
    i                        U d Z ddlZddlZddlZddlZddlZddlZddlZddl	Z	ddl
Z
ddlZddlmZ ddlZddlmZ ddlmZ ddlmZmZmZmZ ddlmZ  ej        e          Z ee          j        j        Zedz  Zed	z  d
z  Z edz  Z! e            dz  dz  Z"d Z#ddddddddddddddddddddgd d!d"d#dd$d%d%d&Z$ e%e$&                    d'i           '                                          Z(e G d( d)                      Z)e G d* d+                      Z*g a+ee)         e,d,<   da-ee.         e,d-<   i a/ee.ef         e,d.<   i Z0ee.ee.ee.ef         f         f         e,d/<   i Z1ee.e*f         e,d0<   i Z2ee.e3f         e,d1<   d2Z4d3ee)         fd4Z5d5e.d3ee.ee.ef         f         fd6Z6d7 Z7d8e*d9efd:Z8d8e*fd;Z9d8e*fd<Z:d3e.fd=Z;d>e.d3e.fd?Z<d3e.fd@Z=dAe.dBed3e.fdCZ>d3e.fdDZ?dEe.d3e.fdFZ@dEe.d3e.fdGZAdEe.d3e.fdHZBd3e.fdIZCdJdKdLdMdNdOdPdMdQdRdSdMgZDdZEdTZFeEeFdfdUeGdVeGdWeee.                  d3e.fdXZHd3eIfdYZJd3eIfdZZKd3ee.         fd[ZLdd\lMmNZN d]d^d_i g d`daZOdbdcd_d>dddedfid>gd`daZPdgdhd_i g d`daZQdidjd_dddkdfdldmidndAdBgd`daZRdodpd_i g d`daZSdqdrd_dEdddsdfidEgd`daZTdtdud_dEdddvdfidEgd`daZUdwdxd_dEdddydfidEgd`daZVdzd{d_i g d`daZWd|d}d_d~dddd~ddTdddddidddg d`daZXddgZY eNjZ        d]ddeOd eKeYd            eNjZ        dbddePd eKeYd            eNjZ        dgddeQd eKeYd            eNjZ        diddeRd eKeYd            eNjZ        doddeSd eKeYd            eNjZ        dqddeTd eKeYd            eNjZ        dtddeUd eKeYd            eNjZ        dwddeVd eKeYd            eNjZ        dzddeWd eKeYd            eNjZ        d|ddeXd eKeYd           dS )a$  
RL Training Tools Module

This module provides tools for running RL training through Tinker-Atropos.
Directly manages training processes without requiring a separate API server.

Features:
- Environment discovery (AST-based scanning for BaseEnv subclasses)
- Configuration management with locked infrastructure settings
- Training run lifecycle via subprocess management
- WandB metrics monitoring

Required environment variables:
- TINKER_API_KEY: API key for Tinker service
- WANDB_API_KEY: API key for Weights & Biases metrics

Usage:
    from tools.rl_training_tool import (
        rl_list_environments,
        rl_select_environment,
        rl_get_current_config,
        rl_edit_config,
        rl_start_training,
        rl_check_status,
        rl_stop_training,
        rl_get_results,
    )
    N)datetime)	dataclass)Path)AnyDictListOptional)get_hermes_homeztinker-atropostinker_atroposenvironmentsconfigslogsrl_trainingc                  r    t                                           rt                              d           dS dS )zNLazily create logs directory on first use (avoid side effects at import time).Texist_okN)TINKER_ATROPOS_ROOTexistsLOGS_DIRmkdir     ;/home/ubuntu/.hermes/hermes-agent/tools/rl_training_tool.py_ensure_logs_dirr   >   s:    !!## &%%%%%& &r   zQwen/Qwen3-8Bzhttp://localhost:8000Ti    i   i  i	        g      ?g?)tokenizer_namerollout_server_url	use_wandbmax_token_lengthmax_num_workersworker_timeouttotal_stepssteps_per_evalmax_batches_offpolicyinference_weighteval_limit_ratiozhttp://localhost:8001/v1x   sglang)
model_namebase_urlapi_keyweightnum_requests_for_evaltimeoutserver_type    gh㈵?i(#  z./temp/)	lora_ranklearning_ratemax_token_trainer_lengthcheckpoint_dirsave_checkpoint_intervalF)envopenaitinkerslurmtestingr8   c                   N    e Zd ZU dZeed<   eed<   eed<   dZeed<   dZeed<   d	S )
EnvironmentInfoz+Information about a discovered environment.name
class_name	file_path descriptionBaseEnvConfigconfig_classN)__name__
__module____qualname____doc__str__annotations__rC   rE   r   r   r   r>   r>   s   sS         55
IIIOOONNNK'L#'''''r   r>   c                       e Zd ZU dZeed<   eed<   eeef         ed<   dZeed<   dZ	eed<   dZ
eed	<   dZeed
<   dZeed<   dZeej                 ed<   dZeej                 ed<   dZeej                 ed<   dS )RunStatezState for a training run.run_idenvironmentconfigpendingstatusrB   error_messagewandb_projectwandb_run_nameg        
start_timeNapi_processtrainer_processenv_process)rF   rG   rH   rI   rJ   rK   r   r   rR   rS   rT   rU   rV   floatrW   r	   
subprocessPopenrX   rY   r   r   r   rM   rM   }   s         ##KKKcNFCM3M3NCJ.2K**+22226OXj./666.2K**+22222r   rM   _environments_current_env_current_config_env_config_cache_active_runs_last_status_checki  returnc                     g } t                                           s| S t                               d          D ]}|j                            d          r	 t          |d          5 }t          j        |                                          }ddd           n# 1 swxY w Y   t          j	        |          D ]}t          |t          j                  r|j        D ]}d}t          |t          j                  r|j        }n!t          |t          j                  r|j        }|dk    r|j        }d}d}	|j        D ]?}
t          |
t          j                  r|
j        D ]}t          |t          j                  rm|j        dk    r,t          |
j        t          j                  r|
j        j        }S|j        d	k    r+t          |
j        t          j                  r|
j        j        }	t          |
t          j                  rwt          |
j        t          j                  rXt          |
j        j        t0                    r9|s7|
j        j                            d
          d                                         }A|                     t9          ||j        t1          |          |p	d|j         |	                      n# t:          $ r'}t<                              d||           Y d}~d}~ww xY w| S )zK
    Scan the environments directory for BaseEnv subclasses using AST.
    z*.py_rNrB   BaseEnvrD   r?   env_config_cls
r   zEnvironment from )r?   r@   rA   rC   rE   zCould not parse %s: %s) ENVIRONMENTS_DIRr   globr?   
startswithopenastparsereadwalk
isinstanceClassDefbasesNameid	AttributeattrstembodyAssigntargetsvalueConstantExprrJ   splitstripappendr>   	Exceptionloggerwarning)r   py_fileftreenodebase	base_nameenv_namerC   rE   itemtargetes                r   _scan_environmentsr      sT    L""$$ #((00 /A /A<""3'' 	+	Ags## +qy**+ + + + + + + + + + + + + + +  %" %"dCL11 $" $
 "" ""$&	%dCH55 2(,II'cm<< 2(,	I$	11'.|H*,K+:L(,	 ^ ^#-dCJ#?#? !M26, %M %M+5fch+G+G )M/5yF/B/BzRVR\^a^jGkGk/B;?:;K17>N1N1NS]^b^hjmjrSsSs1N?Cz} $.dCH#=#= !^*TZY\YeBfBf !^'1$*2BC'H'H %^Q\ %^6:j6F6L6LT6R6RST6U6[6[6]6](//%-+/9*-g,,,7,];]w|;];]-91 1 1    "E7 2%"L  	A 	A 	ANN3Wa@@@@@@@@	A s=   K''BKB	K!B	"H'K
K<K77K<env_file_pathc                    	 t           j                            d|           }t           j                            |          }|t          j        d<   |j                            |           d}t          |          	                                D ]Q\  }}t          |t                    r7|dk    r1t          |d          r!t          t          |d                    r|} nR|si S d}	 |                                \  }}t          |          }nS# t           $ rF}	t"                              d|	           	 ddlm}
 |
}n# t*          $ r
 i cY cY d}	~	S w xY wY d}	~	nd}	~	ww xY w|si S d }i }|j        	                                D ]\  }}|j        } ||j                  }|j        pd	}|t4          v }t          |d
t7          |                    }t          |d          rt7          |          }t8                              di                               ||          }|r ||          n|}|||||d||<   |S # t           $ r'}t"                              d|           i cY d}~S d}~ww xY w)z
    Dynamically import an environment and extract its config fields.
    
    Uses config_init() to get the actual config class, with fallback to
    directly importing BaseEnvConfig if config_init fails.
    
env_moduleNrg   config_initz5config_init failed (%s), using BaseEnvConfig defaultsr   )rD   c                     | d S t          | d          r| j        S t          | d          r5t          | d          r%dt          t          |                     v r| j        S | S )Nr}   r?   	__class__Enum)hasattrr}   rJ   typer?   )vals    r   make_serializablez1_get_env_config_fields.<locals>.make_serializable  sr    {tsG$$ !y sF##  [(A(A  fPSTXY\T]T]P^P^F^F^xJr   rB   rF   
__origin__r8   )r   defaultrC   lockedcurrent_valuez+Could not introspect environment config: %s)	importlibutilspec_from_file_locationmodule_from_specsysmodulesloaderexec_modulevarsitemsrr   r   r   callablegetattrr   r   r   infoatroposlib.envs.baserD   ImportErrormodel_fields
annotationr   rC   LOCKED_FIELD_NAMESrJ   LOCKED_FIELDSgetr   )r   specmodule	env_classr?   objrE   
env_configserver_configsconfig_errorrD   r   fields
field_name
field_info
field_typer   rC   	is_locked	type_namelocked_valuer   r   s                          r   _get_env_config_fieldsr      s    J~55lMRR0066$*L!''' 	f++-- 	 	ID##t$$ ):):3.. 8GC<W<W3X3X  #IE 	I 
	)2)>)>)@)@&J
++LL 	 	 	KKOQ]^^^>>>>>>,   								 	  	I	 	 	 &2&?&E&E&G&G 	 	"J
#.J''
(:;;G$06BK"&88I  
JJHHIz<00 ,
OO	(,,UB77;;JPPL?HU--l;;;gM ""*#!." "F:    DaHHH						s   C H9 #H9 &&D H9 
EE3D<;E<EE	E
H9 EEH9 EH9 $CH9 9
I*I%I*%I*c                  4    t           st                      a dS dS )z)Initialize environment list on first use.N)r]   r   r   r   r   _initialize_environmentsr   /  s%      -*,,- -r   	run_stateconfig_pathc                 	  K   | j         }t                       t          d| dz  }t          d| dz  }t          d| dz  }	 t                              d|           t          |d          }|| _        t          j        dg|t          j	        t          t                              | _        t          j        d	           d
{V  | j                                        /d| _        d| j        j         d| | _        t'          |            d
S t                              d|           t                              d||           t          |d          }|| _        t          j        t*          j        ddt          |          g|t          j	        t          t                    i t.          j        dt/          j        dd          i          | _        t                              d|           t          j        d           d
{V  | j                                        /d| _        d| j        j         d| | _        t'          |            d
S t                              d|           t                              d|           t          j        d           d
{V  d
}t6          D ]}	|	j        | j        k    r|	} n|s(d| _        d| j         d| _        t'          |            d
S t                              d||j                   t          |d          }
|
| _        t          j        t*          j        t          |j                  ddt          |          g|
t          j	        t          t                              | _         t          j        d           d
{V  | j                                         /d| _        d | j         j         d| | _        t'          |            d
S d!| _        tC          j!                    | _"        t                              d"|           t          j#        tI          |                      d
S # tJ          $ r5}d| _        t          |          | _        t'          |            Y d
}~d
S d
}~ww xY w)#z
    Spawn the three processes needed for training:
    1. run-api (Atropos API server)
    2. launch_training.py (Tinker trainer + inference server)
    3. environment.py serve (the Atropos environment)
    api_.logtrainer_env_z-[%s] Starting Atropos API server (run-api)...wzrun-apistdoutstderrcwd   NfailedzAPI server exited with code z. Check z[%s] Atropos API server startedz<[%s] Starting Tinker trainer: launch_training.py --config %szlaunch_training.pyz--configTINKER_API_KEYrB   )r   r   r   r8   z4[%s] Waiting 30 seconds for trainer to initialize...   zTrainer exited with code z3[%s] Trainer started, inference server on port 8001z;[%s] Waiting 90 more seconds before starting environment...Z   Environment '' not foundz#[%s] Starting environment: %s serveserve
   zEnvironment exited with code runningz'[%s] Training run started successfully!)&rN   r   r   r   r   rm   api_log_filer[   r\   STDOUTrJ   r   rW   asynciosleeppollrR   
returncoderS   _stop_training_runtrainer_log_filer   
executableosenvirongetenvrX   r]   r?   rO   rA   env_log_filerY   timerV   create_task_monitor_training_runr   )r   r   rN   api_logtrainer_logenv_logr   r   env_infor8   r   r   s               r   _spawn_training_runr   :  s      F ,,,,,G44444K,,,,,Gb&CVLLL GS))!-	 * 0K$'((	!
 !
 !
	 mA %%''3'I&xYEZEe&x&xov&x&xI#y)))F5v>>> 	RTZ\ghhhS11%5	"$.$4^1:s;?O?OP#$'((Q2:Q/;KR1P1PQQ%
 %
 %
	! 	JFSSSmB$))++7'I&})B[Bf&}&}p{&}&}I#y)))FI6RRR 	QSYZZZmB   	 	Cx9000 1  	'I&Xi6K&X&X&XI#y)))F968CUVVVGS))!-	 * 0^S!344gz3{K[K[\$'((	!
 !
 !
	 mB %%''3'I&yiF[Ff&y&ypw&y&yI#y)))F$	#y{{	=vFFF 	1)<<===== & & &#	"%a&&	9%%%%%%%%%&s5   CQ EQ BQ .C?Q /AQ 
R*RRc                 `  K   | j         dk    rt          j        d           d{V  | j        rU| j                                        <| j        j        }|dk    rd| _         nd| _         d| | _        t          |            dS | j        rU| j                                        <| j        j        }|dk    rd| _         nd| _         d| | _        t          |            dS | j	        r8| j	                                        d| _         d	| _        t          |            dS | j         dk    dS dS )
z*Background task to monitor a training run.r   r   Nr   	completedr   z%Environment process exited with code z!Trainer process exited with code zAPI server exited unexpectedly)
rR   r   r   rY   r   r   rS   r   rX   rW   )r   	exit_codes     r   r   r     si     

i
'
'mB   	Y%:%?%?%A%A%M!-8IA~~#.	  #+	 *]R[*]*]	'y)))E$ 	)B)G)G)I)I)U!1<IA~~#.	  #+	 *Yi*Y*Y	'y)))E  	Y%:%?%?%A%A%M'I&FI#y)))E9 
i
'
'
'
'
'
'r   c                    | j         r| j                                         t                              d| j                   | j                                          	 | j                             d           n.# t          j        $ r | j         	                                 Y nw xY w| j
        r| j
                                        t                              d| j                   | j
                                         	 | j
                            d           n.# t          j        $ r | j
        	                                 Y nw xY w| j        r| j                                        t                              d| j                   | j                                         	 | j                            d           n.# t          j        $ r | j        	                                 Y nw xY w| j        dk    rd| _        d	D ]L}t          | |d          }|7	 |                                 n# t          $ r Y nw xY wt!          | |d           MdS )
z&Stop all processes for a training run.Nz$[%s] Stopping environment process...r   r0   z [%s] Stopping trainer process...z[%s] Stopping API server...r   stopped)r   r   r   )rY   r   r   r   rN   	terminatewaitr[   TimeoutExpiredkillrX   rW   rR   r   closer   setattr)r   rx   fhs      r   r   r     s     )!6!;!;!=!=!E:I<LMMM'')))	)!&&r&2222( 	) 	) 	)!&&(((((	)   -Y%>%C%C%E%E%M6	8HIII!++---	-%**2*6666( 	- 	- 	-%**,,,,,	-  )!6!;!;!=!=!E193CDDD'')))	)!&&r&2222( 	) 	) 	)!&&(((((	) 9$$$	 E + +Yd++>



   ItT***+ +sH   A7 7(B"!B"?D (EE#F? ?(G*)G*H,,
H98H9c                     K   t                       d t          D             t          t                    g dd} t          j        | d          S )a  
    List all available RL environments.
    
    Scans tinker-atropos/tinker_atropos/environments/ for Python files
    containing classes that inherit from BaseEnv.
    
    Returns information about each environment including:
    - name: Environment identifier
    - class_name: Python class name
    - file_path: Path to the environment file
    - description: Brief description if available
    
    TIP: To create or modify RL environments:
    1. Use terminal/file tools to inspect existing environments
    2. Study how they load datasets, define verifiers, and structure rewards
    3. Inspect HuggingFace datasets to understand data formats
    4. Copy an existing environment as a template
    
    Returns:
        JSON string with list of environments
    c                 D    g | ]}|j         |j        |j        |j        d S )r?   r@   rA   rC   r   ).0r8   s     r   
<listcomp>z(rl_list_environments.<locals>.<listcomp>  sE     
 
 
  !n ]"	 
 
 
r   )z8Use rl_select_environment(name) to select an environmentzKRead the file_path with file tools to understand how each environment worksz@Look for load_dataset(), score_answer(), get_next_item() methods)r   counttips   indent)r   r]   lenjsondumps)responses    r   rl_list_environmentsr    sq      , 
 
 %
 
 
 ]##
 
 
 H$ :hq))))r   r?   c                 @  K   t                       d}t          D ]}|j        | k    r|} n|s,t          j        d|  dd t          D             dd          S | at          |j                  }|t          | <   i a	|
                                D ]8\  }}|                    dd	          s|                    d
          t          |<   9t          j                                        d          }|  d| t          d<   t          j        d|  | |j        dd          S )a  
    Select an RL environment for training.
    
    This loads the environment's configuration fields into memory.
    After selecting, use rl_get_current_config() to see all configurable options
    and rl_edit_config() to modify specific fields.
    
    Args:
        name: Name of the environment to select (from rl_list_environments)
    
    Returns:
        JSON string with selection result, file path, and configurable field count
    
    TIP: Read the returned file_path to understand how the environment works.
    Nr   r   c                     g | ]	}|j         
S r   r?   )r   r   s     r   r  z)rl_select_environment.<locals>.<listcomp>G  s    888Q!&888r   )error	availabler  r  r   Fr   z%Y%m%d-%H%M%S-
wandb_namezSelected environment: )messagerO   rA   )r   r]   r?   r  r	  r^   r   rA   r`   r_   r   r   r   nowstrftime)r?   r   r8   config_fieldsr   r   	timestamps          r   rl_select_environmentr  *  s     $ H  8tHE   z6T66688-888
 
    	
 L +8+=>>M+d O"/"5"5"7"7 D D
J~~h.. 	D*4..*C*COJ' ''88I'+$9$9i$9$9OL!:2D22'  	   r   c                    K   t           st          j        ddid          S t                              t           i           } g }g }|                                 D ]\  }}||                    dd          |                    d          |                    dd	          t                              ||                    d                    d
}|                    dd          rGt                              di                               |          |d<   |                    |           |                    |           t          j        t           ||ddd          S )a  
    Get the current environment configuration.
    
    Returns all configurable fields for the selected environment.
    Each environment may have different configuration options.
    
    Fields are divided into:
    - configurable_fields: Can be changed with rl_edit_config()
    - locked_fields: Infrastructure settings that cannot be changed
    
    Returns:
        JSON string with configurable and locked fields
    r  ?No environment selected. Use rl_select_environment(name) first.r  r  r   unknownr   rC   rB   )r?   r   r   rC   r   r   Fr8   r   zBUse rl_edit_config(field, value) to change any configurable field.)rO   configurable_fieldslocked_fieldstip)	r^   r  r	  r`   r   r   r_   r   r   )r  configurabler   r   r   
field_datas         r   rl_get_current_configr!  e  sw       zV
   	 &)),;;MLF"/"5"5"7"7 , ,
JNN6955!~~i00%>>-<<,00Z^^I=V=VWW
 

 >>(E** 	,)6):):5")E)E)I)I*)U)UJ~&MM*%%%%
++++:#+S	 
    r   fieldr}   c                 2  K   t           st          j        ddid          S t                              t           i           }| |vr<t          j        d|  dt          |                                          dd          S ||          }|                    dd	          rIt          j        d
|  dt                              di                               |           dd          S |t          | <   t          j        d|  d| | |t          dd          S )a  
    Update a configuration field.
    
    Use rl_get_current_config() first to see available fields for the
    selected environment. Each environment has different options.
    
    Locked fields (infrastructure settings) cannot be changed.
    
    Args:
        field: Name of the field to update (from rl_get_current_config)
        value: New value for the field
    
    Returns:
        JSON string with updated config or error message
    r  r  r  r  zUnknown field '')r  available_fieldsr   FzField 'z!' is locked and cannot be changedr8   )r  r   zUpdated z = )r  r"  r}   rP   )	r^   r  r	  r`   r   listkeysr   r_   )r"  r}   r  r   s       r   rl_edit_configr(    so        zV
   	 &)),;;MM!!z/u/// $]%7%7%9%9 : :
 
    	
 u%J~~h&& zGuGGG)--eR88<<UCC
 
    	
 #OE:/e////!	 
    r   c                    K   t           st          j        ddid          S t          j        d          st          j        ddid          S d} t
          D ]}|j        t           k    r|}  n| r&t          | j                  	                                s!t          j        ddt            d	id          S t          t          j                              dd
         }t                              d           t          d| dz  }ddl}|                    t"                    }d|vri |d<   t$                                          D ]\  }}||dk    r||d         |<   t$                              dd          }d|vri |d<   ||d         d<   t            d| |d         d<   dt$          v r#t$          d         rt$          d         |d         d<   t+          |d          5 }	t-          j        ||	d           ddd           n# 1 swxY w Y   t1          |t           t$                                          d|t            d|           }
|
t2          |<   t5          j        t9          |
|                     t          j        |dt           t$          |t            d| t          |          t          t:          d| dz            t          t:          d| dz            t          t:          d | dz            d!d"d#	d          S )$av  
    Start a new RL training run with the current environment and config.
    
    Requires an environment to be selected first using rl_select_environment().
    Use rl_edit_config() to adjust configuration before starting.
    
    This spawns three processes:
    1. run-api (Atropos trajectory API)
    2. launch_training.py (Tinker trainer + inference server)
    3. environment.py serve (the selected environment)
    
    WARNING: Training runs take hours. Use rl_check_status() to monitor
    progress (recommended: check every 30 minutes at most).
    
    Returns:
        JSON string with run_id and initial status
    r  r  r  r  r   z0TINKER_API_KEY not set. Add it to ~/.hermes/.envNz Environment file not found for 'r$     Tr   run_z.yamlr   r8   rB   rT   zatropos-tinkerr:   r  rU   r  r   F)default_flow_stylestarting)rN   rO   rP   rR   rT   rU   r   r   r   r   apitrainerr8   zZTraining starting. Use rl_check_status(run_id) to monitor (recommended: every 30 minutes).)	rN   rR   rO   rP   rT   rU   r   r   r  )r^   r  r	  r   r   r]   r?   r   rA   r   rJ   uuiduuid4CONFIGS_DIRr   copydeepcopyr   r_   r   r   rm   yamldumprM   ra   r   r   r   r   )r   r8   rN   r   r4  
run_configr   r}   rT   r   r   s              r   rl_start_trainingr9    s     $  zV
   	
 9%&& zG
   	
 H  8|##HE $  4 233::<< zGGGG
   	
 rr"F t$$$ 4v 4 4 44K KKK}--JJ
5 -2244 2 2
E",1Juj) $''9IJJMz!!!
8,9Jx)0<-G-Gv-G-GJx)*&&?<+H&*9,*G
5,'	k3		 ;1	*aE::::; ; ; ; ; ; ; ; ; ; ; ; ; ; ;  ##%%#&1111  I %L +I{CCDDD:#!&)44F44;''x"5"5"5"55668&=&=&=&==>>x"5"5"5"5566
 

 p     s   !HH	H	rN   c                   K   t          j                     }| t          v rH|t          |          z
  }|t          k     r-t          |z
  }t          j        d| d|dz  dd|dd          S |t          | <   | t
          vrAt          j        d	|  d
t          t
                                                    dd          S t
          |          }|j        r|j        	                                nd|j
        r|j
        	                                nd|j        r|j        	                                ndd}|j        rt          j                     |j        z
  nd}| |j        |j        |dz  d |                                D             |j        |j        t%          t&          d|  dz            t%          t&          d|  dz            t%          t&          d|  dz            dd}|j        r
|j        |d<   	 ddl}|                                }	|	                    t1          j        dd           d|j         d|j        i          }
|
r}|
d         }|j        |d<   |j                            dd          |j                            d          |j                            d          |j                            d          d |d!<   n)# t:          $ r}t%          |          |d"<   Y d}~nd}~ww xY wt          j        |d          S )#aN  
    Get status and metrics for a training run.
    
    RATE LIMITED: For long-running training, this function enforces a
    minimum 30-minute interval between checks for the same run_id.
    
    Args:
        run_id: The run ID returned by rl_start_training()
    
    Returns:
        JSON string with run status and metrics
    Tz&Rate limited. Next check available in <   z.0fz	 minutes.)rate_limitedrN   r  next_check_in_secondsr  r  Run 'r   r  active_runsNr.  r   c                 *    i | ]\  }}||dnd| dS )Nr   zexited ()r   )r   r?   codes      r   
<dictcomp>z#rl_check_status.<locals>.<dictcomp>a  sB     
 
 
d t|))1CD1C1C1C
 
 
r   r   r   r   r   )rN   rR   rO   running_time_minutes	processesrT   rU   r   r  WANDB_ENTITYnousresearch/display_namefilters	wandb_url_stepztrain/reward_meanztrain/percent_correctzeval/percent_correct)stepreward_meanpercent_correcteval_percent_correctmetricswandb_error)r   rb   MIN_STATUS_CHECK_INTERVALr  r	  ra   r&  r'  rW   r   rX   rY   rV   rR   rO   r   rT   rU   rJ   r   rS   wandbApirunsr   r   urlsummaryr   r   )rN   r  elapsed	remainingr   rF  running_timeresultrV  r/  rX  	wandb_runr   s                r   rl_check_statusr`  /  s      )++C###*622...1G;I: $ _IbL____)2	 
     "%v\!!z0V000 1 1 3 344
 
    	
 V$I 09/DNy$))+++$7@7PZ9,11333VZ/8/DNy$))+++$ I :C9MT49;;!555STL " , ,r 1
 
'oo//
 
 
 #0#2x"5"5"5"55668&=&=&=&==>>x"5"5"5"5566
 
 F$  2#1w'iikkxxy88TT9;RTT#Y%=>  
 
  	QI"+-F;!)--gq99(0445HII#,#4#8#89P#Q#Q(1(9(=(=>T(U(U	! !F9  ' ' ' #A}' :fQ''''s   1CK 
K*K%%K*c                 r  K   | t           vrAt          j        d|  dt          t                                                     dd          S t           |          }|j        dvr$t          j        dd|  d|j         d	id          S t          |           t          j        d
|  d| |j        dd          S )z
    Stop a running training job.
    
    Args:
        run_id: The run ID to stop
    
    Returns:
        JSON string with stop confirmation
    r>  r   r?  r  r  )r   r-  r  z' is not running (status: rB  zStopped training run 'r$  )r  rN   rR   )ra   r  r	  r&  r'  rR   r   )rN   r   s     r   rl_stop_trainingrb    s      \!!z0V000 1 1 3 344
 
    	
 V$I666zTvTTAQTTT
   	 y!!!:5F555"  	   r   c                 f  K   | t           vrt          j        dd|  did          S t           |          }| |j        |j        |j        |j        d}	 ddl}|                                }|	                    t          j        d	d
           d|j         d|j        i          }|rL|d         }|j        |d<   t          |j                  |d<   d |                    d          D             |d<   n)# t           $ r}t#          |          |d<   Y d}~nd}~ww xY wt          j        |d          S )z
    Get final results and metrics for a training run.
    
    Args:
        run_id: The run ID to get results for
    
    Returns:
        JSON string with final results
    r  r>  r   r  r  )rN   rR   rO   rT   rU   r   NrG  rH  rI  rJ  rK  rM  final_metricsc                 ,    g | ]}t          |          S r   )dict)r   rows     r   r  z"rl_get_results.<locals>.<listcomp>  s     T T Tsc T T Tr   r   )sampleshistoryrT  )ra   r  r	  rR   rO   rT   rU   rV  rW  rX  r   r   rY  rf  rZ  ri  r   rJ   )rN   r   r^  rV  r/  rX  r_  r   s           r   rl_get_resultsrj    s      \!!z0V000
   	 V$I " ,"0#2 F'iikkxxy88TT9;RTT#Y%=>  
 
  	UQI"+-F;&*9+<&=&=F?# T Ti6G6GPR6G6S6S T T TF9 ' ' ' #A}' :fQ''''s   B!C4 4
D>DDc                     K   g } t                                           D ].\  }}|                     ||j        |j        |j        d           /t          j        | t          |           dd          S )z
    List all training runs (active and completed).
    
    Returns:
        JSON string with list of runs and their status
    )rN   rO   rR   rU   )rX  r  r  r  )	ra   r   r   rO   rR   rU   r  r	  r  )rX  rN   r   s      r   rl_list_runsrl    s       D)//11  	$0&'6	
 
 	 	 	 	 :T     r   zqwen/qwen3-8bzQwen3 8Bsmallrv   r?   scalezz-ai/glm-4.7-flashzGLM-4.7 Flashmediumzminimax/minimax-m2.7zMiniMax M2.7large   	num_steps
group_sizemodelsc                 P  K   t           st          j        ddid          S t          j        d          }|st          j        ddid          S d}t
          D ]}|j        t           k    r|} n|s!t          j        ddt            d	id          S r"fd
t          D             }|sd D             }nt          }| |z  }|t          |          z  }t           |j	        | |||dg d}	t                       t          dz  }
|
                    d           |D ]}|d         }|                    dd          }t          dd            t          d|d          d| d           t          d            |
dt            d| dz  }t          t!          j                              dd         }dt            d| }g t$          j        |j	        ddt          |           d t          |          d!d"d#|d$t          |          d%t(          d&         d'         d(t          t(          d&         d)                   d*t          t(          d&         d+                   d,t          t(          d&         d-                   d.d/d0|d1|d2d3d4d5}d6                    d7 |D                       }|                    |d8          }t          d9|            t          d:t,                      t          d;|            t          d<|  d=| d>| d?           ||d         |d@         |t          |          g dAdAdAdB	}	 t/          j        |t.          j        j        t.          j        j        t          t,                    dC d{V }g }g }|
dt            d| dDz  }d}dF}	 t/          j        t/          j         ||j        |dG           ||j        |dH                    dIJ           d{V  n(# t.          j        $ r |                                   w xY w|!                                 d{V  d                    |          }d                    |          }tE          |dK          5 }|#                    d9| d           |#                    d:t,           d           |#                    dL|j$         d           |#                    dd d           |#                    dMd d           |#                    |pdN           |#                    dd d           |#                    dOd d           |#                    |pdN           ddd           n# 1 swxY w Y   t          dP|            |j$        dAk    rdQ|j$         |d<   |dRd         |dS<   |dRd         |dT<   t          |          |dU<   t          dV|d                     |r.t          dW           |dXd         D ]}t          dY|            nt          dZ           t          d[|            t          d\|%                                            |%                                r(tE          |d]          5 }|D ]}|&                                }|s	 t          j'        |          }|(                    d^g           }|d_xx         d`z  cc<   |daxx         t          |          z  cc<   tS          db |D                       } |dcxx         | z  cc<   |dd         *                    |d_         t          |          | |de           # t          j+        $ r Y w xY w	 ddd           n# 1 swxY w Y   t          df|d_          dg           ndh| |d<   n_# t.          j        $ r di|d<   t          dj           Y n:tX          $ r.}!t          |!          |d<   t          dk|!            Y d}!~!nd}!~!ww xY w|da         dAk    r#t[          |dc         |da         z  dl          |dm<   ndA|dm<   |d_         dAk    rAtS          dn |dd         D                       }"|"|do<   t[          |"|d_         z  dl          |dp<   n
dA|do<   dA|dp<   t          dq|dc          d|da          dr           t          ds|dm         dt           |	du         *                    |           dv |	du         D             }#| t          |          t          |#          |#rt]          |#dw x          dy         nd|#r7t[          tS          dz |#D                       t          |#          z  dl          ndAt_          |#          t          |
          d{|	d|<   t          j        |	d          S )~u  
    Quick inference test for any environment using Atropos's `process` mode.
    
    Runs a few steps of inference + scoring to validate:
    - Environment loads correctly
    - Prompt construction works
    - Inference parsing is robust (tested with multiple model scales)
    - Verifier/scoring logic works
    
    Default: 3 steps × 16 completions = 48 total rollouts per model.
    Tests 3 models = 144 total rollouts. Quick sanity check.
    
    Test models (varying intelligence levels for robustness):
    - qwen/qwen3-8b (small)
    - zhipu-ai/glm-4-flash (medium)
    - minimax/minimax-m1 (large)
    
    Args:
        num_steps: Steps to run (default: 3, max recommended for testing)
        group_size: Completions per step (default: 16, like training)
        models: Optional model IDs to test. If None, uses all 3 test models.
    
    Returns:
        JSON with results per model: steps_tested, accuracy, scores
    r  r  r  r  OPENROUTER_API_KEYz;OPENROUTER_API_KEY not set. Required for inference testing.Nr   r   c                 (    g | ]}|d          v |S )rv   r   )r   mru  s     r   r  z%rl_test_inference.<locals>.<listcomp>2  s'    CCCQ4F1B1Bq1B1B1Br   c                     g | ]}||d d	S )customrn  r   r   ry  s     r   r  z%rl_test_inference.<locals>.<listcomp>4  s"    SSSq!QBBSSSr   )rs  rt  rollouts_per_modeltotal_rollouts)rO   environment_filetest_configmodels_testedinference_testsTr   rv   rI  re   ri   z<============================================================zTesting with r?   z (rB  test_z.jsonlr*  test_inference_RSIAgent_processz--env.total_stepsz--env.group_sizez--env.use_wandbtruez--env.wandb_namez--env.data_path_to_save_groupsz--env.tokenizer_namer8   r   z--env.max_token_lengthr    z--env.max_num_workersr!   z--env.max_batches_offpolicyr%   z--openai.base_urlzhttps://openrouter.ai/api/v1z--openai.api_keyz--openai.model_namez--openai.server_typer9   z--openai.health_checkfalse c              3   4   K   | ]}t          |          V  d S N)rJ   )r   cs     r   	<genexpr>z$rl_test_inference.<locals>.<genexpr>w  s(      //a3q66//////r   z***API_KEY***z	Command: zWorking dir: zWandB run:   u
    steps × z completions = z	 rolloutsro  r   )	modelr?   ro  r_  output_filestepssteps_testedtotal_completionscorrect_completionsr   r   rB   c                   K   	 |                                   d{V }|sdS |                                                                |                               t	          fddD                       rt          d|             )z0Read stream line by line and print in real-time.TNc              3   D   K   | ]}|                                 v V  d S r  )lower)r   kwdecodeds     r   r  z9rl_test_inference.<locals>.read_stream.<locals>.<genexpr>  s0      yyR20yyyyyyr   )
processinggrouprO  progress%r   r  )readlinedecoderstripr   anyprint)stream
lines_listprefixliner  s       @r   read_streamz&rl_test_inference.<locals>.read_stream  s      6!'!2!2222222D "kkmm2244G%%g...yyyy;xyyyyy 64647445556r   u   📊 u   ⚠️ iX  r   r   zReturn code: zSTDOUT:
z(empty)
zSTDERR:
z  Log file: zProcess exited with code ir   r   log_fileu   
  ❌ Error: z  Last errors:z    u%   
  ✅ Process completed successfullyz  Output file: z  File exists: rf   scoresr     r  c              3   &   K   | ]}|d k    dV  dS )r   r  Nr   r   ss     r   r  z$rl_test_inference.<locals>.<genexpr>  s&      -I-IA1q55a5555-I-Ir   r  r  )rO  completionscorrectr  z  Completed z stepszOutput file not created: z"Process timed out after 10 minutesz
  Timeout!z	  Error: r   accuracyc              3   N   K   | ] }|                     d d          dk    dV  !dS )r  r   r  Nr   r  s     r   r  z$rl_test_inference.<locals>.<genexpr>  s=      $b$b1!%%PY[\J]J]`aJaJaQJaJaJaJa$b$br   steps_with_correctstep_success_ratez  Results: z correctz  Accuracy: z.1%r  c                 F    g | ]}|                     d d          dk    |S )r  r   r  r|  s     r   r  z%rl_test_inference.<locals>.<listcomp>  s2    ZZZAQUU>ST=U=UXY=Y=Ya=Y=Y=Yr   c                 .    |                      dd          S )Nr  r   r  )r(   s    r   <lambda>z#rl_test_inference.<locals>.<lambda>  s    j!8L8L r   )keyr  c              3   B   K   | ]}|                     d d          V  dS )r  r   Nr  r|  s     r   r  z$rl_test_inference.<locals>.<genexpr>  s0      ==j!$$======r   )steps_requestedr  models_succeeded
best_modelavg_accuracyenvironment_workingoutput_directoryrZ  )rB   )0r^   r  r	  r   r   r]   r?   TEST_MODELSr  rA   r   r   r   replacer  rJ   r1  r2  r   r   r   joinr   r   create_subprocess_execr[   PIPEwait_forgatherr   r   TimeoutErrorr   r   rm   writer   r   r   loadsr   sumr   JSONDecodeErrorr   roundmaxbool)$rs  rt  ru  r-   r   r8   test_modelstotal_rollouts_per_modelr~  resultstest_output_dir
model_infomodel_idmodel_safe_namer  test_run_idrU   cmdcmd_strcmd_displaymodel_resultsr  stdout_linesstderr_linesr  r  stdout_textstderr_textr   r  r   r  r  r   r  working_modelss$     `                                 r   rl_test_inferencer    su     <  zV
   	 i,--G zR
   	
 H  8|##HE $  z>\>>>
   	
  "CCCC+CCC 	TSSFSSSK!  ):5-K0@0@@N $$."$":,	
 
 
 
G !22O4(((! B7 B7
d#"**344m6mm?j0??H???@@@k &(V(V(V(V(V(VV $*,,''+PLPP;PP

N
$.
09
  
 "%Y
 	
 !$J	

 

  &
 
 !/
 -
 /2+.>.>
 #
 %2%$89I$J
 %
 '*-*>?Q*R&S&S
 $
 &)u)=>O)P%Q%Q
 *
 ,/}U/CD[/\+]+]
   !
  "@!
" #
" !(#
$ "%
$ $,%
& #'
& %-'
( $)
( &-)
0 ((//3/////oog??'+''(((3133444,N,,---f9ff
ffC[fffggg v&('{++!"#$

 

m	##:).).+,,	        G LL&)U)U)U)U)U)UUH
6 
6 
6 
6
&N#GNL'JJ#GNL)LL              '    ,,..        ))L11K))L11K h$$ 	43K333444?(;???@@@>(:>>>???V(((.F...///2{333V(((.F...///2{333	4 	4 	4 	4 	4 	4 	4 	4 	4 	4 	4 	4 	4 	4 	4 +++,,,!Q&&)YWEW)Y)Yg&*5eff*=h'*5eff*=h',/MMj)@g(>@@AAA -*+++ ,RSS 1 - -mTmm,,,,>???555666>(:(:(<(<>>??? %%'' Wk3// )1$% ) )D#'::<<D#' ) ()'+z$'7'7)-(B)?)? -n = = = B = = = -.A B B Bc&kk Q B B B*--I-I-I-I-I*I*I -.C D D D O D D D -g 6 = =,9.,I36v;;/6.4	?" ?" !# !# !# !# $(#7 ) ) ) ()%)) ) ) ) ) ) ) ) ) ) ) ) ) ) ), N~)FNNNOOOO-V-V-VM'*# 	  	  	 %IM'", 	# 	# 	#%(VVM'"/a//""""""""	#
 ,-11(-34}EX7YY[\) )M*%% )*M*%(1,,!$$b$bg0F$b$b$b!b!b2DM./16"]>%BBA2 2M-.. 34M./12M-.oM*?@oo=QdCeoooppp<]:6<<<=== ''6666 [Z!9ZZZN %[))//Zhrc..L.LMMMgVVnr #==n=====N@S@SSUV
 
 
!"#N3300
 
GI :ga((((s   A"^*4AP^*%P))A^* C(U4(^*4U8	8^*;U8	<D^*	]<&B3]]<],	)]<+],	,]<0^*<^ 	 ^*^ 	%^**#`	`$``c                  "    t           j        dk    S )z
    Check if Python version meets the minimum for RL tools.
    
    tinker-atropos depends on the 'tinker' package which requires Python >= 3.11.
    )r      )r   version_infor   r   r   check_rl_python_versionr  '  s     w&&r   c                      t                      sdS t          j        d          } t          j        d          }t          |           ot          |          S )z
    Check if required API keys and Python version are available.
    
    RL training requires:
    - Python >= 3.11 (tinker package requirement)
    - TINKER_API_KEY for the Tinker training API
    - WANDB_API_KEY for Weights & Biases metrics
    Fr   WANDB_API_KEY)r  r   r   r  )
tinker_key	wandb_keys     r   check_rl_api_keysr  0  sO     #$$ u+,,J	/**I
/Y/r   c                  >   g } t                      s:|                     dt          j        j         dt          j        j         d           t          j        d          s|                     d           t          j        d          s|                     d           | S )zV
    Get list of missing requirements for RL tools (API keys and Python version).
    zPython >= 3.11 (current: .rB  r   r  )r  r   r   r  majorminorr   r   )missings    r   get_missing_keysr  @  s     G"$$ ge33C3IeeCL\Lbeeefff9%&& )'(((9_%% ('''Nr   )registryr  zList all available RL environments. Returns environment names, paths, and descriptions. TIP: Read the file_path with file tools to understand how each environment works (verifiers, data loading, rewards).object)r   
propertiesrequired)r?   rC   
parametersr  zSelect an RL environment for training. Loads the environment's default configuration. After selecting, use rl_get_current_config() to see settings and rl_edit_config() to modify them.stringz=Name of the environment to select (from rl_list_environments))r   rC   r!  zGet the current environment configuration. Returns only fields that can be modified: group_size, max_token_length, total_steps, steps_per_eval, use_wandb, wandb_name, max_num_workers.r(  zUpdate a configuration field. Use rl_get_current_config() first to see all available fields for the selected environment. Each environment has different configurable options. Infrastructure settings (tokenizer, URLs, lora_rank, learning_rate) are locked.zMName of the field to update (get available fields from rl_get_current_config)rC   zNew value for the fieldr"  r}   r9  zStart a new RL training run with the current environment and config. Most training parameters (lora_rank, learning_rate, etc.) are fixed. Use rl_edit_config() to set group_size, batch_size, wandb_project before starting. WARNING: Training takes hours.r`  zGet status and metrics for a training run. RATE LIMITED: enforces 30-minute minimum between checks for the same run. Returns WandB metrics: step, state, reward_mean, loss, percent_correct.z#The run ID from rl_start_training()rb  zrStop a running training job. Use if metrics look bad, training is stagnant, or you want to try different settings.zThe run ID to stoprj  znGet final results and metrics for a completed training run. Returns final metrics and path to trained weights.zThe run ID to get results forrl  z@List all training runs (active and completed) with their status.r  a8  Quick inference test for any environment. Runs a few steps of inference + scoring using OpenRouter. Default: 3 steps x 16 completions = 48 rollouts per model, testing 3 models = 144 total. Tests environment loading, prompt construction, inference parsing, and verifier logic. Use BEFORE training to catch issues.integerz@Number of steps to run (default: 3, recommended max for testing))r   rC   r   z1Completions per step (default: 16, like training)arrayr   zgOptional list of OpenRouter model IDs. Default: qwen/qwen3-8b, z-ai/glm-4.7-flash, minimax/minimax-m2.7)r   r   rC   rs  rt  ru  r   r  u   🧪rlc                     t                      S r  )r  argsr  s     r   r  r  a  s    355 r   )r?   emojitoolsetschemahandlercheck_fnrequires_envis_asyncc                 J    t          |                     dd                    S )Nr?   rB   r  )r  r   r  s     r   r  r  c  s     4$((62:N:NOOO r   c                     t                      S r  )r!  r  s     r   r  r  e  s    466 r   c                 r    t          |                     dd          |                     d                    S )Nr"  rB   r}   r  )r(  r   r  s     r   r  r  g  s1    ~DHHWb4I4IQUQYQYZaQbQbccc r   c                     t                      S r  )r9  r  s     r   r  r  i  s    022 r   c                 J    t          |                     dd                    S NrN   rB   )rN   )r`  r   r  s     r   r  r  k  s    dhhx6L6LMMM r   c                 J    t          |                     dd                    S r
  )rb  r   r  s     r   r  r  m  s     /txx"7M7MNNN r   c                 J    t          |                     dd                    S r
  )rj  r   r  s     r   r  r  o  s    ~TXXh5K5KLLL r   c                     t                      S r  )rl  r  s     r   r  r  q  s
    |~~ r   c                     t          |                     dd          |                     dd          |                     d                    S )Nrs  r   rt  rr  ru  r  )r  r   r  s     r   r  r  s  si    0488KQR;S;S`d`h`hiuwy`z`z  DH  DL  DL  MU  DV  DV   W   W   W r   )[rI   rn   r   importlib.utilr   r  r   r[   r   r   r1  loggingr   r6  dataclassesr   pathlibr   typingr   r   r   r	   hermes_constantsr
   	getLoggerrF   r   __file__parentHERMES_ROOTr   rj   r3  r   r   r   setr   r'  r   r>   rM   r]   rK   r^   rJ   r_   r`   ra   rb   rZ   rU  r   r   r   r   r   r   r  r  r!  r(  r9  r`  rb  rj  rl  r  DEFAULT_NUM_STEPSDEFAULT_GROUP_SIZEintr  r  r  r  r  tools.registryr  RL_LIST_ENVIRONMENTS_SCHEMARL_SELECT_ENVIRONMENT_SCHEMARL_GET_CURRENT_CONFIG_SCHEMARL_EDIT_CONFIG_SCHEMARL_START_TRAINING_SCHEMARL_CHECK_STATUS_SCHEMARL_STOP_TRAINING_SCHEMARL_GET_RESULTS_SCHEMARL_LIST_RUNS_SCHEMARL_TEST_INFERENCE_SCHEMA_rl_envregisterr   r   r   <module>r*     sX    : 


       				     



           ! ! ! ! ! !       , , , , , , , , , , , , , , , , , ,		8	$	$ d8nn#*!$44 &)99NJ !I-?v%5& & & *5 !"  *2%(#	
 	

  $(#$&  C" "H S**5"55::<<==  ( ( ( ( ( ( ( ( 3 3 3 3 3 3 3 3" (*tO$ ) ) )"hsm " " ""$c3h $ $ $:< 4T#tCH~"5667 < < <$&d3=! & & &') De$ ) ) ) $ :D1 : : : :zQ# Q$sDcN7J2K Q Q Q Qh- - -r& r& r& r& r& r&j8    B&+( &+ &+ &+ &+Z**C ** ** ** **Z4c 4c 4 4 4 4v,S , , , ,^+ +C +C + + + +df f f f fRV(# V(# V( V( V( V(r3 3    @)( )( )( )( )( )(XC    : JAA8LL!>GLL    '("&e) e)e)e) T#Ye) 		e) e) e) e)X	' ' ' ' '04 0 0 0 0 $s)    " $ # # # # #'=  O]  v~  NP  ^`  ma  ma  b  b (?  QJ  ck  |B  MU  fe  Df  Df  {g  v|  u}  Z~  Z~       (?  QJ  ck  {}  KM  ZN  ZN   O   O !1  CC  \d  GO  `o  ~p  ~p  |I  Kd  {e  tf  tf  u|  ~E	  tF	  SG	  SG	  H	  H	 $7  IF  _g  wy  GI  VJ  VJ  K  K "3  EC  \d  u}  HP  aF  G  G  tH  W_  V`  Sa  Sa  b  b #5  G{  T\  mu  @H  Ym  wn  wn  lo  ~F  }G  KH  KH  I  I !1  Cs  LT  em  x@  Qp  oq  oq  dr  AI  @J  CK  CK  L  L -  ?A  Zb  rt  BD  QE  QE  F  F $7  IC  \d  KT  eg  tu  Bv  Bv  O	X	  i	\
  i
k
  F	l
  F	l
  AH  TZ  \d  Se  v_  x
`  x
`  ta  ta  oq  Sr  Sr  s  s _
-  -VTRm55@Q`grvx x x x  .fdSoOOZk  {B  MQR R R R  .fdSo66ARahswy y y y  'vtLaccn  OV  aef f f f  *&$Og22=N]dosu u u u  (McMMXix  KOP P P P  )NeNNYj  zA  LPQ Q Q Q  'vtLaLLWhw~  JNO O O O  ~VTJ]--8IX_jnp p p p  *&$Og W  WWtE E E E E Er   