From a2a3a2d336d46a13663667ffa8d56bbfc85b3c6b Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 19 Jun 2026 22:10:34 +0000
Subject: [PATCH] cerebellum: make the LLM seam async; add OpenHermesClient
 skeleton

A real model call is network/process IO, so the LLM seam is now a behaviour
with a response callback (ResponseFn), not a synchronous function. StubLLM
answers immediately; OpenHermesClient + OpenHermesConfig (OpenAI-compatible
url/model) are the real plug, swappable for the stub. CerebellumHarness chains
its passes through the async callback. Compiles on ponyc 0.64; demo green.
Next: wire the actual chat/completions call behind OpenHermesClient.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
---
 src/ichor/cerebellum.pony | 77 ++++++++++++++++++++++++++++-----------
 1 file changed, 55 insertions(+), 22 deletions(-)

diff --git a/src/ichor/cerebellum.pony b/src/ichor/cerebellum.pony
index 7f54520..860110a 100644
--- a/src/ichor/cerebellum.pony
+++ b/src/ichor/cerebellum.pony
@@ -4,19 +4,51 @@
 // the synthesis back through the bus (Brain-bound, so it crosses D1). This is
 // the C1 integration spine ("plumbing + sequencing, not an organ").
 //
-// Plug-and-play: the model is the `LLM` seam below. StubLLM stands in now; drop
-// a real OpenHermes client (HTTP / subprocess) in its place and nothing else
-// on the bus changes.
+// Plug-and-play: the model is the `LLM` seam. A real model call is network /
+// process IO, so the seam is ASYNCHRONOUS -- `infer` is a behaviour that hands
+// its result to a callback. StubLLM answers immediately; OpenHermesClient is the
+// real plug. Swap which one you register; nothing else on the bus changes.
 
-// --- the model seam -------------------------------------------------------
-interface val LLM
-  """One inference call. Implement this to plug a model into the cerebellum."""
-  fun infer(prompt: String): String
+// --- the model seam (async) ----------------------------------------------
+type ResponseFn is {(String)} val
+  """Callback the model invokes with its output."""
 
-class val StubLLM is LLM
-  """Stand-in until a real OpenHermes client is wired."""
-  fun infer(prompt: String): String =>
-    "[stub-openhermes] " + prompt
+interface tag LLM
+  be infer(prompt: String, respond: ResponseFn)
+
+// --- stub -----------------------------------------------------------------
+actor StubLLM is LLM
+  new create() => None
+
+  be infer(prompt: String, respond: ResponseFn) =>
+    respond("[stub-openhermes] " + prompt)
+
+// --- the OpenHermes plug --------------------------------------------------
+class val OpenHermesConfig
+  """Where to reach an OpenAI-compatible OpenHermes server."""
+  let url: String       // base, e.g. http://localhost:8080/v1
+  let model: String
+
+  new val create(
+    url': String = "http://localhost:8080/v1",
+    model': String = "openhermes")
+  =>
+    url = url'
+    model = model'
+
+actor OpenHermesClient is LLM
+  let _cfg: OpenHermesConfig
+
+  new create(cfg: OpenHermesConfig) =>
+    _cfg = cfg
+
+  be infer(prompt: String, respond: ResponseFn) =>
+    // TODO(loop next): POST an OpenAI-compatible chat/completions request to
+    // `_cfg.url`/chat/completions via the process seam (curl) or a Pony TCP
+    // client, parse choices[0].message.content, then call respond() with it.
+    // Until that IO is wired and testable in this env, surface intent rather
+    // than perform an untested network call.
+    respond("[openhermes " + _cfg.model + " @ " + _cfg.url + "] " + prompt)
 
 // --- the harness ----------------------------------------------------------
 actor CerebellumHarness is OrganReceiver
@@ -29,18 +61,19 @@ actor CerebellumHarness is OrganReceiver
     _out = out'
     _bus = bus
     _llm = llm
-    // C1 L3: pass count is the responsiveness knob (energy-gated later), the
-    // 4+4 ordering itself is static. Clamp to >=1 so a turn always runs once.
+    // C1 L3: pass count is the responsiveness knob (energy-gated later); the
+    // 4+4 ordering is static. Clamp to >=1 so a turn always runs once.
     _passes = if passes' < 1 then 1 else passes' end
 
   be receive(envl: Envelope) =>
-    // A turn arrives off the bus. Sequence the passes through the model, then
-    // emit the synthesis back onto the bus toward the Brain.
-    var ctx: String = envl.payload
-    var pass: USize = 1
-    while pass <= _passes do
-      ctx = _llm.infer(ctx)
-      pass = pass + 1
+    // A turn arrives off the bus: sequence the passes, then emit.
+    _pass(envl.payload, 1)
+
+  be _pass(ctx: String, n: USize) =>
+    if n > _passes then
+      _out.print("[cerebellum] " + _passes.string() + " passes -> " + ctx)
+      _bus.route(Envelope(Cerebellum, Brain, OrganSecretion, ctx))
+    else
+      let self: CerebellumHarness tag = this
+      _llm.infer(ctx, {(out: String)(self, n) => self._pass(out, n + 1) } val)
     end
-    _out.print("[cerebellum] " + _passes.string() + " passes -> " + ctx)
-    _bus.route(Envelope(Cerebellum, Brain, OrganSecretion, ctx))