Recover non-progressing cast loops and bound host model prompt cache

ale/azerothai_recovery.lua

@@ -1,6 +1,11 @@
 -- Conservative recovery for managed residents with fresh quest/explore intent.
 -- Native position/progress is sampled independently of dashboard summaries.
 local intents, observations, pendingResults = {}, {}, {}
+local progress={}
+-- Successful gameplay, rather than merely starting a cast, resets the idle clock.
+for _,event in ipairs({32,37,54}) do
+ RegisterPlayerEvent(event,function(_,p) if p:IsPlayerbot() then progress[p:GetGUIDLow()]=os.time() end end)
+end
 local LIMIT = 600
 local COOLDOWN = 1800
 local function sqltext(s)
@@ -13,6 +18,22 @@ AuthDBExecute([[CREATE TABLE IF NOT EXISTS azerothai_recovery_log (
 AuthDBExecute([[CREATE TABLE IF NOT EXISTS azerothai_recovery_results (
  attempt_id BIGINT UNSIGNED PRIMARY KEY,status VARCHAR(24) NOT NULL,
  checked_at TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP,detail VARCHAR(255) NOT NULL)]])
+AuthDBExecute([[CREATE TABLE IF NOT EXISTS azerothai_recovery_leases (
+ guid INT UNSIGNED PRIMARY KEY,name VARCHAR(64),account_id INT UNSIGNED,
+ restore_at BIGINT UNSIGNED,gathering TINYINT,looting TINYINT)]])
+local function restoreStrategies(now)
+ local q=AuthDBQuery('SELECT guid,name,account_id,gathering,looting FROM azerothai_recovery_leases WHERE restore_at<='..now)
+ if not q then return end
+ repeat
+  local guid,name,account=q:GetUInt32(0),q:GetString(1),q:GetUInt32(2)
+  local p=GetPlayerByName(name)
+  if p and p:IsPlayerbot() and p:GetGUIDLow()==guid and p:GetAccountId()==account and p:GetBotAI() and not p:GetBotAI():HasGameClientMaster() then
+   if q:GetUInt32(3)==1 then p:GetBotAI():ChangeStrategy('+gather',1) end
+   if q:GetUInt32(4)==1 then p:GetBotAI():ChangeStrategy('+loot',1) end
+   AuthDBExecute('DELETE FROM azerothai_recovery_leases WHERE guid='..guid)
+  end
+ until not q:NextRow()
+end
 local function checkResults(now)
  for guid,r in pairs(pendingResults) do
   local p=GetPlayerByName(r.name);local status,detail
@@ -39,17 +60,55 @@ function AzerothAISetRecoveryIntents(rows)
  return 'Recovery intentions refreshed'
 end
 local function safe(p,a)
- if p:IsDead() or p:IsInCombat() or p:IsCasting() or p:IsTaxi() or a:HasGameClientMaster()
+ if p:IsDead() or p:IsInCombat() or p:IsTaxi() or a:HasGameClientMaster()
   or p:IsRooted() or a:IsMovementImpaired() then return false end
  local g=p:GetGroup()
  if g then
   if #g:GetMembers()~=g:GetMembersCount() then return false end
-  for _,m in pairs(g:GetMembers()) do if not m:IsPlayerbot() then return false end end
+  for _,m in pairs(g:GetMembers()) do if not m:IsPlayerbot() or not m:GetBotAI() or m:GetBotAI():HasGameClientMaster() then return false end end
  end
  return true
 end
+local function recover(p,a,r,guid,now,followup)
+ local x,y,z,map,level=p:GetX(),p:GetY(),p:GetZ(),p:GetMapId(),p:GetLevel()
+     local q=AuthDBQuery('SELECT guid FROM azerothai_lifecycle WHERE guid='..guid..' AND account_id='..r.account..' AND BINARY name='..sqltext(r.name).." AND state='active'")
+     local recent=AuthDBQuery('SELECT id FROM azerothai_recovery_log WHERE guid='..guid..' AND at>NOW()-INTERVAL '..COOLDOWN..' SECOND LIMIT 1')
+     if recent and followup then
+      local failed=AuthDBQuery("SELECT l.id FROM azerothai_recovery_log l JOIN azerothai_recovery_results r ON r.attempt_id=l.id WHERE l.guid="..guid.." AND r.status='no_progress' AND l.at>NOW()-INTERVAL 1800 SECOND AND (SELECT COUNT(*) FROM azerothai_recovery_log x WHERE x.guid=l.guid AND x.at>NOW()-INTERVAL 1800 SECOND)=1 LIMIT 1")
+      if failed then recent=nil end
+     end
+     if q and not recent then
+      local interrupted=p:IsCasting()
+      if interrupted then p:StopSpellCast() end
+      local staleFlight=p:GetMovementType()==7 and not p:IsTaxi()
+      if staleFlight then p:MoveExpire();p:MoveClear() end
+      a:Reset(true)
+      local gathering,looting=a:HasStrategy('gather',1),a:HasStrategy('loot',1)
+      if gathering or looting then
+       AuthDBExecute('INSERT INTO azerothai_recovery_leases VALUES ('..guid..','..sqltext(r.name)..','..r.account..','..(now+120)..','..(gathering and 1 or 0)..','..(looting and 1 or 0)..') ON DUPLICATE KEY UPDATE restore_at=VALUES(restore_at),gathering=GREATEST(gathering,VALUES(gathering)),looting=GREATEST(looting,VALUES(looting))')
+       a:ChangeStrategy('-gather',1);a:ChangeStrategy('-loot',1)
+       a:DoSpecificAction('release loot',true)
+      end
+      local moved=a:DoSpecificAction('move random',true)
+      local detail=(interrupted and 'Interrupted non-progressing cast; ' or '')..(staleFlight and 'Cleared stale flight; ' or '')..'AI replanned; gathering paused 120s; walking request='..tostring(moved)
+      AuthDBExecute('INSERT INTO azerothai_recovery_log(guid,name,detail) VALUES ('..guid..','..sqltext(r.name)..','..sqltext(detail)..')')
+      pendingResults[guid]={name=r.name,account=r.account,at=now,x=x,y=y,z=z,map=map,level=level}
+      PrintInfo('[AzerothAI recovery] '..r.name..': '..detail)
+      return true
+     end
+ return false
+end
+-- Operator-only entry point for a stall already verified through repeated observations.
+function AzerothAIRecoverObservedStall(name,guid,account)
+ local p=GetPlayerByName(name)
+ if not p or not p:IsPlayerbot() or p:GetGUIDLow()~=guid or p:GetAccountId()~=account then return 'Identity unavailable' end
+ local a=p:GetBotAI()
+ if not a or not safe(p,a) then return 'Unsafe to recover now' end
+ return recover(p,a,{name=name,account=account},guid,os.time(),true) and 'Recovery requested; awaiting observed progress' or 'Inactive identity or cooldown'
+end
 local function sample()
  local now=os.time()
+ restoreStrategies(now)
  checkResults(now)
  for guid,r in pairs(intents) do
   local p=GetPlayerByName(r.name)
@@ -62,22 +121,11 @@ local function sample()
     local x,y,z,map=p:GetX(),p:GetY(),p:GetZ(),p:GetMapId()
     local level,hp=p:GetLevel(),p:GetHealth()
     local old=observations[guid]
-    if not old or old.map~=map or old.level~=level or old.hp~=hp or (x-old.x)^2+(y-old.y)^2+(z-old.z)^2>25 then
+    if not old or old.map~=map or old.level~=level or old.hp~=hp or (progress[guid] or 0)>old.since or (x-old.x)^2+(y-old.y)^2+(z-old.z)^2>25 then
      observations[guid]={x=x,y=y,z=z,map=map,level=level,hp=hp,since=now}
     elseif now-old.since>=LIMIT then
      -- Persistent cooldown survives ALE reloads; exact identity must still be active.
-     local q=AuthDBQuery('SELECT guid FROM azerothai_lifecycle WHERE guid='..guid..' AND account_id='..r.account..' AND BINARY name='..sqltext(r.name).." AND state='active'")
+     recover(p,a,r,guid,now)
-     local recent=AuthDBQuery('SELECT id FROM azerothai_recovery_log WHERE guid='..guid..' AND at>NOW()-INTERVAL '..COOLDOWN..' SECOND LIMIT 1')
-     if q and not recent then
-      local staleFlight=p:GetMovementType()==7 and not p:IsTaxi()
-      if staleFlight then p:MoveExpire();p:MoveClear() end
-      a:Reset(true)
-      local moved=a:DoSpecificAction('move random',true)
-      local detail=(staleFlight and 'Cleared stale flight; ' or '')..'AI replanned; walking request='..tostring(moved)
-      AuthDBExecute('INSERT INTO azerothai_recovery_log(guid,name,detail) VALUES ('..guid..','..sqltext(r.name)..','..sqltext(detail)..')')
-      pendingResults[guid]={name=r.name,account=r.account,at=now,x=x,y=y,z=z,map=map,level=level}
-      PrintInfo('[AzerothAI recovery] '..r.name..': '..detail)
-     end
      old.since=now
     end
    end
@@ -85,6 +133,7 @@ local function sample()
  end
  -- Forget residents whose intent changed to resting/socializing or who went offline.
  for guid in pairs(observations) do if not intents[guid] then observations[guid]=nil end end
+ for guid,at in pairs(progress) do if now-at>3600 then progress[guid]=nil end end
 end
 CreateLuaEvent(function()local ok,err=pcall(sample);if not ok then PrintError('[AzerothAI recovery] '..tostring(err)) end end,30000,0)
 PrintInfo('[AzerothAI] Conservative stuck recovery enabled: 10 minutes idle, 30 minute cooldown')

deploy/pc-lab/limit-model-memory.ps1

@@ -0,0 +1,27 @@
+# Bound the bundled llama-server prompt cache; preserve model and context size.
+param([switch]$Restart)
+$ErrorActionPreference='Stop'
+$settings=@{LLAMA_ARG_CACHE_RAM='512';LLAMA_ARG_CTX_CHECKPOINTS='4'}
+foreach($key in $settings.Keys){[Environment]::SetEnvironmentVariable($key,$settings[$key],'User');[Environment]::SetEnvironmentVariable($key,$settings[$key],'Process')}
+$root=Join-Path $env:USERPROFILE 'Tools\AzerothAI'
+$supervisor=Join-Path $root 'recovery.ps1'
+if(Test-Path $supervisor){
+ $text=Get-Content -Raw $supervisor
+ if($text -notmatch 'LLAMA_ARG_CACHE_RAM'){
+  Copy-Item $supervisor "$supervisor.before-cache-limit" -Force
+  $prefix='$env:LLAMA_ARG_CACHE_RAM="512"; $env:LLAMA_ARG_CTX_CHECKPOINTS="4"'
+  Set-Content $supervisor ($prefix+"`r`n"+$text)
+ }
+}
+if($Restart){
+ $task=Get-ScheduledTask -TaskName 'AzerothAI-Recovery' -ErrorAction SilentlyContinue
+ if($task){Stop-ScheduledTask -InputObject $task}
+ try{Invoke-RestMethod -Method Post -Uri http://127.0.0.1:11434/api/generate -ContentType 'application/json' -Body '{"model":"qwen3.5:4b","keep_alive":0}' -TimeoutSec 45 | Out-Null}catch{Write-Warning $_.Exception.Message}
+ $binary=Join-Path $env:USERPROFILE 'Tools\Ollama\ollama.exe'
+ Get-Process ollama -ErrorAction SilentlyContinue | Where-Object {$_.Path -eq $binary} | Stop-Process
+ $logs=Split-Path $binary
+ foreach($file in @('server.out.log','server.err.log')){if(Test-Path "$logs\$file"){Copy-Item "$logs\$file" "$logs\$file.before-cache-limit" -Force}}
+ Start-Process $binary -ArgumentList 'serve' -WindowStyle Hidden -RedirectStandardOutput "$logs\server.out.log" -RedirectStandardError "$logs\server.err.log"
+ if($task){Start-ScheduledTask -InputObject $task}
+}
+Write-Output 'Prompt cache capped at 512 MiB; at most four context checkpoints per slot.'

docs/survival-cooperation.md

@@ -28,3 +28,8 @@ The native guard now samples every two seconds. Retreat is requested below 50% h
 Healthy followers can regroup over 12–200 yards on the same map/phase. Healthy noncombat leaders pause for at most 20 seconds per minute when a noncombat companion is 25–200 yards behind. Recovery requests now extend to 80% health after recent danger. No teleport, resurrection, population or stat changes are made.
 
 The auth table `azerothai_survival_audit` records request acceptance and observed outcomes: out of combat, movement while still in combat, no movement after 30 seconds, or death during observation. Movement means at least five yards from the starting position (or a map change); it does not prove successful escape. Audit records expire after seven days in bounded hourly batches. Snapshots continue to expose the latest action and timestamp.
+
+
+Stall recovery keeps its ten-minute idle clock across casts. Real movement, level/health changes, successful loot or quest completion reset the clock. It can interrupt a noncombat cast only after that threshold; combat, taxi, human control, movement restrictions, fresh intentions and the persistent thirty-minute recovery cooldown remain enforced. Run `deploy/pc-lab/limit-model-memory.ps1 -Restart` on the Windows model host to cap prompt cache at 512 MiB and checkpoints at four. This trades cache reuse for host memory without changing model or context size.
+
+Recovery briefly disables existing gather/loot strategies for two minutes after releasing loot, so an invalid gathering target cannot immediately cancel the walking request. A durable lease restores only previously enabled strategies, including after a Lua reload. An operator may make one follow-up attempt within the cooldown only when the sole recent attempt has a verified no-progress result; automatic attempts retain their thirty-minute cooldown.

tests/recovery-lua.test.py

@@ -4,8 +4,9 @@ code=(Path(__file__).resolve().parents[1]/'ale/azerothai_recovery.lua').read_tex
 setup='''
 now=1000;os.time=function()return now end;attempts=0;expired=0;last=nil;active=true;player={x=0};ai={};captured={}
 function AuthDBExecute(sql) table.insert(captured,sql);if string.find(sql,'INSERT INTO azerothai_recovery_log',1,true) then last=now end end
-function AuthDBQuery(sql) if string.find(sql,'recovery_log',1,true) then return last and now-last<1800 and {} or nil end;return active and {} or nil end
+function AuthDBQuery(sql) if string.find(sql,'recovery_leases',1,true) then return nil end;if string.find(sql,'recovery_log',1,true) then return last and now-last<1800 and {} or nil end;return active and {} or nil end
 function GetPlayerByName(n) return n=='Test' and player or nil end
+hooks={};function RegisterPlayerEvent(n,fn)hooks[n]=fn end
 function CreateLuaEvent(fn) tick=fn end
 function PrintInfo(s) end
 function PrintError(s) error(s) end
@@ -16,6 +17,7 @@ function player:GetBotAI()return ai end
 function player:IsDead()return self.dead end
 function player:IsInCombat()return self.combat end
 function player:IsCasting()return self.casting end
+function player:StopSpellCast()self.casting=false;self.interrupted=true end
 function player:IsTaxi()return self.taxi end
 function player:IsRooted()return self.root end
 function player:GetGroup()return self.group end
@@ -31,7 +33,9 @@ function player:MoveClear()end
 function ai:HasGameClientMaster()return player.master end
 function ai:IsMovementImpaired()return player.impaired end
 function ai:Reset(full)attempts=attempts+1 end
-function ai:DoSpecificAction(action,silent)assert(action=='move random');return true end
+function ai:HasStrategy(s,state)return false end
+function ai:ChangeStrategy()end
+function ai:DoSpecificAction(action,silent)assert(action=='move random' or action=='release loot');return true end
 function step(seconds,action)
  now=now+seconds
  AzerothAISetRecoveryIntents({{guid=9,account=2,name='Test',action=action or 'quest'}})
@@ -42,7 +46,7 @@ def new():
  l=LuaRuntime();l.execute(setup);l.execute(code);return l
 l=new();l.execute('step(0);step(599);assert(attempts==0);step(1);assert(attempts==1);step(600);assert(attempts==1);step(1200);assert(attempts==2)')
 print('Threshold and persistent cooldown passed')
-for field in ['dead','combat','casting','taxi','root','master','impaired']:
+for field in ['dead','combat','taxi','root','master','impaired']:
  l=new();l.execute(f'player.{field}=true;step(0);step(601);assert(attempts==0)')
 print('Death, combat, cast, taxi, root, master and impaired exclusions passed')
 for action in ['rest','socialize','trade']:
@@ -60,3 +64,13 @@ l=new();l.execute("step(0);step(600);player.x=8;step(30);assert(string.find(capt
 l=new();l.execute("step(0);step(600);step(180);assert(string.find(captured[#captured], '6E6F5F70726F6772657373'))")
 l=new();l.execute("step(0);step(600);player.dead=true;step(30);assert(string.find(captured[#captured], '756E7665726966696564'))")
 print('Recovery outcomes distinguish native progress, no progress, and unavailable residents')
+
+l=new();l.execute('player.casting=true;step(0);step(599);assert(attempts==0);step(1);assert(attempts==1 and player.interrupted)')
+l=new();l.execute('player.casting=true;step(0);now=now+500;hooks[32](32,player);step(0);step(101);assert(attempts==0)')
+l=new();l.execute('player.casting=true;step(0);player.combat=true;step(601);assert(attempts==0 and not player.interrupted)')
+print('Repeated casts cannot reset idle evidence; productive casts and combat stay protected')
+
+l=new();l.execute("function ai:HasStrategy()return true end;local changes={};function ai:ChangeStrategy(s)table.insert(changes,s)end;step(0);step(600);assert(changes[1]=='-gather' and changes[2]=='-loot');assert(attempts==1)")
+print('Recovery pauses gathering before walking, with a durable restoration lease')
+l=new();l.execute("local query=AuthDBQuery;local changes={};function ai:ChangeStrategy(s)table.insert(changes,s)end;function AuthDBQuery(sql)if string.find(sql,'SELECT guid,name,account_id,gathering,looting',1,true) then return {GetUInt32=function(_,i)return ({[0]=9,[2]=2,[3]=1,[4]=0})[i] end,GetString=function()return 'Test' end,NextRow=function()return false end} end;return query(sql)end;step(0);assert(changes[1]=='+gather' and #changes==1)")
+print('Lease restoration restores only originally enabled strategies')