[isc-klu]

This PR integrates ascot 0.4.2, the latest release as of 2026-10-29. ascot aims to find errors the IRIS compiler can't, and to give more detail on the errors the compiler already finds.

My test set has 579 programs that IRIS rejects, at compile time or at runtime, and 543 that run without error:

• The compiler and ascot both report an error in 436 of the 579.
• ascot also reports an error in the other 143, which IRIS only rejects at runtime.
• ascot reports no error on any of the 543 that run without error.

A preliminary agent study used 47 code-generation tasks written by an independent author. With ascot, Haiku 4.5 used 33% fewer tool calls. Opus 4.8 saw no change, since it already needs about one call per task on average.

How much ascot helps in iris-agentic-dev's own setting is still unknown. Along with the integration, this PR includes benchmark results on iris-agentic-dev's tests/e2e/tasks/benchmark/corpus task set.

[isc-klu]

I ran iris_check (ascot 0.4.2) on the erroneous and clean version of each of the 42 tasks in tests/e2e/tasks/benchmark/corpus, against IRIS Community 2026.2. No model was involved. The erroneous version is the fixture the agent is given; the clean version is the task's reference solution.

Summary numbers

On the 42 erroneous programs:

• ascot reports errors on none, and warnings on 13, of which 2 are relevant to the bug. See the discussion after the table about the irrelevant warning.
• The compiler reports errors on none.

On the 42 clean programs:

• ascot reports errors on none, and warnings on 10.
• The compiler reports errors on none.

In short:

• Recall: 2 of 42. ascot reports a warning relevant to the bug on CORPUS-20 and CORPUS-37.
• Precision: 2 of 13. ascot flags 13 erroneous programs, and on 2 of them a warning points at the bug.
• Latency per file: median 876 ms cold (a fresh process per file) and 406 ms warm (one running server), from a debug build of iad. Both include the IRIS lookups ascot makes for the file.

Per-task results

Every diagnostic ascot reported is a warning. Line numbers are 1-based within each class. The other 28 tasks got no diagnostics in either version.

Task	Erroneous version	Clean version	Relevant to the bug
CORPUS-03	L7, L8, L9: This call returns %Status, which Do discards -- consider Set to capture it, or checking it directly	Same	No
CORPUS-05	L7, L8: This call returns %Status, which Do discards -- consider Set to capture it, or checking it directly L9: Too few arguments: 'sc' is required. Only arguments with default values and Output arguments are considered optional. L9: The parameter 'len' is declared as a by-ref parameter but got a normal argument	Same, except the 'sc' warning is at L11 and the 'len' warning is gone	No
CORPUS-09	L9: Too few arguments: 'deletecount' is required. Only arguments with default values and Output arguments are considered optional.	Same	No
CORPUS-10	None	L8: Variable 'pInsert' is assigned but never read	No
CORPUS-17	L9: Too few arguments: 'deletecount' is required. Only arguments with default values and Output arguments are considered optional. L21: Too few arguments: 'pSelectMode' is required. Only arguments with default values and Output arguments are considered optional.	Same	No
CORPUS-18	L8: Too few arguments: 'deletecount' is required. Only arguments with default values and Output arguments are considered optional.	Same	No
CORPUS-20	L14: Variable 'tSC' is assigned but never read	None	Yes
CORPUS-21	L14: This call returns %Status, which Do discards -- consider Set to capture it, or checking it directly	Same	No
CORPUS-23	L4: Variable 'pN' is assigned but never read	None	No
CORPUS-26	L9, L28: Too few arguments: 'deletecount' is required. Only arguments with default values and Output arguments are considered optional.	Same, the second one at L29	No
CORPUS-28	L8: Too few arguments: 'deletecount' is required. Only arguments with default values and Output arguments are considered optional.	Same	No
CORPUS-31	L4: Variable 'pClass' is assigned but never read	None	No
CORPUS-37	L6: ObjectScript has no operator precedence and evaluates strictly left-to-right; this may not compute what you expect	None	Yes
CORPUS-40	L6: Variable 'pInsert' is assigned but never read	Same	No

Warnings not relevant to the bug

They fall into three groups.

"Do discards %Status" (CORPUS-03, 05, 21). These are correct. SetAt and Rewind return a status that the code throws away. That is a bad pattern worth flagging, even though it isn't the bug here. CORPUS-20's bug is the same mistake: a returned status that is never checked.

"Too few arguments" (CORPUS-05, 09, 17, 18, 26, 28). These are false positives caused by imprecise library method signatures:

• An argument is used as Output but declared as ByRef:
  • deletecount of %Library.Persistent.%DeleteExtent, in CORPUS-09, 17, 18, 26 and 28
  • sc of %Stream.TmpCharacter.Read, in CORPUS-05
• An optional argument effectively defaults to "":
  • pSelectMode of %SQL.Statement.%OnNew, reached through %New() in CORPUS-17, which reads it only through $get(pSelectMode)

"Assigned but never read" (CORPUS-10, 23, 31, 40). In the stubs in CORPUS-23 and 31, the parameter really is unused. In the %OnBeforeSave and %OnDelete callbacks in CORPUS-10 and 40, the callback signature requires the parameter, so the warning is more debatable. Here ascot trades some precision for recall.

[isc-klu]

I added 15 tasks to the corpus, CORPUS-43 to CORPUS-57 (commits b6f4f25 and 0ae3ca7): bugs the IRIS compiler accepts and the program trips on at runtime, drawn from the runtime-proven tests of my type-checking corpus, plus the misspelled macro you asked for. All 15 are holdout from the day they were written. I then ran iris_check (ascot 0.4.2) on the erroneous and clean version of each, against IRIS Community 2026.2, no model involved.

On the 15 erroneous programs:

• ascot reports errors on 10 and warnings on the other 5; in all 15 the diagnostic is on the planted line.
• The compiler reports an error on one, CORPUS-57 (misspelled macro); see below the table.

On the 15 clean programs:

• ascot reports nothing.
• The compiler reports errors on none.

In short:

• Recall: 15 of 15.
• Precision: 15 of 15.
• Latency per file: median 864 ms cold (a fresh process per file) and 416 ms warm (one running server), from a debug build of iad. Both include the IRIS lookups ascot makes for the file.

These 15 are tasks I wrote around defects ascot detects, so they show what the tool covers.

Per-task results

Task	Bug	Compiler, erroneous version	ascot, erroneous version	ascot, clean version
CORPUS-43	variable set in one If branch, returned on both	None	L9: This '%String' value may be undefined	None
CORPUS-44	accumulator set only inside a For that may run zero times	None	L9: This '%DataType' value may be undefined	None
CORPUS-45	$DATA answering 10 treated as a value	None	L14: This 'Any' value may be undefined	None
CORPUS-46	Quit with a value inside For	None	L9: Quit with %String inside control structure	None
CORPUS-47	%Status method quits with no value	None	L9: Expected return value of type '%Status' but got nothing	None
CORPUS-48	one argument too many to an own method	None	L6: Too many arguments	None
CORPUS-49	utility declared Method, called through ##class	None	L20: Cannot reference instance-only member 'ToFahrenheit' in a class method	None
CORPUS-50	wrong property name set through an object variable	None	L18: Cannot find member 'Turtle' in Bench.Card; did you mean: Title, Owner, %%OID, %GetParameter?	None
CORPUS-51	misspelled %SYSTEM.Version method	None	L6: Cannot find member 'GetMajorVersion' in Class '%SYSTEM.Version'; did you mean: GetMajor, GetVersion, GetCompilerMajorVersion, GetCompilerVersion, GetMinor?	None
CORPUS-52	JSON text returned where %DynamicObject is declared	None	L6: %String is incompatible with %DynamicObject	None
CORPUS-53	collection property subscripted	None	L13: Property 'Items' is not MultiDimensional, so it takes no subscripts	None
CORPUS-54	$CASE with no default	None	L6: This '%String' value may be undefined	None
CORPUS-55	Output parameter never assigned	None	L4: Output parameter 'pCount' is never assigned; its caller will get an undefined value L6: Variable 'tCount' is assigned but never read	None
CORPUS-56	$ORDER loop with its empty check inverted	None	L23: This value is undefined: tQty	None
CORPUS-57	misspelled macro	3 errors, see below	L7: The macro 'ISERROR' is undefined; did you mean: ISERR, ...	None

The erroneous program in CORPUS-57 is the only one the compiler rejects:

ERROR:  Bench.Guard.1(4) : MPP5610 : Referenced macro not defined: 'ISERROR'
ERROR: Bench.Guard.cls(Save+3) #1002: Invalid character in tag : '$$$ISERROR(tSC)' : Offset:10
ERROR: Bench.Guard.cls(Save+6) #1026: Invalid command : '1 }' : Offset:11

Only the first line is the bug and no fix is suggested. ascot reports one error at the exact class line and suggests the right macro, ISERR.

[isc-klu]

The last two commits improves how ascot interact with humans and AI agents: it now takes source code rather than a path to a file system. The new way allows users to check ObjectScript sources without touching the file system. The second commit is a simple change to make a CI happy.

They don't change the diagnostics that ascot provides.