Spaces:
Running
Running
File size: 13,071 Bytes
399b80c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 | """Conversation log formatting for execution analysis.
Converts ``conversations.jsonl`` entries into a priority-based text block
suitable for LLM analysis prompts. All functions are pure (stateless).
Priority levels (lower = more important):
0 β CRITICAL : User instruction (never truncated)
1 β CRITICAL : Final iteration assistant response (never truncated)
2 β HIGH : Tool calls (name + args) AND tool errors β kept together
3 β HIGH : Non-final assistant reasoning; tool results with embedded summary
4 β MEDIUM : Tool success results (try to preserve)
5 β LOW : System guidance messages between iterations
SKIP : Skill injection text, verbose system prompts (not included;
skill & tool info are provided separately in the prompt)
"""
from __future__ import annotations
import re
from typing import Any, Dict, List, Optional
# Per-section truncation limits (kept in sync with analyzer constants)
TOOL_ERROR_MAX_CHARS = 1000
TOOL_SUCCESS_MAX_CHARS = 800
TOOL_ARGS_MAX_CHARS = 500
TOOL_SUMMARY_MAX_CHARS = 1500
def format_conversations(
conversations: List[Dict[str, Any]],
budget: int,
) -> str:
"""Format ``conversations.jsonl`` entries into a readable text block.
Uses priority-based truncation instead of simple tail-truncation.
When total exceeds *budget*:
1. Include all priority β€ 3 (CRITICAL + HIGH) segments in full.
2. Add MEDIUM + LOW segments until budget is exhausted, truncating
if possible.
3. If even HIGH content exceeds budget, keep priority 0-1 in full,
budget-allocate priority 2, and summarize priority 3.
"""
# Count total iterations for priority assignment
total_iters = sum(
1 for c in conversations if c.get("type") == "iteration"
)
# Phase 1: Collect all segments in chronological order with priority
segments: List[Dict[str, Any]] = []
for conv in conversations:
conv_type = conv.get("type", "")
if conv_type == "setup":
_collect_setup_segments(conv, segments)
elif conv_type == "iteration":
_collect_iteration_segments(conv, total_iters, segments)
# Phase 2: Assemble with budget management
return _assemble_with_budget(segments, budget)
def _collect_setup_segments(
conv: Dict[str, Any],
segments: List[Dict[str, Any]],
) -> None:
"""Extract segments from a ``type: "setup"`` conversation entry.
Only the user instruction is extracted. System prompts (including skill
injection text and tool descriptions) are skipped β they are provided in
dedicated sections of the analysis prompt.
"""
for msg in conv.get("messages", []):
role = msg.get("role", "")
content = msg.get("content", "")
if not isinstance(content, str):
content = str(content)
if role == "user":
segments.append({
"priority": 0, # CRITICAL β always keep
"text": f"[USER INSTRUCTION]\n{content}",
"iteration": 0,
"role": "user",
"truncatable_to": None,
})
def _collect_iteration_segments(
conv: Dict[str, Any],
total_iters: int,
segments: List[Dict[str, Any]],
) -> None:
"""Extract segments from a ``type: "iteration"`` conversation entry.
Key design decisions:
- Tool calls and tool errors share the SAME high priority (2)
- Tool success results get MEDIUM priority (4)
- Shell agent results with embedded "Execution Summary" get HIGH (3).
"""
iteration = conv.get("iteration", "?")
is_last = (iteration == total_iters) if isinstance(iteration, int) else False
# Process delta_messages in order
for msg in conv.get("delta_messages", []):
role = msg.get("role", "")
content = msg.get("content", "")
if not isinstance(content, str):
content = str(content)
if role == "assistant":
# Assistant reasoning
if content:
priority = 1 if is_last else 3
segments.append({
"priority": priority,
"text": f"[Iter {iteration}] ASSISTANT: {content}",
"iteration": iteration,
"role": "assistant",
"truncatable_to": None,
})
# Tool calls
for tc in msg.get("tool_calls", []):
fn = tc.get("function", {})
fn_name = fn.get("name", "?")
fn_args = fn.get("arguments", "")
if isinstance(fn_args, str) and len(fn_args) > TOOL_ARGS_MAX_CHARS:
fn_args = fn_args[:TOOL_ARGS_MAX_CHARS] + "..."
segments.append({
"priority": 2, # HIGH β paired with tool results/errors
"text": f"[Iter {iteration}] TOOL_CALL: {fn_name}({fn_args})",
"iteration": iteration,
"role": "tool_call",
"truncatable_to": None,
})
elif role == "tool":
# Tool result
is_error = _is_error_result(content)
if is_error:
truncated = content[:TOOL_ERROR_MAX_CHARS]
if len(content) > TOOL_ERROR_MAX_CHARS:
truncated += f"... [truncated, total {len(content)} chars]"
segments.append({
"priority": 2, # HIGH β errors are critical, same tier as tool calls
"text": f"[Iter {iteration}] TOOL_ERROR: {truncated}",
"iteration": iteration,
"role": "tool_error",
"truncatable_to": None,
})
else:
# Check if result contains a self-generated summary
# (e.g. shell_agent produces "Execution Summary (N steps):")
summary = _extract_embedded_summary(content)
if summary:
# Show the embedded summary (high value, compact)
segments.append({
"priority": 3, # HIGH β self-generated summaries are informative
"text": f"[Iter {iteration}] TOOL_RESULT (with summary):\n{summary}",
"iteration": iteration,
"role": "tool_result",
"truncatable_to": 500,
})
else:
truncated = content[:TOOL_SUCCESS_MAX_CHARS]
if len(content) > TOOL_SUCCESS_MAX_CHARS:
truncated += f"... [truncated, total {len(content)} chars]"
segments.append({
"priority": 4, # MEDIUM β try to preserve success results
"text": f"[Iter {iteration}] TOOL_RESULT: {truncated}",
"iteration": iteration,
"role": "tool_result",
"truncatable_to": 300,
})
elif role == "system":
# System guidance between iterations (e.g. "Iteration N complete...")
if content:
segments.append({
"priority": 5, # LOW β guidance messages
"text": f"[Iter {iteration}] SYSTEM: {content}",
"iteration": iteration,
"role": "system",
"truncatable_to": 150,
})
def _assemble_with_budget(
segments: List[Dict[str, Any]],
budget: int,
) -> str:
"""Assemble segments into final text respecting the character budget.
Strategy:
1. Include all segments with priority β€ 3 (CRITICAL + HIGH) in full.
2. Add MEDIUM + LOW segments in chronological order until budget is hit.
3. If even HIGH-priority content exceeds budget, progressively truncate
older iterations while preserving user instruction and final iteration.
"""
# Calculate essential (priority β€ 3) size
essential = [s for s in segments if s["priority"] <= 3]
essential_chars = sum(len(s["text"]) for s in essential)
remaining_budget = budget - essential_chars
if remaining_budget < 0:
# Essential content alone exceeds budget β need to reduce
# Keep priority 0-1 (user instruction + final iteration) in full
# Truncate priority 2-3 (tool calls/errors + older assistant content)
return _assemble_essential_only(segments, budget)
# Build output in chronological order
output_parts: List[str] = []
used_chars = 0
skipped_count = 0
for seg in segments:
text = seg["text"]
priority = seg["priority"]
if priority <= 3:
# Essential β always include
output_parts.append(text)
used_chars += len(text) + 1
elif used_chars + len(text) + 1 <= budget:
# Within budget β include
output_parts.append(text)
used_chars += len(text) + 1
else:
# Over budget β try truncation
truncatable_to = seg.get("truncatable_to")
if truncatable_to and len(text) > truncatable_to:
truncated = text[:truncatable_to] + "... [budget-truncated]"
if used_chars + len(truncated) + 1 <= budget:
output_parts.append(truncated)
used_chars += len(truncated) + 1
continue
skipped_count += 1
if skipped_count > 0:
output_parts.append(
f"\n[... {skipped_count} lower-priority segment(s) omitted due to length ...]"
)
return "\n\n".join(output_parts)
def _assemble_essential_only(
segments: List[Dict[str, Any]],
budget: int,
) -> str:
"""Fallback: even essential content exceeds budget.
Keep:
- User instruction (priority 0) β never truncated
- Final iteration (priority 1) β never truncated
- Tool calls + tool errors (priority 2) β budget-allocated, truncated if needed
- Non-final assistant reasoning (priority 3) β heavily summarized
"""
output_parts: List[str] = []
used_chars = 0
# Pass 1: priority 0 and 1 (user instruction + final iteration)
for seg in segments:
if seg["priority"] <= 1:
output_parts.append(seg["text"])
used_chars += len(seg["text"]) + 1
remaining = budget - used_chars
# Pass 2: priority 2 (tool calls + tool errors) β budget-allocated
tool_segments = [s for s in segments if s["priority"] == 2]
if tool_segments:
per_segment_budget = max(400, remaining // (len(tool_segments) + 1))
for seg in tool_segments:
text = seg["text"]
if len(text) > per_segment_budget:
text = text[:per_segment_budget] + "... [budget-truncated]"
if used_chars + len(text) + 1 <= budget:
output_parts.append(text)
used_chars += len(text) + 1
# Pass 3: priority 3 (non-final assistant reasoning) β one-line summaries
assistants = [s for s in segments if s["priority"] == 3]
if assistants and used_chars < budget:
output_parts.append("\n--- Older iteration summaries ---")
for seg in assistants:
first_line = seg["text"].split("\n", 1)[0][:200]
if used_chars + len(first_line) + 1 > budget:
output_parts.append("[... remaining iterations omitted ...]")
break
output_parts.append(first_line)
used_chars += len(first_line) + 1
return "\n\n".join(output_parts)
def _is_error_result(content: str) -> bool:
"""Detect if a tool result represents an error."""
if not content:
return False
# Check common error patterns in the first 200 chars
head = content[:200].lower()
return (
content.startswith("[ERROR]")
or content.startswith("ERROR")
or "error" in head[:50]
or "task failed" in head
or "connection refused" in head
or "timed out" in head
or "traceback" in head
)
def _extract_embedded_summary(content: str) -> Optional[str]:
"""Extract self-generated summary from tool result content.
Shell agent results often contain an ``Execution Summary (N steps):``
block that provides a compact view of what happened internally.
This is more informative than the raw output.
"""
# Look for "Execution Summary (N steps):" pattern
match = re.search(
r"(Execution Summary \(\d+ steps?\):.*?)(?:={10,}|$)",
content,
re.DOTALL,
)
if match:
summary = match.group(1).strip()
# Also capture any "Summary:" line after the steps
summary_match = re.search(r"\nSummary:\s*(.+)", content)
if summary_match:
summary += f"\nConclusion: {summary_match.group(1).strip()}"
return summary[:TOOL_SUMMARY_MAX_CHARS]
return None
|