@misterhuydo/sentinel 1.1.6 → 1.1.7

package/.cairn/session.json

@@ -1,6 +1,6 @@
 {
-  "message": "Auto-checkpoint at 2026-03-23T08:59:04.141Z",
-  "checkpoint_at": "2026-03-23T08:59:04.142Z",
+  "message": "Auto-checkpoint at 2026-03-23T09:08:47.973Z",
+  "checkpoint_at": "2026-03-23T09:08:47.974Z",
   "active_files": [],
   "notes": [],
   "mtime_snapshot": {}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@misterhuydo/sentinel",
-  "version": "1.1.6",
+  "version": "1.1.7",
   "description": "Sentinel — Autonomous DevOps Agent installer and manager",
   "bin": {
     "sentinel": "./bin/sentinel.js"

package/python/sentinel/sentinel_boss.py

@@ -181,6 +181,12 @@ When to act vs. when to ask:
 - If a tool call will take a moment (search, fetch, pull), prefix your reply with a brief "working" line ending in "..." before the results, e.g. "Searching SSOLWA for TryDig activity..." then the actual output.
   Never just say a working line and stop — always follow it with the results in the same message.
 
+Session context — critical rules:
+- Loaded conversation history is prior-session background only. It may be hours or days old.
+- NEVER say "the previous search", "I already fetched", "as I found earlier", or any phrase implying you already did part of the current task — unless a tool result appears in THIS response's tool calls.
+- When handling a new request, call the tools fresh. Do not assume any prior tool result is still current or that any prior step "counts" toward the current task.
+- The only exception: if the user explicitly asks about something from the history ("what did you find earlier?"), you may reference it — but note it is from a prior session.
+
 Issue identification — before calling create_issue:
 1. Determine if the message is a REAL issue/task (bug report, feature request, investigation ask)
    vs. a status question, tool query, or casual chat. If not an issue, just answer normally.

package/python/sentinel/slack_bot.py

@@ -371,12 +371,32 @@ async def _run_turn(session: _Session, message: str, client, cfg_loader, store,
     channel = session.channel
 
     # Load persisted history from DB on the first turn of a new session.
-    # Clean it to strip any orphaned tool_use turns from a previous crashed session.
+    # - _clean_history strips orphaned tool_use turns from a previous crashed session.
+    # - Trim to last 6 exchanges (12 messages) to prevent stale tool results from bleeding
+    #   into the current session and causing the model to treat old work as already done.
+    # - Inject a session boundary marker so the model clearly sees where prior context ends.
     if not session.history_loaded:
-        session.history = _clean_history(store.load_conversation(session.user_id))
+        loaded = _clean_history(store.load_conversation(session.user_id))
+        if loaded:
+            # Keep only the most recent 6 exchanges from prior session
+            _PRIOR_TURNS = 6
+            trimmed = loaded[-(_PRIOR_TURNS * 2):]
+            # Prepend a boundary pair so the model treats everything before it as old context
+            session.history = [
+                {
+                    "role": "user",
+                    "content": "[system: new session started — the following is prior conversation context only]",
+                },
+                {
+                    "role": "assistant",
+                    "content": [{"type": "text", "text": "Understood. I'll treat the prior context as reference only and handle your new request fresh."}],
+                },
+            ] + trimmed
+        else:
+            session.history = []
         session.history_loaded = True
 
-    # Trim history to avoid context overflow on long conversations
+    # Trim history to avoid context overflow on long active conversations
     if len(session.history) > _MAX_HISTORY_TURNS * 2:
         session.history = session.history[-(_MAX_HISTORY_TURNS * 2):]