Files
Cortex-Inara/cortex/routers/distill.py
Scott Idem f8f7cd75da feat: audit log, usage tracking UI, OpenAI orchestrator compaction, onboarding + docs
Tool audit log:
- Every orchestrator tool call logged to home/{user}/tool_audit/YYYY-MM-DD.jsonl
- Files panel sidebar: audit log group (collapsed), date-linked read-only table
- Admin endpoints: /api/audit/files, /api/audit/day, /api/audit/recent, /api/audit/stats
- Engine and model name recorded per entry

OpenAI orchestrator improvements:
- Context budget enforcement: 75% of model context_k (min 16k)
- Message compaction: truncates old tool results when approaching budget
- max_rounds respected per model config (intersected with server cap)

OpenRouter onboarding (setup.html, onboarding.py, app.js, settings.html):
- Step 3 of 3: /setup/model with curated model picker
- Chat banner for users on server-default model (informational, not alarmist)
- Settings quick-link card; /setup/model works standalone for existing users

Model registry + session store:
- set_role_config / get_role_config for per-role tool lists and system_append
- session_store: session rename, session name backfill endpoint

UI updates (app.js, index.html, style.css, local_llm.html):
- Role toggle in context panel
- Off-the-record mode
- Agent notes read-only viewer
- OPERATIONS.md loaded at T2+ in context

Documentation:
- HELP.md: full tool table, per-role tool sets, Agent Notes, usage tracking
- TOOLS.md: Agent Notes section, count corrected to 44
- ARCH__SYSTEM.md, ARCH__BACKENDS.md, MASTER.md updated to match reality
- CLAUDE.md: onboarding flow, documentation philosophy sections
- README.md: stack in practice, DeepSeek TUI mention, architecture diagram updated
- TODO__Agents.md: onboarding task completed with deviation notes

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-05-08 21:26:43 -04:00

239 lines
7.8 KiB
Python

"""
Manual memory distillation endpoints.
POST /distill/short — roll session logs → MEMORY_SHORT.md (no LLM)
POST /distill/mid — summarize short → MEMORY_MID.md (LLM)
POST /distill/long — integrate mid → MEMORY_LONG.md (LLM)
POST /distill/all — run all three in sequence
POST /distill/rebuild — wipe mid + long, then run all three from scratch
All endpoints require ?user=<username>&persona=<name> query params.
Concurrency: one distillation at a time per persona. A second request while one
is running returns 409 immediately — no silent queuing.
"""
import asyncio
from datetime import datetime, timedelta
from fastapi import APIRouter, HTTPException, Query
from memory_distiller import distill_short, distill_mid, distill_long
from persona import validate as validate_persona, set_context, persona_path as _persona_path
import scheduler
router = APIRouter(prefix="/distill")
# Per-persona asyncio lock. Key: (user, persona)
_LOCKS: dict[tuple, asyncio.Lock] = {}
_LOCKS_META: dict[tuple, str] = {} # key → which step is currently running
# Minimum time between successive runs of each endpoint, per persona.
# Prevents accidental rapid-fire runs and token waste.
_COOLDOWNS: dict[tuple, timedelta] = {
"short": timedelta(minutes=1),
"mid": timedelta(minutes=30),
"long": timedelta(hours=6),
"all": timedelta(hours=1),
"rebuild": timedelta(hours=6),
}
_LAST_RUN: dict[tuple, datetime] = {} # key: (user, persona, endpoint)
def _get_lock(user: str, persona: str) -> asyncio.Lock:
key = (user, persona)
if key not in _LOCKS:
_LOCKS[key] = asyncio.Lock()
return _LOCKS[key]
def _resolve(user: str, persona: str) -> tuple[str, str]:
try:
u, p = validate_persona(user, persona)
except Exception:
raise HTTPException(status_code=404, detail=f"Persona not found: {user}/{persona}")
set_context(u, p)
return u, p
def _check_lock(user: str, persona: str) -> asyncio.Lock:
"""Return the lock if free, raise 409 if already held."""
lock = _get_lock(user, persona)
if lock.locked():
step = _LOCKS_META.get((user, persona), "distillation")
raise HTTPException(
status_code=409,
detail=f"A {step} is already running for {persona} — please wait for it to finish.",
)
return lock
def _check_cooldown(user: str, persona: str, endpoint: str) -> None:
"""Raise 429 if the endpoint was run too recently for this persona."""
cooldown = _COOLDOWNS.get(endpoint)
if not cooldown:
return
key = (user, persona, endpoint)
last = _LAST_RUN.get(key)
if last:
elapsed = datetime.now() - last
if elapsed < cooldown:
remaining = cooldown - elapsed
mins = int(remaining.total_seconds() // 60)
secs = int(remaining.total_seconds() % 60)
wait = f"{mins}m {secs}s" if mins else f"{secs}s"
raise HTTPException(
status_code=429,
detail=f"{endpoint} was just run — please wait {wait} before running again.",
)
def _record_run(user: str, persona: str, endpoint: str) -> None:
_LAST_RUN[(user, persona, endpoint)] = datetime.now()
@router.get("/status")
async def distill_status() -> dict:
from config import settings
# Include which personas are currently distilling
active = [f"{u}/{p}" for (u, p), lock in _LOCKS.items() if lock.locked()]
return {
"enabled": settings.auto_distill,
"jobs": scheduler.status(),
"active": active,
"config": {
"short": settings.auto_distill_short,
"mid": settings.auto_distill_mid,
"long": settings.auto_distill_long,
},
}
@router.post("/short")
async def do_distill_short(
user: str = Query(...),
persona: str = Query(...),
) -> dict:
u, p = _resolve(user, persona)
_check_cooldown(u, p, "short")
lock = _check_lock(u, p)
async with lock:
_LOCKS_META[(u, p)] = "short distill"
try:
result = distill_short(u, p)
_record_run(u, p, "short")
return {"ok": True, **result}
finally:
_LOCKS_META.pop((u, p), None)
@router.post("/mid")
async def do_distill_mid(
user: str = Query(...),
persona: str = Query(...),
) -> dict:
u, p = _resolve(user, persona)
_check_cooldown(u, p, "mid")
lock = _check_lock(u, p)
async with lock:
_LOCKS_META[(u, p)] = "mid distill"
try:
result = await distill_mid(u, p)
if "error" not in result:
_record_run(u, p, "mid")
return {"ok": "error" not in result, **result}
finally:
_LOCKS_META.pop((u, p), None)
@router.post("/long")
async def do_distill_long(
user: str = Query(...),
persona: str = Query(...),
) -> dict:
u, p = _resolve(user, persona)
_check_cooldown(u, p, "long")
lock = _check_lock(u, p)
async with lock:
_LOCKS_META[(u, p)] = "long distill"
try:
result = await distill_long(u, p)
if "error" not in result:
_record_run(u, p, "long")
return {"ok": "error" not in result, **result}
finally:
_LOCKS_META.pop((u, p), None)
@router.post("/all")
async def do_distill_all(
user: str = Query(...),
persona: str = Query(...),
) -> dict:
u, p = _resolve(user, persona)
_check_cooldown(u, p, "all")
lock = _check_lock(u, p)
async with lock:
_LOCKS_META[(u, p)] = "full distill"
try:
short_result = distill_short(u, p)
mid_result = await distill_mid(u, p)
if "error" in mid_result:
return {"ok": False, "short": short_result, "mid": mid_result}
long_result = await distill_long(u, p)
ok = "error" not in long_result
if ok:
_record_run(u, p, "all")
return {
"ok": ok,
"short": short_result,
"mid": mid_result,
"long": long_result,
}
finally:
_LOCKS_META.pop((u, p), None)
@router.post("/rebuild")
async def do_distill_rebuild(
user: str = Query(...),
persona: str = Query(...),
) -> dict: # noqa: E501
"""Wipe MEMORY_MID and MEMORY_LONG (with backups), then run short → mid → long.
Use when memories have drifted, been corrupted, or you want a clean slate
rebuilt purely from session logs. Hand-edited content will be replaced.
"""
u, p = _resolve(user, persona)
_check_cooldown(u, p, "rebuild")
lock = _check_lock(u, p)
async with lock:
_LOCKS_META[(u, p)] = "memory rebuild"
try:
from memory_distiller import _rotate_backup, _read
inara_dir = _persona_path(u, p)
# Back up then wipe mid and long before rebuilding
for name in ("MEMORY_MID.md", "MEMORY_LONG.md"):
path = inara_dir / name
if path.exists():
_rotate_backup(path)
path.write_text(
f"# {name}\n\n*Cleared for rebuild — {__import__('datetime').datetime.now().strftime('%Y-%m-%d %H:%M')}.*\n"
)
short_result = distill_short(u, p)
mid_result = await distill_mid(u, p)
if "error" in mid_result:
return {"ok": False, "short": short_result, "mid": mid_result, "rebuilt": True}
long_result = await distill_long(u, p)
ok = "error" not in long_result
if ok:
_record_run(u, p, "rebuild")
return {
"ok": ok,
"short": short_result,
"mid": mid_result,
"long": long_result,
"rebuilt": True,
}
finally:
_LOCKS_META.pop((u, p), None)