Merge pull request #53 from mattmezza/feat/reply-decision

feat: reply-decision gate to stop infinite loops in group chats (#36)

Author: mattmezza

README.md

@@ -9,6 +9,7 @@ A self-hosted personal AI agent that runs in a single Docker container. MPA acts
 - **Calendar** — CalDAV integration (Google Calendar, iCloud, etc.)
 - **Contacts** — CardDAV providers via the built-in contacts CLI
 - **Personae** — Swappable agent identities (own character, skill/tool scope, voice). Bind one per chat — and, on Telegram, per forum topic — so several run concurrently, each with its own isolated context. Give a persona its own bot token and it becomes a separate Telegram contact (bot-per-persona)
+- **Reply decision** — In shared/group chats the agent decides per message whether to reply at all, staying quiet for messages aimed at another bot or caught in a bot-to-bot reaction loop, with a hard rate cap that guarantees runaway loops end (off by default)
 - **Memory** — Two-tier system: permanent long-term facts and expiring short-term context, both extracted automatically from conversations
 - **Scheduled tasks** — Cron-based jobs for morning briefings, email checks, contact sync, and custom tasks
 - **Subagents** — Delegate a scoped subtask to a sub-loop under a chosen persona, on demand or scheduled. Runs sync (result returned in-turn) or in the background; a finished background batch is distilled by a summary inference into a one-line chat notification + a concise context digest (raw output never reaches the user or the agent's context). The agent sizes each run (steps / token budget / thinking effort) and defaults the persona to its own; scope is a subset of the caller's (inherit-never-widen). Monitor and cancel from Telegram (`/jobs`) or the admin UI

api/admin.py

@@ -1058,6 +1058,15 @@ async def partial_llm() -> HTMLResponse:
         tr_model = await config_store.get("task_reflection.model") or "deepseek-v4-flash"
         gd_thinking_level = await config_store.get("goal_decomposition.thinking_level") or ""
         tr_thinking_level = await config_store.get("task_reflection.thinking_level") or ""
+        rd_enabled = await config_store.get("reply_decision.enabled")
+        rd_enabled = rd_enabled if rd_enabled is not None else "false"
+        rd_provider = await config_store.get("reply_decision.provider") or "deepseek"
+        rd_model = await config_store.get("reply_decision.model") or "deepseek-v4-flash"
+        rd_thinking_level = await config_store.get("reply_decision.thinking_level") or ""
+        rd_group_only = await config_store.get("reply_decision.group_only")
+        rd_group_only = rd_group_only if rd_group_only is not None else "true"
+        rd_max_replies = await config_store.get("reply_decision.max_replies_per_window") or "6"
+        rd_window_seconds = await config_store.get("reply_decision.window_seconds") or "120"
         compaction_provider = await config_store.get("compaction.provider") or "deepseek"
         compaction_model = await config_store.get("compaction.model") or "deepseek-v4-flash"
         compaction_thinking_level = await config_store.get("compaction.thinking_level") or ""
@@ -1104,6 +1113,13 @@ async def partial_llm() -> HTMLResponse:
             tr_provider=tr_provider,
             tr_model=tr_model,
             tr_thinking_level=tr_thinking_level,
+            rd_enabled=rd_enabled,
+            rd_provider=rd_provider,
+            rd_model=rd_model,
+            rd_thinking_level=rd_thinking_level,
+            rd_group_only=rd_group_only,
+            rd_max_replies=rd_max_replies,
+            rd_window_seconds=rd_window_seconds,
             compaction_provider=compaction_provider,
             compaction_model=compaction_model,
             compaction_thinking_level=compaction_thinking_level,

api/templates/base.html

@@ -82,7 +82,7 @@
     }
     window.restartAgentNow = restartAgentNow;
 
-    function llmTab(provider, apiKey, model, openaiKey, openaiBaseUrl, googleKey, googleBaseUrl, grokKey, grokBaseUrl, deepseekKey, deepseekBaseUrl, extractionProvider, extractionModel, consolidationProvider, consolidationModel, gdEnabled, gdProvider, gdModel, trEnabled, trProvider, trModel, promptToolUsageOverride, promptHistoryOverride, defaultToolUsage, defaultHistoryHandling, promptCaptureEnabled, compactionProvider, compactionModel, thinkingLevel, extractionThinkingLevel, consolidationThinkingLevel, gdThinkingLevel, trThinkingLevel, compactionThinkingLevel, visionEnabled, visionProvider, visionModel) {
+    function llmTab(provider, apiKey, model, openaiKey, openaiBaseUrl, googleKey, googleBaseUrl, grokKey, grokBaseUrl, deepseekKey, deepseekBaseUrl, extractionProvider, extractionModel, consolidationProvider, consolidationModel, gdEnabled, gdProvider, gdModel, trEnabled, trProvider, trModel, promptToolUsageOverride, promptHistoryOverride, defaultToolUsage, defaultHistoryHandling, promptCaptureEnabled, compactionProvider, compactionModel, thinkingLevel, extractionThinkingLevel, consolidationThinkingLevel, gdThinkingLevel, trThinkingLevel, compactionThinkingLevel, visionEnabled, visionProvider, visionModel, rdEnabled, rdProvider, rdModel, rdThinkingLevel, rdGroupOnly, rdMaxReplies, rdWindowSeconds) {
       const currentProvider = provider || 'anthropic';
       return {
         providerOptions: [
@@ -146,6 +146,15 @@
         gdResultOk: false,
         trResult: '',
         trResultOk: false,
+        rdEnabled: rdEnabled === 'true' || rdEnabled === true,
+        rdProvider: rdProvider || 'deepseek',
+        rdModel: rdModel || 'deepseek-v4-flash',
+        rdThinkingLevel: rdThinkingLevel || '',
+        rdGroupOnly: rdGroupOnly === 'true' || rdGroupOnly === true,
+        rdMaxReplies: rdMaxReplies || '6',
+        rdWindowSeconds: rdWindowSeconds || '120',
+        rdResult: '',
+        rdResultOk: false,
         compactionProvider: compactionProvider || 'anthropic',
         compactionModel: compactionModel || 'claude-haiku-4-5',
         compactionResult: '',
@@ -403,6 +412,12 @@
             if (list.length) { this.gdModel = list[0].value; }
             return;
           }
+          if (kind === 'rd') {
+            if (this.rdModel) return;
+            const list = this.modelOptions(this.rdProvider);
+            if (list.length) { this.rdModel = list[0].value; }
+            return;
+          }
           if (this.trModel) return;
           const list = this.modelOptions(this.trProvider);
           if (list.length) { this.trModel = list[0].value; }

api/templates/partials/llm.html

@@ -63,7 +63,14 @@
   {{ compaction_thinking_level|default('', true)|tojson|forceescape }},
   {{ vision_enabled|default('false', true)|tojson|forceescape }},
   {{ vision_provider|default('anthropic', true)|tojson|forceescape }},
-  {{ vision_model|default('claude-haiku-4-5', true)|tojson|forceescape }}
+  {{ vision_model|default('claude-haiku-4-5', true)|tojson|forceescape }},
+  {{ rd_enabled|default('false', true)|tojson|forceescape }},
+  {{ rd_provider|default('deepseek', true)|tojson|forceescape }},
+  {{ rd_model|default('deepseek-v4-flash', true)|tojson|forceescape }},
+  {{ rd_thinking_level|default('', true)|tojson|forceescape }},
+  {{ rd_group_only|default('true', true)|tojson|forceescape }},
+  {{ rd_max_replies|default('6', true)|tojson|forceescape }},
+  {{ rd_window_seconds|default('120', true)|tojson|forceescape }}
 )">
   <div class="card">
     <h2 class="text-base mb-1">System Prompt Controls</h2>
@@ -480,6 +487,89 @@ <h2 class="text-base mb-1">Task Reflection</h2>
     </template>
   </div>
 
+  <div class="card">
+    <h2 class="text-base mb-1">Reply Decision</h2>
+    <p class="text-muted text-xs mb-4">In shared/group chats with several bots and people, decide per message whether to reply at all — staying quiet for messages aimed at another bot, caught in a bot-to-bot reaction loop, or that the agent can't usefully add to. Off by default; uses a cheap/fast model for the yes/no call.</p>
+
+    <form class="space-y-4" @submit.prevent>
+      <div class="flex items-center gap-2">
+        <label class="label mb-0">Enabled</label>
+        <input type="checkbox" x-model="rdEnabled" class="rounded border-border">
+      </div>
+
+      <div>
+        <label class="label">Provider</label>
+        <select class="input-sm" style="max-width:500px" x-model="rdProvider" @change="resetBackgroundModel('rd')" x-effect="$el.value = rdProvider" :disabled="!rdEnabled">
+          <template x-for="item in providerOptions" :key="'rd-' + item.value">
+            <option :value="item.value" :selected="item.value === rdProvider" x-text="item.label"></option>
+          </template>
+        </select>
+      </div>
+      <div>
+        <label class="label">Model</label>
+        <input type="text" class="input-sm" style="max-width:500px" x-model="rdModel" :disabled="!rdEnabled"
+               list="dl-model-rd" placeholder="Type or pick a model">
+        <datalist id="dl-model-rd">
+          <template x-for="item in modelOptions(rdProvider)" :key="'rd-m-' + item.value">
+            <option :value="item.value"></option>
+          </template>
+        </datalist>
+        <p class="text-muted text-xs mt-1">Model used to decide whether a group-chat message warrants a reply.</p>
+      </div>
+{{ think('rdThinkingLevel', 'rdProvider', 'rdModel', 'dl-think-rd', true) }}
+
+      <div class="flex items-center gap-2">
+        <label class="label mb-0">Group chats only</label>
+        <input type="checkbox" x-model="rdGroupOnly" class="rounded border-border" :disabled="!rdEnabled">
+        <span class="text-muted text-xs">When on, 1:1 chats always get a reply; only group chats are gated.</span>
+      </div>
+      <div class="flex flex-wrap gap-4">
+        <div>
+          <label class="label">Max replies per window</label>
+          <input type="number" min="1" class="input-sm" style="max-width:160px" x-model="rdMaxReplies" :disabled="!rdEnabled">
+        </div>
+        <div>
+          <label class="label">Window (seconds)</label>
+          <input type="number" min="1" class="input-sm" style="max-width:160px" x-model="rdWindowSeconds" :disabled="!rdEnabled">
+        </div>
+      </div>
+      <p class="text-muted text-xs">Hard backstop: never send more than this many auto-replies into one chat per rolling window — guarantees a runaway loop terminates even if the model keeps voting to reply.</p>
+    </form>
+
+    <div class="flex items-center gap-2 mt-4">
+      <button class="btn-primary btn-sm"
+              @click="
+                fetch('/config', {
+                  method: 'PATCH',
+                  headers: {
+                    'Content-Type': 'application/json',
+                    'Authorization': 'Bearer ' + (localStorage.getItem('admin_api_key') || '')
+                  },
+                  body: JSON.stringify({values: {
+                    'reply_decision.enabled': rdEnabled ? 'true' : 'false',
+                    'reply_decision.provider': rdProvider,
+                    'reply_decision.model': rdModel,
+                    'reply_decision.thinking_level': rdThinkingLevel,
+                    'reply_decision.group_only': rdGroupOnly ? 'true' : 'false',
+                    'reply_decision.max_replies_per_window': String(rdMaxReplies),
+                    'reply_decision.window_seconds': String(rdWindowSeconds)
+                  }})
+                })
+                .then(r => { rdResultOk = r.ok; return r.json(); })
+                .then(d => {
+                  rdResult = rdResultOk ? 'Reply decision settings saved' : (d.detail || 'Error');
+                  if (rdResultOk && window.showToast) { window.showToast('Reply decision settings saved'); }
+                })
+                .catch(e => { rdResultOk = false; rdResult = e.message; })
+              ">
+        Save reply decision
+      </button>
+    </div>
+    <template x-if="rdResult">
+      <div :class="rdResultOk ? 'alert-success' : 'alert-error'" class="mt-2" x-text="rdResult"></div>
+    </template>
+  </div>
+
   <div class="card">
     <h2 class="text-base mb-1">History Compaction</h2>
     <p class="text-muted text-xs mb-4">Model used to summarize older conversation turns when a session's context grows large. Enable compaction and set the trigger threshold in the <strong>History</strong> tab.</p>

channels/whatsapp.py

@@ -125,7 +125,10 @@ async def handle_webhook(self, payload: dict) -> dict:
             user_id=sender,
             chat_id=chat_id,
         )
-        await self.send(sender, response.text)
+        # Empty text means "no reply" (e.g. the reply-decision gate, #36) — stay
+        # silent rather than sending a blank WhatsApp message.
+        if response.text:
+            await self.send(sender, response.text)
         if getattr(response, "system_notice", None):
             await self.send(sender, response.system_notice)
         return {"ok": True}

config.yml.example

@@ -150,3 +150,16 @@ subagent_summary:
   provider: "deepseek"
   model: "deepseek-v4-flash"   # fast + cheap is ideal for this distillation
   thinking_level: ""           # "" (off) | "low" | "medium" | "high"
+
+# Reply decision — in shared/group chats with several bots and people, decide
+# per message whether to reply at all. Stays quiet for messages aimed at another
+# bot, caught in a bot-to-bot reaction loop, or that the agent can't usefully add
+# to. Off by default (1:1 chats always reply, and it costs one extra LLM call).
+# Enable it for group chats that mix multiple bots.
+reply_decision:
+  enabled: false
+  provider: "deepseek"
+  model: "deepseek-v4-flash"  # fast + cheap model for the yes/no reply decision
+  group_only: true            # only gate group chats; DMs always get a reply
+  max_replies_per_window: 6   # hard backstop: cap auto-replies per chat per window
+  window_seconds: 120         # rolling window for the cap — guarantees loops end