teststand-mcp: Instructions file for Claude Code

TestStandMCP — Behavior Rules for Claude

User-facing rules (imported from the plugin skill)

The rules a plugin user needs live in the teststand-rules skill, which is their single source: edit them there, never here. The imports below load them into every session in this repository. build/Test-SkillSources.ps1 fails when a moved heading reappears in this file or an import dangles; it cannot see a rule restated in a sentence, so keep only the developer fact (measurement, history, test ID, source pointer) in the sections below and point to the skill file for the rule.

@plugin/skills/teststand-rules/SKILL.md @plugin/skills/teststand-rules/references/rebuild.md @plugin/skills/teststand-rules/references/sequence-design.md @plugin/skills/teststand-rules/references/agents.md @plugin/skills/teststand-rules/references/headless-and-tools.md @plugin/skills/teststand-rules/references/tool-contracts.md

Developer notes behind the imported rules

The skill states each rule for a plugin user: no dates, no test IDs, no class names. The facts below are the developer side of the same rules: measurements, history, the suites that pin them and the source that implements them. Each bullet names the skill file that holds the rule.

Export/import and rebuild fidelity (rebuild.md)

• Measured 2026-07-29 on 8 sequences of TFW_MDC_com_Python.seq (47 steps, 13 object-oriented Python steps, 5 LabVIEW .lvlibp steps, 1 cross-file SequenceCall): 3 MCP calls and ZERO FileDiffer differences inside the imported scope (the only rows left are the 22 sequences not imported). The same rebuild with the granular tools took ~700 calls, 3 diff iterations and left 224 differences.
• A WHOLE-file rebuild reaches identical: true, verified 2026-07-29 on TFW_Symphony_DutCom.seq (17 sequences, 79 steps, 43 variables, 18 .lvlibp panes from TWO packed libraries, 4 cross-file calls, 2 ActiveX steps, 49 types) in 4 MCP calls with 0 warnings.
• An import measured 5.5 min before the passes were made cheap; <dest>.import.json exists because that outlives the MCP transport window.
• Subset import: the in-memory type list still lists a type the save just dropped, so the loss is detected AFTER a save+reload. attach='preserve' (what a 1:1 rebuild needs) attaches almost nothing, so a type survives only while still referenced.
• The safe defaults (labview_panes='copy', cross_file_prototypes='copy', variables='copy', keep_unused_types) are locked down by tests in T35. The worker route for the cross-file cache is not viable: it starts its own engine and opens every callee, and one 3 MB callee (Easy.Log.seq) ran it into a 300 s timeout; 'copy' reproduced the same 6 Prototype members in about a second.
• DiffReportShaper.DefaultMaxResults (150) sits on Options, so compare_sequence_files in native mode inherits the cap; diff_sequence_files and compare_sequence_files mode=native are the same diff.
• Connector-pane binding through set_module_parameter, measured: 31 differences became 68 (write everything), 41 (non-empty only), 39 (flag-aware), while writing ViCall.Parms[i].ArgVal and UseDefaultValues INDEPENDENTLY gave 9. A SequenceCall argument's UseDef has the same asymmetry.
• The two prototype-load kinds are mutually exclusive per server process (measured 2026-07-29): a LabVIEW pane load first makes the next SeqCall load return prototypeLoaded:false ("LoadPrototype could not resolve the target/module"), and the reverse order fails the same way.
• Residuals closed 2026-07-29, each by cloning from the source file instead of loading: the LabVIEW connector panes (pass 5), the cross-file Prototype cache (pass 7), and a step's authored TS.AdditionalResultsHints / CustomResults / ErrorDialogOptions for EVERY step (pass 6b). copy_typedefs reproduces every type correctly (0 types differences); the declarative route had lost the type INSTANCE, not the type.
• A named-type instance's ENUM member reading as explicitly set was long recorded as an irreducible API limit; it is not. Instantiating the type (insert_local_variable dataType:"LogEvent") writes the default enumerator NAME out, so the member reads [Debug] where the editor-authored original has {Debug} (get_property_tree: symbolicName:"Debug", isDefault:false vs symbolicName:"", isDefault:true). Not writing the value does not help: the import already skips it (isDefault:true in the model, and WriteEnumLeafExplicit returns early on null ordinal+value). variables='copy' clones the variable and reproduces the state exactly.
• NI.Analyzer.IgnoredMessages is invisible to the engine API; only FileDiffer's raw reader sees it, as a lone File Properties > Attributes delete.
• Typed PropertyObject (2026-07-29): Clone(path, CopyAllFlags) then SetPropertyObject on a subtree whose node carries a NAMED TYPE re-registers a second instance of the type; the fix writes the leaf scalars (SetValString/SetValNumber/SetValBoolean) on the existing node. Measured: modulesCloned dropped 79 -> 47 and the dialog disappeared. The decision lives in StepCopyPolicy (scalar compared and written by value, empty list-like node skipped, object copy last resort), pinned by tests in T35. The FileDiffer reported identical: true for the file that raised the dialog. audit_type_consistency (2026-07-30) reads the file's TypeUsageList RAW, duplicates included, which EnumerateFileTypeDefs deliberately hides.
• File format (2026-07-31): MEDELA_TFW's real files are XML, so a rebuild used to come out binary: content-identical yet 25 KB against 3.4 MB on TFW_MDC_com_Python (x133, pure serialization). TestStandService ParseFileWritingFormat / ApplyFileWritingFormat; pinned by T36_FileFormatTests (tests asserting the RAW bytes TOF1 vs a UTF-8 BOM + <?xml). An unknown format name THROWS rather than falling back to binary.
• .lvlibp guard (2026-07-30): InputGuards.IsPackedLibraryModulePath plus the refusal message; force_unsafe_inprocess:true overrides. The in-process load raised the MSVC delay-load SEH 0xC06D007E (the LabVIEW Run-Time lvrt.dll) with LabVIEW 2026 32-bit already started and responsive; it escapes managed try/catch, so the server PROCESS DIED and the NI Error Reporter appeared. isolate:true is crash-safe but cannot bind the running LabVIEW ADE, so it only times out (8 steps x 120 s, all timeouts; it also reads the file from DISK). An earlier note recommended isolate:false after one apparently good run; that advice caused a crash and is retracted.
• Per-sequence clone recipe: memory teststand-whole-sequence-clone-rebuild-2026-07-08.
• set_step_property_flags (measured 2026-10-24, TestStand 2026 Q3): a plain SetFlags REPLACES the bitfield. 0x4 then 0x80 (no exact) reads back 0x400080, so 0x4 is gone; exact=true gives the same result; the engine keeps its own 0x400000 bit in both modes. The earlier "SetFlags only ever turns bits ON" descriptions and comments were wrong and are fixed; the exact path (SetExactFlags) stays as the verified write (read-back check, then SetInstanceOverrideFlags).