fix: inject compacted summary back into message_history and sharpen compaction prompt

After compaction, message_history.clear() wiped all context and the stored
summary in COMPACTED_SUMMARIES was never re-injected, so the next runner turn
started completely blank and the agent would repeat already-tried approaches.

Fix: immediately after clearing, push a user+assistant exchange containing the
summary into message_history so it flows through history_context on the next
iteration as normal conversation context.

Also: add an explicit 'Exhausted Approaches — DO NOT RETRY' section (§9) and
'Recommended Next Steps' section (§10) to the compaction prompt so the support
model produces a checklist of dead ends the main agent must not revisit.

src/cai/repl/commands/memory.py

@@ -1221,7 +1221,8 @@ After the analysis, provide a structured summary with these sections:
 6. **All User Messages**: Complete list of user messages in order
 7. **Pending Tasks**: What still needs to be done
 8. **Current Work**: What was being worked on when the conversation ended
-9. **Optional Next Step**: If there's a clear next action, mention it
+9. **Exhausted Approaches — DO NOT RETRY**: Every technique, command, path, or attack vector that was attempted and failed. Format each as a bullet starting with ❌. Be specific (include exact commands, URLs, usernames, ports). This section is CRITICAL — the agent will use it to avoid wasting time on dead ends.
+10. **Recommended Next Steps**: Concrete actions NOT yet tried, ordered by likelihood of success.
 
 ## Important Guidelines
 
@@ -1232,6 +1233,7 @@ After the analysis, provide a structured summary with these sections:
 - Maintain technical accuracy - don't paraphrase technical terms
 - The summary will be used as the primary context for resuming work, so completeness is crucial
 - When the conversation is resumed, it should feel like a natural continuation
+- Section 9 (Exhausted Approaches) is the most important section for offensive/hacking tasks: list every failed attempt so the agent doesn't loop.
 
 This session is being continued from a previous conversation that ran out of context. The conversation is summarized below:"""
         

src/cai/sdk/agents/models/openai_chatcompletions.py

@@ -3525,6 +3525,28 @@ class OpenAIChatCompletionsModel(Model):
                                     APPLIED_MEMORY_IDS[self.agent_name] = []
                                 COMPACTED_SUMMARIES[self.agent_name] = [_summary]
                                 self.message_history.clear()
+                                # Re-inject the summary as the first exchange so
+                                # the next Runner turn has full context and won't
+                                # repeat work that was already attempted.
+                                self.message_history.append({
+                                    "role": "user",
+                                    "content": (
+                                        "<previous_session_memory>\n"
+                                        + _summary
+                                        + "\n</previous_session_memory>\n\n"
+                                        "This is your memory from the previous context window. "
+                                        "Use it to continue your work. "
+                                        "Do NOT retry any approach already marked as failed or exhausted."
+                                    ),
+                                })
+                                self.message_history.append({
+                                    "role": "assistant",
+                                    "content": (
+                                        "Understood. I have reviewed my previous session memory. "
+                                        "I will continue the task using only new approaches "
+                                        "and will not repeat anything already attempted."
+                                    ),
+                                })
                                 os.environ["CAI_CONTEXT_USAGE"] = "0.0"
                                 _console.print(
                                     "[bold green]✓ Memory summary applied — "