| 1 | """Core compaction logic for the compaction plugin.""" |
| 2 | import os |
| 3 | from collections import deque |
| 4 | |
| 5 | import models as models_module |
| 6 | from agent import Agent |
| 7 | from helpers import files, tokens |
| 8 | from helpers.history import History, clear_responses_provider_state, output_text |
| 9 | from helpers.persist_chat import ( |
| 10 | export_json_chat, |
| 11 | get_chat_folder_path, |
| 12 | save_tmp_chat, |
| 13 | remove_msg_files, |
| 14 | ) |
| 15 | from helpers.state_monitor_integration import mark_dirty_all |
| 16 | from helpers.localization import Localization |
| 17 | |
| 18 | MIN_COMPACTION_TOKENS = 1000 |
| 19 | COMPACTION_CHUNK_TARGET_RATIO = 0.9 |
| 20 | COMPACTION_CHUNK_VERIFY_RATIO = 0.98 |
| 21 | |
| 22 | from plugins._model_config.helpers.model_config import ( |
| 23 | get_chat_model_config, |
| 24 | get_utility_model_config, |
| 25 | get_preset_by_name, |
| 26 | build_model_config, |
| 27 | build_chat_model, |
| 28 | build_utility_model, |
| 29 | ) |
| 30 | |
| 31 | |
| 32 | def _save_pre_compaction_backup(context, full_text: str) -> dict[str, str]: |
| 33 | """Save the original chat as JSON and plain text before compaction. |
| 34 | |
| 35 | Returns dict with 'json' and 'txt' absolute file paths. |
| 36 | """ |
| 37 | timestamp = Localization.get().now().strftime("%Y%m%d-%H%M%S") |
| 38 | backup_dir = os.path.join(get_chat_folder_path(context.id), "backups") |
| 39 | os.makedirs(backup_dir, exist_ok=True) |
| 40 | |
| 41 | json_path = os.path.join(backup_dir, f"pre-compact-{timestamp}.json") |
| 42 | txt_path = os.path.join(backup_dir, f"pre-compact-{timestamp}.txt") |
| 43 | |
| 44 | json_content = export_json_chat(context) |
| 45 | files.write_file(json_path, json_content) |
| 46 | files.write_file(txt_path, full_text) |
| 47 | |
| 48 | return {"json": json_path, "txt": txt_path} |
| 49 | |
| 50 | |
| 51 | def _build_model(use_chat_model: bool, preset_name: str | None, agent): |
| 52 | """Build the LLM model for compaction based on user selection. |
| 53 | |
| 54 | If preset_name is given, builds from that preset's config. |
| 55 | Otherwise falls back to the agent's currently configured model. |
| 56 | """ |
| 57 | if preset_name: |
| 58 | preset = get_preset_by_name(preset_name) |
| 59 | if preset: |
| 60 | model_key = "chat" if use_chat_model else "utility" |
| 61 | cfg = preset.get(model_key, {}) |
| 62 | if cfg.get("provider") or cfg.get("name"): |
| 63 | mc = build_model_config(cfg, models_module.ModelType.CHAT) |
| 64 | return cfg, models_module.get_chat_model( |
| 65 | mc.provider, mc.name, model_config=mc, **mc.build_kwargs() |
| 66 | ) |
| 67 | |
| 68 | if use_chat_model: |
| 69 | cfg = get_chat_model_config(agent) |
| 70 | return cfg, build_chat_model(agent) |
| 71 | else: |
| 72 | cfg = get_utility_model_config(agent) |
| 73 | return cfg, build_utility_model(agent) |
| 74 | |
| 75 | |
| 76 | async def run_compaction( |
| 77 | context, |
| 78 | use_chat_model: bool = True, |
| 79 | preset_name: str | None = None, |
| 80 | ) -> None: |
| 81 | """ |
| 82 | Compact the chat history into a single summarized message. |
| 83 | |
| 84 | This function: |
| 85 | 1. Extracts the full conversation text |
| 86 | 2. Estimates token count and checks against model context window |
| 87 | 3. If needed, splits history and summarizes iteratively |
| 88 | 4. Calls the LLM to generate a comprehensive summary |
| 89 | 5. Replaces the history with a single context message containing the summary |
| 90 | 6. Resets the log and creates a response log item |
| 91 | 7. Persists the changes |
| 92 | |
| 93 | The function streams progress to the frontend via the log system. |
| 94 | If any error occurs, the original history is preserved. |
| 95 | """ |
| 96 | agent = context.agent0 |
| 97 | |
| 98 | try: |
| 99 | # Step 1: Extract full conversation text |
| 100 | history_output = agent.history.output() |
| 101 | full_text = output_text(history_output, ai_label="assistant", human_label="user") |
| 102 | |
| 103 | if not full_text.strip(): |
| 104 | raise ValueError("No conversation content to compact") |
| 105 | |
| 106 | # Step 2: Estimate tokens, resolve model, and compute context budget |
| 107 | token_count = tokens.approximate_tokens(full_text) |
| 108 | |
| 109 | resolved_cfg, model = _build_model(use_chat_model, preset_name, agent) |
| 110 | ctx_length = int(resolved_cfg.get("ctx_length", 128000)) if resolved_cfg else 128000 |
| 111 | max_input_tokens = int(ctx_length * 0.7) |
| 112 | |
| 113 | # Step 3: Create progress log item (count user-visible messages only) |
| 114 | visible_types = {"user", "response"} |
| 115 | visible_count = sum(1 for item in context.log.logs if item.type in visible_types) |
| 116 | log_item = context.log.log( |
| 117 | type="info", |
| 118 | heading="Compacting chat history...", |
| 119 | content=f"Analyzing {visible_count} messages (~{token_count} tokens)...", |
| 120 | ) |
| 121 | |
| 122 | # Step 4: Handle large histories by chunking if necessary |
| 123 | if token_count > max_input_tokens: |
| 124 | summary = await _compact_large_history( |
| 125 | agent, full_text, token_count, max_input_tokens, log_item, model |
| 126 | ) |
| 127 | else: |
| 128 | summary = await _compact_single_pass( |
| 129 | agent, full_text, log_item, model |
| 130 | ) |
| 131 | |
| 132 | if not summary or not summary.strip(): |
| 133 | raise ValueError("Compaction produced empty summary") |
| 134 | |
| 135 | # Step 5: Save pre-compaction backup before destroying history |
| 136 | backup_paths = _save_pre_compaction_backup(context, full_text) |
| 137 | |
| 138 | # Step 6: Replace history with compacted version |
| 139 | backup_note = ( |
| 140 | f"\n\n---\n" |
| 141 | f"*Pre-compaction backup of the full original conversation:*\n" |
| 142 | f"- `{backup_paths['txt']}`" |
| 143 | ) |
| 144 | compacted_content = f"## Context compacted\n\n{summary}{backup_note}" |
| 145 | |
| 146 | agent.history = History(agent=agent) |
| 147 | # History summaries are context, not orphaned assistant turns. |
| 148 | agent.history.add_message(ai=False, content=compacted_content) |
| 149 | clear_responses_provider_state(agent) |
| 150 | agent.data.pop(Agent.DATA_NAME_CTX_WINDOW, None) |
| 151 | |
| 152 | # Clear subordinate chain |
| 153 | agent.data.pop(Agent.DATA_NAME_SUBORDINATE, None) |
| 154 | context.streaming_agent = None |
| 155 | |
| 156 | # Step 7: Reset log and create response |
| 157 | context.log.reset() |
| 158 | context.log.log( |
| 159 | type="response", |
| 160 | heading="Context compacted", |
| 161 | content=compacted_content, |
| 162 | update_progress="none", |
| 163 | ) |
| 164 | |
| 165 | # Step 8: Persist and notify |
| 166 | save_tmp_chat(context) |
| 167 | remove_msg_files(context.id) |
| 168 | |
| 169 | # Step 9: Force progress bar to inactive state LAST |
| 170 | # This must happen after all log operations and persist |
| 171 | context.log.set_progress("Waiting for input", 0, False) |
| 172 | mark_dirty_all(reason="plugins.compaction.compact_chat") |
| 173 | |
| 174 | except Exception as e: |
| 175 | # Log error but don't modify history |
| 176 | context.log.log( |
| 177 | type="error", |
| 178 | heading="Compaction Failed", |
| 179 | content=str(e), |
| 180 | ) |
| 181 | mark_dirty_all(reason="plugins.compaction.compact_chat_error") |
| 182 | raise |
| 183 | |
| 184 | |
| 185 | async def _compact_single_pass(agent, full_text: str, log_item, model) -> str: |
| 186 | """Compact history in a single LLM call using the provided model.""" |
| 187 | system_prompt = agent.read_prompt("compact.sys.md") |
| 188 | user_prompt = agent.read_prompt("compact.msg.md", conversation=full_text) |
| 189 | |
| 190 | async def stream_cb(chunk: str, total: str): |
| 191 | if chunk: |
| 192 | log_item.stream(content=chunk) |
| 193 | |
| 194 | summary, _ = await model.unified_call( |
| 195 | system_message=system_prompt, |
| 196 | user_message=user_prompt, |
| 197 | response_callback=stream_cb, |
| 198 | ) |
| 199 | return summary |
| 200 | |
| 201 | |
| 202 | async def _compact_large_history( |
| 203 | agent, full_text: str, token_count: int, max_input_tokens: int, log_item, model |
| 204 | ) -> str: |
| 205 | """Handle large histories by splitting into chunks and summarizing iteratively.""" |
| 206 | chunks = _split_text_for_compaction(agent, full_text, token_count, max_input_tokens) |
| 207 | log_item.update( |
| 208 | content=f"History is large (~{token_count} tokens). Splitting into {len(chunks)} chunks...", |
| 209 | ) |
| 210 | |
| 211 | summaries = [] |
| 212 | for i, chunk in enumerate(chunks, 1): |
| 213 | log_item.update(content=f"Summarizing part {i}/{len(chunks)}...") |
| 214 | |
| 215 | system_prompt = agent.read_prompt("compact.sys.md") |
| 216 | user_prompt = agent.read_prompt("compact.msg.md", conversation=chunk) |
| 217 | |
| 218 | chunk_summary, _ = await model.unified_call( |
| 219 | system_message=system_prompt, |
| 220 | user_message=user_prompt, |
| 221 | ) |
| 222 | summaries.append(chunk_summary) |
| 223 | |
| 224 | combined = "\n\n---\n\n".join(summaries) |
| 225 | log_item.update(content="Creating final summary from parts...") |
| 226 | |
| 227 | final_prompt = agent.read_prompt("compact.sys.md") |
| 228 | final_user = agent.read_prompt( |
| 229 | "compact.msg.md", |
| 230 | conversation=f"This is a multi-part conversation. Here are summaries of each part:\n\n{combined}", |
| 231 | ) |
| 232 | |
| 233 | async def stream_cb(chunk: str, total: str): |
| 234 | if chunk: |
| 235 | log_item.stream(content=chunk) |
| 236 | |
| 237 | final_summary, _ = await model.unified_call( |
| 238 | system_message=final_prompt, |
| 239 | user_message=final_user, |
| 240 | response_callback=stream_cb, |
| 241 | ) |
| 242 | return final_summary |
| 243 | |
| 244 | |
| 245 | def _split_text_for_compaction( |
| 246 | agent, full_text: str, token_count: int, max_input_tokens: int |
| 247 | ) -> list[str]: |
| 248 | """Split large compaction input into prompt-safe chunks. |
| 249 | |
| 250 | The previous line-midpoint split left a single-line payload as one empty |
| 251 | chunk plus one still-oversized chunk. This splitter derives a conservative |
| 252 | character target from the measured token density, then verifies each prompt |
| 253 | and keeps splitting any chunk that still exceeds the model input budget. |
| 254 | """ |
| 255 | text = full_text or "" |
| 256 | if not text: |
| 257 | return [] |
| 258 | |
| 259 | prompt_overhead = _compaction_input_tokens(agent, "") |
| 260 | usable_tokens = max(max_input_tokens - prompt_overhead, 1) |
| 261 | target_tokens = max(int(usable_tokens * COMPACTION_CHUNK_TARGET_RATIO), 1) |
| 262 | |
| 263 | if token_count <= target_tokens: |
| 264 | return [text] |
| 265 | |
| 266 | chars_per_token = max(len(text) / max(token_count, 1), 0.01) |
| 267 | target_chars = max(int(target_tokens * chars_per_token), 1) |
| 268 | chunks = _split_text_by_chars(text, target_chars) |
| 269 | |
| 270 | verified: list[str] = [] |
| 271 | max_verified_tokens = max(int(max_input_tokens * COMPACTION_CHUNK_VERIFY_RATIO), 1) |
| 272 | pending = deque(chunk for chunk in chunks if chunk) |
| 273 | |
| 274 | while pending: |
| 275 | chunk = pending.popleft() |
| 276 | if not chunk: |
| 277 | continue |
| 278 | |
| 279 | if ( |
| 280 | len(chunk) <= 1 |
| 281 | or _compaction_input_tokens(agent, chunk) <= max_verified_tokens |
| 282 | ): |
| 283 | verified.append(chunk) |
| 284 | continue |
| 285 | |
| 286 | split_chunks = _split_text_by_chars(chunk, max(len(chunk) // 2, 1)) |
| 287 | if len(split_chunks) <= 1: |
| 288 | verified.append(chunk) |
| 289 | else: |
| 290 | pending.extendleft(reversed(split_chunks)) |
| 291 | |
| 292 | return verified |
| 293 | |
| 294 | |
| 295 | def _compaction_input_tokens(agent, conversation: str) -> int: |
| 296 | system_prompt = agent.read_prompt("compact.sys.md") |
| 297 | user_prompt = agent.read_prompt("compact.msg.md", conversation=conversation) |
| 298 | return tokens.approximate_tokens(system_prompt) + tokens.approximate_tokens( |
| 299 | user_prompt |
| 300 | ) |
| 301 | |
| 302 | |
| 303 | def _split_text_by_chars(text: str, target_chars: int) -> list[str]: |
| 304 | if not text: |
| 305 | return [] |
| 306 | |
| 307 | target_chars = max(int(target_chars), 1) |
| 308 | chunks: list[str] = [] |
| 309 | start = 0 |
| 310 | length = len(text) |
| 311 | |
| 312 | while start < length: |
| 313 | end = min(start + target_chars, length) |
| 314 | if end < length: |
| 315 | floor = start + max((end - start) // 2, 1) |
| 316 | split_at = text.rfind("\n", floor, end) |
| 317 | if split_at == -1: |
| 318 | split_at = text.rfind(" ", floor, end) |
| 319 | if split_at > start: |
| 320 | end = split_at + 1 |
| 321 | |
| 322 | chunk = text[start:end] |
| 323 | if chunk: |
| 324 | chunks.append(chunk) |
| 325 | start = end |
| 326 | |
| 327 | return chunks |
| 328 | |
| 329 | |
| 330 | async def get_compaction_stats(context) -> dict: |
| 331 | """ |
| 332 | Get statistics about the current chat for the confirmation modal. |
| 333 | |
| 334 | Returns: |
| 335 | dict with message_count, token_count, model_name |
| 336 | """ |
| 337 | agent = context.agent0 |
| 338 | |
| 339 | # Count user-visible conversation turns only |
| 340 | # 'user' = user sent a message, 'response' = agent final response |
| 341 | # Other types (agent, tool, code_exe, etc.) are intermediate processing steps |
| 342 | visible_types = {"user", "response"} |
| 343 | message_count = sum( |
| 344 | 1 for item in context.log.logs |
| 345 | if item.type in visible_types |
| 346 | ) |
| 347 | |
| 348 | # Estimate tokens |
| 349 | history_output = agent.history.output() |
| 350 | full_text = output_text(history_output, ai_label="assistant", human_label="user") |
| 351 | token_count = tokens.approximate_tokens(full_text) if full_text else 0 |
| 352 | |
| 353 | # Get model names for both chat and utility |
| 354 | chat_cfg = get_chat_model_config(agent) |
| 355 | utility_cfg = get_utility_model_config(agent) |
| 356 | chat_model_name = chat_cfg.get("name", "Default") if chat_cfg else "Default" |
| 357 | utility_model_name = utility_cfg.get("name", "Default") if utility_cfg else "Default" |
| 358 | |
| 359 | return { |
| 360 | "message_count": message_count, |
| 361 | "token_count": token_count, |
| 362 | "model_name": chat_model_name, |
| 363 | "chat_model_name": chat_model_name, |
| 364 | "utility_model_name": utility_model_name, |
| 365 | } |