Split out of #13831 so the question survives PR #13931's landing. #13931 delivers only the route-independent half — the drift pin #13831 says is owed either way, plus a docblock recording the measured decomposition. It changes the constant's contents not at all. The route itself is this card.
The fork changed shape, so the original framing no longer applies
#13831 posed it as complete registry vs. deliberately advertised subset. Measurement (full detail in #13931) makes the complete-registry reading unreachable in the shape the constant has:
72 registered names
= 39 callable BARE, as fn(x)
+ 33 callable ONLY on a receiver, as x.fn()
39 bare-callable
= 27 added by registerStdLib -> all 27 advertised
+ 8 cel-js built-ins -> advertised
+ 4 cel-js built-ins -> WITHHELD: bytes, dyn, type, uint
Two findings reshape the decision:
- The card's sub-premise was falsified. The 27
registerFunction sites yield exactly 27 names; 50 of the 72 come from cel-js's own global registrations. "Derive the constant from the registration site" is impossible as stated. - Every consumer spends an entry as a bare call, so a flat list of all 72 is not merely unreachable but wrong:
objectui's Studio predicate editor (CelPredicateField.tsx) inserts a chosen suggestion verbatim as `label(`, and the pre-existing runtime pin probes each entry with a bare-call expression. Advertising split or map would autocomplete them into an author's predicate, which faults no matching overload.
⇒ The catalog's real membership rule — never previously written down — is bare-callability. That makes the 37-name gap deliberate rather than stale, and moves the live fork onto the 39 bare-callables, not the 72.
Worth noting for grading: #13831's own example gap list is mostly a category error. Of type map filter exists split substring getFullYear base64 json orValue, nine are receiver-only; only type is bare-callable.
The options
A. Keep 35, and keep it explicitly a subset. What #13931's docblock and pin already assume. The four withheld names are CEL type primitives with no measured demand.
B. Widen to the 39 bare-callables. A four-name addition (bytes, dyn, type, uint), mechanically derivable from Environment.getDefinitions(). Cheap — but it hands an AI author dyn(x) and uint(x).
C. Rename the symbol so its name stops promising completeness. ⛔ A published-surface break on an exported symbol; deliberately not improvised.
Recommendation: A
The implementing seat recommends A and the reasoning holds up:
- Demand. The only measured consumers are objectui's autocomplete and the MCP
validate_expression tool, both served correctly today. Nothing asks for bytes/dyn/type/uint. - The defect was an unstated contract, not a wrong number. A stated rule plus an executable pin closes the class; widening closes one instance.
- Asymmetric cost for AI authoring. Under-advertising costs expressiveness; over-advertising emits faulting code.
dyn and uint are precisely the primitives that let a model produce something which type-checks and is unusable. - Scope discipline. B is capability expansion with zero measured pull.
C is a maintainer call regardless of A/B.
The second half of the same decision
skills/objectstack-formula/SKILL.md currently says the constant "is the canonical list" and that "Only the functions above are callable." The second claim is measurably false — type(x) resolves and evaluates, as do 33 receiver methods. It is false in the safe direction (it under-promises), so nothing is broken by it today.
It is deliberately not corrected in #13931, for two reasons: the correct wording depends on which route is picked, and skills/** is a governed surface. Recommendation is to fold it into whichever route is chosen rather than file it separately — a separate card could not be graded before this one is.
Refs: #13831 (the finding this splits from) · #13931 (the route-independent pin) · #13594 (the ruling pinning the lint gate's oracle to the engine's own check(), unaffected by any route here)
Split out of #13831 so the question survives PR #13931's landing. #13931 delivers only the route-independent half — the drift pin #13831 says is owed either way, plus a docblock recording the measured decomposition. It changes the constant's contents not at all. The route itself is this card.
The fork changed shape, so the original framing no longer applies
#13831 posed it as complete registry vs. deliberately advertised subset. Measurement (full detail in #13931) makes the complete-registry reading unreachable in the shape the constant has:
Two findings reshape the decision:
registerFunctionsites yield exactly 27 names; 50 of the 72 come from cel-js's own global registrations. "Derive the constant from the registration site" is impossible as stated.objectui's Studio predicate editor (CelPredicateField.tsx) inserts a chosen suggestion verbatim as`label(`, and the pre-existing runtime pin probes each entry with a bare-call expression. Advertisingsplitormapwould autocomplete them into an author's predicate, which faultsno matching overload.⇒ The catalog's real membership rule — never previously written down — is bare-callability. That makes the 37-name gap deliberate rather than stale, and moves the live fork onto the 39 bare-callables, not the 72.
Worth noting for grading: #13831's own example gap list is mostly a category error. Of
type map filter exists split substring getFullYear base64 json orValue, nine are receiver-only; onlytypeis bare-callable.The options
A. Keep 35, and keep it explicitly a subset. What #13931's docblock and pin already assume. The four withheld names are CEL type primitives with no measured demand.
B. Widen to the 39 bare-callables. A four-name addition (
bytes,dyn,type,uint), mechanically derivable fromEnvironment.getDefinitions(). Cheap — but it hands an AI authordyn(x)anduint(x).C. Rename the symbol so its name stops promising completeness. ⛔ A published-surface break on an exported symbol; deliberately not improvised.
Recommendation: A
The implementing seat recommends A and the reasoning holds up:
validate_expressiontool, both served correctly today. Nothing asks forbytes/dyn/type/uint.dynanduintare precisely the primitives that let a model produce something which type-checks and is unusable.C is a maintainer call regardless of A/B.
The second half of the same decision
skills/objectstack-formula/SKILL.mdcurrently says the constant "is the canonical list" and that "Only the functions above are callable." The second claim is measurably false —type(x)resolves and evaluates, as do 33 receiver methods. It is false in the safe direction (it under-promises), so nothing is broken by it today.It is deliberately not corrected in #13931, for two reasons: the correct wording depends on which route is picked, and
skills/**is a governed surface. Recommendation is to fold it into whichever route is chosen rather than file it separately — a separate card could not be graded before this one is.Refs: #13831 (the finding this splits from) · #13931 (the route-independent pin) · #13594 (the ruling pinning the lint gate's oracle to the engine's own
check(), unaffected by any route here)