kavinkrisher commited on
Commit
3ea5812
Β·
1 Parent(s): a46052d

LLM: show full script+bugs in initial prompt, clearer error feedback

Browse files
Files changed (1) hide show
  1. gradio_app.py +22 -10
gradio_app.py CHANGED
@@ -1051,24 +1051,34 @@ CRITICAL RULES:
1051
 
1052
  def _obs_to_prompt(obs):
1053
  parts = [f"Stage {obs.current_stage} | step {obs.stage_step_number}"]
 
 
1054
  if obs.last_run_error:
1055
- parts.append(f"\nERROR:\n{obs.last_run_error[:500]}")
1056
  if obs.last_run_output:
1057
- parts.append(f"\nOUTPUT:\n{obs.last_run_output[:300]}")
1058
  if obs.actor_feedback:
1059
- parts.append(f"\nREVIEWER FEEDBACK:\n{obs.actor_feedback[:300]}")
1060
- if obs.script_content:
1061
- parts.append(f"\nCURRENT SCRIPT:\n{obs.script_content[:800]}")
1062
  return "\n".join(parts)
1063
 
1064
  _env = DataEngEnvironment()
1065
  init_obs = _env.reset()
1066
  yield emit(" βœ“ Private env initialised β€” Stage 1\n")
1067
 
 
 
 
 
 
 
 
 
 
 
1068
  LABELS = {1:"Data Repair",2:"Training Monitor",3:"Eval Validation",4:"Deploy Gate"}
1069
  msgs = [
1070
  {"role": "system", "content": _SYS},
1071
- {"role": "user", "content": _obs_to_prompt(init_obs) + "\n\nWhat is your first action?"},
1072
  ]
1073
  cur_stage = 1
1074
  step = 0
@@ -1142,13 +1152,15 @@ CRITICAL RULES:
1142
  msgs.append({"role": "assistant", "content": json.dumps(action)})
1143
  feedback = _obs_to_prompt(obs)
1144
  if atype == "run_script" and obs.last_run_error:
1145
- feedback += "\n\nScript errored. Use edit_script with a FULL replacement to fix all bugs."
1146
  elif atype == "run_script" and not obs.last_run_error:
1147
- feedback += "\n\nClean run β€” Accuracy printed. Call submit now."
1148
  elif atype == "edit_script":
1149
- feedback += "\n\nScript replaced. Now call run_script to verify."
1150
  elif atype == "query_actor":
1151
- feedback += "\n\nReviewer replied above. Now edit_script to apply the fix, then run_script."
 
 
1152
  msgs.append({"role": "user", "content": feedback})
1153
  if len(msgs) > 16:
1154
  msgs = [msgs[0]] + msgs[1:3] + msgs[-10:]
 
1051
 
1052
  def _obs_to_prompt(obs):
1053
  parts = [f"Stage {obs.current_stage} | step {obs.stage_step_number}"]
1054
+ if obs.script_content:
1055
+ parts.append(f"\nCURRENT SCRIPT:\n{obs.script_content[:1000]}")
1056
  if obs.last_run_error:
1057
+ parts.append(f"\nRUN ERROR:\n{obs.last_run_error[:500]}")
1058
  if obs.last_run_output:
1059
+ parts.append(f"\nRUN OUTPUT:\n{obs.last_run_output[:300]}")
1060
  if obs.actor_feedback:
1061
+ parts.append(f"\nREVIEWER:\n{obs.actor_feedback[:300]}")
 
 
1062
  return "\n".join(parts)
1063
 
1064
  _env = DataEngEnvironment()
1065
  init_obs = _env.reset()
1066
  yield emit(" βœ“ Private env initialised β€” Stage 1\n")
1067
 
1068
+ # Build the initial user message with full broken script visible
1069
+ _init_prompt = _obs_to_prompt(init_obs)
1070
+ # Hint: show exactly what needs to change so LLM doesn't hallucinate
1071
+ _init_prompt += (
1072
+ "\n\nThe script above has TWO bugs:\n"
1073
+ "1. Line uses 'age_years' but the column is named 'age' β€” fix the column name\n"
1074
+ "2. No missing-value handling β€” add df.dropna() before feature selection\n"
1075
+ "Write a FULL replacement script fixing BOTH bugs, then call run_script."
1076
+ )
1077
+
1078
  LABELS = {1:"Data Repair",2:"Training Monitor",3:"Eval Validation",4:"Deploy Gate"}
1079
  msgs = [
1080
  {"role": "system", "content": _SYS},
1081
+ {"role": "user", "content": _init_prompt},
1082
  ]
1083
  cur_stage = 1
1084
  step = 0
 
1152
  msgs.append({"role": "assistant", "content": json.dumps(action)})
1153
  feedback = _obs_to_prompt(obs)
1154
  if atype == "run_script" and obs.last_run_error:
1155
+ feedback += "\n\nScript errored (see RUN ERROR above). Use edit_script with a FULL script replacement β€” include ALL imports and fix EVERY bug shown in the error."
1156
  elif atype == "run_script" and not obs.last_run_error:
1157
+ feedback += "\n\nClean run β€” script printed output with no error. Call submit NOW."
1158
  elif atype == "edit_script":
1159
+ feedback += "\n\nScript replaced. Call run_script to verify the fix."
1160
  elif atype == "query_actor":
1161
+ feedback += "\n\nReviewer responded above. Use edit_script with full script to apply the fix, then run_script."
1162
+ elif atype == "submit":
1163
+ feedback += "\n\nStage not yet passing. Review the current script and error carefully, then edit_script to fix remaining bugs."
1164
  msgs.append({"role": "user", "content": feedback})
1165
  if len(msgs) > 16:
1166
  msgs = [msgs[0]] + msgs[1:3] + msgs[-10:]