fix: Ollama keep_alive + context continuity rule

- ollama.py: keep_alive=-1 on all /api/chat calls. Without this Ollama
  unloads the model after 5 min of inactivity, forcing a full reload and
  re-prefill that causes apparent context loss between slow turns.
- runner.py: CONTINUITY RULE added to system prompt. Prevents the agent
  from responding with 'I'm not sure what you're referring to' when the
  user sends a follow-up like 'fix it' or 'fix the issue' after the agent
  just described a finding. Instructs the agent to look back at the
  conversation history and act immediately.

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

Author: tbitcs

src/specsmith/agent/providers/ollama.py

@@ -129,6 +129,10 @@ def _complete_native(self, messages: list[Message], max_tokens: int) -> Completi
             "model": self.model,
             "messages": [m.to_dict() for m in messages],
             "stream": False,
+            # keep_alive=-1: keep the model loaded indefinitely between turns.
+            # Without this Ollama unloads after 5 min of inactivity, forcing a
+            # full reload + re-prefill which causes apparent context loss.
+            "keep_alive": -1,
             "options": {"num_predict": max_tokens, "num_ctx": self._num_ctx},
         }
         if self._think is not None:
@@ -161,6 +165,7 @@ def _complete_with_tools(
             "messages": [m.to_dict() for m in messages],
             "stream": False,
             "tools": [t.to_openai_schema() for t in tools],
+            "keep_alive": -1,
             "options": {"num_predict": max_tokens, "num_ctx": self._num_ctx},
         }
         data = self._post("/api/chat", payload)
@@ -208,6 +213,7 @@ def stream(
             "model": self.model,
             "messages": [m.to_dict() for m in messages],
             "stream": True,
+            "keep_alive": -1,
             "options": {"num_predict": max_tokens, "num_ctx": self._num_ctx},
         }
         if tools:

src/specsmith/agent/runner.py

@@ -223,6 +223,18 @@ def build_system_prompt(
 Example good: "Audit found 3 issues: LEDGER.md is missing and 2 requirements lack tests."
 Example bad: "The tool returned: [\u2717] LEDGER.md MISSING, [\u2717] REQ-001 uncovered..."
 
+## CONTINUITY RULE — CRITICAL (never break context):
+When the user sends a follow-up like "fix it", "do it", "implement that",
+"fix the issue", "yes", "continue", or any message that refers to something
+you described in a PREVIOUS message in this conversation:
+  1. LOOK BACK at your own previous messages to find what "it" refers to.
+  2. ACT on it immediately — propose the fix or make the change.
+  3. NEVER say "I'm not sure what you're referring to" if you just described
+     a finding or issue. That response is a CRITICAL FAILURE.
+  4. If genuinely ambiguous across multiple findings, quote the most recent
+     one: "In my last response I found X — fixing that now."
+You have the full conversation history. Use it.
+
 You are an AEE-integrated specsmith agent for this project.
 
 ## Project Governance