fix(agent/run): silent no-response + provider visibility (#187)

* fix(agent/run): silent no-response + provider visibility — closes #186-followup

Root causes of specsmith run returning nothing:
1. DEFAULT_OLLAMA_MODEL=qwen2.5:7b not installed; Ollama 404 silently swallowed
2. run_chat returning None gave user zero feedback
3. EventEmitter always wrote JSONL even in interactive terminal mode

Fixes:
- chat_runner: _pick_ollama_model() queries /api/tags and selects the first
  installed model from a preference list (lighter models first); falls back
  to DEFAULT_OLLAMA_MODEL only when the API is unreachable or list is empty.
  SPECSMITH_OLLAMA_MODEL env var still wins unconditionally.
- runner: _handle_command() prints an actionable hint when run_chat returns
  None — explains whether Ollama is running (with model name) or no provider
  is available at all.
- events: PlainTextEmitter subclass — token() writes raw text to stdout,
  emit() is a no-op. Used by AgentRunner when json_events=False so LLM
  responses render as readable prose instead of JSONL blobs.
- runner: check_providers() probes all four providers (Ollama, Anthropic,
  OpenAI, Gemini) and returns ProviderStatus(name, available, model, note).
- runner: _print_banner() shows a provider status table in interactive mode
  so the user knows upfront which model will respond before typing.
- cli: specsmith run --check validates providers and exits 0/1 without
  starting the REPL.
- run_interactive: adds a trailing newline after each streamed response so
  the next prompt doesn't bleed onto the last output line.

19 new tests in test_agent_run_feedback.py (854 total, 0 failures).

Co-Authored-By: Oz <oz-agent@warp.dev>

* style: ruff format cli.py

Co-Authored-By: Oz <oz-agent@warp.dev>

* style: fix import sort in test_agent_run_feedback.py

Co-Authored-By: Oz <oz-agent@warp.dev>

---------

Co-authored-by: Oz <oz-agent@warp.dev>

Author: tbitcs

src/specsmith/agent/chat_runner.py

@@ -34,6 +34,24 @@
 
 DEFAULT_OLLAMA_HOST = "http://127.0.0.1:11434"
 DEFAULT_OLLAMA_MODEL = os.environ.get("SPECSMITH_OLLAMA_MODEL", "qwen2.5:7b")
+
+# Ordered preference list used by _pick_ollama_model() when the default
+# model is not installed.  Lighter / faster models first so the REPL
+# stays responsive on developer hardware.
+_OLLAMA_MODEL_PREFERENCE = [
+    "qwen2.5:7b",
+    "qwen2.5-coder:7b-instruct",
+    "qwen3:8b",
+    "mistral:7b-instruct-q4_0",
+    "llama3:8b-instruct-q4_K_M",
+    "qwen2.5:14b",
+    "qwen2.5-coder:14b",
+    "qwen3:14b",
+    "mistral-nemo:12b",
+    "phi4:14b-q4_K_M",
+    "deepseek-r1:14b",
+    "qwen3:30b-a3b",
+]
 SYSTEM_PROMPT = (
     "You are Nexus, the local-first agentic developer assistant inside "
     "Specsmith. Always end your response with the canonical contract:\n"
@@ -239,19 +257,54 @@ def run_single_prompt(prompt: str, *, max_tokens: int = 500) -> str | None:  # n
 # ---------------------------------------------------------------------------
 
 
+def _ollama_alive(host: str) -> bool:
+    try:
+        with urlopen(f"{host}/api/tags", timeout=2):  # noqa: S310
+            return True
+    except (URLError, TimeoutError, OSError):
+        return False
+
+
+def _pick_ollama_model(host: str) -> str:
+    """Return the best available Ollama model for this machine.
+
+    Checks the installed model list from ``/api/tags`` and walks
+    ``_OLLAMA_MODEL_PREFERENCE`` to find the first match.  Falls back
+    to ``DEFAULT_OLLAMA_MODEL`` when the API is unreachable or the list
+    is empty.  An explicit ``SPECSMITH_OLLAMA_MODEL`` env var always wins.
+    """
+    env_override = os.environ.get("SPECSMITH_OLLAMA_MODEL", "").strip()
+    if env_override:
+        return env_override
+    try:
+        with urlopen(f"{host}/api/tags", timeout=2) as resp:  # noqa: S310
+            data = json.loads(resp.read())
+        installed = {m["name"] for m in data.get("models", []) if m.get("name")}
+        for candidate in _OLLAMA_MODEL_PREFERENCE:
+            if candidate in installed:
+                return candidate
+        # None of the preferred models found — use whatever is first
+        if installed:
+            return sorted(installed)[0]
+    except Exception:  # noqa: BLE001 — best-effort; fall through to default
+        pass
+    return DEFAULT_OLLAMA_MODEL
+
+
 def _run_ollama(
     messages: list[dict[str, str]],
     emitter: EventEmitter,
     block_id: str,
 ) -> tuple[str | None, _UsageDelta]:
     """Stream from a local Ollama daemon using only stdlib."""
     host = os.environ.get("OLLAMA_HOST", DEFAULT_OLLAMA_HOST).rstrip("/")
-    model = os.environ.get("SPECSMITH_OLLAMA_MODEL", DEFAULT_OLLAMA_MODEL)
     usage = _UsageDelta()
 
     if not _ollama_alive(host):
         return None, usage
 
+    model = _pick_ollama_model(host)
+
     payload = json.dumps({"model": model, "messages": messages, "stream": True}).encode("utf-8")
     req = Request(  # noqa: S310 - URL is a hardcoded localhost default
         f"{host}/api/chat",
@@ -284,14 +337,6 @@ def _run_ollama(
     return ("".join(pieces) if pieces else None), usage
 
 
-def _ollama_alive(host: str) -> bool:
-    try:
-        with urlopen(f"{host}/api/tags", timeout=2):  # noqa: S310
-            return True
-    except (URLError, TimeoutError, OSError):
-        return False
-
-
 def _run_anthropic(
     messages: list[dict[str, str]],
     emitter: EventEmitter,

src/specsmith/agent/events.py

@@ -231,4 +231,24 @@ def task_complete(
         )
 
 
-__all__ = ["EventEmitter"]
+class PlainTextEmitter(EventEmitter):
+    """Human-readable variant of EventEmitter for interactive terminal sessions.
+
+    Writes LLM token chunks directly to the stream (no JSON wrapping) so
+    ``specsmith run`` without ``--json-events`` produces readable output.
+    All non-token protocol events are silently dropped — system messages
+    are handled separately by ``AgentRunner._default_emit_event``.
+    """
+
+    def token(self, block_id: str, text: str) -> None:  # noqa: ARG002
+        self.stream.write(text)
+        with contextlib.suppress(Exception):
+            self.stream.flush()
+
+    def emit(self, event: dict[str, Any]) -> None:
+        # Drop JSONL protocol frames in plain-text mode; token() handles
+        # the only output that matters for interactive sessions.
+        pass
+
+
+__all__ = ["EventEmitter", "PlainTextEmitter"]