|
| 1 | +#!/usr/bin/env python3 |
| 2 | +"""mios-compact -- compact recent agent + system state into a single |
| 3 | +markdown digest that can be ingested as an OWUI knowledge artifact. |
| 4 | +
|
| 5 | +Operator directive 2026-05-17: "make sure there's tools to compact |
| 6 | +all this and artifact it natively for OWUI knowledge/database". The |
| 7 | +agent stack generates a lot of latent state (recent chats, hermes |
| 8 | +session decisions, daemon classifications, launch verifier failures, |
| 9 | +git commits) that's useful for the agent to RAG against on later |
| 10 | +turns. This helper pulls the lot, summarizes via the local CPU |
| 11 | +model, and writes a versioned markdown file under |
| 12 | +/var/lib/mios/compacted/. Pair with mios-knowledge-add to register |
| 13 | +the file as a Knowledge collection in OWUI. |
| 14 | +
|
| 15 | +Sections in the rendered digest: |
| 16 | + 1. Recent operator chats (last N user turns + agent responses) |
| 17 | + 2. Launch verifier failures (from daemon) |
| 18 | + 3. Recent hermes tool-call patterns |
| 19 | + 4. Daemon classify summaries (system log roll-up) |
| 20 | + 5. Git commits this session |
| 21 | +
|
| 22 | +Output: /var/lib/mios/compacted/<utc-iso>.md (timestamped, never |
| 23 | +overwrites previous digests; mios-knowledge-add picks the newest). |
| 24 | +
|
| 25 | +Usage: |
| 26 | + mios-compact # default: last 24h of activity |
| 27 | + mios-compact --since "12 hours ago" # parseable by `date` |
| 28 | + mios-compact --chats 10 # cap chat count |
| 29 | + mios-compact --out <path> # override output path |
| 30 | + mios-compact --stdout # print to stdout instead of file |
| 31 | + mios-compact --no-llm # skip the CPU summarization step |
| 32 | + # (raw section dumps only) |
| 33 | +
|
| 34 | +Exit codes: |
| 35 | + 0 = digest written (or printed) |
| 36 | + 1 = required state unreachable (OWUI db missing, ollama down) |
| 37 | + 64 = bad args |
| 38 | +""" |
| 39 | +from __future__ import annotations |
| 40 | + |
| 41 | +import argparse |
| 42 | +import datetime |
| 43 | +import json |
| 44 | +import os |
| 45 | +import sqlite3 |
| 46 | +import subprocess |
| 47 | +import sys |
| 48 | +import urllib.error |
| 49 | +import urllib.request |
| 50 | +from pathlib import Path |
| 51 | + |
| 52 | +OWUI_DB = Path(os.environ.get( |
| 53 | + "MIOS_OWUI_DB", "/var/lib/mios/open-webui/webui.db")) |
| 54 | +DAEMON_STATE = Path(os.environ.get( |
| 55 | + "MIOS_DAEMON_STATE", "/var/lib/mios/daemon/state.json")) |
| 56 | +LAUNCH_FAILURES = Path(os.environ.get( |
| 57 | + "MIOS_DAEMON_LAUNCH_FAILURES", "/var/lib/mios/daemon/launch_failures.json")) |
| 58 | +COMPACTED_DIR = Path(os.environ.get( |
| 59 | + "MIOS_COMPACTED_DIR", "/var/lib/mios/compacted")) |
| 60 | +REPO_ROOT = Path(os.environ.get("MIOS_REPO_ROOT", "/mnt/c/MiOS")) |
| 61 | +OLLAMA_URL = os.environ.get("MIOS_OLLAMA_URL", "http://127.0.0.1:11434") |
| 62 | +SUMMARY_MODEL = os.environ.get("MIOS_COMPACT_MODEL", "qwen2.5-coder:7b") |
| 63 | + |
| 64 | + |
| 65 | +def _ts() -> str: |
| 66 | + return datetime.datetime.utcnow().strftime("%Y-%m-%dT%H-%M-%SZ") |
| 67 | + |
| 68 | + |
| 69 | +def _since_to_epoch(since: str) -> int: |
| 70 | + """Parse a date(1)-compatible time string. Returns epoch seconds.""" |
| 71 | + try: |
| 72 | + out = subprocess.run( |
| 73 | + ["date", "-d", since, "+%s"], |
| 74 | + capture_output=True, text=True, check=True, timeout=4, |
| 75 | + ) |
| 76 | + return int(out.stdout.strip()) |
| 77 | + except Exception: |
| 78 | + return int(datetime.datetime.utcnow().timestamp()) - 86400 |
| 79 | + |
| 80 | + |
| 81 | +# --- 1. recent OWUI chats ---------------------------------------- |
| 82 | + |
| 83 | +def section_recent_chats(since_epoch: int, max_chats: int) -> str: |
| 84 | + if not OWUI_DB.is_file(): |
| 85 | + return "_OWUI db not present at " + str(OWUI_DB) + "_" |
| 86 | + rows = [] |
| 87 | + try: |
| 88 | + c = sqlite3.connect(str(OWUI_DB)) |
| 89 | + rows = c.execute( |
| 90 | + "SELECT id, chat, updated_at FROM chat " |
| 91 | + "WHERE updated_at > ? ORDER BY updated_at DESC LIMIT ?", |
| 92 | + (since_epoch, max_chats), |
| 93 | + ).fetchall() |
| 94 | + except Exception as e: |
| 95 | + return f"_chat scan failed: {type(e).__name__}: {e}_" |
| 96 | + if not rows: |
| 97 | + return "_no chats in window_" |
| 98 | + out = [] |
| 99 | + for cid, raw, updated_at in rows: |
| 100 | + try: |
| 101 | + d = json.loads(raw or "{}") |
| 102 | + except Exception: |
| 103 | + continue |
| 104 | + title = (d.get("title") or "(untitled)")[:80] |
| 105 | + messages = d.get("messages") or [] |
| 106 | + # Pair user+assistant pairs (last 3 of each) |
| 107 | + user_msgs = [m for m in messages if isinstance(m, dict) |
| 108 | + and m.get("role") == "user"][-3:] |
| 109 | + asst_msgs = [m for m in messages if isinstance(m, dict) |
| 110 | + and m.get("role") == "assistant"][-3:] |
| 111 | + out.append(f"### {title}") |
| 112 | + out.append(f"_chat_id: {cid[:12]}, updated: " |
| 113 | + f"{datetime.datetime.utcfromtimestamp(updated_at).isoformat()}Z_") |
| 114 | + for m in user_msgs: |
| 115 | + content = str(m.get("content") or "").strip()[:300] |
| 116 | + out.append(f"- **user**: {content}") |
| 117 | + for m in asst_msgs: |
| 118 | + content = str(m.get("content") or "").strip() |
| 119 | + # Strip <details> block (raw reasoning); keep operator-facing |
| 120 | + if "<details" in content: |
| 121 | + content = content.split("</details>", 1)[-1].strip() |
| 122 | + out.append(f"- **agent**: {content[:400]}") |
| 123 | + out.append("") |
| 124 | + return "\n".join(out) |
| 125 | + |
| 126 | + |
| 127 | +# --- 2. launch verifier failures --------------------------------- |
| 128 | + |
| 129 | +def section_launch_failures() -> str: |
| 130 | + if not LAUNCH_FAILURES.is_file(): |
| 131 | + return "_no launch_failures.json (mios-daemon launch_verifier not yet ticked)_" |
| 132 | + try: |
| 133 | + data = json.loads(LAUNCH_FAILURES.read_text()) |
| 134 | + except Exception as e: |
| 135 | + return f"_failures file unreadable: {e}_" |
| 136 | + if not data: |
| 137 | + return "_no false-success launches recorded_" |
| 138 | + out = [] |
| 139 | + for f in data[-20:]: |
| 140 | + out.append(f"- `{f.get('ts','?')}` app=**{f.get('app','?')}** " |
| 141 | + f"verdict=`{f.get('verifier_summary','?')}` " |
| 142 | + f"prompt: _{(f.get('user_prompt') or '')[:120]}_") |
| 143 | + return "\n".join(out) |
| 144 | + |
| 145 | + |
| 146 | +# --- 3. daemon classify summaries -------------------------------- |
| 147 | + |
| 148 | +def section_daemon_state() -> str: |
| 149 | + if not DAEMON_STATE.is_file(): |
| 150 | + return "_no daemon state.json_" |
| 151 | + try: |
| 152 | + d = json.loads(DAEMON_STATE.read_text()) |
| 153 | + except Exception as e: |
| 154 | + return f"_state.json unreadable: {e}_" |
| 155 | + out = [] |
| 156 | + cls = d.get("classify") or {} |
| 157 | + if cls: |
| 158 | + out.append(f"- **classify** ({cls.get('severity','?')}): " |
| 159 | + f"{cls.get('summary','?')}") |
| 160 | + ref = d.get("refusal") or {} |
| 161 | + if ref: |
| 162 | + out.append(f"- **refusal**: model=`{ref.get('model','?')}` " |
| 163 | + f"phrase=_{(ref.get('phrase') or '')[:120]}_") |
| 164 | + sg = d.get("suggestions") or {} |
| 165 | + if sg: |
| 166 | + out.append(f"- **suggestions** ts={sg.get('ts','?')} " |
| 167 | + f"count={sg.get('count','?')}") |
| 168 | + lv = d.get("launch_verifier") or {} |
| 169 | + if lv: |
| 170 | + out.append(f"- **launch_verifier** ts={lv.get('ts','?')} " |
| 171 | + f"scanned={lv.get('claims_scanned',0)} " |
| 172 | + f"false_success={lv.get('false_success_count',0)}") |
| 173 | + return "\n".join(out) or "_state empty_" |
| 174 | + |
| 175 | + |
| 176 | +# --- 4. recent git commits (session work record) ----------------- |
| 177 | + |
| 178 | +def section_git_commits(since_epoch: int) -> str: |
| 179 | + if not (REPO_ROOT / ".git").is_dir(): |
| 180 | + return f"_repo not at {REPO_ROOT}_" |
| 181 | + try: |
| 182 | + since_iso = datetime.datetime.utcfromtimestamp(since_epoch).isoformat() |
| 183 | + out = subprocess.run( |
| 184 | + ["git", "-C", str(REPO_ROOT), "log", |
| 185 | + f"--since={since_iso}", "--oneline", "--no-decorate"], |
| 186 | + capture_output=True, text=True, timeout=8, |
| 187 | + ) |
| 188 | + lines = out.stdout.strip().splitlines() |
| 189 | + if not lines: |
| 190 | + return "_no commits in window_" |
| 191 | + return "\n".join(f"- `{line}`" for line in lines[:30]) |
| 192 | + except Exception as e: |
| 193 | + return f"_git query failed: {e}_" |
| 194 | + |
| 195 | + |
| 196 | +# --- LLM compaction ---------------------------------------------- |
| 197 | + |
| 198 | +def _llm_summarize(raw_digest: str, timeout_s: int = 90) -> str: |
| 199 | + """Send the raw digest to the CPU model with a tight system |
| 200 | + prompt, return a compacted summary. Returns the raw digest on |
| 201 | + failure (best-effort).""" |
| 202 | + system = ( |
| 203 | + "You compact a multi-section MiOS session digest into a " |
| 204 | + "structured markdown brief. Preserve every fact from the " |
| 205 | + "input. NO prose padding, NO 'In summary', NO 'I hope this " |
| 206 | + "helps'. Output sections in this order:\n" |
| 207 | + "1. Headline (1 line: most important state change since the window started)\n" |
| 208 | + "2. Operator asks (bulleted)\n" |
| 209 | + "3. Agent actions taken (bulleted; cite tool names)\n" |
| 210 | + "4. Failures + verifier verdicts (bulleted)\n" |
| 211 | + "5. System state snapshot (1-3 lines)\n" |
| 212 | + "6. Carry-forward (1-3 lines: open follow-ups or pending state)\n" |
| 213 | + "Mirror the operator's language." |
| 214 | + ) |
| 215 | + payload = { |
| 216 | + "model": SUMMARY_MODEL, |
| 217 | + "messages": [ |
| 218 | + {"role": "system", "content": system}, |
| 219 | + {"role": "user", "content": raw_digest[:32000]}, |
| 220 | + ], |
| 221 | + "options": { |
| 222 | + "num_gpu": 0, "num_thread": 8, |
| 223 | + "num_predict": 700, "temperature": 0.0, |
| 224 | + }, |
| 225 | + "stream": False, "keep_alive": -1, |
| 226 | + } |
| 227 | + try: |
| 228 | + req = urllib.request.Request( |
| 229 | + f"{OLLAMA_URL}/api/chat", |
| 230 | + data=json.dumps(payload).encode("utf-8"), |
| 231 | + headers={"Content-Type": "application/json"}, |
| 232 | + method="POST", |
| 233 | + ) |
| 234 | + with urllib.request.urlopen(req, timeout=timeout_s) as r: |
| 235 | + body = json.loads(r.read()) |
| 236 | + msg = body.get("message") or {} |
| 237 | + out = (msg.get("content") or "").strip() |
| 238 | + if not out: |
| 239 | + out = (msg.get("thinking") or msg.get("reasoning") or "").strip() |
| 240 | + return out or raw_digest |
| 241 | + except (urllib.error.URLError, OSError, json.JSONDecodeError) as e: |
| 242 | + sys.stderr.write(f"[mios-compact] llm-summarize failed: {e}\n") |
| 243 | + return raw_digest |
| 244 | + |
| 245 | + |
| 246 | +# --- main --------------------------------------------------------- |
| 247 | + |
| 248 | +def main() -> int: |
| 249 | + ap = argparse.ArgumentParser(prog="mios-compact", |
| 250 | + description=__doc__.splitlines()[0]) |
| 251 | + ap.add_argument("--since", default="24 hours ago", |
| 252 | + help='date(1) string, e.g. "12 hours ago"') |
| 253 | + ap.add_argument("--chats", type=int, default=12, |
| 254 | + help="cap N most recent chats (default 12)") |
| 255 | + ap.add_argument("--out", default=None, |
| 256 | + help="explicit output path (default: timestamped)") |
| 257 | + ap.add_argument("--stdout", action="store_true", |
| 258 | + help="print to stdout instead of writing a file") |
| 259 | + ap.add_argument("--no-llm", action="store_true", |
| 260 | + help="skip the CPU summarization step") |
| 261 | + args = ap.parse_args() |
| 262 | + |
| 263 | + since_epoch = _since_to_epoch(args.since) |
| 264 | + since_iso = datetime.datetime.utcfromtimestamp(since_epoch).isoformat() |
| 265 | + now_iso = datetime.datetime.utcnow().isoformat() |
| 266 | + |
| 267 | + raw_sections = [ |
| 268 | + f"# MiOS session digest", |
| 269 | + f"_window: {since_iso}Z → {now_iso}Z_", |
| 270 | + "", |
| 271 | + "## Recent operator chats", |
| 272 | + section_recent_chats(since_epoch, args.chats), |
| 273 | + "", |
| 274 | + "## Launch verifier failures", |
| 275 | + section_launch_failures(), |
| 276 | + "", |
| 277 | + "## Daemon state snapshot", |
| 278 | + section_daemon_state(), |
| 279 | + "", |
| 280 | + "## Git commits this window", |
| 281 | + section_git_commits(since_epoch), |
| 282 | + ] |
| 283 | + raw_text = "\n".join(raw_sections) |
| 284 | + |
| 285 | + if args.no_llm: |
| 286 | + out_text = raw_text |
| 287 | + else: |
| 288 | + compacted = _llm_summarize(raw_text) |
| 289 | + out_text = ( |
| 290 | + f"# MiOS session digest (compacted)\n\n" |
| 291 | + f"_compacted by `{SUMMARY_MODEL}` at {now_iso}Z; " |
| 292 | + f"raw digest covers {since_iso}Z → {now_iso}Z._\n\n" |
| 293 | + f"{compacted}\n\n---\n\n" |
| 294 | + f"<details><summary>Raw section dumps (uncompacted)</summary>\n\n" |
| 295 | + f"{raw_text}\n\n</details>" |
| 296 | + ) |
| 297 | + |
| 298 | + if args.stdout: |
| 299 | + sys.stdout.write(out_text + "\n") |
| 300 | + return 0 |
| 301 | + |
| 302 | + COMPACTED_DIR.mkdir(parents=True, exist_ok=True) |
| 303 | + out_path = Path(args.out) if args.out else ( |
| 304 | + COMPACTED_DIR / f"digest-{_ts()}.md") |
| 305 | + out_path.write_text(out_text, encoding="utf-8") |
| 306 | + try: |
| 307 | + os.chmod(out_path, 0o644) |
| 308 | + except Exception: |
| 309 | + pass |
| 310 | + print(f"[mios-compact] wrote {out_path} ({len(out_text)} chars)") |
| 311 | + return 0 |
| 312 | + |
| 313 | + |
| 314 | +if __name__ == "__main__": |
| 315 | + sys.exit(main()) |
0 commit comments