[BUG] The InfiniteAIAssistant command type is unreachable, unconfigurable, and silently skipped at run time

[chhoumann]

Found while bringing the four imperative format previews under the #1560 inert-preview mechanism (#1565) - AIAssistantInfiniteCommandSettingsModal.ts is one of the four, and it turns out nothing can reach it.

Describe the bug

CommandType.InfiniteAIAssistant exists, has a full settings modal
(src/gui/MacroGUIs/AIAssistantInfiniteCommandSettingsModal.ts, 275 lines: result
joiner, chunk separator, max chunk tokens, merge chunks, sampling params), a
IInfiniteAIAssistantCommand interface, package-disclosure support
(src/services/packagePreview.ts:637), model-ref pinning
(src/ai/modelRefPinning.ts:15) and a migration
(src/migrations/refreshStaleDefaultModelSeeds.ts:95).

But:

1. You cannot create one. The macro builder's add-command menu offers "Add AI
   Assistant command" only - there is no InfiniteAIAssistant entry anywhere in
   CommandList.svelte or MacroBuilder.ts.
2. You cannot configure one. CommandList.svelte:243 only renders a configure
   affordance for CommandType.AIAssistant. Verified live: seeding a macro with
   an InfiniteAIAssistant command renders the row with Delete and Reorder
   buttons but no Configure button.
3. It does nothing at run time. MacroChoiceEngine.executeCommands
   (src/engine/MacroChoiceEngine.ts:292-306) is a flat chain of if (isXCommand(command)) guards with no InfiniteAIAssistant branch and no
   else. The command is silently skipped - no error, no notice, no log.
4. The modal is only ever constructed from the devMode-only testQuickAdd
   command (src/main.ts:194), which additionally passes
   type: CommandType.AIAssistant for an IInfiniteAIAssistantCommand - so even
   the dev harness builds a mistyped object.

So the code is dead, but not inertly dead: a package importing a macro that
contains one will show "AI assistant" in the import disclosure
(MacroDisclosure.svelte:16) and then do nothing when the macro runs.

Expected behavior

Either wire the type up (creatable in the macro builder, configurable, executed
by the engine) or remove it and migrate any InfiniteAIAssistant commands found
in existing data - but not the current state, where the disclosure promises a
step that the engine drops on the floor.

Given nobody can have created one through the UI in the currently-supported
versions, removal + a migration that converts them to a plain AIAssistant
command (or drops them with a notice) looks like the smaller, honest fix. Worth
checking whether an older QuickAdd could create them before deciding.

Additional context

Deliberately not folded into #1565's diff, which stays at the preview/call-site
layer. #1565 still fixes this modal's preview so the two AI prompt modals do not
diverge, whichever way this issue is resolved.

[chhoumann]

Investigated end to end, including a live repro. Recommendation below, with three corrections to the issue body.

Corrections

1. It shipped in 1.2.0, not 1.10.0. ae90edeb landed 2023-06-20 19:31Z and tag 1.2.0 was cut six minutes later. Every release since has carried it.
2. It was never creatable in any released version, not just "the currently-supported ones". Across all 202 tags, git log --all -S"InfiniteAIAssistant" --name-only shows the identifier only ever touched CommandType.ts, IAIAssistantCommand.ts, the modal (+test), main.ts, packagePreview.ts, MacroDisclosure.svelte, modelRefPinning.ts, refreshStaleDefaultModelSeeds.ts, systemPromptFields.test.ts. Never a builder, engine or API file. No InfiniteAIAssistantCommand class has ever existed; the only add-AI path is CommandSequenceEditor.ts:516 -> new AIAssistantCommand(), which hardcodes CommandType.AIAssistant. It was abandoned deliberately: chunked prompting shipped as a user script (discussion [BETA FEATURE] Chunked Prompts - Prompt OpenAI with text way beyond the input limits #535, "I'm kind of on the fence about whether it should be a feature") and later as quickAddApi.ai.chunkedPrompt().
3. "Silently skipped" understates it. Repro in an isolated vault, macro = Wait -> InfiniteAIAssistant(outputVariableName: "out") -> Capture with format AI said: {{VALUE:out}}. The skipped step never sets out, so the capture's {{VALUE:out}} hits ensureValueVariableResolved -> promptForVariable and a modal titled "out" opens mid-macro asking the user to type the AI result by hand. Non-interactive runs abort with ChoiceAbortError. Also confirmed live: no Configure button (Delete + Reorder only), survives load and all 15 migrations intact, macro reports success.

Recommendation: two PRs, and no data migration.

PR 1 - stop dropping steps. MacroChoiceEngine.executeCommands (:280-306) is a flat if (isX(command)) chain with no else. InfiniteAIAssistant is one instance of a class. Any unrecognised type from a hand-edited data.json, a hand-authored package (QuickAddPackage.ts:41-47 validates only that id/name/type are strings), or a data.json written by a newer QuickAdd and read by an older one (which our own downgrade recipe produces) is dropped the same way, including inside Conditional branches. Use the switch + const exhaustiveCheck: never idiom already in this file at :615-618, so adding a CommandType without a handler becomes a tsc error. Two caveats: add if (!command) continue first (the loop has no null guard, and packageImportService guards for exactly this at :294/:945/:1121), otherwise a corrupt data.json produces red notices right after the #1583 resilience work; and continue-and-shout rather than abort, matching executeObsidianCommand at :529-538.

PR 2 - remove the type. Enum member, IInfiniteAIAssistantCommand, the 268-line modal (not 275) + its 114-line test, the systemPromptFields.test.ts entry #1581 just added, the devMode testQuickAdd command (its entire body is constructing this modal; touches commandLabels.ts and the strict toEqual in main.commandLabels.test.ts), the packagePreview case, the MacroDisclosure label, and the isAICommand clauses in modelRefPinning.ts:15 and refreshStaleDefaultModelSeeds.ts:95. Collateral: getMaxChunkTokensUpperBound / getLargestModelMaxTokens (aiHelpers.ts:250,272) lose their only production caller, along with aiHelpers.maxChunkTokens.test.ts.

Two things fall out for free. KNOWN_COMMAND_TYPES is Object.values(CommandType) (packagePreview.ts:245), so removal flips a stray command from the false "AI assistant / sends note content to your AI provider" row to the honest "Unknown capability. Review it manually." row. Option (E) needs no separate work. And pinAiCommandModelRefs stops classifying it as an AI command at import time (packageImportService.ts:658) - a behaviour change worth naming in the PR body rather than filing under "2 LOC".

Why no migration, despite what I wrote in the issue:

• Converting to a plain AIAssistant is not inert. With promptTemplate.enable: false, executeAIAssistant -> runAIAssistant opens a prompt-template picker and fires a billable network request. It turns a no-op into an interactive, billable step the user never configured. Under disableOnlineFeatures (the default, settings.ts:122) it throws instead.
• Dropping the command deletes user data to fix a hazard nobody demonstrably has.
• A migration runs once and is flagged forever, but hand-editing and package import still work afterwards. PR 1 covers the hazard permanently and generically.
• chunkSeparator is a string here and a RegExp in chunkedPrompt, so a stored "\n" is a literal backslash-n. A conversion could not carry the setting faithfully anyway.

Rejecting "wire it up". It is not a missing if. IInfiniteAIAssistantCommand has no promptTemplate and no text source, and ChunkedPrompt throws without a chunk reference (AIAssistant.ts:886-892). Wiring it up means designing where the text comes from plus a template picker: a new feature with zero recorded demand, competing against quickAddApi.ai.chunkedPrompt, which is a strict superset (documented, 588-line test file, auto-split on provider rejection, 500-prompt cap). If chunking ever returns to the builder, the right shape is a "chunk long input" toggle on the existing AI Assistant command reusing its promptTemplate, not a second command type. Recording that here so it does not get re-proposed.

Docs, in PR 2. AIAssistant.md:203 is an anchor-pinned ### Max chunk tokens {#max-chunk-tokens} sitting under "Model settings and token budgets" next to real settings. The only "Max chunk tokens" control in the plugin is the slider inside the unreachable modal (AIAssistantInfiniteCommandSettingsModal.ts:227). A reader will hunt for a setting that does not exist. Anchors are frozen, so this moves to the chunkedPrompt API docs with the anchor preserved.

Acquisition surface, for the record: hand-edited data.json, a script mutating settings.choices, or a hand-authored package. Zero hits in issues, zero in discussions, zero public JSON on GitHub, zero of the 6 public packages. That is "no evidence anyone has one", not "impossible".