LocalAI/docs/design/realtime-state-machines.md

Realtime API state machines — map & re-architecture research

Status: research / design (compaction phase). No code changes implied yet.

The realtime API (core/http/endpoints/openai/realtime*.go) grew feature-by-feature (server_vad → semantic_vad/EOU, streaming pipeline, tool turns, compaction, voice gate, sound detection, WebRTC). The result is several implicit state machines whose states and transitions are scattered across goroutine-local variables, shared Session/Conversation fields under five different mutexes, raw channels, and context cancellation. State is inferred from variable combinations rather than stored; several illegal/inconsistent states are reachable.

This document (1) inventories the implicit machines, (2) catalogues the cross-cutting failure modes, (3) researches how to re-implement them explicitly and verifiably, and (4) lists the invariants a correct implementation must guarantee.

All line numbers are against the current feat/realtime-semantic-vad-eou branch and will drift; treat them as anchors.

---

Part 1 — Inventory of the implicit state machines

There is no state/status field anywhere in Session or Conversation. Every machine below is reconstructed from variable combinations.

M1. Connection / transport lifecycle

Two transports implement one Transport interface; their lifecycles differ sharply.

• WebSocket (realtime_transport_ws.go): essentially stateless — a *websocket.Conn plus a write sync.Mutex. No send queue, no send goroutine, no closed flag. "Closed" = ReadEvent returns an error.
• WebRTC (realtime_transport_webrtc.go): an explicit-ish machine built from raw channels — dcReady (closed by dcDone sync.OnceFunc), closed (closed by closeDone sync.OnceFunc from either OnConnectionStateChange or Close()), flushed, sessionCh (cap 1), inEvents/outEvents (cap 256), plus a sendLoop goroutine and RTP counters under rtpMu.

Conceptual states (connecting → data-channel-open → session-created → active → closing → closed) are not stored; the only persisted membership state is the sessions[sessionID] map entry (exists realtime.go:631→:1009). session-created and session-updated are events, not states.

Teardown order (realtime.go:989-1010): cancelActiveResponse → close(decodeDone) → close(done) (if VAD running) → close(soundWindowDone) → wg.Wait() → delete(sessions,…). Then, WebRTC only, defer transport.Close() → closeDone() → <-flushed → pc.Close().

M2. Audio-input / turn-detection (server_vad + semantic_vad + EOU)

One handleVAD goroutine (realtime.go:1322) on a 300 ms ticker. Mode is re-evaluated every tick under sessionLock (:1350-1357) so it can flip mid-turn.

• server_vad states are encoded by the goroutine-local speechStarted bool (:1337) plus silence measured (not timed) as audioLength - segEndTime > silenceThreshold recomputed each tick (:1461). States: idle → inspecting → speech-detected → awaiting-commit → committing → transcribing/responding. "Holdback" is a byte count (noSpeechHoldbackSec*rate*2), not a timer.
• semantic_vad adds the liveTurnState struct (realtime_semantic_vad.go): live (nil = closed), unavailable (sticky degrade → behaves as server_vad), eouAtSec, parts, itemID (allocated at turn open so captions can stream), deltasSent. Extra states: closed, open/streaming-ASR, EOU-pending, EOU-fallback (dynamic silence threshold 0 s when EOU pending, else eagerness 8/4/2 s), retranscribe-gate, EOU-rejected, finished, discarded. The one cross-goroutine edge: the backend recv callback pushes onto events (buffered 64, non-blocking — drops on overflow, :116-117); drainEvents reads it on the tick.
• Voice gate (realtime_voicegate.go) runs inside the commit goroutine: resolving → authorized/rejected, with a sticky voiceVerified (under gateMu) for when:first.

M3. Response lifecycle (+ synchronous tool-turn recursion)

A response is "active" iff Session.activeResponseDone is non-nil and unclosed (responseMu, :172). One goroutine owns it; its lifetime == that channel's. State is observable only through the response.* event stream and ItemStatus* on the assistant item. Logical states: idle → starting → generating-text → generating-audio → tool-call-pending → tool-executing → awaiting-next-tool-turn → cancelling → done(completed|cancelled) | failed.

• Cancellation is cooperative at discrete checkpoints (ctx.Err() at :2172,2364,2394, realtime_stream.go:193,202,241,259).
• The tool loop is synchronous recursion on the same goroutine, bounded by maxAssistantToolTurns = 10; each level mints a fresh responseID and emits a full response.created … response.done{Completed} cycle — so one user turn can emit several response.done{Completed} events under different IDs.
• Terminal events are not exactly-once: failed paths return with no response.done; cancelled paths emit done{Cancelled}; the completed terminal is unconditional at the tail of emitToolCallItems.

M4. Conversation / compaction

Conversation: Items + Memory (rolling summary) under Lock; compacting atomic.Bool. States: normal ↔ compacting. Compaction (realtime_compaction.go) snapshots overflow under Lock, summarizes unlocked, re-locks and commits guarded by an optimistic head-prefixMatches check. It is launched only by turn-0 triggerResponse (:1963), off the response path — so a long agentic turn (recursion calls triggerResponseAtTurn directly) can append many tool items and never compact until the next user turn (compaction starvation).

M5. Streaming sub-machines (transcription, chunker, TTS)

Backend LLM/TTS/transcription streams are synchronous callback recv loops on the caller's goroutine — no internal goroutines/channels. The only true concurrent FSM is:

• TTS pipeline (realtime_tts_pipeline.go): one worker goroutine, an unbounded mutex-guarded queue, a coalesced wake chan (cap 1), a closed flag, a done chan closed once by the worker's defer, a lock-free failed atomic.Bool, and worker-owned audio/firstErr that are safe to read only after wait() joins via done. Idempotent wait(); deferred wait() backstop guarantees no worker leak.
• Chunker (realtime_chunker.go): a pure single-buffer FSM (buffering ↔ emitting, flush = hard boundary). No concurrency guard — correctness depends entirely on push/flush being called from one goroutine (the LLM recv loop). On cancel the flush is skipped, so the buffered partial clause is intentionally dropped.
• Transcription (realtime_transcription.go): stateless straight-line function; "streaming" is just repeated synchronous callbacks.

---

Part 2 — Cross-cutting failure modes (why it's a mess)

1. Shared mutable Session config with inconsistent locking (the core problem). updateSession/updateTransSession mutate Voice, Instructions, Tools, OutputModalities, ModelConfig, ModelInterface, sample rates, and the shared InputAudioTranscription pointer under sessionLock. But in-flight response/speech/transcription goroutines read those same fields without any lock (realtime_speech.go:72-79, realtime_stream.go:228, semantic_vad :110). Reloading ModelInterface mid-response is a data race against a running Predict/TTS/Transcribe, and the swapped-out model is dropped without Close. sessionLock actually guards the global sessions map; it only mutually excludes the handful of other sites that happen to also take it (handleVAD tick, the commit branch). Response goroutines never take it.

2. Two writers of the active-response pair. startResponse/cancelActiveResponse are called from both the main read loop (:836,973,981,990) and the VAD goroutine (barge-in :1429, end-of-speech :1543). responseMu guards only the field swap; the <-done wait is outside the lock. A read-loop ResponseCreate racing a VAD speech_stopped can have both read the same prior pair, both overwrite, and briefly leave two live response goroutines both appending to conv.Items. The "never overlapping" guarantee holds only under the unstated assumption that responses are driven from a single goroutine — which is false.

3. State is inferred, not stored. Whether a response is active, whether a turn is open, whether audio is being buffered — all are derived from combinations of booleans, nil-checks, channel state, and context error. No single source of truth; no place to assert an invariant.

4. Reachable inconsistent states. e.g. after a semantic-VAD discardTurn, speechStarted stays true while lts is closed, so they disagree and the next onset suppresses SpeechStarted. Mid-stream cancel leaves the client having seen output_item.added/content_part.added with no matching …done. events-channel overflow silently drops an EOU, degrading EOU-pending to the 2–8 s fallback.

5. Lifecycle/ownership gaps. decodeOpusLoop is a bare go (not in wg) and can run after delete(sessions,…). handleIncomingAudioTrack (pion OnTrack goroutine) has no shutdown signal — it appends to OpusFrames until ReadRTP errors, unjoined by wg. WebRTC outEvents enqueued before the DC opens are lost on early failure.

6. The done-channel/vadServerStarted toggle dance. A single done local (:655) is reassigned to a fresh channel on each VAD start (:662) and closed at toggle-off (:670) and teardown (:999). Safe today only because one goroutine owns it — one variable name meaning different channels over time is a structural fragility, not an explicit lifecycle.

---

Part 3 — Research: explicit, verifiable re-implementation

The goal the user stated: transitions cannot lead to an inconsistent state, and we can verify that. Four layered techniques, from architecture down to runtime.

3.1 Architecture: single-writer session actor (share by communicating)

The root cause of (1) and (2) is shared mutable state across goroutines. The most effective, idiomatic-Go fix is to give each session one owning goroutine that holds all session state with no locks, and have every other goroutine communicate with it over channels:

            ┌────────── inbound events ──────────┐
 transport ─┤  client events (ReadEvent)         │
   VAD     ─┤  vad: speech_started/stopped, EOU   ├─►  session actor  ──► outbound
 model I/O ─┤  llm/tts/asr results, errors        │   (owns ALL state,    events
 timers    ─┤  ticks, deadlines                   │    single goroutine)
            └────────────────────────────────────┘

• All state mutation happens in one place; sessionLock, responseMu, gateMu, AudioBufferLock, OpusFramesLock, Conversation.Lock collapse into "the actor owns it." Worker goroutines (Predict/TTS/ASR, opus decode, RTP read) become stateless effects that take an immutable snapshot in and send results back as events.
• ModelInterface reload becomes an event the actor sequences relative to responses (e.g. drain/cancel the active response first), eliminating the mid-call swap race.
• Cancellation stays context-based but the actor is the only thing that starts/stops responses, killing the dual-writer race (2).

This is the actor / CSP model. It does not by itself prove correctness — that's what 3.2–3.4 add — but it makes the state centralized and explicit, which is the precondition for verification.

3.2 Make illegal states unrepresentable (type-level)

Inside the actor, model each machine as an explicit state with a pure transition function next(state, event) (state, []effect, error):

• Represent states as a Go sealed sum type (interface with an unexported marker method, one struct per state carrying only that state's data) so e.g. EOU-pending data cannot be accessed while Closed. This is the Go equivalent of an ADT and is the single biggest lever for "inconsistent state unrepresentable."
• The transition function is total and pure (no I/O, no goroutines): it returns the next state plus a list of effects (send event, start Predict, arm timer) that the actor executes. Pure transition functions are trivially unit-testable and property-testable.
• An unexpected (state, event) pair returns an explicit error / stays put and logs — never a silent half-transition.

The four machines are hierarchical (a statechart): Connection ⊃ Turn(M2) and Response(M3) ⊃ Tool-turn; Conversation(M4) and the TTS sub-machine(M5) are largely orthogonal regions. Model them as nested states rather than one flat enum.

Library options (all guard logic, none give concurrency safety — that's 3.1's job):

• qmuntal/stateless — declarative, hierarchical, guard/entry/exit actions; closest fit.
• looplab/fsm — simpler, flat, event-callback based.
• Hand-rolled transition tables — most control, no dep; recommended here given the hierarchy and the desire to keep transitions auditable. go.mod currently pulls no FSM lib.

3.3 Design-time formal verification (prove the protocol)

Before/while coding, model the protocol (not the Go) in a model checker to prove the hard concurrency properties exhaustively:

• FizzBee (the adopted tool) to specify the actor's event/state space and check: no two concurrent active responses; barge-in + ResponseCancel + speech_stopped interleavings never deadlock or drop a turn; every response.created is eventually followed by exactly one terminal; teardown joins all goroutines. The cancel/startResponse/barge-in interplay (failure mode 2) is exactly the kind of liveness/safety property model checkers exist for.
• Keep the spec small and focused on the M2↔M3 boundary (turn detection ↔ response), which is where the real races live.

3.4 Implementation-time & runtime verification

• Exhaustive table-driven transition tests: since transitions are a pure function, enumerate (state × event) and assert the result for every cell, including the illegal cells (assert they error / no-op). This is the practical stand-in for a proof that "no transition leads to inconsistent state."
• Property-based testing: feed random event sequences into the actor and assert global invariants hold after every step (Part 4). This catches reachable-bad-state bugs the example tests miss. (Implemented as Ginkgo/Gomega seeded random-walk specs — see Part 6.2 for why not rapid.)
• Race detector under load: run the property tests with -race; with 3.1 there should be zero shared mutable state, so -race cleanliness becomes a meaningful signal rather than noise.
• Runtime invariant assertions + structured transition logging: log every state --event--> state with the session ID; assert invariants in dev builds. Replace today's silent degradations (dropped EOU, suppressed SpeechStarted) with explicit, observable transitions.

3.5 Recommended path for LocalAI

1. Specify the M2↔M3 protocol in FizzBee; nail the cancel/barge-in invariants.
2. Introduce a per-session actor (3.1) that owns existing state behind the current Transport interface — incremental, keeps the event types.
3. Replace each implicit machine with an explicit sealed-state transition function (3.2), one at a time: Response first (highest-risk dual-writer), then Turn/VAD, then Connection, then leave TTS/Chunker/Compaction (already mostly self-contained) for last.
4. Land the table-driven + property-based test suites alongside each machine; gate on -race.

---

Part 4 — Invariants a correct implementation must guarantee

These are the "cannot reach inconsistent state" properties to encode as assertions, property-test oracles, and FizzBee invariants:

1. At most one active response per session at any instant (no overlapping response goroutines; no two appenders to conv.Items from response logic).
2. Exactly one terminal per response.created: every emitted response.created is followed by exactly one of response.done{completed|cancelled} or a defined failure terminal — never zero, never two. (Decide whether agentic tool turns are one response or many; make it explicit either way.)
3. No response.* content events after that response's terminal. No output_item.added/content_part.added without a matching …done (even on cancel).
4. Turn/response coupling: speechStarted ⟺ a live turn is open; barge-in cancels the active response before a new turn's commit starts.
5. No config field is read by a worker while being mutated (reload is sequenced against in-flight work; a response uses an immutable snapshot of model/voice/tools).
6. Audio buffer monotonic & consistent: commit/clear/append/VAD-drop never lose or double-consume bytes; clear resets all turn state (including lts).
7. No dropped control events: an EOU/Final is never silently lost (no overflow-drop on a bounded channel that changes turn outcome).
8. Clean teardown: every spawned goroutine (incl. decodeOpusLoop, handleIncomingAudioTrack) is signalled and joined before the session is deleted; no sends after transport close.
9. Compaction safety & liveness: compaction never races a reader into a torn Items; and it actually runs when the trigger is exceeded, including inside long agentic turns.
10. Idempotent close: every channel/resource closed exactly once on every path.

---

Implementation status

• M3 (response coordination) — first vertical slice landed. Explicit machine in core/http/endpoints/openai/respcoord/ (sealed State/Event/Effect sum types, a total pure Next, a single-writer Coordinator); transition-table + Ginkgo/Gomega seeded-property + concurrent conformance tests (green under -race); a deterministic characterization test pinning the legacy dual-writer race. Authoritative spec: formal-verification/response_lifecycle.fizz. Gate: scripts/realtime-conformance.sh (Go layer always; FizzBee when pinned) wired as make test-realtime-conformance and .github/workflows/realtime-conformance.yml. See formal-verification/README.md.
• Gate is fail-closed and pinned (done). fizzbee.sha256 pins all four platforms; the gate hard-fails without FizzBee; CI installs+caches the verified binary with no skip; pre-commit runs the gate on respcoord/** or formal-verification/** changes.
• M3 wired into the live session (done). realtime_respcoord.go adds responseSink (the respcoord.Coordinator + a goroutine-spawning effect sink) to Session. The legacy startResponse/cancelActiveResponse and the dual-writer activeResponse*/responseMu fields are gone; all six call sites (manual commit, response.create, VAD speech-stopped, response.cancel, barge-in, teardown) route through it. Barge-in/cancel are now non-blocking (removes the legacy ~300 ms VAD stall); teardown stops input goroutines, then cancels + wait()s all response goroutines before deleting the session. EmitTerminal is a no-op for now (the response body still emits its own response.done) — coordination is fixed without changing wire behavior. Verified: builds, go vet clean, all 300 openai specs pass under -race, and make test-realtime (the mock-backend realtime e2e suite, 12 specs over WS + WebRTC) passes.
• Single authoritative terminal + populated Output/Usage (done). One response.created and one response.done per response.create, even across the server-side agentic tool loop (which is now internal turns of one response, not one terminal each). A liveResponse accumulator threads through triggerResponse→triggerResponseAtTurn→emitToolCallItems/streamLLMResponse, collecting output items as they complete and summing token usage; triggerResponse emits the one terminal (completed/cancelled; failed still emits none, matching legacy) with Output + Usage filled in (both were always empty before). Verified: 301 openai specs under -race (incl. a new triggerResponse terminal test) + make test-realtime. Design note: emission is hoisted to triggerResponse (the body owns it) rather than the coordinator's EmitTerminal effect — at cancel/supersede time the coordinator doesn't yet have the body's partial Output, so the body, which does, is the natural emitter. The coordinator still guarantees one body run per response.create, so "exactly one terminal" holds transitively; EmitTerminal remains the spec's logical marker (no-op in the sink).
• M2 (turn detection) — model + spec landed AND wired into the live session. Explicit machine in core/http/endpoints/openai/turncoord/ (sealed State = Idle | Speaking{Turn}, Event = Onset | Silence | Abort{Reason}, Effect = BargeIn | OpenTurn | EmitSpeechStarted | EmitSpeechStopped | CommitTurn | DiscardTurn, a total pure Next, a single-writer Coordinator); transition-table + Ginkgo/Gomega seeded-property + concurrent conformance tests (green under -race). The fix it encodes: "speech detected" and "a turn is open" — the two legacy variables (speechStarted and lts.open()) that a discardTurn could desync (failure mode 4) — become ONE state, so the next-onset suppression bug is unrepresentable. Authoritative spec: formal-verification/turn_lifecycle.fizz, with an always assertion Coupled (speech ⟺ turn-open), verified non-vacuous (deleting self.speech = 0 in Abort makes the checker report Coupled violated). The gate (scripts/realtime-conformance.sh, pre-commit, CI) covers turncoord and the spec. Wired (done): realtime_turncoord.go adds turnSink (the turncoord.Coordinator + a loop-local effect sink) to handleVAD. The legacy speechStarted bool is gone; onset/no-speech-clear/commit/teardown route through coord.Apply(Onset|Abort{NoSpeech}|Silence|Abort{Teardown}). The turn id is minted at onset and carried by the coordinator to the committed event (so it matches the live captions); liveTurnState.openTurn now takes that id instead of minting its own. A semantic→server mode switch mid-turn is deliberately NOT an abort (it only drops the orphaned live stream and lets the turn continue under server_vad), so it stays inline. Verified: builds, go vet/gofmt/golangci-lint clean, all openai specs under -race, and make test-realtime (12 e2e specs over WS + WebRTC) pass.
• M1 (connection lifecycle) — model + spec landed AND wired. Explicit machine in core/http/endpoints/openai/conncoord/ (sealed State = Live{VADRunning} | Torn, Event = SetVAD | Close, Effect = StartVAD | StopVAD | Teardown, a total pure Next, a single-writer Coordinator); transition-table + Ginkgo/Gomega seeded-property + concurrent conformance tests (green under -race). It replaces the legacy vadServerStarted bool + the done channel reassigned on every turn-detection toggle and closed from two sites (failure mode 6): the coordinator owns whether the VAD goroutine runs, so its done channel is closed exactly once and never resurrected after teardown; Close moves to Torn, which absorbs every later event so teardown runs exactly once even from multiple exit paths (invariants #8, #10). Spec: formal-verification/conn_lifecycle.fizz (always assertion TeardownOnce + NoRunAfterTorn), verified non-vacuous (deleting self.torn = 1 in Close fails TeardownOnce). Wired (done): realtime_conncoord.go adds connSink; the handler's setup/toggleVAD/teardown now route through conn.setVAD(...)/conn.close(); the done/vadServerStarted locals and the manual ordered-teardown block are gone (the Teardown effect performs that sequence). Verified: builds, vet/gofmt/golangci-lint clean, openai specs under -race, make test-realtime (12 e2e WS+WebRTC), full conformance gate green (3 Go packages + 3 fizz specs PASSED).
• M4 (conversation compaction) — model + spec landed AND wired. Explicit machine in core/http/endpoints/openai/compactcoord/ (sealed State = Idle | Running, Event = Trigger | Finished, Effect = StartCompaction, a total pure Next, a single-writer Coordinator); transition-table + Ginkgo/Gomega seeded-property + concurrent (effect-spawns-work-reports-Finished) conformance tests (green under -race). It makes the legacy compacting atomic.Bool single-flight guard explicit: a Trigger while Running is dropped (not superseded — compaction is idempotent work on the same overflow), so at most one summarize+evict runs per conversation (invariant #9). Spec: formal-verification/compaction.fizz (always assertion SingleFlight), verified non-vacuous (deleting the if self.active == 0 guard fails SingleFlight). Wired (done): realtime_compactcoord.go adds compactionSink; the Conversation.compacting atomic.Bool is replaced by Conversation.compaction *compactionSink (built at conversation creation with the summarize+evict run closure); maybeCompact now calls conv.compaction.trigger(). The summarizer resolution + compact() stay in the sink's spawned goroutine (off the response path); compact() itself (snapshot/summarize-unlocked/optimistic-commit) is unchanged. Verified: builds, vet/gofmt/golangci-lint clean, openai specs under -race, make test-realtime (12 e2e), full conformance gate green (4 Go packages + 4 fizz specs PASSED).
• M5 (TTS pipeline lifecycle) — model + spec landed AND wired. Explicit machine in core/http/endpoints/openai/ttscoord/ (sealed State = Open | Closing | Closed, Event = Close | WorkerExited, Effect = Wake, a total pure Next, a single-writer Coordinator); transition-table + Ginkgo/Gomega seeded-property + two-writer conformance tests (green under -race). It is a genuine two-writer machine (producer Close from wait() vs worker WorkerExited); it makes the legacy closed bool lifecycle explicit and monotonic, fixes the latent enqueue-after-close silent drop (enqueue is now gated on Open), and guarantees idempotent wait() (one wake / one worker join). The poison failed latch stays a lock-free atomic.Bool (orthogonal, read per clause on the worker's hot path). Spec: formal-verification/tts_pipeline.fizz (always assertion WakeOnce + Monotonic), verified non-vacuous (deleting the if self.phase == 0 guard in Close fails WakeOnce). Wired (done): realtime_tts_pipeline.go's ttsPipeline embeds the coordinator (and is its effect sink — Wake → signal()); closed bool is gone; the worker checks closing() and raises WorkerExited on drain, enqueue rejects once not Open, wait() raises Close. The wake/done channel mechanics are unchanged. Verified: builds, vet/gofmt/golangci-lint clean, openai specs under -race, make test-realtime (12 e2e), full conformance gate green (5 Go packages + 5 fizz specs PASSED).
• All five mapped machines (M1–M5) are now explicit, wired, and verified. The realtime-conformance gate model-checks all .fizz specs and runs all five Go conformance suites under -race, fail-closed.
• The machines form a hierarchy, and that relationship is now modeled and enforced. M1 (connection) is the parent region; when it tears down, every child must be terminal. Previously this was only an imperative side effect of conncoord's teardown ordering, with a real gap (M4 compaction was fire-and-forget and could outlive the torn session). Now:
  • formal-verification/session_lifecycle.fizz is a composition spec that models conn + its direct children (vad/M2, resp/M3, compaction/M4) as one statechart and asserts ChildrenDieWithParent (conn torn ⟹ all children terminal) plus "no child starts after teardown". Its non-vacuity reproduces the exact M4 gap (drop the compaction-terminate line → assertion fails).
  • respcoord (M3) and compactcoord (M4) gained an absorbing Terminated state + a Shutdown event, so a response/compaction cannot start after teardown (structural "no resurrection").
  • conncoord's Teardown effect now explicitly drives the children terminal: stop+join the VAD goroutine (M2), respSink.shutdown() (M3 → Terminated, joins response goroutines and their M5 pipelines), and compaction.shutdown() for every conversation (M4: cancel the in-flight summary via a session-scoped context, then join — closing the gap). compact now takes a context so teardown can bound the join. M2's terminal is realized by the goroutine join and M5's by its existing Closed; the persistent coordinators (M3/M4) carry the explicit Terminated state.

Part 5 — Library vs hand-rolled (Go ecosystem, verified 2026-06)

Researched against live GitHub/pkg.go.dev data. Verdict: hand-roll a typed transition table over sealed sum-type states for the per-connection machines. No Go library gives the two properties we most want — compile-time-illegal states and a pure next(state,event)->(state,[]effect,error); every library models states as string/int/any and fires side-effecting callbacks mid-transition. And since the actor (Part 3.1) drives everything from one goroutine, the libraries' main value-add — internal locking — is dead weight.

Library landscape:

Option	Stars / status	Hierarchy	Typed states	Illegal-transition	Viz	Fit
hand-rolled table + sealed sum types	—	DIY (parent field / nested switch)	yes (sealed iface)	explicit default:	~30 LOC Mermaid emitter	best
qmuntal/stateless (port of .NET Stateless)	1.36k, v1.8.0 2026-02, maintained	yes (substates, guards, entry/exit, internal/ignored)	any	error + OnUnhandledTrigger + PermittedTriggers	DOT	best library fallback if hierarchy grows
looplab/fsm	3.4k, v1.0.3 2025-05, maintained	flat	strings	typed errors	DOT+Mermaid	only for flat machines wanting free diagrams
cocoonspace/fsm	89, dormant 2021	flat	int	bool no-op	—	lock-free but dead; DIY beats it
true Harel statecharts (gstate, statechartx)	≤10, <1yr, single-author	parallel+history	varies	varies	varies	only if we truly need parallel regions; unproven
Temporal / Cadence	large, maintained	n/a	n/a	n/a	n/a	overkill — external cluster+DB, durable replay, wrong latency class

Decision: hand-roll; keep qmuntal/stateless as the fallback if one machine grows deep hierarchy/guards faster than we want to hand-maintain (its error-on-illegal-trigger and PermittedTriggers() are the most useful library features for our "reject illegal transitions" requirement, at the cost of any-typed states). Add a tiny Mermaid emitter over the hand-rolled table so we keep the visualization the libraries advertise.

Part 6 — Formal design tied to code, and making it authoritative

The user requirement: the formal design is authoritative — a coding agent should be unable to silently change implementation behavior without it being caught against the spec; the default path is "update the spec and re-verify," not "edit the code and ignore the spec." This is a conformance + enforcement problem, in three layers.

6.1 The source of truth & design-time check

Write the concurrency-critical core — the M2↔M3 boundary (turn detection ↔ response: barge-in, ResponseCancel, speech_stopped, the dual-writer race) — as a FizzBee spec and model-check it in CI. Keep the spec small and focused on M2↔M3; that is where the real safety/liveness properties (Part 4 invariants 1–4) live. (FizzBee is the adopted model checker — see Part 6.4.)

6.2 The conformance bridge (code ↔ spec)

The honest finding: design-time model checking is well-supported; the Go conformance bridge is thin everywhere and needs per-spec glue. Two layers, adopted together:

1. FizzBee MBT — the authoritative layer. The .fizz spec is model-checked, and fizz mbt-scaffold --lang go generates Go interfaces + a go test harness; you implement adapters mapping model actions→code and StateGetter→state. Conformance runs as plain go test — the cleanest CI fit. Risk: pre-1.0, essentially one maintainer (pin a version + sha256, vendor examples).
2. Ginkgo/Gomega seeded property tests — the Go-native floor. A small Go model (the test's open/registered shadow) is the oracle; a fixed-seed random walk drives random event sequences against the Coordinator, asserting the Part-4 invariants after each step / per seed. It checks the implementation against a Go oracle — it complements, but does not replace, the FizzBee check of the design. (We originally specced pgregory.net/rapid here for its (*T).Repeat driver and automatic shrinking, but LocalAI mandates Ginkgo/Gomega for all tests — its forbidigo lint forbids stdlib testing assertions — and rapid.Check needs a concrete *testing.T/*rapid.T that cannot run inside a Ginkgo It. Rather than weaken the lint gate with an exclusion, the property layer is hand-rolled seeded walks: fixed seeds make every failure reproducible, at the cost of rapid's automatic shrinking. rapid is consequently not a direct dependency.)

These compose: model-check the design (6.1) for "the design is right"; conformance-test the code (6.2) for "the code matches the design." Add go test -race (with -cpu=1,2,4, repeated runs) over the stateful tests for interleaving-bug discovery, and Go native fuzzing over the same harness for coverage-guided sequence exploration + a committable regression corpus. (testing/quick is frozen — do not use.)

There is no viable single-source-of-truth codegen (one spec compiled into both the runtime Go and the model) for retrofitting existing Go — the candidates are research-grade and greenfield-only. Our practical substitute is the CI gate below plus a single Go transition table that emits both the diagram and the test action set.

6.3 Enforcement — making the design un-ignorable for agents

Structural enforcement, leveraging this repo's existing non-bypassable gate culture (pre-commit + monotonic ratchets; --no-verify is forbidden, baselines never lowered):

1. Add a realtime-conformance gate to the pre-commit/CI pipeline that runs (a) the model check (6.1) and (b) the conformance bridge (6.2). A behavior change that does not conform turns the gate red; the only green paths are make the code conform or update the spec — and updating the spec re-triggers the model check, so an illegal design is rejected too. This is the actual mechanism that makes "update the design and verify" the default rather than optional.
2. Treat the spec as a ratchet artifact like coverage: the gate must not be weakened, the spec not deleted, the build tag not silently disabled.
3. Write an .agents/realtime-state-machines.md guide (indexed from CLAUDE.md) stating the spec is the source of truth: change the spec first, re-run the gate, then implement. The doc is secondary; the gate is what enforces it.

6.4 Decided stack

• Implementation: hand-rolled sealed-state transition functions + single-writer actor (Parts 3.1–3.2).
• Design-time + conformance: FizzBee (decided). .fizz spec is model-checked, and fizz's Go MBT generator (mbt/generator/templates/go → interfaces/adapters/test; driven via a gRPC plugin in mbt/lib/go) produces a go test conformance harness whose adapters map model actions → our actor and StateGetter → our state. Go is a first-class MBT target (Go + Rust are the only two). Verified 2026-06: Apache-2.0, v0.5.2, prebuilt linux/macos×x86/arm binaries, ships Claude Code skills (/fizz-spec|check|debug|mbt) for the spec-authoring loop.
• Go-native layer: Ginkgo/Gomega seeded property tests run alongside — they check the implementation, complementing (not substituting for) the FizzBee check of the design. Skipping FizzBee is NOT "degrading to the Go layer": the design authority would be gone. The gate is therefore fail-closed (see Enforcement). (Originally specced as rapid; switched to Ginkgo/Gomega to satisfy LocalAI's Ginkgo-only forbidigo lint without weakening that gate — see Part 6.2.)
• Enforcement: the realtime-conformance pre-commit/CI gate + .agents/ guide (Part 6.3).

FizzBee risk mitigations (decided):

• The gate is fail-closed: a missing FizzBee is a hard failure, never a silent skip. The only bypass is the explicit, loud REALTIME_CONFORMANCE_SKIP_FIZZBEE=1 (local only; CI never sets it; pre-commit runs the gate on any respcoord/** or formal-verification/** change so a pure .fizz edit still re-verifies).
• CI pins the FizzBee release binary by version + sha256 (formal-verification/fizzbee.sha256, all four platforms, digests from the GitHub release; installer verifies before extract, CI caches it). Not go-gettable: pkg/modelchecker imports the Bazel-internal fizz/proto with no committed .pb.go, so a plain go get won't build — hence the pinned binary.
• Keep the .fizz model portable (no exotic features) so it stays re-expressible in another model checker if FizzBee is ever abandoned — lock-in is at the tooling layer only, not the design.

Open questions (decide before implementing)

• Scope of the actor refactor: full single-writer per session, or incrementally migrate one machine at a time behind the existing locks? (Suggest: M3 response coordination first — it has the load-bearing dual-writer bug.)

Resolved: FSM library vs hand-rolled → hand-rolled sealed-state tables, qmuntal/stateless fallback (Part 5). Conformance bridge → FizzBee (model-check + Go MBT) with a Ginkgo/Gomega seeded-property Go-native floor as hedge (Part 6.4). Single-source-of-truth codegen (PGo/MPCal) → not viable (research-grade, greenfield-only); substitute is the CI conformance gate (Part 6.3).

Agentic turn semantics → invariant #2 is one response.done per response.create (OpenAI-faithful); the server-side AssistantExecutor tool loop becomes internal sub-states of a single response rather than emitting one terminal per turn. Verified safe in-tree: the current response.done carries only {id, object, status} (Output/Usage never populated), the React UI (Talk.jsx:330) reads only status, every unit test already asserts ResponseDone == 1 for tool turns, no test expects multiplicity, and the server-side recursion is untested. Collapsing also fixes a latent "Listening…" flicker mid-agentic-loop. The client-driven tool loop (fresh response.create per round-trip) legitimately keeps one terminal each — unaffected. Follow-up: actually populate Output + Usage in the single terminal (currently always empty).