TypeGlish 0.7.2: the checker takes your word for it

TL;DR 0.7.2 ships $DEFINE word, a command that teaches the compiler one of your nouns so the wallboard stops asking to be introduced, with five new blocking codes so a malformed declaration can never quietly do nothing. Then it takes findings away: the cross-file pass can now retract a local warning the composition disproves and prints a receipt for it, the importer repairs a no-space heading typo that was turning into 6 blocking errors, and two advisories that were simply wrong are gone. Monaco embedders have one thing to do: the editor surface is now the host's by default, so pass { chrome: true } if you want the brand background back. Upgrade with npm i -D typeglish@latest.

The interesting releases are usually the ones that add something. This one is mostly the opposite, and it is the better kind of work. Four of the five changes below delete a finding that should never have fired: a warning about a variable that a sibling file was using all along, an advisory that swallowed half a sentence and offered to name it, two nags at a list that was already well formed, and a cascade of six errors that traced back to a single missing space. The fifth, $DEFINE word, is new language surface, and it exists for the same reason: your product nouns are real, and the compiler kept treating them as strangers.

§1Teach it a noun it has never met

TypeGlish reads the as a lookup instruction. That is the whole basis of the Self-Containedness Law we pulled apart in the policy-doc teardown: when you write the wallboard, you are telling the model to retrieve a specific wallboard, and if the document never introduced one, the model resolves the pointer from its own priors instead of from your prompt. Hence prompt/unintroduced-definite, which fires on every definite the file has not established.

Correct, and for a genuinely ambiguous noun, useful. But some nouns are not ambiguous at all: they are your product. There is exactly one wallboard, everyone on the floor knows what it is, and no amount of "You manage a wallboard." makes the prompt better. Until now the only honest answers were to write a throwaway introduction line or live with the finding forever. 0.7.2 adds a third: declare it.

<$CONFIG>
  $DEFINE word wallboard
    - class:: noun
    - unique:: true
</$CONFIG>

# Role

You are a support agent.

# Instructions

Log every resolved ticket in the wallboard.

Run the identical file on 0.7.1 and you can see how new the surface is. The parser had never heard of the command, so the whole block fell through to the generic rule that content does not belong inside <$CONFIG>:

wallboard.tg:2:3  error  structure/config-section-content  Content inside <$CONFIG> never reaches the model — only $-commands ($IMPORT, $CONFIG, $TOOL, $DEFINE, $TEST) belong here. Move prompt content outside the section.
wallboard.tg:3:5  warn   structure/missing-period  Unterminated statement — end it with a period (or ! ?; a lead-in may end with ":"). Statement boundaries are a compile contract.
wallboard.tg:13:1  info   prompt/unintroduced-definite  "the wallboard" retrieves something this document never introduces - a model must guess which wallboard is meant.

✗ 1 file — 1 error, 4 warning, 1 info

The reason a new command is worth five new diagnostic codes is the failure mode it would otherwise have. A config block that accepts unknown keys is a config block that lies to you: you write a row, the compiler ignores it, and nothing anywhere says so. The params here are a closed schema of exactly two, class and unique, and stepping outside it is a blocking error that states the consequence in the message.

<$CONFIG>
  $DEFINE word wallboard
    - class:: noun
    - plural:: wallboards
</$CONFIG>

# Role

You are a support agent.

# Instructions

Log every resolved ticket in the wallboard.

The other four are structure/define-word-class (a class outside noun, verb, adjective), structure/define-word-unique (unique on a verb, which the message calls a noun feature), structure/define-word-form (a lemma that is not a single word), and structure/duplicate-define-word (the same lemma declared twice). All five block compilation, and all five are in typeglish reference and typeglish --explain as of this version. One collision worth knowing: $DEFINE word AS @{tier} is equal to gold is still a named state called word. The AS decides which command you wrote.

§2The composition can now clear a finding

0.7.0 made a directory of prompts one program: files joined by $IMPORT compose before the proofs run, so a rule here and a rule in a file you transclude are checked against each other. That pass only ever added findings. 0.7.2 lets it subtract one, which turns out to matter for the most ordinary cross-file arrangement there is.

Here is the shape. A root file declares an input and hands the actual decision to a shared module:

<$CONFIG>
  $REQUIRE variable minute_parity: one of even, odd
  $IMPORT file "identity.tg" as identity
</$CONFIG>

# Instructions

@<identity.identity>

Answer the customer question in two sentences.

Check copilot.tg by itself and the single-file pass says something reasonable and wrong: you declared a variable and never used it. It cannot see the sibling. Before 0.7.2 that warning survived the program pass too, so the only way to clear it was to delete a declaration the composed prompt genuinely needs.

prompts/copilot.tg:2:21  warn   clarity/unused-variable  $REQUIRE variable "minute_parity" is never used — no @{minute_parity} reference fills it. Remove it, or reference it in the prompt.
prompts/identity.tg:6:3  info   prompt/implied-doer  Nothing in this document says who follows its instructions - every imperative implies a doer the model must infer.

✓ 2 files — 0 error, 1 warning, 1 info
program: 1 root — prompts/copilot.tg (2 files)

0.7.2 composes the two files, finds that the composed document does not reproduce the warning, and withdraws it. What it does not do is let it disappear quietly:

prompts/identity.tg:6:3  info   prompt/implied-doer  Nothing in this document says who follows its instructions - every imperative implies a doer the model must infer.

✓ 2 files — 0 error, 0 warning, 1 info
program: 1 root — prompts/copilot.tg (2 files)
program: retracted 1 finding the composition disproves — prompts/copilot.tg clarity/unused-variable (line 2)

The conservatism here is worth reading, because it is what keeps the feature from becoming a hole. A finding is retracted only if two things hold. First, its code has a row in a declared registry of retractable codes, each with a written reason, and there is no catch-all: an unregistered code never retracts however thoroughly a composition fails to reproduce it. Today that registry has exactly one row, clarity/unused-variable. Second, the finding's own line has to have been on trial, meaning it was actually present in the composed document, because an imported module's manifest rows are stripped at the splice and the composition never judged them. A file that belongs to no multi-file closure gets zero retraction, so if you check single files nothing about your output changed.

§3One missing space, six blocking errors

This one is a good argument for reading your worklist skeptically. Take a plain markdown prompt of the kind typeglish import exists for, and give it the typo everybody has typed: #Instructions, no space after the hash. CommonMark says that is not a heading, and the checker agrees, which means the document has no section structure at all and every numbered step in it lands in one scope.

#Instructions

1. Greet the customer by name.
2. Ask for the order number.
3. Look up the order.
4. Confirm the shipping address.

#Constraints

1. Never promise a refund date.
2. Never share internal notes.

On 0.7.1 that imports into a wall of findings, none of which mention the cause:

  6 errors · 1 warn · 3 info (was 6 · 3 · 3)

the worklist:
  grammar
    L3 error structure/duplicate-step: Duplicate step "1" — 2 steps carry this number, so a reference to it is ambiguous. Renumber one of them. (conflicts with line 7)
    L10 error structure/duplicate-step: Duplicate step "1" — 2 steps carry this number. (conflicts with line 2)
    L10 warn structure/unordered-step: Step 1 appears after step 4 — steps are out of order.
    L1 error structure/dangling-step-ref: Reference to step instructions, which is not defined.
    L8 error structure/dangling-step-ref: Reference to step constraints, which is not defined.

✗ 6 errors block build — causes: structure/duplicate-step x4, structure/dangling-step-ref x2

0.7.2 repairs the typo at the door, and reports the repair rather than performing it silently:

  2 headings normalized (#Instructions -> # Instructions) - a no-space hash is a heading typo, not prose
  0 errors · 0 warns · 3 info (was 6 · 3 · 3 — 6 errors resolved by the repairs and lifts above)

the worklist:
  optimization
    L3 info prompt/implied-doer: Nothing in this document says who follows its instructions…
    L4 info prompt/unintroduced-definite: "the order number" retrieves something this document never introduces…
    L6 info prompt/unintroduced-definite: "the shipping address" retrieves something this document never introduces…

✓ builds clean — next: typeglish build support.tg

§4Two advisories that were simply wrong

Both of these are false positives, and both were the kind that erode trust fastest, because the advice attached to them was visibly nonsense.

The first is a boundary problem. When a sentence opens with the and runs several words before its verb, the compiler cannot prove where the name ends, so it declines to bind and offers you the explicit %name literal% form. Fine in principle. In practice the match ran greedily to the last declaration verb on the line and could swallow an entire earlier clause:

# Role

You are a support agent.

# Context

The user no longer wants to speak or has asked twice.

0.7.2 stops the walk at the first finite verb, on the reasoning that a name never swallows a verb, and the file checks at 0 error, 0 warning, 0 info. The advisory still fires on spans the compiler genuinely cannot bound, which is what it was for.

The second is a pattern the checker did not recognize. English distributes one conditional across a lead and a list all the time: the action lives in the lead line, the conditions are the members. Every member looked like a guard with no THEN:

# Role

You are a support agent.

# Instructions

Call the escalation playbook if:
1. When a customer asks for a manager more than once.
2. When a refund exceeds your approval cap.

Worth saying what did not happen: this is a recognition fix, not a suppression. The playbook.tg above still reports prompt/unintroduced-definite on the escalation playbook, because nothing in the file introduced one. Two warnings went away and the finding that was right stayed. If you want that last one gone too, you now know the command: it is the one in §1.

§5The editor surface belongs to the host

The single entry under Features in this changelog is for people embedding typeglish/monaco, and it is the one change here that can alter what you see. setupTypeGlish takes an options object now:

setupTypeGlish(monaco);                        // tokens only; your IDE keeps its chrome
setupTypeGlish(monaco, { chrome: true });      // the brand surface, both themes
setupTypeGlish(monaco, { chrome: myColors });  // your own Monaco colors map

The reasoning is a nice statement of what a language package is for: its job is to say what a token means and paint it accordingly, and deciding that every editor embedding a .tg file must sit on a particular background is the host's call, not the language's. The token colors are untouched, so the hardness and directness weights that the language server renders still read the same. If your app wanted the brand surface, it is one argument. If you never noticed the background was not yours, this is the release where it becomes yours.

npm i -D typeglish@latest              # 0.7.2
npx typeglish@latest check prompts/    # or no install at all