Phase 13. LLM (cassette playback)

Field	Value
MEP	MEP-49 §Phases · Phase 13
Status	LANDED
Started	2026-05-28 13:40 (GMT+7)
Landed	2026-05-28 14:15 (GMT+7)
Tracking issue	—
Tracking PR	—

Gate

TestPhase13LLM: 5 fixtures green on Swift 6.0+, macOS 15, with cassette playback via MOCHI_LLM_CASSETTE_DIR. Gate builds each fixture, sets the cassette dir env var, runs the binary, and compares stdout to .out.

Goal-alignment audit

LLM calls are non-deterministic at runtime, so the gate uses pre-recorded cassette files (matching MEP-46 Phase 13 for the BEAM backend). The v1 implementation is synchronous: mochiLLMGenerate reads a .txt cassette file keyed by a DJB2 XOR hash of provider + "\0" + model + "\0" + prompt. This matches the BEAM backend's cassette format exactly. Apple FoundationModels on-device inference and cloud fallback are deferred.

Sub-phases

#	Scope	Status	Commit
13.0	LLMExpr → mochiLLMGenerate(provider, model, prompt)	LANDED	mep/0049-phase-13
13.1	Cassette lookup: DJB2 XOR hash → MOCHI_LLM_CASSETTE_DIR/<hash>.txt	LANDED	mep/0049-phase-13
13.2	Apple FoundationModels on-device inference path	DEFERRED	—
13.3	Cloud fallback via URLSession for non-Apple platforms	DEFERRED	—
13.4	Structured output: @Generable protocol + @Guide annotations	DEFERRED	—

Sub-phase 13.0 -- LLM expression lowering

Decisions made (13.0)

LLMExpr: the aotir IR node for an LLM call, carrying Provider, Model, and Prompt fields.

Lowering: the lowerer emits:

mochiLLMGenerate("anthropic", model, prompt)

as a RawSwiftExpr. The provider string is embedded as a string literal; model and prompt are lowered expressions.

Return type: String. The function is synchronous; no async/await needed.

Sub-phase 13.1 -- Cassette lookup

Decisions made (13.1)

Hash function: DJB2 XOR variant: h = (h &* 33) ^ UInt64(byte), initialised at 5381, over the UTF-8 bytes of provider + "\0" + model + "\0" + prompt. This matches the BEAM backend's cassette hash exactly, so cassette files are portable between backends.

Cassette file path: $MOCHI_LLM_CASSETTE_DIR/<hash>.txt. The file contains the raw LLM response text (no JSON envelope). The result is trimmed of trailing newlines before being returned.

mochiLLMGenerate implementation:

public func mochiLLMGenerate(_ provider: String, _ model: String, _ prompt: String) -> String {
    if let cassetteDir = ProcessInfo.processInfo.environment["MOCHI_LLM_CASSETTE_DIR"] {
        let key = mochiDJB2Key(provider, model, prompt)
        let path = "\(cassetteDir)/\(key).txt"
        if let content = try? String(contentsOfFile: path, encoding: .utf8) {
            return content.trimmingCharacters(in: .newlines)
        }
        fputs("mochi_llm: cassette not found: \(path)\n", stderr)
        return ""
    }
    fputs("mochi_llm: MOCHI_LLM_CASSETTE_DIR not set\n", stderr)
    return ""
}

Fixture structure: each LLM fixture is a subdirectory (not a flat .mochi file) containing <name>.mochi, <name>.out, and a cassette/ subdirectory with <hash>.txt files.

Files changed

File	Purpose
transpiler3/swift/lower/lower.go	LLMExpr lowering to mochiLLMGenerate(provider, model, prompt)
transpiler3/swift/runtime/Sources/MochiRuntime/LLM.swift	mochiLLMGenerate, mochiDJB2Key
transpiler3/swift/build/phase13_test.go	TestPhase13LLM: 5 fixtures with cassette playback
tests/transpiler3/swift/fixtures/phase13-llm/	5 fixture subdirectories (each with cassette/ dir)

Test set

• TestPhase13LLM -- 5 fixtures: generate_anthropic, generate_concat, generate_in_var, generate_multiple, generate_text.

Deferred work

• Apple FoundationModels on-device inference (LanguageModelSession, #if canImport(FoundationModels)). Deferred to Phase 13.2.
• Cloud fallback via URLSession for Linux/Windows. Deferred to Phase 13.3.
• Structured output: @Generable macro + @Guide field annotations. Deferred to Phase 13.4.
• Streaming LLM responses (@llm(stream: true)). Deferred.
• Multi-turn conversation history. Deferred.