy name and its replacement
   rg -n -A4 '^### Invoice( \(|$)' docs/THESAURUS.md  # the entry itself

The only trap: * and | are regex metacharacters — use -F for **Term**, and never search for table pipes (there are none).

1. Act on the shape of the line you hit:

1. Check the bounded context if Index lines carry ctx: — the same word may be

canonical in one context and banned in another.

ALWAYS run this before naming: classes, interfaces, types, enums, aggregates, entities, value objects, functions, methods, commands, queries, domain events, variables, constants, fields, parameters, DB tables/columns/collections, API endpoints and response fields, files, directories, modules, packages, feature flags, config keys, environment variables, and commit messages or PR titles that reference domain concepts.

If the Concept Is New

Before minting a new term, try four levers (from FPF F.14 "Name less, express more"):

1. Reuse — does an existing term already cover this? Maybe the concept is a variant, not a new thing
2. Compose — can you combine existing terms? OrderLineItem reuses Order + LineItem
3. Qualify — is this the same concept in a different state/window? Don't create NightOperator — use Operator with a time qualifier
4. Ask — if still unclear: "I need to name [concept]. The thesaurus doesn't have a term for this. What does the domain call it?"

If the user doesn't have an answer either — that's a white spot, not a dead end. Building a ubiquitous language is co-creation, not extraction. Add it to ## Unresolved with a [WHITE-SPOT] tag in the header. Don't force a name for an undefined concept.

Only after all four fail, mint a new term:

1. Name what the invariants make true (FPF F.5) — don't name aspirationally. If the code doesn't enforce "Premium", don't call it PremiumCustomer
2. Use minimal generality — choose the narrowest name whose rules you actually enforce. Don't upgrade Task to Activity to sound universal
3. Keep it to 1-3 words — no rhetorical adjectives ("robust", "optimal", "advanced")
4. Add it to the thesaurus: an Index line (Term, Identifier, kind:, avoid: —

omit avoid: when empty) and a ### Term entry with at least a Definition. Keep both sorted

1. Then use the term in code

If You Find an Inconsistency

When existing code uses a term that contradicts the thesaurus:

• Flag it: "Found fetchPurchases() but the Index line says Order, with Purchase under avoid:"
• Suggest a rename if scope is small
• For large-scale renames, note as tech debt and ask user how to proceed

If the Ambiguity Won't Resolve Itself

When a name lands in ## Unresolved — two spellings for one concept, one spelling for two concepts, no obvious winner — the current tree can't settle it, but the repository's history often can: which identifier was born first, which commit removed one while adding the other, which one is dying.

Offer it, don't run it silently. Whenever you present ## Unresolved items — after generation, after an audit, or when a naming question hits one — offer once:

"I can mine the git history for these — when each name was born, which replaced which, which is growing vs dying. Temporary index outside the repo, deleted afterwards. Want me to try that before you answer them by hand?"

Skip the offer if there is no .git, history is shorter than ~50 commits or squashed from an import, or the items are [WHITE-SPOT] tags (an unnamed concept leaves no trace).

If the user accepts, read references/git-history-mining.md and follow it. The short version:

S=<skill-dir>/scripts/git_term_index.py
python3 $S build --repo-dir . --content   # throwaway SQLite index in $TMPDIR, never in the repo
python3 $S query Account Customer         # birth, dormancy, trajectory, renames, messages
python3 $S pair User Customer             # competing names: birth order + swap commits
python3 $S contexts Account               # where the name lives — the polysemy check
python3 $S search 'rename account'        # BM25 search over commit messages
python3 $S clean                          # delete the index when done

The index covers the whole diff history, so "born" means born. Use --pathspec src/ on large repos — it cuts build time and sharpens the signal at once.

pair ends in a labelled verdict — RENAME (strong / probable / possible), DRIFT, NOT A RENAME, COEXISTENCE — with the direction inferred from evidence, not argument order. Report the label as given; do not upgrade it. "RENAME — strong" means commits exchange the names and a subject announces it; everything weaker needs git show first.

For polysemy (one word, two meanings, two modules) use the path split: pair prints files: A in N, B in M, both in K, and contexts <name> gives the directory breakdown for a single word. both in 0 — no file ever contained both — is real evidence for two bounded contexts; shared files mean synonym drift, which history cannot settle. Only claim a path split when the tool printed one. File-only renames (the file moved, the identifier did not) appear in query's file-renames section, not in pair — run both.

History is evidence, not authority. It ranks candidates and cites commits; the user decides. Report proposals in one batch with confidence levels, apply only what is approved, and record the commit behind each applied decision (— renamed in a41f2c9 on the Legacy line, or a - **History**: line on the entry).

Naming Rules by DDD Construct

The Index Kind column selects the rule.

Aggregates & Aggregate Roots (aggregate)

Use the business domain term. Singular. No technical suffixes.

GOOD: Order, Invoice, UserAccount, ShoppingCart
BAD:  OrderAggregate, OrderRoot, OrderAggregateImpl, OrderEntity

Entities (entity)

Singular noun from the domain. Something with identity.

GOOD: OrderLineItem, PaymentTransaction, Customer
BAD:  OrderLineItemEntity, OrderLineItemImpl, OrderLineItemObj

Value Objects (value)

Singular noun describing an immutable concept. Describes what it is, not what it does.

GOOD: Money, Email, PhoneNumber, Address, DateRange
BAD:  MoneyValue, EmailValidator, PriceInfo, AmountData

Domain Events (event)

Past tense verb + noun. Something that happened.

GOOD: OrderPlaced, PaymentCaptured, InvoiceSent, InventoryReserved
BAD:  OrderEvent, OnOrderPlaced, CreateOrder (that's a command)

Commands (command)

Imperative verb + noun. An action requested.

GOOD: CreateOrder, CancelInvoice, ProcessRefund, ReserveInventory
BAD:  OrderCreated (that's an event), NewOrder, OrderCommand

Queries (query)

Question or retrieval. Verb + object or descriptive name.

GOOD: GetOrderById, FindInvoicesByCustomer, ListPendingOrders
BAD:  RetrieveOrderData, OrderQuery, GetterForOrder

Domain Services (service)

Named after business activities the domain expert recognizes.

GOOD: InvoiceCalculator, OrderFulfillment, NotificationSender
BAD:  OrderManager, GenericService, HelperService

Repositories

Repository suffix is acceptable — it's an infrastructure pattern. Repositories are not thesaurus terms; they take the name of the aggregate they store.

GOOD: OrderRepository, InvoiceRepository, CustomerRepository
BAD:  OrderStorage, OrderPersistence, OrderFinder, OrderDao

Methods on Aggregates

Commands (change state): Imperative verb, no "Get" prefix.

GOOD: order.Cancel(), order.AddLineItem(product, quantity), order.Recalculate()
BAD:  order.CancelOrderMethod(), order.GetCancelled(), order.DoCancelOrder()

Queries (read-only): Start with Get, Is, Has, Can, or a domain verb.

GOOD: order.GetTotal(), order.IsExpired(), order.CanBeShipped()
BAD:  order.FetchInfo(), order.CheckData()

Naming Anti-Patterns to Detect and Flag

Lexical Firewall: the ## Forbidden section

The domain layer must be protected from transient jargon, vague terms, and implementation details. The thesaurus's ## Forbidden section lists words that MUST NOT appear in domain names and must always be replaced with a specific domain term.

Weasel Words (never use in domain layer)

Technical Jargon in Domain Layer

Domain code must be free of implementation details:

BAD:  MongoOrder, SqlUserRepository, HttpOrderService, OrderDto, OrderEntity
GOOD: Order, OrderRepository (interface), PaymentGateway, Order (just Order)

Technical prefixes/suffixes belong ONLY in the infrastructure layer, and even there the domain role should lead:

INFRASTRUCTURE LAYER (OK): MongoOrderRepository, RedisSessionCache, HttpPaymentClient
DOMAIN LAYER (NEVER):      MongoOrder, RedisSession, HttpPayment
NAME THE ROLE, NOT THE TECH: SessionStore not RedisCache, EventPublisher not KafkaProducer

Framework caveat: In frameworks that intentionally blend domain and persistence (Active Record pattern, ORM-centric frameworks), the model IS the domain entity. Keep the domain noun clean and let framework coupling live in inheritance, annotations, or metadata — not in the class name. Flag technical jargon only when it becomes part of the business-facing name or leaks outside its boundary.

Synonym Drift

Same concept called different things in different parts of code:

PROBLEM: "Customer" in auth, "User" in API, "Account" in billing — all mean the same thing
FIX:     Pick ONE canonical term per bounded context. Put the others in that line's `avoid:` list.

Abbreviation Boundary

Ban abbreviations in durable, domain-bearing names: types, exported functions, modules, API fields, DB columns, events, config keys.

Allow conventional short-lived local identifiers when meaning is obvious in scope: i, j, ctx, req, res, err, tx, db, e for events.

Allow industry-standard acronyms when they are the dominant term: SKU, VAT, URL, ID, OAuth. Do NOT force unnatural expansions if experts use the acronym.

PROBLEM: usr, user, account, acct — competing abbreviations for the same durable concept
FIX:     Pick ONE canonical form for domain-bearing names. Short-lived locals are exempt.

Translation Chain ("Telephone Game")

When different artifacts use different terms for the same concept across the knowledge chain, information is lost at each translation:

SMELL: Domain expert says "Campaign" → PM writes "Promotion" in spec →
       Dev codes `marketing_push` → QA tests "advertising effort"
FIX:   Same term everywhere: expert, PM, dev, QA all say and write "Campaign"

This is worse than synonym drift because each translation also loses nuance and business rules. How to detect: compare terms in requirements/specs/tickets against code names. If they don't match, the ubiquitous language has a translation gap — adopt the domain expert's term everywhere.

Casing Conventions

The Index Identifier is PascalCase. Every other form is derived from it mechanically using the project's conventions — never re-worded:

Key rules:

• Use the EXACT Identifier — don't abbreviate (ord), don't expand (orderObject), don't synonym (purchase)
• Compound names combine Identifiers: OrderLineItem, not PurchaseLineItem
• Technical suffixes for infrastructure roles are fine: OrderRepository, OrderDTO (in infra layer only)
• Multi-word Identifiers keep all words in every casing: ProcessingStage → processing_stage, never proc_stage
• Variables and parameters use full descriptive names: totalAmount not amt, customerEmail not cEmail

Updating the Thesaurus

When changing terms, use the least strong relation that tells the truth (from FPF F.13):

Key test: Can you point to the same concept before and after the change?

• Yes, same concept, better wording → Rename (keep as legacy alias for reading old code)
• No, the concept actually changed → Split or Merge (not a rename)

Alias parsimony: keep at most 1 legacy alias per term — the one readers will most likely encounter in old code. Registry invariant still holds after every edit: a name is under avoid: or in Legacy, never both.

Old layout? No thesaurus-format frontmatter key (or 1.x) means the pre-index prose layout — see "Migrating an Existing Thesaurus" in references/generating-thesaurus.md.

Quick Checklist Before Naming Anything

1. Did I grep the thesaurus for this name and its synonyms? Where did the hit land?
2. Would a domain expert recognize this name?
3. Does it contain a weasel word (Manager, Service, Handler, Info, Data, Item, Base, Util)?
4. Is it too generic (could mean multiple things in different contexts)?
5. Does it reveal infrastructure details (Mongo, Sql, Http, Dto, Entity, Model)?
6. Is it consistent with other uses of this term across the codebase?
7. Am I using the EXACT Identifier from the Index, or a synonym from its avoid: list?
8. Am I in the right bounded context for this term?
9. Does the name match its Kind — past tense for event, imperative for command?
10. Can I explain what this name represents in one sentence using domain language?
11. If the concept is new — did I add the Index line and the entry before using it?

If any answer raises a concern — stop and fix before proceeding.