feat(orchestrator): W4 resources view and host samples
A resources timer samples host CPU, RAM, swap and per-session tokens into samples.jsonl; the hub serves them and a Resources view charts them. Committed below the bar on the operator's word 2026-11-12 (no further per-bundle tests, one gate after the merge): the W4 e2e never ran in a real browser. 🎯 Fix: 92 % 💥 Break: 6 % 💸 Cost: 100 % ⏱️ Time: 100 % 🛡️ Security: 100 % 🏆 Score: 86 🚦 Gate: ⛔ break 6 % > 5 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Kevin Veen-BirkenbachandClaude Opus 5.5 committed
1 parent
fd8fc08603
commit
35c270e120
21 files changed
+1937
-18
No files matched your search
@@ -739,6 +739,8 @@ flowchart TD
|
||||
|
||||
### Phase 6: Resources (S7)
|
||||
|
||||
Replaced by [003 W4 Resources](003-operator-workbench.md#w4-resources).
|
||||
|
||||
- [ ] The tab shows tokens per agent and accumulated, and CPU, RAM, disk and
|
||||
load, on the shared time axis, several values per chart.
|
||||
- [ ] Token counts equal the sum of the transcript usage fields of a fixture
|
||||
|
||||
@@ -248,22 +248,22 @@ flowchart TD
|
||||
|
||||
### W4 Resources
|
||||
|
||||
- [ ] `resources.py sample`, run every 60 s by a user timer that
|
||||
- [x] `resources.py sample`, run every 60 s by a user timer that
|
||||
`make orchestrator-install` enables, appends one row: used and total per
|
||||
mount, CPU, RAM and swap percent, load, GPU utilization and VRAM per
|
||||
GPU, waiting and active queue jobs.
|
||||
- [ ] An NVIDIA GPU whose PCI `runtime_status` is not `active` is recorded as
|
||||
- [x] An NVIDIA GPU whose PCI `runtime_status` is not `active` is recorded as
|
||||
asleep without running `nvidia-smi`.
|
||||
- [ ] `resources.py tokens` adds input, output, cache-read and cache-write
|
||||
- [x] `resources.py tokens` adds input, output, cache-read and cache-write
|
||||
tokens per agent and model from transcript usage fields since the last
|
||||
byte offset; no sampler or parser calls an LLM.
|
||||
- [ ] Rows older than 7 days fold into 15-minute means; means older than 90
|
||||
- [x] Rows older than 7 days fold into 15-minute means; means older than 90
|
||||
days are dropped.
|
||||
- [ ] The Resources entry shows on one time axis disks, CPU, RAM, swap, GPU,
|
||||
VRAM and stacked tokens per agent; bands mark idle (< 30 %) and
|
||||
saturated (> 90 %); a marker flags spans where jobs waited while CPU and
|
||||
RAM stayed under 50 %; the view reads only the sampler files.
|
||||
- [ ] 001 S7 points to this section.
|
||||
- [x] 001 S7 points to this section.
|
||||
|
||||
### Contracts
|
||||
|
||||
@@ -289,7 +289,7 @@ The work packages meet only through these interfaces:
|
||||
used, total}], cpu, ram, swap, load, gpus: [{name, util, vram_used,
|
||||
vram_total} | {name, asleep: true}], queue: {waiting, active}}`),
|
||||
`tokens.jsonl` (`{t, agent, model, input, output, cache_read,
|
||||
cache_write}`, t the hour) and `offsets.json` under
|
||||
cache_write}`, t the hour), `offsets.json` and `cpu.json` under
|
||||
`~/.local/state/agents-council/resources/`.
|
||||
- **Mounts**: the hub mounts `~/.local/state/agents-council` read-only and
|
||||
opens `gates.db` read-only.
|
||||
@@ -339,6 +339,10 @@ The work packages meet only through these interfaces:
|
||||
- [ ] Resources sample parsing on fixtures (df, meminfo, nvidia-smi, asleep
|
||||
GPU); token offsets; retention fold; diff cap and masking; feedback
|
||||
payload.
|
||||
- [x] W4 part of the item above: sample parsing on fixtures (df, meminfo,
|
||||
CPU ticks between runs, nvidia-smi, asleep and unread GPU); token
|
||||
offsets, a message counted once across fork transcripts; retention
|
||||
fold.
|
||||
- [ ] Bubble schedule: scale per queue position, word reveal times from a
|
||||
duration, reading-pace duration, one bubble per agent.
|
||||
- Integration:
|
||||
@@ -348,7 +352,7 @@ The work packages meet only through these interfaces:
|
||||
- [ ] The host executor applies a signed board batch to a scratch register
|
||||
and refuses a forged one; deletes a local merged branch without a
|
||||
token and refuses a remote one.
|
||||
- [ ] The sampler writes rows in a scratch HOME.
|
||||
- [x] The sampler writes rows in a scratch HOME.
|
||||
- E2E, headless Chromium and Firefox at 1920, 1366 and 390 px:
|
||||
- [ ] Tasks badge, no Inbox entry, operator cards with bulk actions, review
|
||||
diff expanded, feedback by right-click and `f` reaching the radio
|
||||
|
||||
Executable
+590
@@ -0,0 +1,590 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Args:
|
||||
sample: append one row to <state>/samples.jsonl: {t, disks: [{mount,
|
||||
used, total}], cpu, ram, swap, load, gpus: [{name, util, vram_used,
|
||||
vram_total} | {name, asleep: true}], queue: {waiting, active}}; t in
|
||||
epoch seconds, used, total and vram in bytes, cpu, ram, swap and util
|
||||
in percent, load the 1-minute load average, a GPU named by its PCI
|
||||
address. cpu is the busy share of the /proc/stat ticks since those
|
||||
the run before kept in <state>/cpu.json as [t, idle, total], None on
|
||||
the first run and after a gap of more than 15 minutes. An NVIDIA GPU
|
||||
whose PCI runtime_status is not active is recorded asleep, and
|
||||
nvidia-smi is not run for it; an awake one nvidia-smi gives no
|
||||
reading for (no driver, a driver mismatch) keeps its name with util
|
||||
and vram None. In the first minute of a 15-minute bucket, rows older
|
||||
than 7 days fold into 15-minute means and means older than 90 days
|
||||
are dropped.
|
||||
tokens: add the input, output, cache-read and cache-write tokens of the
|
||||
usage fields in the Claude Code transcripts under --transcripts since
|
||||
each file's byte offset to <state>/tokens.jsonl as {t, agent, model,
|
||||
input, output, cache_read, cache_write}, t the hour; agent is the
|
||||
name the council claim of the session carries, else the session id's
|
||||
first 8 characters. A message counts once by its message id and
|
||||
request id, also when written over several lines or copied into a
|
||||
fork's transcript. <state>/offsets.json holds {files: {path: offset},
|
||||
seen: {hour: [key digest]}}, seen keeping the hours this run read and
|
||||
the last two, and is replaced only when it changes. In the first
|
||||
minute of an hour, rows of the same hour, agent and model merge and
|
||||
rows older than 90 days are dropped.
|
||||
--state: the directory of these files, the unit passes
|
||||
~/.local/state/agents-council/resources.
|
||||
--council: the council directory with queue.json and the claims.
|
||||
--transcripts (tokens): Claude Code's projects directory.
|
||||
--proc, --sys (sample): where /proc and /sys are mounted.
|
||||
|
||||
Returns:
|
||||
Runs on the host every minute through kvb-orchestrator-resources.timer,
|
||||
because the hub's container sees neither the host's disks, memory and
|
||||
GPUs nor its transcripts' offsets; nothing here calls a language model.
|
||||
The exit code is 0.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import base64
|
||||
import datetime
|
||||
import hashlib
|
||||
import json
|
||||
import os
|
||||
import shutil
|
||||
import stat
|
||||
import subprocess
|
||||
import tempfile
|
||||
import time
|
||||
from collections.abc import Callable
|
||||
from pathlib import Path
|
||||
|
||||
INTERVAL = 60
|
||||
BUCKET = 900
|
||||
HOUR = 3600
|
||||
FINE = 7 * 86400
|
||||
COARSE = 90 * 86400
|
||||
COMMAND_SECONDS = 10
|
||||
KEY_BYTES = 9
|
||||
KIB = 1 << 10
|
||||
MIB = 1 << 20
|
||||
JSON_BYTES = 1 << 24
|
||||
BUDGET = 64 << 20
|
||||
NVIDIA = "0x10de"
|
||||
DISPLAY = "0x03"
|
||||
NVIDIA_QUERY = "pci.bus_id,utilization.gpu,memory.used,memory.total"
|
||||
UNREAD = {"util": None, "vram_used": None, "vram_total": None}
|
||||
COUNTS = (("input", "input_tokens"), ("output", "output_tokens"), ("cache_read", "cache_read_input_tokens"), ("cache_write", "cache_creation_input_tokens"))
|
||||
NAMES = tuple(name for name, _ in COUNTS)
|
||||
HOST_DIGITS = (("cpu", 1), ("ram", 1), ("swap", 1), ("load", 2))
|
||||
CLAIMS = ("*/claim.json", "*/*/claim.json")
|
||||
SAMPLES = "samples.jsonl"
|
||||
TOKENS = "tokens.jsonl"
|
||||
OFFSETS = "offsets.json"
|
||||
TICKS = "cpu.json"
|
||||
|
||||
|
||||
def read_text(path: Path) -> str:
|
||||
try:
|
||||
return path.read_text(encoding="utf-8")
|
||||
except (OSError, UnicodeDecodeError):
|
||||
return ""
|
||||
|
||||
|
||||
def read_json(path: Path) -> object:
|
||||
"""Returns:
|
||||
The parsed JSON of path, read without following a symlink and only
|
||||
from a regular file of at most JSON_BYTES; None when that fails.
|
||||
"""
|
||||
try:
|
||||
with os.fdopen(os.open(path, os.O_RDONLY | os.O_NOFOLLOW | os.O_NONBLOCK), "rb") as handle:
|
||||
if not stat.S_ISREG(os.fstat(handle.fileno()).st_mode):
|
||||
return None
|
||||
data = handle.read(JSON_BYTES + 1)
|
||||
return None if len(data) > JSON_BYTES else json.loads(data.decode("utf-8"))
|
||||
except (OSError, ValueError, RecursionError):
|
||||
return None
|
||||
|
||||
|
||||
def number(text: str) -> float | None:
|
||||
try:
|
||||
return float(text)
|
||||
except ValueError:
|
||||
return None
|
||||
|
||||
|
||||
def percent(part: float, whole: float | None) -> float | None:
|
||||
return round(100 * part / whole, 1) if whole else None
|
||||
|
||||
|
||||
def command(argv: list[str]) -> str | None:
|
||||
if shutil.which(argv[0]) is None:
|
||||
return None
|
||||
try:
|
||||
done = subprocess.run(argv, capture_output=True, text=True, timeout=COMMAND_SECONDS, check=False)
|
||||
except (OSError, subprocess.SubprocessError):
|
||||
return None
|
||||
return done.stdout if done.returncode == 0 else None
|
||||
|
||||
|
||||
def parse_df(text: str) -> list[dict]:
|
||||
"""Returns:
|
||||
One {mount, used, total} in bytes per block device in the output of
|
||||
df -P -k, at the first mount df lists for it; loop devices and
|
||||
filesystems without a device are left out.
|
||||
"""
|
||||
found, seen = [], set()
|
||||
for line in text.splitlines()[1:]:
|
||||
parts = line.split(None, 5)
|
||||
if len(parts) < 6 or not parts[0].startswith("/dev/") or parts[0].startswith("/dev/loop") or parts[0] in seen:
|
||||
continue
|
||||
total, used = number(parts[1]), number(parts[2])
|
||||
if total is None or used is None:
|
||||
continue
|
||||
seen.add(parts[0])
|
||||
found.append({"mount": parts[5], "used": int(used) * KIB, "total": int(total) * KIB})
|
||||
return found
|
||||
|
||||
|
||||
def parse_meminfo(text: str) -> dict[str, int]:
|
||||
values = {}
|
||||
for line in text.splitlines():
|
||||
key, _, rest = line.partition(":")
|
||||
fields = rest.split()
|
||||
if fields and fields[0].isdigit():
|
||||
values[key.strip()] = int(fields[0])
|
||||
return values
|
||||
|
||||
|
||||
def memory(values: dict[str, int]) -> tuple[float | None, float | None]:
|
||||
"""Returns:
|
||||
RAM in use ((MemTotal - MemAvailable) / MemTotal) and swap in use in
|
||||
percent; swap is 0 without a swap device, None when unknown.
|
||||
"""
|
||||
total, available = values.get("MemTotal"), values.get("MemAvailable")
|
||||
ram = percent(total - available, total) if total and available is not None else None
|
||||
swap_total, swap_free = values.get("SwapTotal"), values.get("SwapFree")
|
||||
if swap_total == 0:
|
||||
return ram, 0.0
|
||||
return ram, percent(swap_total - swap_free, swap_total) if swap_total and swap_free is not None else None
|
||||
|
||||
|
||||
def cpu_ticks(text: str) -> tuple[int, int] | None:
|
||||
"""Returns:
|
||||
(idle, total) ticks of the summary line of /proc/stat: idle counts
|
||||
idle and iowait, total the first eight fields, which already hold
|
||||
guest time.
|
||||
"""
|
||||
fields = text.split("\n", 1)[0].split()
|
||||
if len(fields) < 6 or fields[0] != "cpu" or not all(field.isdigit() for field in fields[1:9]):
|
||||
return None
|
||||
values = [int(field) for field in fields[1:9]]
|
||||
return values[3] + values[4], sum(values)
|
||||
|
||||
|
||||
def cpu_percent(before: tuple[int, int] | None, after: tuple[int, int] | None) -> float | None:
|
||||
if before is None or after is None:
|
||||
return None
|
||||
idle, total = after[0] - before[0], after[1] - before[1]
|
||||
return percent(total - idle, total) if total > 0 else None
|
||||
|
||||
|
||||
def parse_loadavg(text: str) -> float | None:
|
||||
fields = text.split()
|
||||
return number(fields[0]) if fields else None
|
||||
|
||||
|
||||
def parse_nvidia(text: str) -> dict[str, dict]:
|
||||
"""Returns:
|
||||
Per PCI address (domain of 4 hex digits) {util, vram_used,
|
||||
vram_total} from the CSV nvidia-smi prints for NVIDIA_QUERY with
|
||||
noheader,nounits; vram in bytes, an unknown value None.
|
||||
"""
|
||||
found = {}
|
||||
for line in text.splitlines():
|
||||
parts = [part.strip() for part in line.split(",")]
|
||||
if len(parts) != 4:
|
||||
continue
|
||||
util, used, total = (number(part) for part in parts[1:])
|
||||
found[parts[0].lower()[-12:]] = {"util": util, "vram_used": None if used is None else int(used * MIB), "vram_total": None if total is None else int(total * MIB)}
|
||||
return found
|
||||
|
||||
|
||||
def nvidia_gpus(sys_root: Path) -> list[tuple[str, bool]]:
|
||||
"""Returns:
|
||||
The PCI address of every NVIDIA display controller under sys_root and
|
||||
whether its runtime_status reads active.
|
||||
"""
|
||||
found = []
|
||||
for device in sorted((sys_root / "bus" / "pci" / "devices").glob("*")):
|
||||
if read_text(device / "vendor").strip() == NVIDIA and read_text(device / "class").strip().startswith(DISPLAY):
|
||||
found.append((device.name, read_text(device / "power" / "runtime_status").strip() == "active"))
|
||||
return found
|
||||
|
||||
|
||||
def gpus(sys_root: Path, run: Callable[[list[str]], str | None]) -> list[dict]:
|
||||
devices = nvidia_gpus(sys_root)
|
||||
awake = [address for address, active in devices if active]
|
||||
readings = parse_nvidia(run(["nvidia-smi", f"--id={','.join(awake)}", f"--query-gpu={NVIDIA_QUERY}", "--format=csv,noheader,nounits"]) or "") if awake else {}
|
||||
return [{"name": address, **readings.get(address, UNREAD)} if active else {"name": address, "asleep": True} for address, active in devices]
|
||||
|
||||
|
||||
def queue(council: Path) -> dict:
|
||||
data = read_json(council / "queue.json")
|
||||
entries = data.get("entries") if isinstance(data, dict) else None
|
||||
statuses = [entry.get("status") for entry in entries if isinstance(entry, dict)] if isinstance(entries, list) else []
|
||||
return {"waiting": statuses.count("waiting"), "active": statuses.count("active")}
|
||||
|
||||
|
||||
def integers(value: object, size: int) -> bool:
|
||||
return isinstance(value, list) and len(value) == size and all(isinstance(item, int) and not isinstance(item, bool) for item in value)
|
||||
|
||||
|
||||
def cpu_since(proc: Path, state: Path, t: int) -> float | None:
|
||||
"""Returns:
|
||||
The busy share of the ticks in proc's stat since those the run
|
||||
before kept in <state>/cpu.json at most BUCKET seconds before t,
|
||||
else None; this run's ticks replace them.
|
||||
"""
|
||||
ticks = cpu_ticks(read_text(proc / "stat"))
|
||||
stored = read_json(state / TICKS)
|
||||
before = (stored[1], stored[2]) if integers(stored, 3) and t - BUCKET <= stored[0] <= t else None
|
||||
if ticks is not None:
|
||||
replace(state / TICKS, json.dumps([t, *ticks]))
|
||||
return cpu_percent(before, ticks)
|
||||
|
||||
|
||||
def sample(proc: Path, sys_root: Path, council: Path, run: Callable[[list[str]], str | None], state: Path) -> dict:
|
||||
t = int(time.time())
|
||||
ram, swap = memory(parse_meminfo(read_text(proc / "meminfo")))
|
||||
return {
|
||||
"t": t,
|
||||
"disks": parse_df(run(["df", "-P", "-k", "-l"]) or ""),
|
||||
"cpu": cpu_since(proc, state, t),
|
||||
"ram": ram,
|
||||
"swap": swap,
|
||||
"load": parse_loadavg(read_text(proc / "loadavg")),
|
||||
"gpus": gpus(sys_root, run),
|
||||
"queue": queue(council),
|
||||
}
|
||||
|
||||
|
||||
def average(values: list[object], digits: int) -> float | None:
|
||||
found = [value for value in values if isinstance(value, int | float) and not isinstance(value, bool)]
|
||||
return round(sum(found) / len(found), digits) if found else None
|
||||
|
||||
|
||||
def whole(values: list[object]) -> int | None:
|
||||
found = average(values, 0)
|
||||
return None if found is None else int(found)
|
||||
|
||||
|
||||
def listed(rows: list[dict], key: str) -> list[dict]:
|
||||
return [item for row in rows if isinstance(row.get(key), list) for item in row[key] if isinstance(item, dict)]
|
||||
|
||||
|
||||
def grouped(items: list[dict], key: str) -> dict[object, list[dict]]:
|
||||
found: dict[object, list[dict]] = {}
|
||||
for item in items:
|
||||
found.setdefault(item.get(key), []).append(item)
|
||||
return found
|
||||
|
||||
|
||||
def gpu_mean(name: object, readings: list[dict]) -> dict:
|
||||
awake = [reading for reading in readings if reading.get("asleep") is not True]
|
||||
if not awake:
|
||||
return {"name": name, "asleep": True}
|
||||
return {"name": name, "util": average([reading.get("util") for reading in awake], 1), "vram_used": whole([reading.get("vram_used") for reading in awake]), "vram_total": whole([reading.get("vram_total") for reading in awake])}
|
||||
|
||||
|
||||
def disk_means(rows: list[dict]) -> list[dict]:
|
||||
return [{"mount": mount, "used": whole([disk.get("used") for disk in disks]), "total": whole([disk.get("total") for disk in disks])} for mount, disks in grouped(listed(rows, "disks"), "mount").items()]
|
||||
|
||||
|
||||
def queue_mean(rows: list[dict]) -> dict:
|
||||
queues = [row["queue"] for row in rows if isinstance(row.get("queue"), dict)]
|
||||
return {key: average([entry.get(key) for entry in queues], 2) for key in ("waiting", "active")}
|
||||
|
||||
|
||||
def mean_row(start: int, rows: list[dict]) -> dict:
|
||||
"""Returns:
|
||||
The 15-minute mean of rows at start: per mount and per GPU the mean
|
||||
of the rows that list it, a GPU asleep in every row stays asleep;
|
||||
a single row already at start comes back unchanged.
|
||||
"""
|
||||
if len(rows) == 1 and rows[0]["t"] == start:
|
||||
return rows[0]
|
||||
host = {key: average([row.get(key) for row in rows], digits) for key, digits in HOST_DIGITS}
|
||||
gpu_rows = [gpu_mean(name, readings) for name, readings in grouped(listed(rows, "gpus"), "name").items()]
|
||||
return {"t": start, "disks": disk_means(rows), **host, "gpus": gpu_rows, "queue": queue_mean(rows)}
|
||||
|
||||
|
||||
def timed(row: object) -> bool:
|
||||
return isinstance(row, dict) and isinstance(row.get("t"), int | float) and not isinstance(row.get("t"), bool)
|
||||
|
||||
|
||||
def retain(rows: list[dict], now: float) -> list[dict]:
|
||||
"""Returns:
|
||||
rows without those older than 90 days, every 15-minute bucket that
|
||||
ended more than 7 days ago folded into one mean at its start, the
|
||||
newer rows unchanged and in order.
|
||||
"""
|
||||
buckets: dict[int, list[dict]] = {}
|
||||
recent = []
|
||||
for row in rows:
|
||||
if not timed(row) or row["t"] < now - COARSE:
|
||||
continue
|
||||
start = int(row["t"] // BUCKET * BUCKET)
|
||||
if start + BUCKET <= now - FINE:
|
||||
buckets.setdefault(start, []).append(row)
|
||||
else:
|
||||
recent.append(row)
|
||||
return [mean_row(start, buckets[start]) for start in sorted(buckets)] + recent
|
||||
|
||||
|
||||
def read_rows(path: Path) -> list[dict]:
|
||||
rows = []
|
||||
for line in read_text(path).splitlines():
|
||||
try:
|
||||
row = json.loads(line)
|
||||
except (ValueError, RecursionError):
|
||||
continue
|
||||
if isinstance(row, dict):
|
||||
rows.append(row)
|
||||
return rows
|
||||
|
||||
|
||||
def encoded(rows: list[dict]) -> str:
|
||||
return "".join(json.dumps(row, separators=(",", ":")) + "\n" for row in rows)
|
||||
|
||||
|
||||
def append_rows(path: Path, rows: list[dict]) -> None:
|
||||
if rows:
|
||||
with path.open("a", encoding="utf-8") as handle:
|
||||
handle.write(encoded(rows))
|
||||
|
||||
|
||||
def replace(path: Path, text: str) -> None:
|
||||
with tempfile.NamedTemporaryFile("w", encoding="utf-8", dir=path.parent, prefix=f".{path.name}.", delete=False) as handle:
|
||||
try:
|
||||
handle.write(text)
|
||||
except BaseException:
|
||||
os.unlink(handle.name)
|
||||
raise
|
||||
os.replace(handle.name, path)
|
||||
|
||||
|
||||
def rewrite(path: Path, keep: Callable[[list[dict], float], list[dict]], now: float) -> None:
|
||||
rows = read_rows(path)
|
||||
kept = keep(rows, now)
|
||||
if kept != rows:
|
||||
replace(path, encoded(kept))
|
||||
|
||||
|
||||
def record(state: Path, row: dict, now: float) -> None:
|
||||
path = state / SAMPLES
|
||||
append_rows(path, [row])
|
||||
if now % BUCKET < INTERVAL:
|
||||
rewrite(path, retain, now)
|
||||
|
||||
|
||||
def claim_names(council: Path) -> dict[str, str]:
|
||||
names = {}
|
||||
for pattern in CLAIMS:
|
||||
for path in sorted(council.glob(pattern)):
|
||||
data = read_json(path)
|
||||
if isinstance(data, dict) and isinstance(data.get("session"), str) and isinstance(data.get("name"), str) and data["name"].strip():
|
||||
names[data["session"]] = data["name"].strip()
|
||||
return names
|
||||
|
||||
|
||||
def epoch(value: object) -> float | None:
|
||||
if not isinstance(value, str):
|
||||
return None
|
||||
try:
|
||||
parsed = datetime.datetime.fromisoformat(value)
|
||||
except ValueError:
|
||||
return None
|
||||
return (parsed if parsed.tzinfo else parsed.replace(tzinfo=datetime.UTC)).timestamp()
|
||||
|
||||
|
||||
def count(value: object) -> int:
|
||||
return value if isinstance(value, int) and not isinstance(value, bool) and value > 0 else 0
|
||||
|
||||
|
||||
def assistant(raw: bytes) -> tuple[dict, dict, dict] | None:
|
||||
"""Returns:
|
||||
(row, message, usage) of an assistant transcript line whose message
|
||||
carries a usage map and whose row a session id, else None.
|
||||
"""
|
||||
try:
|
||||
row = json.loads(raw)
|
||||
except (ValueError, RecursionError):
|
||||
return None
|
||||
message = row.get("message") if isinstance(row, dict) and row.get("type") == "assistant" else None
|
||||
fields = message.get("usage") if isinstance(message, dict) else None
|
||||
return (row, message, fields) if isinstance(fields, dict) and isinstance(row.get("sessionId"), str) else None
|
||||
|
||||
|
||||
def message_key(ident: object, request: object) -> str:
|
||||
"""Returns:
|
||||
A short digest of (message id, request id), "" without a message id.
|
||||
"""
|
||||
if not isinstance(ident, str) or not ident:
|
||||
return ""
|
||||
text = f"{ident}\0{request if isinstance(request, str) else ''}"
|
||||
return base64.urlsafe_b64encode(hashlib.blake2b(text.encode("utf-8"), digest_size=KEY_BYTES).digest()).decode("ascii")
|
||||
|
||||
|
||||
def usage(raw: bytes) -> tuple[str, float, str, str, tuple[int, ...]] | None:
|
||||
"""Returns:
|
||||
(message key, time, session, model, counts in COUNTS order) of an
|
||||
assistant transcript line with a usage field, a time and a model,
|
||||
else None.
|
||||
"""
|
||||
found = assistant(raw)
|
||||
if found is None:
|
||||
return None
|
||||
row, message, fields = found
|
||||
ts, model = epoch(row.get("timestamp")), message.get("model")
|
||||
if ts is None or not isinstance(model, str):
|
||||
return None
|
||||
return message_key(message.get("id"), row.get("requestId")), ts, row["sessionId"], model, tuple(count(fields.get(field)) for _, field in COUNTS)
|
||||
|
||||
|
||||
def first_sight(key: str, ts: float, seen: dict[int, set[str]], keys: dict[int, set[str]]) -> bool:
|
||||
"""Returns:
|
||||
Whether the message key at ts is in neither seen nor keys at its hour
|
||||
or a neighbouring one, then files it in keys at its hour; a line
|
||||
without a key is always a first sight. The neighbours catch a copy a
|
||||
fork wrote seconds after the original across the hour.
|
||||
"""
|
||||
if not key:
|
||||
return True
|
||||
hour = int(ts // HOUR * HOUR)
|
||||
known = any(key in found.get(near, ()) for found in (seen, keys) for near in (hour - HOUR, hour, hour + HOUR))
|
||||
keys.setdefault(hour, set()).add(key)
|
||||
return not known
|
||||
|
||||
|
||||
def scan(path: Path, start: int, budget: int, seen: dict[int, set[str]]) -> tuple[int, list[tuple[float, str, str, tuple[int, ...]]], dict[int, set[str]]]:
|
||||
"""Returns:
|
||||
The offset after the last complete line read from start, at most
|
||||
about budget bytes on, the usage of every message seen holds no key
|
||||
of, and per hour the keys of every message read.
|
||||
"""
|
||||
found = []
|
||||
keys: dict[int, set[str]] = {}
|
||||
position = start
|
||||
with path.open("rb") as handle:
|
||||
handle.seek(start)
|
||||
for raw in handle:
|
||||
if not raw.endswith(b"\n") or position - start >= budget:
|
||||
break
|
||||
position += len(raw)
|
||||
read = usage(raw) if b'"usage"' in raw else None
|
||||
if read is not None and first_sight(read[0], read[1], seen, keys) and any(read[4]):
|
||||
found.append(read[1:])
|
||||
return position, found, keys
|
||||
|
||||
|
||||
def token_rows(sums: dict[tuple, list[int]]) -> list[dict]:
|
||||
return [{"t": t, "agent": agent, "model": model, **dict(zip(NAMES, counts, strict=True))} for (t, agent, model), counts in sorted(sums.items())]
|
||||
|
||||
|
||||
def add(sums: dict[tuple, list[int]], key: tuple, counts: tuple[int, ...]) -> None:
|
||||
total = sums.setdefault(key, [0] * len(COUNTS))
|
||||
for index, value in enumerate(counts):
|
||||
total[index] += value
|
||||
|
||||
|
||||
def compact(rows: list[dict], now: float) -> list[dict]:
|
||||
sums: dict[tuple, list[int]] = {}
|
||||
for row in rows:
|
||||
if timed(row) and row["t"] >= now - COARSE and isinstance(row.get("agent"), str) and isinstance(row.get("model"), str):
|
||||
add(sums, (row["t"], row["agent"], row["model"]), tuple(count(row.get(name)) for name in NAMES))
|
||||
return token_rows(sums)
|
||||
|
||||
|
||||
def start_of(entry: object, info: os.stat_result, now: float) -> int:
|
||||
if isinstance(entry, int) and not isinstance(entry, bool) and entry >= 0:
|
||||
return entry if entry <= info.st_size else 0
|
||||
return info.st_size if info.st_mtime < now - COARSE else 0
|
||||
|
||||
|
||||
def read_transcript(path: Path, entry: object, budget: int, now: float, seen: dict[int, set[str]]) -> tuple[int, int, list, dict[int, set[str]]]:
|
||||
"""Returns:
|
||||
(start, end, usages, keys per hour) of path read from the offset
|
||||
entry records, without reading once budget is spent.
|
||||
"""
|
||||
info = path.stat()
|
||||
start = start_of(entry, info, now)
|
||||
if budget <= 0 or start >= info.st_size:
|
||||
return start, start, [], {}
|
||||
return start, *scan(path, start, budget, seen)
|
||||
|
||||
|
||||
def seen_of(value: object) -> dict[int, set[str]]:
|
||||
if not isinstance(value, dict):
|
||||
return {}
|
||||
return {int(hour): {key for key in keys if isinstance(key, str)} for hour, keys in value.items() if hour.isdigit() and isinstance(keys, list)}
|
||||
|
||||
|
||||
def kept_seen(seen: dict[int, set[str]], touched: set[int], now: float) -> dict[str, list[str]]:
|
||||
floor = now // HOUR * HOUR - HOUR
|
||||
return {str(hour): sorted(keys) for hour, keys in sorted(seen.items()) if hour in touched or hour >= floor}
|
||||
|
||||
|
||||
def count_tokens(transcripts: Path, council: Path, state: Path, now: float, budget: int) -> list[dict]:
|
||||
"""Returns:
|
||||
The rows appended to <state>/tokens.jsonl by this run.
|
||||
"""
|
||||
stored = read_json(state / OFFSETS)
|
||||
offsets = stored.get("files") if isinstance(stored, dict) and isinstance(stored.get("files"), dict) else {}
|
||||
seen = seen_of(stored.get("seen") if isinstance(stored, dict) else None)
|
||||
names = claim_names(council)
|
||||
sums: dict[tuple, list[int]] = {}
|
||||
files, touched = {}, set()
|
||||
for path in sorted(transcripts.rglob("*.jsonl")):
|
||||
key = path.relative_to(transcripts).as_posix()
|
||||
try:
|
||||
start, end, found, keys = read_transcript(path, offsets.get(key), budget, now, seen)
|
||||
except OSError:
|
||||
if key in offsets:
|
||||
files[key] = offsets[key]
|
||||
continue
|
||||
files[key] = end
|
||||
budget -= end - start
|
||||
for hour, read in keys.items():
|
||||
seen.setdefault(hour, set()).update(read)
|
||||
touched.add(hour)
|
||||
for ts, session, model, counts in found:
|
||||
add(sums, (int(ts // HOUR * HOUR), names.get(session) or session[:8], model), counts)
|
||||
rows = token_rows(sums)
|
||||
append_rows(state / TOKENS, rows)
|
||||
kept = {"files": files, "seen": kept_seen(seen, touched, now)}
|
||||
if kept != stored:
|
||||
replace(state / OFFSETS, json.dumps(kept, sort_keys=True, separators=(",", ":")))
|
||||
if now % HOUR < INTERVAL:
|
||||
rewrite(state / TOKENS, compact, now)
|
||||
return rows
|
||||
|
||||
|
||||
def main() -> int:
|
||||
parser = argparse.ArgumentParser(description="Sample the host's resources and count the transcripts' tokens for the Agent Orchestrator.")
|
||||
commands = parser.add_subparsers(dest="command", required=True)
|
||||
sampling, counting = commands.add_parser("sample"), commands.add_parser("tokens")
|
||||
for sub in (sampling, counting):
|
||||
sub.add_argument("--state", type=Path, required=True)
|
||||
sub.add_argument("--council", type=Path, required=True)
|
||||
sampling.add_argument("--proc", type=Path, required=True)
|
||||
sampling.add_argument("--sys", type=Path, required=True)
|
||||
counting.add_argument("--transcripts", type=Path, required=True)
|
||||
args = parser.parse_args()
|
||||
args.state.mkdir(parents=True, exist_ok=True)
|
||||
if args.command == "sample":
|
||||
row = sample(args.proc, args.sys, args.council, command, args.state)
|
||||
record(args.state, row, row["t"])
|
||||
else:
|
||||
count_tokens(args.transcripts, args.council, args.state, time.time(), BUDGET)
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -2,12 +2,14 @@ from __future__ import annotations
|
||||
|
||||
import re
|
||||
import secrets
|
||||
import time
|
||||
from collections.abc import Callable
|
||||
from http import HTTPStatus
|
||||
from urllib.parse import urlsplit
|
||||
|
||||
from .engines import STT_LIMIT
|
||||
from .exchange import PLAIN_CSP, Exchange
|
||||
from .resources import RESOURCES, sampled
|
||||
|
||||
BODY_LIMIT = 1 << 20
|
||||
UPLOAD_LIMIT = 50 << 20
|
||||
@@ -107,6 +109,8 @@ class Handler(Exchange):
|
||||
self.answer(lambda: hub.events(self.query()), HTTPStatus.OK)
|
||||
elif path == "/api/tree":
|
||||
self.answer(hub.repositories, HTTPStatus.OK)
|
||||
elif path == "/api/resources":
|
||||
self.answer(lambda: sampled(hub.kept / RESOURCES, hub.span(self.query(), time.time())), HTTPStatus.OK)
|
||||
elif path == "/api/subagents":
|
||||
self.answer(lambda: hub.subagents(self.query()), HTTPStatus.OK)
|
||||
elif path == "/api/runner/jobs":
|
||||
|
||||
@@ -0,0 +1,38 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
from .records import epoch, parse_line, read_regular
|
||||
|
||||
RESOURCES = "resources"
|
||||
SAMPLES = "samples.jsonl"
|
||||
TOKENS = "tokens.jsonl"
|
||||
HOUR = 3600
|
||||
FILE_LIMIT = 64 << 20
|
||||
|
||||
|
||||
def rows_between(path: Path, begin: float, end: float, width: float) -> list[dict]:
|
||||
try:
|
||||
lines = read_regular(path, FILE_LIMIT).splitlines()
|
||||
except (OSError, ValueError):
|
||||
return []
|
||||
found = []
|
||||
for row in map(parse_line, lines):
|
||||
t = None if row is None else epoch(row.get("t"))
|
||||
if t is not None and begin <= t + width and t <= end:
|
||||
found.append(row)
|
||||
return found
|
||||
|
||||
|
||||
def sampled(folder: Path, span: tuple[float, float]) -> dict:
|
||||
"""Args:
|
||||
folder: the host sampler's directory, mounted read-only.
|
||||
span: (begin, end) of the page's time window in epoch seconds.
|
||||
|
||||
Returns:
|
||||
{"since", "until", "samples", "tokens"}: the rows of samples.jsonl
|
||||
within the span and those of tokens.jsonl whose hour overlaps it, as
|
||||
the host unit wrote them; the hub reads nothing of the host itself.
|
||||
"""
|
||||
begin, end = span
|
||||
return {"since": begin, "until": end, "samples": rows_between(folder / SAMPLES, begin, end, 0), "tokens": rows_between(folder / TOKENS, begin, end, HOUR)}
|
||||
@@ -0,0 +1,12 @@
|
||||
[Unit]
|
||||
Description=Agent Orchestrator resources: disks, CPU, RAM, swap, load, GPU, the queue and the transcripts' tokens, into the files the orchestrator reads
|
||||
|
||||
[Service]
|
||||
Type=oneshot
|
||||
Environment=PATH=/usr/bin:/bin
|
||||
EnvironmentFile=%h/.local/share/kvb-orchestrator/.env
|
||||
ExecStart=/usr/bin/env python3 %h/.local/share/kvb-orchestrator/client/resources.py sample --state ${KVB_ORCHESTRATOR_KEPT}/resources --council ${KVB_ORCHESTRATOR_COUNCIL} --proc /proc --sys /sys
|
||||
ExecStart=/usr/bin/env python3 %h/.local/share/kvb-orchestrator/client/resources.py tokens --state ${KVB_ORCHESTRATOR_KEPT}/resources --council ${KVB_ORCHESTRATOR_COUNCIL} --transcripts ${KVB_ORCHESTRATOR_PROJECTS}
|
||||
Nice=10
|
||||
IOSchedulingClass=idle
|
||||
TimeoutStartSec=300
|
||||
@@ -0,0 +1,9 @@
|
||||
[Unit]
|
||||
Description=Agent Orchestrator resources every minute
|
||||
|
||||
[Timer]
|
||||
OnCalendar=*-*-* *:*:00
|
||||
AccuracySec=1s
|
||||
|
||||
[Install]
|
||||
WantedBy=timers.target
|
||||
@@ -406,6 +406,8 @@ pre.notes { margin: 0; padding: 8px 10px; background: var(--sunken); border-radi
|
||||
.float-tooltip-kap { position: absolute; z-index: 9; padding: 4px 8px; border-radius: 6px; background: var(--glass-card); color: var(--text); font-size: var(--fs-sm); white-space: nowrap; pointer-events: none; box-shadow: var(--shadow); }
|
||||
{% include "style-hierarchy.css" %}
|
||||
|
||||
{% include "style-resources.css" %}
|
||||
|
||||
{% include "style-inbox.css" %}
|
||||
|
||||
{% include "style-ui.css" %}
|
||||
@@ -455,6 +457,8 @@ pre.notes { margin: 0; padding: 8px 10px; background: var(--sunken); border-radi
|
||||
|
||||
{% include "bubbles.js" %}
|
||||
|
||||
{% include "resources.js" %}
|
||||
|
||||
{% include "nav.js" %}
|
||||
|
||||
{% include "popup.js" %}
|
||||
@@ -560,6 +564,8 @@ pre.notes { margin: 0; padding: 8px 10px; background: var(--sunken); border-radi
|
||||
|
||||
{% include "page-alerts.js" %}
|
||||
|
||||
{% include "page-resources.js" %}
|
||||
|
||||
{% include "page-rules.js" %}
|
||||
|
||||
{% include "page-keyboard.js" %}
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
const Nav = (() => {
|
||||
"use strict";
|
||||
const TABS = [["stream", "Stream"], ["agents", "Agents"], ["artifacts", "Artifacts"], ["tasks", "Tasks"]];
|
||||
const TABS = [["stream", "Stream"], ["agents", "Agents"], ["artifacts", "Artifacts"], ["tasks", "Tasks"], ["resources", "Resources"]];
|
||||
const SUBS = {stream: [["inbox", "Inbox"], ["rules", "Firewall"]]};
|
||||
const VIEWS = [["orbit", "Orbit"], ["chord", "Chord"], ["swimlane", "Swimlane"], ["tree", "Tree"], ["3d", "3D"]];
|
||||
const TASK_VIEWS = [["list", "List"], ["board", "Board"], ["rasci", "RASCI"], ["gantt", "Gantt"], ["burn", "Burndown and burnup"], ["3d", "3D"]];
|
||||
@@ -9,6 +9,7 @@ const Nav = (() => {
|
||||
stream: "bi-activity", inbox: "bi-inbox", rules: "bi-card-checklist", agents: "bi-people", artifacts: "bi-git", tasks: "bi-list-check",
|
||||
orbit: "bi-globe2", chord: "bi-bezier2", swimlane: "bi-bar-chart-steps", tree: "bi-diagram-3", "3d": "bi-badge-3d", hierarchy: "bi-diagram-2",
|
||||
list: "bi-list-task", board: "bi-kanban", rasci: "bi-grid-3x3", gantt: "bi-calendar-range", burn: "bi-graph-down-arrow",
|
||||
resources: "bi-speedometer2",
|
||||
};
|
||||
const OLD = {graph: ["agents", "orbit"], tree: ["artifacts", "tree"]};
|
||||
const viewsOf = (screen) => (screen === "tasks" ? TASK_VIEWS : [...VIEWS, ...(OWN[screen] || [])]);
|
||||
|
||||
@@ -0,0 +1,174 @@
|
||||
const RESOURCE = {left: 46, right: 10, top: 6, bottom: 6, axis: 18, poll: 5000, stale: 60, tones: 6, hour: 3600, narrow: 600};
|
||||
const RESOURCE_PANELS = [["disks", "bi-hdd-stack", "Disks"], ["host", "bi-cpu", "CPU, RAM and swap"], ["gpu", "bi-gpu-card", "GPU and VRAM"], ["tokens", "bi-coin", "Tokens per agent"]];
|
||||
const RESOURCE_ZONES = {idle: "Idle: below 30 %", saturated: "Saturated: above 90 %"};
|
||||
const RESOURCE_MARKS = {
|
||||
wait: ["#bi-hourglass-split", (mark) => `${mark.peak} jobs waited while CPU and RAM stayed under 50 %`],
|
||||
asleep: ["#bi-moon", (mark) => `${mark.name} asleep`],
|
||||
unknown: ["#bi-question-circle", (mark) => `${mark.name} unknown: no reading`],
|
||||
};
|
||||
const resourceView = {data: null, dom: null, busy: false, key: "", at: 0, timer: 0};
|
||||
const resourceKey = () => JSON.stringify([ui.preset, ui.from, ui.to, ui.asOf]);
|
||||
const compactCount = (value) => (value >= 1e6 ? `${fixed(value / 1e6)}M` : value >= 1e3 ? `${fixed(value / 1e3)}k` : String(Math.round(value)));
|
||||
const resourceLabel = (line) => (line.kind === "vram" ? `${line.name} VRAM` : line.name);
|
||||
|
||||
function resourcesScreen() {
|
||||
const count = h("span", {class: "count", title: "Samples", "aria-label": "Samples"}, "0");
|
||||
const panels = RESOURCE_PANELS.map(([kind, name, label]) => {
|
||||
const legend = h("div", {class: "res-legend"});
|
||||
const plot = h("div", {class: "res-plot"});
|
||||
const panel = h("section", {class: `card res-panel res-${kind}`, "data-panel": kind, "aria-label": label},
|
||||
h("div", {class: "res-head"}, h("span", {class: "res-title", title: label, "aria-label": label}, icon(name)), legend), plot);
|
||||
return {kind, label, legend, plot, panel};
|
||||
});
|
||||
resourceView.dom = {count, panels};
|
||||
if (!resourceView.timer) {
|
||||
resourceView.timer = window.setInterval(pollResources, RESOURCE.poll);
|
||||
window.addEventListener("resize", drawResources);
|
||||
}
|
||||
window.setTimeout(() => {
|
||||
drawResources();
|
||||
fetchResources();
|
||||
}, 0);
|
||||
const reload = iconBtn("bi-arrow-clockwise", "Reload", "btn small soft", fetchResources, "Reload: read the sampler files again");
|
||||
return h("div", {class: "graph resources"},
|
||||
h("div", {class: "card controls"}, h("div", {class: "toolbar", role: "toolbar", "aria-label": "Resources"}, reload, count)),
|
||||
h("div", {class: "res-panels"}, panels.map((entry) => entry.panel)));
|
||||
}
|
||||
|
||||
function pollResources() {
|
||||
const dom = resourceView.dom;
|
||||
if (!dom || !dom.count.isConnected || document.hidden || resourceView.busy) {
|
||||
return;
|
||||
}
|
||||
if (resourceKey() !== resourceView.key || nowSeconds() - resourceView.at >= RESOURCE.stale) {
|
||||
fetchResources();
|
||||
}
|
||||
}
|
||||
|
||||
async function fetchResources() {
|
||||
if (resourceView.busy) {
|
||||
return;
|
||||
}
|
||||
resourceView.busy = true;
|
||||
resourceView.key = resourceKey();
|
||||
try {
|
||||
const response = await fetch(`/api/resources?${windowQuery()}`, {cache: "no-store"});
|
||||
if (response.ok) {
|
||||
resourceView.data = await response.json();
|
||||
}
|
||||
} catch {
|
||||
ui.offline = true;
|
||||
}
|
||||
resourceView.at = nowSeconds();
|
||||
resourceView.busy = false;
|
||||
drawResources();
|
||||
}
|
||||
|
||||
function resourceFrame(entry, span, size) {
|
||||
const axis = entry.kind === "tokens";
|
||||
const bottom = size.height - (axis ? RESOURCE.axis : RESOURCE.bottom);
|
||||
const x = (t) => RESOURCE.left + ((t - span.from) / (span.to - span.from)) * (size.width - RESOURCE.left - RESOURCE.right);
|
||||
const svg = s("svg", {class: `chart res-chart res-${entry.kind}`, "data-panel": entry.kind, "data-from": span.from, "data-to": span.to, viewBox: `0 0 ${size.width} ${size.height}`, role: "img", "aria-label": entry.label});
|
||||
const ticks = size.width < RESOURCE.narrow ? 2 : 4;
|
||||
for (let index = 0; index <= ticks; index += 1) {
|
||||
const t = span.from + ((span.to - span.from) * index) / ticks;
|
||||
svg.append(s("line", {class: "grid", x1: fixed(x(t)), x2: fixed(x(t)), y1: RESOURCE.top, y2: fixed(bottom)}));
|
||||
if (axis) {
|
||||
svg.append(s("text", {class: "axis", x: fixed(x(t)), y: size.height - 4, "text-anchor": index === 0 ? "start" : index === ticks ? "end" : "middle"}, (span.to - span.from > DAY ? absolute : clock).format(stampOf(t))));
|
||||
}
|
||||
}
|
||||
return {svg, x, top: RESOURCE.top, bottom, right: size.width - RESOURCE.right};
|
||||
}
|
||||
|
||||
function resourceScale(frame, values, label) {
|
||||
const y = (value) => frame.top + (1 - value / values[values.length - 1]) * (frame.bottom - frame.top);
|
||||
for (const value of values) {
|
||||
frame.svg.append(s("line", {class: "grid", x1: RESOURCE.left, x2: fixed(frame.right), y1: fixed(y(value)), y2: fixed(y(value))}));
|
||||
frame.svg.append(s("text", {class: "axis", x: RESOURCE.left - 6, y: fixed(y(value)), dy: "0.35em", "text-anchor": "end"}, label(value)));
|
||||
}
|
||||
return y;
|
||||
}
|
||||
|
||||
function resourceMarks(frame, marks, span) {
|
||||
for (const mark of marks) {
|
||||
const left = frame.x(Math.max(mark.from, span.from));
|
||||
const right = frame.x(Math.min(mark.to, span.to));
|
||||
if (right <= left) {
|
||||
continue;
|
||||
}
|
||||
const [href, words] = RESOURCE_MARKS[mark.kind];
|
||||
frame.svg.append(s("rect", {class: `res-mark ${mark.kind}`, "data-from": mark.from, "data-to": mark.to, "data-peak": mark.peak, x: fixed(left), width: fixed(Math.max(right - left, 2)), y: frame.top, height: fixed(frame.bottom - frame.top)}, s("title", {}, words(mark))));
|
||||
frame.svg.append(s("use", {class: `res-icon ${mark.kind}`, href, x: fixed(left + 2), y: frame.top + 2, width: 11, height: 11}));
|
||||
}
|
||||
}
|
||||
|
||||
function resourcePercent(entry, list, marks, span, size) {
|
||||
const frame = resourceFrame(entry, span, size);
|
||||
const y = resourceScale(frame, [0, 50, 100], (value) => `${value} %`);
|
||||
for (const zone of Resources.ZONES) {
|
||||
frame.svg.append(s("rect", {class: `band ${zone.kind}`, "data-zone": zone.kind, x: RESOURCE.left, width: fixed(frame.right - RESOURCE.left), y: fixed(y(zone.to)), height: fixed(y(zone.from) - y(zone.to))}, s("title", {}, RESOURCE_ZONES[zone.kind])));
|
||||
}
|
||||
resourceMarks(frame, marks, span);
|
||||
const tones = new Map();
|
||||
const legend = [];
|
||||
for (const line of list) {
|
||||
if (!tones.has(line.name)) {
|
||||
tones.set(line.name, tones.size % RESOURCE.tones);
|
||||
}
|
||||
const tone = `res-c${tones.get(line.name)} k-${line.kind}`;
|
||||
const last = [...line.points].reverse().find(([, value]) => value !== null);
|
||||
const words = `${resourceLabel(line)} ${dash(last === undefined ? null : last[1])} %`;
|
||||
for (const part of Resources.segments(line.points)) {
|
||||
const d = part.map(([t, value], index) => `${index ? "L" : "M"}${fixed(frame.x(t))},${fixed(y(Math.min(Math.max(value, 0), 100)))}`).join(" ");
|
||||
frame.svg.append(s("path", {class: `res-line ${tone}`, "data-series": line.name, "data-kind": line.kind, "data-points": part.length, d: part.length === 1 ? `${d} h1` : d}, s("title", {}, words)));
|
||||
}
|
||||
legend.push(h("span", {class: `res-key ${tone}`, "data-series": line.name, "data-kind": line.kind}, words));
|
||||
}
|
||||
return {svg: frame.svg, legend};
|
||||
}
|
||||
|
||||
function resourceTokens(entry, stack, span, size) {
|
||||
const frame = resourceFrame(entry, span, size);
|
||||
const y = resourceScale(frame, [0, Math.max(stack.top, 1)], compactCount);
|
||||
const tones = new Map(stack.agents.map((agent, index) => [agent, index % RESOURCE.tones]));
|
||||
for (const bar of stack.bars) {
|
||||
const left = frame.x(Math.max(bar.t, span.from));
|
||||
const right = frame.x(Math.min(bar.t + RESOURCE.hour, span.to));
|
||||
if (right <= left) {
|
||||
continue;
|
||||
}
|
||||
for (const part of bar.parts) {
|
||||
const words = `${part.agent} · ${absolute.format(stampOf(bar.t))}: ${part.total} tokens (input ${part.input}, output ${part.output}, cache read ${part.cache_read}, cache write ${part.cache_write}) · ${part.models.join(", ")}`;
|
||||
frame.svg.append(s("rect", {class: `res-bar res-c${tones.get(part.agent)}`, "data-agent": part.agent, "data-t": bar.t, "data-total": part.total, x: fixed(left), width: fixed(Math.max(right - left - 1, 1)), y: fixed(y(part.to)), height: fixed(Math.max(y(part.from) - y(part.to), 0.5))}, s("title", {}, words)));
|
||||
}
|
||||
}
|
||||
const legend = stack.agents.map((agent) => h("span", {class: `res-key res-c${tones.get(agent)}`, "data-series": agent}, `${agent} ${compactCount(stack.totals[agent])}`));
|
||||
return {svg: frame.svg, legend};
|
||||
}
|
||||
|
||||
function drawResources() {
|
||||
const dom = resourceView.dom;
|
||||
const data = resourceView.data;
|
||||
if (!dom || !dom.count.isConnected || data === null) {
|
||||
return;
|
||||
}
|
||||
const span = {from: data.since, to: Math.max(data.until, data.since + 60)};
|
||||
const lines = Resources.lines(data.samples);
|
||||
const marks = {
|
||||
disks: [],
|
||||
host: Resources.waits(data.samples).map((mark) => ({...mark, kind: "wait"})),
|
||||
gpu: ["asleep", "unknown"].flatMap((kind) => Resources[kind](data.samples).flatMap((gpu) => gpu.spans.map((mark) => ({...mark, kind, name: gpu.name})))),
|
||||
};
|
||||
dom.count.textContent = String(data.samples.length);
|
||||
for (const entry of dom.panels) {
|
||||
const box = entry.plot.getBoundingClientRect();
|
||||
const size = {width: Math.round(box.width), height: Math.round(box.height)};
|
||||
if (size.width < RESOURCE.left + RESOURCE.right + 10 || size.height < RESOURCE.axis + 10) {
|
||||
entry.plot.replaceChildren();
|
||||
continue;
|
||||
}
|
||||
const drawn = entry.kind === "tokens" ? resourceTokens(entry, Resources.stack(data.tokens), span, size) : resourcePercent(entry, lines[entry.kind], marks[entry.kind], span, size);
|
||||
entry.plot.replaceChildren(drawn.svg);
|
||||
entry.legend.replaceChildren(...drawn.legend);
|
||||
}
|
||||
}
|
||||
@@ -55,7 +55,7 @@
|
||||
}
|
||||
|
||||
function screenOf() {
|
||||
return {stream: streamScreen, inbox: tasksScreen, rules: rulesScreen, agents: agentsScreen, artifacts: artifactsScreen, tasks: tasksScreen}[ui.screen]();
|
||||
return {stream: streamScreen, inbox: tasksScreen, rules: rulesScreen, agents: agentsScreen, artifacts: artifactsScreen, tasks: tasksScreen, resources: resourcesScreen}[ui.screen]();
|
||||
}
|
||||
|
||||
function render(force) {
|
||||
@@ -73,7 +73,7 @@
|
||||
const main = document.getElementById("main");
|
||||
main.setAttribute("aria-labelledby", `tab-${ui.screen}`);
|
||||
main.replaceChildren(screenOf());
|
||||
document.body.classList.toggle("fit", ["tasks", "inbox"].includes(ui.screen) || (["agents", "artifacts"].includes(ui.screen) && !["tree", "hierarchy"].includes(ui.views[ui.screen])));
|
||||
document.body.classList.toggle("fit", ["tasks", "inbox", "resources"].includes(ui.screen) || (["agents", "artifacts"].includes(ui.screen) && !["tree", "hierarchy"].includes(ui.views[ui.screen])));
|
||||
drawn.screen = ui.screen;
|
||||
drawn.sig = sig;
|
||||
if (g3.host && !g3.host.isConnected) {
|
||||
|
||||
@@ -0,0 +1,144 @@
|
||||
const Resources = (() => {
|
||||
"use strict";
|
||||
const IDLE = 30;
|
||||
const SATURATED = 90;
|
||||
const CALM = 50;
|
||||
const GAP = 1800;
|
||||
const STEP = 60;
|
||||
const COUNTS = ["input", "output", "cache_read", "cache_write"];
|
||||
const ZONES = [{kind: "idle", from: 0, to: IDLE}, {kind: "saturated", from: SATURATED, to: 100}];
|
||||
const finite = (value) => typeof value === "number" && Number.isFinite(value);
|
||||
const share = (part, whole) => (finite(part) && finite(whole) && whole > 0 ? Math.round((1000 * part) / whole) / 10 : null);
|
||||
const listOf = (row, key) => (Array.isArray(row[key]) ? row[key].filter((item) => item && typeof item === "object") : []);
|
||||
const timed = (samples) => samples.filter((row) => row && finite(row.t));
|
||||
|
||||
function series(samples, pick) {
|
||||
const found = new Map();
|
||||
for (const row of timed(samples)) {
|
||||
for (const [name, kind, value] of pick(row)) {
|
||||
const key = `${kind}\u0000${name}`;
|
||||
if (!found.has(key)) {
|
||||
found.set(key, {name: String(name), kind, points: []});
|
||||
}
|
||||
found.get(key).points.push([row.t, finite(value) ? value : null]);
|
||||
}
|
||||
}
|
||||
return [...found.values()];
|
||||
}
|
||||
|
||||
function lines(samples) {
|
||||
return {
|
||||
disks: series(samples, (row) => listOf(row, "disks").map((disk) => [disk.mount, "disk", share(disk.used, disk.total)])),
|
||||
host: series(samples, (row) => [["CPU", "host", row.cpu], ["RAM", "host", row.ram], ["Swap", "host", row.swap]]),
|
||||
gpu: series(samples, (row) => listOf(row, "gpus").flatMap((gpu) => [[gpu.name, "util", gpu.asleep ? null : gpu.util], [gpu.name, "vram", gpu.asleep ? null : share(gpu.vram_used, gpu.vram_total)]])),
|
||||
};
|
||||
}
|
||||
|
||||
function segments(points) {
|
||||
const found = [];
|
||||
let current = [];
|
||||
let last = null;
|
||||
for (const [t, value] of points) {
|
||||
if (current.length && (value === null || t - last > GAP)) {
|
||||
found.push(current);
|
||||
current = [];
|
||||
}
|
||||
if (value !== null) {
|
||||
current.push([t, value]);
|
||||
}
|
||||
last = t;
|
||||
}
|
||||
if (current.length) {
|
||||
found.push(current);
|
||||
}
|
||||
return found;
|
||||
}
|
||||
|
||||
function spans(samples, test, value) {
|
||||
const rows = timed(samples);
|
||||
const found = [];
|
||||
let open = null;
|
||||
rows.forEach((row, index) => {
|
||||
if (open && index && row.t - rows[index - 1].t > GAP) {
|
||||
found.push(open);
|
||||
open = null;
|
||||
}
|
||||
if (!test(row)) {
|
||||
if (open) {
|
||||
found.push(open);
|
||||
}
|
||||
open = null;
|
||||
return;
|
||||
}
|
||||
const next = rows[index + 1];
|
||||
const end = row.t + (next && next.t - row.t <= GAP ? next.t - row.t : STEP);
|
||||
const peak = value(row);
|
||||
open = open ? {from: open.from, to: end, peak: Math.max(open.peak, peak)} : {from: row.t, to: end, peak};
|
||||
});
|
||||
if (open) {
|
||||
found.push(open);
|
||||
}
|
||||
return found;
|
||||
}
|
||||
|
||||
function waits(samples) {
|
||||
const waiting = (row) => (row.queue && finite(row.queue.waiting) ? row.queue.waiting : 0);
|
||||
return spans(samples, (row) => waiting(row) > 0 && finite(row.cpu) && finite(row.ram) && row.cpu < CALM && row.ram < CALM, waiting);
|
||||
}
|
||||
|
||||
function gpuSpans(samples, test) {
|
||||
const names = [...new Set(timed(samples).flatMap((row) => listOf(row, "gpus").filter(test).map((gpu) => String(gpu.name))))];
|
||||
return names.map((name) => ({name, spans: spans(samples, (row) => listOf(row, "gpus").some((gpu) => String(gpu.name) === name && test(gpu)), () => 1)}));
|
||||
}
|
||||
|
||||
const asleep = (samples) => gpuSpans(samples, (gpu) => gpu.asleep === true);
|
||||
const unknown = (samples) => gpuSpans(samples, (gpu) => gpu.asleep !== true && !finite(gpu.util) && !finite(gpu.vram_used));
|
||||
|
||||
function stack(tokens) {
|
||||
const totals = new Map();
|
||||
const hours = new Map();
|
||||
for (const row of tokens) {
|
||||
if (!row || !finite(row.t) || typeof row.agent !== "string") {
|
||||
continue;
|
||||
}
|
||||
const counts = COUNTS.map((name) => (finite(row[name]) && row[name] > 0 ? row[name] : 0));
|
||||
const sum = counts.reduce((total, value) => total + value, 0);
|
||||
if (!totals.has(row.agent)) {
|
||||
totals.set(row.agent, 0);
|
||||
}
|
||||
totals.set(row.agent, totals.get(row.agent) + sum);
|
||||
if (!hours.has(row.t)) {
|
||||
hours.set(row.t, new Map());
|
||||
}
|
||||
const hour = hours.get(row.t);
|
||||
if (!hour.has(row.agent)) {
|
||||
hour.set(row.agent, {agent: row.agent, total: 0, models: [], ...Object.fromEntries(COUNTS.map((name) => [name, 0]))});
|
||||
}
|
||||
const entry = hour.get(row.agent);
|
||||
COUNTS.forEach((name, index) => {
|
||||
entry[name] += counts[index];
|
||||
});
|
||||
entry.total += sum;
|
||||
if (typeof row.model === "string" && !entry.models.includes(row.model)) {
|
||||
entry.models.push(row.model);
|
||||
}
|
||||
}
|
||||
const agents = [...totals].sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0])).map(([agent]) => agent);
|
||||
const bars = [...hours.keys()].sort((a, b) => a - b).map((t) => {
|
||||
let base = 0;
|
||||
const parts = agents.filter((agent) => hours.get(t).has(agent)).map((agent) => {
|
||||
const entry = hours.get(t).get(agent);
|
||||
const part = {...entry, from: base, to: base + entry.total};
|
||||
base += entry.total;
|
||||
return part;
|
||||
});
|
||||
return {t, total: base, parts};
|
||||
});
|
||||
return {agents, totals: Object.fromEntries(totals), bars, top: Math.max(0, ...bars.map((bar) => bar.total))};
|
||||
}
|
||||
|
||||
return {IDLE, SATURATED, CALM, GAP, COUNTS, ZONES, lines, segments, spans, waits, asleep, unknown, stack};
|
||||
})();
|
||||
if (typeof module === "object" && module.exports) {
|
||||
module.exports = Resources;
|
||||
}
|
||||
@@ -0,0 +1,27 @@
|
||||
body.fit .graph.resources > .res-panels { flex: 1; }
|
||||
.res-panels { display: flex; flex-direction: column; gap: 8px; min-height: 0; }
|
||||
.res-panel { display: flex; flex: 1 1 0; flex-direction: column; gap: 2px; min-height: 72px; padding: 6px 10px; overflow: hidden; }
|
||||
.res-head { display: flex; align-items: center; gap: 4px 12px; min-width: 0; overflow: hidden; color: var(--muted); font-size: var(--fs-xs); white-space: nowrap; }
|
||||
.res-title { display: inline-flex; color: var(--text); font-size: var(--fs-md); }
|
||||
.res-legend { display: flex; gap: 4px 12px; min-width: 0; overflow: hidden; font-variant-numeric: tabular-nums; }
|
||||
.res-key { display: inline-flex; align-items: center; gap: 5px; }
|
||||
.res-key::before { content: ""; width: 12px; height: 3px; border-radius: 2px; background: var(--c); }
|
||||
.res-key.k-vram::before { background: repeating-linear-gradient(90deg, var(--c) 0 4px, transparent 4px 6px); }
|
||||
.res-plot { position: relative; flex: 1; min-height: 0; }
|
||||
.res-plot > svg { position: absolute; inset: 0; width: 100%; height: 100%; }
|
||||
svg .band.idle { fill: var(--ok-soft); }
|
||||
svg .band.saturated { fill: var(--bad-soft); }
|
||||
svg .res-line { fill: none; stroke: var(--c); stroke-width: 1.8; stroke-linecap: round; stroke-linejoin: round; }
|
||||
svg .res-line.k-vram { stroke-dasharray: 5 3; }
|
||||
svg .res-mark.wait { fill: var(--warn-soft); stroke: var(--warn); stroke-dasharray: 3 3; }
|
||||
svg .res-mark.asleep { fill: var(--neutral-soft); }
|
||||
svg .res-mark.unknown { fill: var(--neutral-soft); stroke: var(--muted); stroke-dasharray: 2 3; }
|
||||
svg .res-icon.wait { fill: var(--warn); }
|
||||
svg .res-icon.asleep, svg .res-icon.unknown { fill: var(--muted); }
|
||||
svg .res-bar { fill: var(--c); }
|
||||
.res-c0 { --c: var(--info); }
|
||||
.res-c1 { --c: var(--ok); }
|
||||
.res-c2 { --c: var(--accent); }
|
||||
.res-c3 { --c: var(--warn); }
|
||||
.res-c4 { --c: var(--violet); }
|
||||
.res-c5 { --c: var(--bad); }
|
||||
+22
-2
@@ -4,7 +4,8 @@
|
||||
# write its .env, build and start the container, register the MCP
|
||||
# server for Claude Code, build pkgmgr from the newest signed tag of
|
||||
# PKGMGR_REPO and enable the tree timer and the cleanup path unit,
|
||||
# enable the sessions timer and the offices path unit; re-enable the runner and re-register
|
||||
# enable the sessions timer and the offices path unit, enable the
|
||||
# resources timer; re-enable the runner and re-register
|
||||
# the hook when the migration took them over; an enabled runner gets
|
||||
# the fresh unit and a restart
|
||||
# migrate move every trace of the earlier name to the kvb-orchestrator names
|
||||
@@ -40,6 +41,7 @@ TREE_UNIT="${NAME}-tree"
|
||||
CLEANUP_UNIT="${NAME}-cleanup"
|
||||
OFFICES_UNIT="${NAME}-offices"
|
||||
SESSIONS_UNIT="${NAME}-sessions"
|
||||
RESOURCES_UNIT="${NAME}-resources"
|
||||
INSTALLED_SKILLS="${HOME}/.claude/skills"
|
||||
HOST_SCRIPTS=(kvb-org-organize/scripts/council.py kvb-org-organize/scripts/nofollow.py kvb-org-rasci/scripts/rasci.py)
|
||||
TREE_DIR="${HOME}/.local/state/pkgmgr/tree"
|
||||
@@ -155,7 +157,7 @@ install_hub() {
|
||||
cp -a "${REPO_ROOT}/orchestrator/." "${HUB}/"
|
||||
# Python loads a cached .pyc whose header matches the source mtime and size, and cp -a keeps the mtime: one planted in the sandbox-writable tree would run on the host
|
||||
find "${HUB}" -name __pycache__ -type d -prune -exec rm -rf {} +
|
||||
chmod +x "${HUB}/client/permission_hook.py" "${HUB}/client/runner.py" "${HUB}/client/approval.py" "${HUB}/client/cleanup.py" "${HUB}/client/offices.py" "${HUB}/client/sessions.py"
|
||||
chmod +x "${HUB}/client/permission_hook.py" "${HUB}/client/runner.py" "${HUB}/client/approval.py" "${HUB}/client/cleanup.py" "${HUB}/client/offices.py" "${HUB}/client/sessions.py" "${HUB}/client/resources.py"
|
||||
log "skills: the PermissionRequest hook client is ${HUB}/client/permission_hook.py"
|
||||
log "skills: approvals need a pinned passkey: settings, 'Passkey registrieren', then 'make orchestrator-pin CREDENTIAL=<blob>'; until then every executor refuses"
|
||||
printf 'KVB_ORCHESTRATOR_USER=%s:%s\nKVB_ORCHESTRATOR_STATE=%s\nKVB_ORCHESTRATOR_COUNCIL=%s\nKVB_ORCHESTRATOR_SKILLS=%s\nKVB_ORCHESTRATOR_PROJECTS=%s\nKVB_ORCHESTRATOR_TREE=%s\nKVB_ORCHESTRATOR_TIMERS=%s\nKVB_ORCHESTRATOR_APPROVERS=%s\nKVB_ORCHESTRATOR_KEPT=%s\n' \
|
||||
@@ -243,6 +245,18 @@ sessions_timer() {
|
||||
log "skills: ${SESSIONS_UNIT}.timer writes ${STATE}/agent-sessions.json every 15 seconds while the Agent Orchestrator page polls"
|
||||
}
|
||||
|
||||
resources_timer() {
|
||||
if ! has_systemd; then
|
||||
warn "skills: no systemd user session; the Resources view stays empty."
|
||||
return 0
|
||||
fi
|
||||
cp -f "${HUB}/systemd/${RESOURCES_UNIT}.service" "${HUB}/systemd/${RESOURCES_UNIT}.timer" "${UNITS}/"
|
||||
systemctl --user daemon-reload
|
||||
systemctl --user enable --now "${RESOURCES_UNIT}.timer"
|
||||
systemctl --user restart "${RESOURCES_UNIT}.timer"
|
||||
log "skills: ${RESOURCES_UNIT}.timer samples the host and counts the transcripts' tokens into ${HOME}/.local/state/agents-council/resources every minute"
|
||||
}
|
||||
|
||||
skills_current() {
|
||||
local script
|
||||
for script in "${HOST_SCRIPTS[@]}"; do
|
||||
@@ -271,6 +285,7 @@ install_all() {
|
||||
install_hub
|
||||
tree_timer
|
||||
sessions_timer
|
||||
resources_timer
|
||||
offices_unit
|
||||
if ((TOOK_RUNNER)) || runner_enabled; then runner || true; fi
|
||||
if ((TOOK_HOOK)); then
|
||||
@@ -354,6 +369,11 @@ check() {
|
||||
else
|
||||
todo "${SESSIONS_UNIT}.timer off: agent states wait for make orchestrator-install"
|
||||
fi
|
||||
if [[ "$(systemctl --user is-active "${RESOURCES_UNIT}.timer" 2>/dev/null || true)" == "active" ]]; then
|
||||
ok "${RESOURCES_UNIT}.timer active"
|
||||
else
|
||||
todo "${RESOURCES_UNIT}.timer off: the Resources view waits for make orchestrator-install"
|
||||
fi
|
||||
if [[ "$(systemctl --user is-active "${OFFICES_UNIT}.path" 2>/dev/null || true)" == "active" ]]; then
|
||||
ok "${OFFICES_UNIT}.path active"
|
||||
else
|
||||
|
||||
@@ -35,9 +35,9 @@ OUT = Path(os.environ.get("E2E_OUT", "/tmp/orchestrator-phase4-shots"))
|
||||
WIDTHS = ((1920, 1080), (1366, 768), (390, 844))
|
||||
VIEWS = ("orbit", "chord", "swimlane", "tree", "3d")
|
||||
TASK_VIEWS = ("list", "board", "rasci", "gantt", "burn", "3d")
|
||||
SCREENS = ["stream", "stream/inbox", "stream/rules", *(f"{tab}/{view}" for tab in ("agents", "artifacts") for view in VIEWS), "agents/hierarchy", *(f"tasks/{view}" for view in TASK_VIEWS)]
|
||||
SCREENS = ["stream", "stream/inbox", "stream/rules", *(f"{tab}/{view}" for tab in ("agents", "artifacts") for view in VIEWS), "agents/hierarchy", *(f"tasks/{view}" for view in TASK_VIEWS), "resources"]
|
||||
SUBS = ["Inbox", "Firewall"]
|
||||
TABS = ["Stream", "Agents", "Artifacts", "Tasks"]
|
||||
TABS = ["Stream", "Agents", "Artifacts", "Tasks", "Resources"]
|
||||
HELPERS = """
|
||||
const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
|
||||
const until = async (check, ms) => {
|
||||
@@ -1768,7 +1768,7 @@ class TestOrchestratorE2E(HubCase):
|
||||
self.assertEqual((probe["title"], probe["h1"]), ("(2) Agent Orchestrator", "Agent Orchestrator"))
|
||||
self.assertEqual(probe["header"], ["radio-open", "bubbles-toggle", "palette-open", "settings-open"])
|
||||
self.assertEqual((probe["tabs"], probe["subs"]), (TABS, SUBS))
|
||||
self.assertEqual(probe["badges"], {"tab-stream": None, "tab-inbox": "2", "tab-rules": None, "tab-agents": "1", "tab-artifacts": None, "tab-tasks": None})
|
||||
self.assertEqual(probe["badges"], {"tab-stream": None, "tab-inbox": "2", "tab-rules": None, "tab-agents": "1", "tab-artifacts": None, "tab-tasks": None, "tab-resources": None})
|
||||
self.assertEqual(probe["banner"], "MAYDAY · Docs: disk full")
|
||||
self.assertFalse(probe["rail"])
|
||||
self.assertEqual((probe["time"], probe["replay"]), ("24 h ▾", True))
|
||||
|
||||
@@ -0,0 +1,152 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import shutil
|
||||
import sys
|
||||
import time
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
|
||||
HERE = Path(__file__).resolve().parent
|
||||
sys.path[:0] = [str(HERE.parent), str(HERE)]
|
||||
|
||||
import test_orchestrator_e2e as base
|
||||
from hub_case import HubCase
|
||||
|
||||
OUT = Path(os.environ.get("E2E_OUT", "/tmp/workbench-w4-shots"))
|
||||
GPU = "0000:01:00.0"
|
||||
ROWS = 360
|
||||
WAITED = (120, 180)
|
||||
ASLEEP = 180
|
||||
UNREAD = (300, 320)
|
||||
PANELS = {"disks": "Disks", "host": "CPU, RAM and swap", "gpu": "GPU and VRAM", "tokens": "Tokens per agent"}
|
||||
GUARD = """() => {
|
||||
window.e2eViolations = [];
|
||||
document.addEventListener("securitypolicyviolation", (event) => window.e2eViolations.push(`${event.violatedDirective} ${event.blockedURI}`));
|
||||
}"""
|
||||
CHARTS = """(async () => {@HELPERS@
|
||||
const counted = () => text(document.querySelector('#main [aria-label="Samples"]'));
|
||||
if (!(await until(() => all("#main svg.res-chart").length === 4 && counted() === "@ROWS@", 15000))) {
|
||||
return {error: text(document.getElementById("main")), count: counted()};
|
||||
}
|
||||
await sleep(300);
|
||||
const charts = all("#main svg.res-chart");
|
||||
const number = (node, key) => Number(node.dataset[key]);
|
||||
const attr = (node, key) => Number(node.getAttribute(key));
|
||||
const stacks = {};
|
||||
for (const rect of all("#main rect.res-bar")) {
|
||||
(stacks[rect.dataset.t] = stacks[rect.dataset.t] || []).push({agent: rect.dataset.agent, y: attr(rect, "y"), height: attr(rect, "height")});
|
||||
}
|
||||
const piles = Object.values(stacks).map((parts) => parts.sort((a, b) => b.y - a.y));
|
||||
const labelled = (node) => (node ? [node.getAttribute("title"), node.getAttribute("aria-label")] : null);
|
||||
return {
|
||||
panels: all("#main .res-panel").map((panel) => panel.dataset.panel),
|
||||
axes: [...new Set(charts.map((chart) => `${chart.dataset.from}|${chart.dataset.to}`))],
|
||||
since: number(charts[0], "from"),
|
||||
bands: charts.filter((chart) => chart.dataset.panel !== "tokens").map((chart) => {
|
||||
const [idle, saturated] = [chart.querySelector("rect.band.idle"), chart.querySelector("rect.band.saturated")];
|
||||
return [chart.dataset.panel, attr(idle, "height") / attr(saturated, "height"), attr(idle, "y") > attr(saturated, "y")];
|
||||
}),
|
||||
series: charts.map((chart) => [chart.dataset.panel, [...new Set(all("path.res-line", chart).map((path) => `${path.dataset.series}|${path.dataset.kind}`))]]),
|
||||
waits: all("#main rect.res-mark.wait").map((rect) => [number(rect, "from"), number(rect, "to"), number(rect, "peak")]),
|
||||
asleep: all("#main rect.res-mark.asleep").map((rect) => [number(rect, "from"), number(rect, "to")]),
|
||||
unknown: all("#main rect.res-mark.unknown").map((rect) => [number(rect, "from"), number(rect, "to")]),
|
||||
agents: [...new Set(all("#main rect.res-bar").map((rect) => rect.dataset.agent))].sort(),
|
||||
bottoms: [...new Set(piles.map((parts) => parts[0].agent))],
|
||||
stacked: piles.every((parts) => parts.every((part, index) => !index || Math.abs(part.y + part.height - parts[index - 1].y) < 0.2)),
|
||||
legend: all("#main .res-host .res-key").map(text),
|
||||
labels: {tab: labelled(document.getElementById("tab-resources")), reload: labelled(document.querySelector('#main [aria-label="Reload"]')), titles: all("#main .res-title").map(labelled)},
|
||||
unlabelled: all("#main svg.bi").map((svg) => svg.parentElement).filter((owner) => !text(owner) && !(owner.getAttribute("title") && owner.getAttribute("aria-label"))).length,
|
||||
plots: all("#main .res-plot").map((plot) => plot.getBoundingClientRect().height > 20),
|
||||
tall: document.documentElement.scrollHeight > innerHeight,
|
||||
wide: document.documentElement.scrollWidth > innerWidth + 1,
|
||||
violations: window.e2eViolations || null,
|
||||
};
|
||||
})()"""
|
||||
KEY = """(async () => {@HELPERS@
|
||||
document.body.dispatchEvent(new KeyboardEvent("keydown", {key: "5", bubbles: true}));
|
||||
const opened = Boolean(await until(() => location.hash === "#resources" && document.querySelector("#main svg.res-chart"), 8000));
|
||||
return {opened, current: document.getElementById("tab-resources").getAttribute("aria-current")};
|
||||
})()"""
|
||||
|
||||
|
||||
def probe(template: str) -> str:
|
||||
return template.replace("@HELPERS@", base.HELPERS).replace("@ROWS@", str(ROWS))
|
||||
|
||||
|
||||
def quiet(found: list[str] | None) -> list[str] | None:
|
||||
return None if found is None else [seen for seen in found if not seen.endswith("/favicon.ico")]
|
||||
|
||||
|
||||
@unittest.skipUnless(shutil.which(os.environ.get("CHROMIUM", "chromium")), "no Chromium")
|
||||
class TestWorkbenchResources(HubCase):
|
||||
drive = base.TestOrchestratorE2E.drive
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
OUT.mkdir(parents=True, exist_ok=True)
|
||||
now = int(time.time())
|
||||
self.start = now - now % 60 - ROWS * 60
|
||||
samples = [self.sample(index) for index in range(ROWS)]
|
||||
hour = self.start - self.start % 3600
|
||||
tokens = [{"t": hour + step * 3600, "agent": "Skills", "model": "claude-opus-5-5", "input": 1000 * (step + 1), "output": 500, "cache_read": 20000, "cache_write": 1000} for step in range(7)]
|
||||
tokens += [{"t": hour + step * 3600, "agent": "Stage", "model": "claude-haiku-4-5", "input": 200, "output": 100, "cache_read": 3000, "cache_write": 0} for step in range(0, 7, 2)]
|
||||
folder = self.kept / "resources"
|
||||
folder.mkdir()
|
||||
(folder / "samples.jsonl").write_text(self.jsonl(samples), encoding="utf-8")
|
||||
(folder / "tokens.jsonl").write_text(self.jsonl([row for row in tokens if row["t"] <= now]), encoding="utf-8")
|
||||
self.base = f"http://{self.host}/"
|
||||
|
||||
def sample(self, index: int) -> dict:
|
||||
waited, saturated = WAITED[0] <= index < WAITED[1], WAITED[1] <= index < WAITED[1] + 60
|
||||
cpu, ram, waiting = (25, 35, 2) if waited else (95, 60, 1) if saturated else (10, 20, 0) if index < WAITED[0] else (60, 55, 0)
|
||||
gpu = {"name": GPU, "asleep": True} if index < ASLEEP else {"name": GPU, "util": None, "vram_used": None, "vram_total": None} if UNREAD[0] <= index < UNREAD[1] else {"name": GPU, "util": 50, "vram_used": 2048 << 20, "vram_total": 8192 << 20}
|
||||
root = 950 if index >= ROWS - 60 else 400
|
||||
disks = [{"mount": "/", "used": root << 30, "total": 1000 << 30}, {"mount": "/boot/efi", "used": 3 << 20, "total": 100 << 20}]
|
||||
return {"t": self.start + index * 60, "disks": disks, "cpu": cpu, "ram": ram, "swap": 5, "load": cpu / 25, "gpus": [gpu], "queue": {"waiting": waiting, "active": 1}}
|
||||
|
||||
def steps(self, browser: str) -> list[dict]:
|
||||
steps = [{"width": width, "height": height, "url": f"{self.base}?{width}#resources", "wait": 1200, "probe": probe(CHARTS), "shot": str(OUT / f"{browser}-{width}-resources.png")} for width, height in base.WIDTHS]
|
||||
steps.append({"width": 1366, "height": 768, "url": f"{self.base}?key#stream", "wait": 1200, "probe": probe(KEY), "shot": None})
|
||||
return steps
|
||||
|
||||
def check(self, data: dict, browser: str) -> None:
|
||||
results = data["results"]
|
||||
for (width, _), found in zip(base.WIDTHS, results, strict=False):
|
||||
with self.subTest(browser=browser, width=width):
|
||||
self.assertNotIn("error", found, found)
|
||||
self.assertEqual(found["panels"], list(PANELS))
|
||||
self.assertEqual(len(found["axes"]), 1, found["axes"])
|
||||
self.assertLess(abs(found["since"] - (time.time() - 86400)), 600)
|
||||
self.assertEqual([(panel, lower) for panel, _, lower in found["bands"]], [("disks", True), ("host", True), ("gpu", True)])
|
||||
for panel, ratio, _ in found["bands"]:
|
||||
self.assertAlmostEqual(ratio, 3, delta=0.2, msg=panel)
|
||||
self.assertEqual(found["series"], [["disks", ["/|disk", "/boot/efi|disk"]], ["host", ["CPU|host", "RAM|host", "Swap|host"]], ["gpu", [f"{GPU}|util", f"{GPU}|vram"]], ["tokens", []]])
|
||||
self.assertEqual(found["waits"], [[self.start + WAITED[0] * 60, self.start + WAITED[1] * 60, 2]])
|
||||
self.assertEqual(found["asleep"], [[self.start, self.start + ASLEEP * 60]])
|
||||
self.assertEqual(found["unknown"], [[self.start + UNREAD[0] * 60, self.start + UNREAD[1] * 60]])
|
||||
self.assertEqual((found["agents"], found["bottoms"], found["stacked"]), (["Skills", "Stage"], ["Skills"], True))
|
||||
self.assertEqual(found["legend"], ["CPU 60 %", "RAM 55 %", "Swap 5 %"])
|
||||
self.assertEqual(found["labels"], {"tab": ["Resources (key 5)", "Resources"], "reload": ["Reload: read the sampler files again", "Reload"], "titles": [[name, name] for name in PANELS.values()]})
|
||||
self.assertEqual((found["unlabelled"], found["plots"]), (0, [True] * 4))
|
||||
self.assertEqual((found["tall"], found["wide"]), (False, False), "the page scrolls")
|
||||
self.assertEqual(quiet(found["violations"]), [])
|
||||
self.assertEqual(results[len(base.WIDTHS)], {"opened": True, "current": "page"})
|
||||
|
||||
def test_chromium_draws_the_resources_charts_from_the_sampler_files_at_three_widths(self):
|
||||
data = self.drive(self.steps("chromium"), None, init=GUARD)
|
||||
(OUT / "chromium-resources.json").write_text(json.dumps(data, indent=1), encoding="utf-8")
|
||||
self.assertEqual(data["errors"], [])
|
||||
self.check(data, "chromium")
|
||||
|
||||
@unittest.skipUnless(shutil.which(os.environ.get("FIREFOX", "firefox")), "no Firefox")
|
||||
def test_firefox_draws_the_resources_charts_from_the_sampler_files_at_three_widths(self):
|
||||
data = self.drive(self.steps("firefox"), None, "bidi.mjs", os.environ.get("FIREFOX", "firefox"), init=GUARD)
|
||||
(OUT / "firefox-resources.json").write_text(json.dumps(data, indent=1), encoding="utf-8")
|
||||
self.assertEqual([error for error in data["errors"] if "/favicon.ico" not in error], [])
|
||||
self.check(data, "firefox")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -64,9 +64,9 @@ class TestNavModule(unittest.TestCase):
|
||||
done = subprocess.run(["node", "-e", NAV_PROBE, str(NAV)], capture_output=True, text=True, timeout=60, check=True)
|
||||
cls.data = json.loads(done.stdout)
|
||||
|
||||
def test_the_navigation_has_four_entries_and_stream_has_inbox_and_rules(self):
|
||||
def test_the_navigation_has_five_entries_and_stream_has_inbox_and_rules(self):
|
||||
stream = ["stream", "Stream", [["inbox", "Inbox"], ["rules", "Firewall"]]]
|
||||
self.assertEqual(self.data["tabs"], [stream, ["agents", "Agents", []], ["artifacts", "Artifacts", []], ["tasks", "Tasks", []]])
|
||||
self.assertEqual(self.data["tabs"], [stream, ["agents", "Agents", []], ["artifacts", "Artifacts", []], ["tasks", "Tasks", []], ["resources", "Resources", []]])
|
||||
|
||||
def test_agents_and_artifacts_offer_five_views_in_order_and_tasks_six(self):
|
||||
shared = [["orbit", "Orbit"], ["chord", "Chord"], ["swimlane", "Swimlane"], ["tree", "Tree"], ["3d", "3D"]]
|
||||
|
||||
@@ -0,0 +1,229 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import subprocess
|
||||
import time
|
||||
import unittest
|
||||
|
||||
from hub_case import HubCase
|
||||
from hub_support import HUB, NAV
|
||||
|
||||
MODULE = HUB / "templates" / "resources.js"
|
||||
PAGE = HUB / "templates" / "page-resources.js"
|
||||
STYLE = HUB / "templates" / "style-resources.css"
|
||||
T = 1_760_000_000
|
||||
FIXTURE = """
|
||||
const asleep = {name: "0000:01:00.0", asleep: true};
|
||||
const awake = (util, used) => ({name: "0000:01:00.0", util, vram_used: used, vram_total: 1000});
|
||||
const unread = {name: "0000:01:00.0", util: null, vram_used: null, vram_total: null};
|
||||
const row = (t, cpu, ram, waiting, gpu, disks) => ({t, cpu, ram, swap: 0, load: 1, disks, gpus: [gpu], queue: {waiting, active: 1}});
|
||||
const root = [{mount: "/", used: 250, total: 1000}];
|
||||
const samples = [
|
||||
row(T, 20, 30, 0, asleep, root),
|
||||
row(T + 60, 40, 45, 2, asleep, root),
|
||||
row(T + 120, 45, 40, 3, awake(80, 500), [...root, {mount: "/mnt", used: 1, total: 4}]),
|
||||
row(T + 180, 70, 40, 3, awake(90, 600), root),
|
||||
row(T + 240, 30, 30, 1, awake(10, 100), root),
|
||||
row(T + 300, 35, 30, 0, unread, root),
|
||||
row(T + 5000, 10, 10, 1, asleep, root),
|
||||
row(T + 5060, 10, 10, 0, asleep, root),
|
||||
];
|
||||
const tokens = [
|
||||
{t: T, agent: "Skills", model: "opus", input: 10, output: 20, cache_read: 300, cache_write: 40},
|
||||
{t: T, agent: "Stage", model: "opus", input: 1, output: 2, cache_read: 3, cache_write: 4},
|
||||
{t: T, agent: "Skills", model: "haiku", input: 5, output: 0, cache_read: 0, cache_write: 0},
|
||||
{t: T + 3600, agent: "Stage", model: "opus", input: 500, output: 0, cache_read: 0, cache_write: 0},
|
||||
{t: "x", agent: "Bad", model: "opus", input: 1, output: 0, cache_read: 0, cache_write: 0},
|
||||
];
|
||||
"""
|
||||
PROBE = """
|
||||
const R = require(process.argv[1]);
|
||||
const T = Number(process.argv[2]);
|
||||
@FIXTURE@
|
||||
const lines = R.lines(samples);
|
||||
const pick = (list) => list.map((entry) => [entry.name, entry.kind, entry.points]);
|
||||
console.log(JSON.stringify({
|
||||
zones: R.ZONES,
|
||||
disks: pick(lines.disks),
|
||||
host: pick(lines.host),
|
||||
gpu: pick(lines.gpu),
|
||||
segments: R.segments([[T, 1], [T + 60, null], [T + 120, 3], [T + 180, 4], [T + 4000, 5]]),
|
||||
waits: R.waits(samples),
|
||||
asleep: R.asleep(samples),
|
||||
unknown: R.unknown(samples),
|
||||
stack: R.stack(tokens),
|
||||
}));
|
||||
""".replace("@FIXTURE@", FIXTURE)
|
||||
PAGE_PROBE = """
|
||||
const fs = require("fs");
|
||||
const vm = require("vm");
|
||||
const Resources = require(process.argv[1]);
|
||||
const code = fs.readFileSync(process.argv[2], "utf8");
|
||||
const T = Number(process.argv[3]);
|
||||
@FIXTURE@
|
||||
const make = (tag, attrs, children) => {
|
||||
const node = {tag, attrs: {}, children: [], textContent: "", isConnected: true,
|
||||
append(...items) { this.children.push(...items); },
|
||||
replaceChildren(...items) { this.children = items; },
|
||||
getBoundingClientRect() { return {width: 800, height: 120}; }};
|
||||
for (const [key, value] of Object.entries(attrs)) {
|
||||
if (key !== "on" && value !== null && value !== undefined && value !== false) {
|
||||
node.attrs[key] = String(value);
|
||||
}
|
||||
}
|
||||
node.children = children.flat(Infinity).filter((child) => child !== null && child !== undefined && child !== false);
|
||||
return node;
|
||||
};
|
||||
const h = (tag, attrs, ...children) => make(tag, attrs, children);
|
||||
const icon = (name) => make("svg", {class: "bi", "data-icon": name}, []);
|
||||
const fetched = [];
|
||||
const context = {
|
||||
Resources, h, s: h, icon, DAY: 86400, ui: {preset: 86400, from: null, to: null, asOf: null, offline: false},
|
||||
iconBtn: (name, label, cls, run, hint) => make("button", {title: hint, "aria-label": label}, [icon(name)]),
|
||||
fixed: (value) => value.toFixed(1), dash: (value) => (value === null ? "–" : value), absolute: {format: () => "d"}, clock: {format: () => "c"}, stampOf: (t) => new Date(t * 1000),
|
||||
window: {setTimeout: (run) => run(), setInterval: () => 1, addEventListener: () => null}, document: {hidden: false},
|
||||
windowQuery: () => "from=1", nowSeconds: () => 0,
|
||||
fetch: async (url) => {
|
||||
fetched.push(url);
|
||||
return {ok: true, json: async () => ({since: T - 600, until: T + 6000, samples, tokens})};
|
||||
},
|
||||
};
|
||||
vm.runInNewContext(`${code}\\nscreen = resourcesScreen();`, context);
|
||||
const walk = (node, found) => {
|
||||
if (node && typeof node === "object") {
|
||||
found.push(node);
|
||||
node.children.forEach((child) => walk(child, found));
|
||||
}
|
||||
return found;
|
||||
};
|
||||
const has = (node, name) => (node.attrs.class || "").split(" ").includes(name);
|
||||
(async () => {
|
||||
for (let round = 0; round < 5; round += 1) {
|
||||
await new Promise((resolve) => setImmediate(resolve));
|
||||
}
|
||||
const all = walk(context.screen, []);
|
||||
const charts = all.filter((node) => has(node, "res-chart"));
|
||||
const marks = (kind) => all.filter((node) => has(node, "res-mark") && has(node, kind)).map((node) => [Number(node.attrs["data-from"]), Number(node.attrs["data-to"])]);
|
||||
console.log(JSON.stringify({
|
||||
fetched,
|
||||
count: all.find((node) => node.attrs["aria-label"] === "Samples").textContent,
|
||||
charts: charts.map((chart) => [chart.attrs["data-panel"], chart.attrs["data-from"], chart.attrs["data-to"]]),
|
||||
bands: charts.map((chart) => walk(chart, []).filter((node) => has(node, "band")).map((node) => node.attrs["data-zone"])),
|
||||
lines: [...new Set(all.filter((node) => has(node, "res-line")).map((node) => `${node.attrs["data-series"]}|${node.attrs["data-kind"]}`))],
|
||||
waits: marks("wait"),
|
||||
asleep: marks("asleep"),
|
||||
unknown: marks("unknown"),
|
||||
words: all.filter((node) => has(node, "res-mark") && has(node, "unknown")).map((node) => node.children[0].children.join("")),
|
||||
bars: all.filter((node) => has(node, "res-bar")).map((node) => [node.attrs["data-agent"], Number(node.attrs["data-t"])]),
|
||||
legends: all.filter((node) => has(node, "res-key")).map((node) => node.children.join("")),
|
||||
titles: all.filter((node) => has(node, "res-title")).map((node) => [node.attrs.title, node.attrs["aria-label"], node.children[0].attrs["data-icon"]]),
|
||||
}));
|
||||
})();
|
||||
""".replace("@FIXTURE@", FIXTURE)
|
||||
|
||||
|
||||
class TestResourcesModule(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
done = subprocess.run(["node", "-e", PROBE, str(MODULE), str(T)], capture_output=True, text=True, timeout=60, check=True)
|
||||
cls.data = json.loads(done.stdout)
|
||||
|
||||
def test_bands_mark_idle_below_thirty_and_saturated_above_ninety_percent(self):
|
||||
self.assertEqual(self.data["zones"], [{"kind": "idle", "from": 0, "to": 30}, {"kind": "saturated", "from": 90, "to": 100}])
|
||||
|
||||
def test_disks_cpu_ram_swap_gpu_and_vram_become_percent_lines_on_one_time_axis(self):
|
||||
times = [T, T + 60, T + 120, T + 180, T + 240, T + 300, T + 5000, T + 5060]
|
||||
self.assertEqual(self.data["disks"], [["/", "disk", [[t, 25] for t in times]], ["/mnt", "disk", [[T + 120, 25]]]])
|
||||
self.assertEqual([name for name, _, _ in self.data["host"]], ["CPU", "RAM", "Swap"])
|
||||
self.assertEqual(self.data["host"][0][2], [[t, cpu] for t, cpu in zip(times, (20, 40, 45, 70, 30, 35, 10, 10), strict=True)])
|
||||
gpu, vram = self.data["gpu"]
|
||||
self.assertEqual((gpu[0], gpu[1], vram[0], vram[1]), ("0000:01:00.0", "util", "0000:01:00.0", "vram"))
|
||||
self.assertEqual(gpu[2], [[T, None], [T + 60, None], [T + 120, 80], [T + 180, 90], [T + 240, 10], [T + 300, None], [T + 5000, None], [T + 5060, None]])
|
||||
self.assertEqual(vram[2][2:5], [[T + 120, 50], [T + 180, 60], [T + 240, 10]])
|
||||
|
||||
def test_a_line_breaks_at_a_missing_value_and_at_a_gap_of_more_than_half_an_hour(self):
|
||||
self.assertEqual(self.data["segments"], [[[T, 1]], [[T + 120, 3], [T + 180, 4]], [[T + 4000, 5]]])
|
||||
|
||||
def test_a_marker_spans_rows_where_jobs_waited_while_cpu_and_ram_stayed_under_half(self):
|
||||
self.assertEqual(self.data["waits"], [{"from": T + 60, "to": T + 180, "peak": 3}, {"from": T + 240, "to": T + 300, "peak": 1}, {"from": T + 5000, "to": T + 5060, "peak": 1}])
|
||||
|
||||
def test_an_asleep_gpu_gets_its_own_spans(self):
|
||||
self.assertEqual(self.data["asleep"], [{"name": "0000:01:00.0", "spans": [{"from": T, "to": T + 120, "peak": 1}, {"from": T + 5000, "to": T + 5120, "peak": 1}]}])
|
||||
|
||||
def test_an_awake_gpu_without_a_reading_gets_unknown_spans_apart_from_asleep_ones(self):
|
||||
self.assertEqual(self.data["unknown"], [{"name": "0000:01:00.0", "spans": [{"from": T + 300, "to": T + 360, "peak": 1}]}])
|
||||
|
||||
def test_tokens_stack_per_hour_with_the_biggest_agent_at_the_bottom(self):
|
||||
stack = self.data["stack"]
|
||||
self.assertEqual((stack["agents"], stack["top"]), (["Stage", "Skills"], 500))
|
||||
first, second = stack["bars"]
|
||||
self.assertEqual((first["t"], first["total"]), (T, 385))
|
||||
self.assertEqual([(part["agent"], part["from"], part["to"], part["models"]) for part in first["parts"]], [("Stage", 0, 10, ["opus"]), ("Skills", 10, 385, ["opus", "haiku"])])
|
||||
self.assertEqual({key: first["parts"][1][key] for key in ("input", "output", "cache_read", "cache_write")}, {"input": 15, "output": 20, "cache_read": 300, "cache_write": 40})
|
||||
self.assertEqual((second["t"], [(part["agent"], part["to"]) for part in second["parts"]]), (T + 3600, [("Stage", 500)]))
|
||||
|
||||
|
||||
class TestResourcesPage(unittest.TestCase):
|
||||
def test_the_view_draws_four_panels_on_one_time_axis_from_the_api_rows_alone(self):
|
||||
done = subprocess.run(["node", "-e", PAGE_PROBE, str(MODULE), str(PAGE), str(T)], capture_output=True, text=True, timeout=60, check=True)
|
||||
found = json.loads(done.stdout)
|
||||
self.assertEqual((found["fetched"], found["count"]), (["/api/resources?from=1"], "8"))
|
||||
axis = [str(T - 600), str(T + 6000)]
|
||||
self.assertEqual(found["charts"], [[panel, *axis] for panel in ("disks", "host", "gpu", "tokens")])
|
||||
self.assertEqual(found["bands"], [["idle", "saturated"]] * 3 + [[]])
|
||||
self.assertEqual(found["lines"], ["/|disk", "/mnt|disk", "CPU|host", "RAM|host", "Swap|host", "0000:01:00.0|util", "0000:01:00.0|vram"])
|
||||
self.assertEqual(found["waits"], [[T + 60, T + 180], [T + 240, T + 300], [T + 5000, T + 5060]])
|
||||
self.assertEqual(found["asleep"], [[T, T + 120], [T + 5000, T + 5120]])
|
||||
self.assertEqual((found["unknown"], found["words"]), ([[T + 300, T + 360]], ["0000:01:00.0 unknown: no reading"]))
|
||||
self.assertEqual(found["bars"], [["Stage", T], ["Skills", T], ["Stage", T + 3600]])
|
||||
self.assertEqual(found["legends"], ["/ 25 %", "/mnt 25 %", "CPU 10 %", "RAM 10 %", "Swap 0 %", "0000:01:00.0 10 %", "0000:01:00.0 VRAM 10 %", "Stage 510", "Skills 375"])
|
||||
names = [("Disks", "bi-hdd-stack"), ("CPU, RAM and swap", "bi-cpu"), ("GPU and VRAM", "bi-gpu-card"), ("Tokens per agent", "bi-coin")]
|
||||
self.assertEqual(found["titles"], [[name, name, icon] for name, icon in names])
|
||||
|
||||
|
||||
class TestHubResources(HubCase):
|
||||
def write(self, name: str, rows: list[dict], tail: str = "") -> None:
|
||||
folder = self.kept / "resources"
|
||||
folder.mkdir(exist_ok=True)
|
||||
(folder / name).write_text(self.jsonl(rows) + tail, encoding="utf-8")
|
||||
|
||||
def fetch(self, query: str) -> tuple[int, bytes]:
|
||||
status, _, body = self.get(f"/api/resources?{query}")
|
||||
return status, body
|
||||
|
||||
def test_the_view_reads_only_the_sampler_files_within_the_window(self):
|
||||
samples = [{"t": T + step, "cpu": step / 60, "ram": 1, "swap": 0, "load": 0, "disks": [], "gpus": [], "queue": {"waiting": 0, "active": 0}} for step in (0, 60, 120, 180)]
|
||||
tokens = [{"t": T - 3600, "agent": "Skills", "model": "opus", "input": 1, "output": 0, "cache_read": 0, "cache_write": 0}, {"t": T + 3600, "agent": "Skills", "model": "opus", "input": 2, "output": 0, "cache_read": 0, "cache_write": 0}]
|
||||
self.write("samples.jsonl", samples, "{broken\n[1]\n")
|
||||
self.write("tokens.jsonl", tokens)
|
||||
status, body = self.fetch(f"from={T + 60}&to={T + 120}")
|
||||
self.assertEqual(status, 200, body)
|
||||
self.assertEqual(json.loads(body), {"since": T + 60, "until": T + 120, "samples": samples[1:3], "tokens": []})
|
||||
status, body = self.fetch(f"from={T - 1800}&to={T + 3600}")
|
||||
self.assertEqual(json.loads(body)["tokens"], tokens)
|
||||
self.assertEqual(json.loads(body)["samples"], samples)
|
||||
|
||||
def test_missing_files_answer_empty_rows_and_a_reversed_window_is_refused(self):
|
||||
status, body = self.fetch("")
|
||||
self.assertEqual(status, 200, body)
|
||||
found = json.loads(body)
|
||||
self.assertEqual((found["samples"], found["tokens"]), ([], []))
|
||||
self.assertAlmostEqual(found["until"] - found["since"], 86400)
|
||||
self.assertLess(abs(found["until"] - time.time()), 30)
|
||||
self.assertEqual(self.fetch(f"from={T + 10}&to={T}")[0], 400)
|
||||
self.assertEqual(self.fetch("from=x")[0], 400)
|
||||
|
||||
def test_the_page_carries_the_resources_tab_module_view_and_icons(self):
|
||||
html = self.index()
|
||||
self.assertIn(MODULE.read_text(encoding="utf-8").strip(), html)
|
||||
self.assertIn(PAGE.read_text(encoding="utf-8").strip(), html)
|
||||
self.assertIn(STYLE.read_text(encoding="utf-8").strip(), html)
|
||||
for name in ("speedometer2", "hdd-stack", "cpu", "gpu-card", "coin", "hourglass-split", "moon", "question-circle"):
|
||||
with self.subTest(icon=name):
|
||||
self.assertIn(f'<symbol id="bi-{name}"', html)
|
||||
done = subprocess.run(["node", "-e", "const N = require(process.argv[1]); console.log(JSON.stringify([N.TABS.at(-1), N.ICONS.resources, N.route('#resources')]));", str(NAV)], capture_output=True, text=True, timeout=60, check=True)
|
||||
self.assertEqual(json.loads(done.stdout), [["resources", "Resources"], "bi-speedometer2", {"screen": "resources", "view": None, "redirect": None}])
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -123,6 +123,7 @@ class TestOrchestratorMigration(unittest.TestCase):
|
||||
self.assertEqual(((config / NEW).stat().st_mode & 0o777, (config / NEW / "approvers.json").stat().st_mode & 0o777), (0o700, 0o600))
|
||||
self.assertTrue((units / f"{NEW}-runner.service").is_file())
|
||||
self.assertTrue((units / f"{NEW}-sessions.timer").is_file())
|
||||
self.assertTrue((units / f"{NEW}-resources.timer").is_file())
|
||||
self.assertIn(f"KVB_ORCHESTRATOR_STATE={local / NEW}", (share / NEW / ".env").read_text(encoding="utf-8"))
|
||||
settings = self.home / ".claude" / "settings.json"
|
||||
data = json.loads(settings.read_text(encoding="utf-8"))
|
||||
@@ -131,7 +132,7 @@ class TestOrchestratorMigration(unittest.TestCase):
|
||||
self.assertIn(f"{share / NEW}/client/permission_hook.py", hooks[1]["hooks"][0]["command"])
|
||||
self.assertEqual(settings.stat().st_mode & 0o777, 0o640)
|
||||
self.assertIn(f"{OLD}/client", (self.home / ".claude" / f"settings.json.before-{NEW}-migration").read_text(encoding="utf-8"))
|
||||
self.assertEqual((first["images"], first["mcp"], first["enabled"]), ([f"{NEW}:local"], {NEW: "http://127.0.0.1:8737/mcp"}, [f"{NEW}-offices.path", f"{NEW}-runner.service", f"{NEW}-sessions.timer"]))
|
||||
self.assertEqual((first["images"], first["mcp"], first["enabled"]), ([f"{NEW}:local"], {NEW: "http://127.0.0.1:8737/mcp"}, [f"{NEW}-offices.path", f"{NEW}-resources.timer", f"{NEW}-runner.service", f"{NEW}-sessions.timer"]))
|
||||
self.assertIn(["docker", "compose", "--file", str(share / OLD / "compose.yml"), "down", "--remove-orphans"], first["calls"])
|
||||
|
||||
before = snapshot(self.home)
|
||||
|
||||
@@ -0,0 +1,294 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
from unittest import mock
|
||||
|
||||
import hub_support
|
||||
import resources as sampler
|
||||
|
||||
SCRIPT = hub_support.HUB / "client" / "resources.py"
|
||||
UNITS = hub_support.HUB / "systemd"
|
||||
NOW = 1_800_000_000
|
||||
DAY = 86400
|
||||
MIB = 1 << 20
|
||||
DF = """Filesystem 1024-blocks Used Available Capacity Mounted on
|
||||
/dev/nvme0n1p2 976762584 512000000 460000000 53% /
|
||||
devtmpfs 8000000 0 8000000 0% /dev
|
||||
tmpfs 8000000 1000 7999000 1% /dev/shm
|
||||
/dev/nvme0n1p2 976762584 512000000 460000000 53% /home
|
||||
/dev/nvme0n1p1 523248 12000 511248 3% /boot/efi
|
||||
/dev/loop3 65536 65536 0 100% /snap/core/1
|
||||
/dev/sda1 1953514584 100000000 1853514584 6% /mnt/Backup Disk
|
||||
overlay 976762584 512000000 460000000 53% /var/lib/docker/overlay2/x/merged
|
||||
"""
|
||||
DISKS = [
|
||||
{"mount": "/", "used": 512000000 * 1024, "total": 976762584 * 1024},
|
||||
{"mount": "/boot/efi", "used": 12000 * 1024, "total": 523248 * 1024},
|
||||
{"mount": "/mnt/Backup Disk", "used": 100000000 * 1024, "total": 1953514584 * 1024},
|
||||
]
|
||||
MEMINFO = """MemTotal: 32000000 kB
|
||||
MemFree: 2000000 kB
|
||||
MemAvailable: 24000000 kB
|
||||
Buffers: 500000 kB
|
||||
SwapTotal: 8000000 kB
|
||||
SwapFree: 6000000 kB
|
||||
HugePages_Total: 0
|
||||
"""
|
||||
STAT_BEFORE = "cpu 1000 0 500 8000 500 0 0 0 0 0\ncpu0 500 0 250 4000 250 0 0 0 0 0\n"
|
||||
STAT_AFTER = "cpu 1600 0 700 8400 600 0 0 0 0 0\ncpu0 800 0 350 4200 300 0 0 0 0 0\n"
|
||||
STAT_LATER = "cpu 2000 0 900 8600 700 0 0 0 0 0\n"
|
||||
LOADAVG = "1.25 0.80 0.50 2/1234 56789\n"
|
||||
NVIDIA_SMI = "00000000:01:00.0, 37, 2048, 8188\n00000000:03:00.0, [N/A], [N/A], 4096\n"
|
||||
AWAKE, ASLEEP = "0000:01:00.0", "0000:02:00.0"
|
||||
STUB = """#!/usr/bin/env python3
|
||||
import json, os, sys
|
||||
from pathlib import Path
|
||||
with Path(os.environ["STUB_LOG"]).open("a", encoding="utf-8") as handle:
|
||||
handle.write(json.dumps([Path(sys.argv[0]).name, *sys.argv[1:]]) + "\\n")
|
||||
sys.stdout.write(os.environ["STUB_" + Path(sys.argv[0]).name.replace("-", "_").upper()])
|
||||
"""
|
||||
|
||||
|
||||
def device(root: Path, address: str, vendor: str, kind: str, status: str | None) -> None:
|
||||
folder = root / "bus" / "pci" / "devices" / address
|
||||
(folder / "power").mkdir(parents=True)
|
||||
(folder / "vendor").write_text(f"{vendor}\n", encoding="utf-8")
|
||||
(folder / "class").write_text(f"{kind}\n", encoding="utf-8")
|
||||
if status is not None:
|
||||
(folder / "power" / "runtime_status").write_text(f"{status}\n", encoding="utf-8")
|
||||
|
||||
|
||||
def host_sys(root: Path) -> Path:
|
||||
device(root, AWAKE, "0x10de", "0x030000", "active")
|
||||
device(root, "0000:01:00.1", "0x10de", "0x040300", "active")
|
||||
device(root, ASLEEP, "0x10de", "0x030200", "suspended")
|
||||
device(root, "0000:00:02.0", "0x8086", "0x030000", "active")
|
||||
return root
|
||||
|
||||
|
||||
def row(t: int, cpu: float, used: int, waiting: int, gpu: dict) -> dict:
|
||||
return {"t": t, "disks": [{"mount": "/", "used": used, "total": 1000}], "cpu": cpu, "ram": cpu / 2, "swap": 0.0, "load": cpu / 10, "gpus": [gpu], "queue": {"waiting": waiting, "active": 1}}
|
||||
|
||||
|
||||
class TestParsing(unittest.TestCase):
|
||||
def test_df_keeps_each_local_block_device_once_with_its_first_mount_in_bytes(self):
|
||||
self.assertEqual(sampler.parse_df(DF), DISKS)
|
||||
self.assertEqual(sampler.parse_df(""), [])
|
||||
|
||||
def test_meminfo_gives_ram_and_swap_in_percent_and_no_swap_as_zero(self):
|
||||
self.assertEqual(sampler.memory(sampler.parse_meminfo(MEMINFO)), (25.0, 25.0))
|
||||
without = MEMINFO.replace("SwapTotal: 8000000", "SwapTotal: 0").replace("SwapFree: 6000000", "SwapFree: 0")
|
||||
self.assertEqual(sampler.memory(sampler.parse_meminfo(without)), (25.0, 0.0))
|
||||
self.assertEqual(sampler.memory({}), (None, None))
|
||||
|
||||
def test_cpu_percent_is_the_busy_share_of_the_ticks_between_two_reads(self):
|
||||
before, after = sampler.cpu_ticks(STAT_BEFORE), sampler.cpu_ticks(STAT_AFTER)
|
||||
self.assertEqual((before, after), ((8500, 10000), (9000, 11300)))
|
||||
self.assertEqual(sampler.cpu_percent(before, after), 61.5)
|
||||
self.assertIsNone(sampler.cpu_percent(after, after))
|
||||
self.assertIsNone(sampler.cpu_ticks("intr 1 2 3\n"))
|
||||
|
||||
def test_loadavg_is_the_one_minute_load(self):
|
||||
self.assertEqual(sampler.parse_loadavg(LOADAVG), 1.25)
|
||||
self.assertIsNone(sampler.parse_loadavg(""))
|
||||
|
||||
def test_nvidia_smi_csv_gives_util_and_vram_in_bytes_per_pci_address(self):
|
||||
self.assertEqual(sampler.parse_nvidia(NVIDIA_SMI), {
|
||||
AWAKE: {"util": 37.0, "vram_used": 2048 * MIB, "vram_total": 8188 * MIB},
|
||||
"0000:03:00.0": {"util": None, "vram_used": None, "vram_total": 4096 * MIB},
|
||||
})
|
||||
|
||||
def test_the_queue_counts_waiting_and_active_entries(self):
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
council = Path(tmp)
|
||||
self.assertEqual(sampler.queue(council), {"waiting": 0, "active": 0})
|
||||
entries = [{"status": "waiting"}, {"status": "waiting"}, {"status": "active"}, {"status": "done"}, "junk"]
|
||||
(council / "queue.json").write_text(json.dumps({"entries": entries}), encoding="utf-8")
|
||||
self.assertEqual(sampler.queue(council), {"waiting": 2, "active": 1})
|
||||
|
||||
|
||||
class TestCpu(unittest.TestCase):
|
||||
def test_cpu_is_the_busy_share_since_the_ticks_the_run_before_kept_and_the_sampler_never_sleeps(self):
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
proc, state = Path(tmp) / "proc", Path(tmp) / "state"
|
||||
proc.mkdir()
|
||||
state.mkdir()
|
||||
(proc / "stat").write_text(STAT_BEFORE, encoding="utf-8")
|
||||
with mock.patch.object(sampler.time, "sleep", side_effect=AssertionError("slept")), mock.patch.object(sampler.time, "time", return_value=NOW - 60):
|
||||
self.assertIsNone(sampler.sample(proc, Path(tmp) / "sys", Path(tmp), lambda argv: None, state)["cpu"])
|
||||
(proc / "stat").write_text(STAT_AFTER, encoding="utf-8")
|
||||
self.assertEqual(sampler.cpu_since(proc, state, NOW), 61.5)
|
||||
(proc / "stat").write_text(STAT_LATER, encoding="utf-8")
|
||||
self.assertIsNone(sampler.cpu_since(proc, state, NOW + sampler.BUCKET + 1))
|
||||
self.assertEqual(json.loads((state / "cpu.json").read_text(encoding="utf-8")), [NOW + sampler.BUCKET + 1, 9300, 12200])
|
||||
|
||||
|
||||
class TestGpus(unittest.TestCase):
|
||||
def setUp(self):
|
||||
tmp = tempfile.TemporaryDirectory()
|
||||
self.addCleanup(tmp.cleanup)
|
||||
self.root = Path(tmp.name)
|
||||
self.calls = []
|
||||
|
||||
def run_smi(self, argv: list[str]) -> str:
|
||||
self.calls.append(argv)
|
||||
return NVIDIA_SMI
|
||||
|
||||
def test_an_asleep_gpu_is_recorded_asleep_and_nvidia_smi_queries_only_the_awake_one(self):
|
||||
found = sampler.gpus(host_sys(self.root), self.run_smi)
|
||||
self.assertEqual(found, [{"name": AWAKE, "util": 37.0, "vram_used": 2048 * MIB, "vram_total": 8188 * MIB}, {"name": ASLEEP, "asleep": True}])
|
||||
self.assertEqual(len(self.calls), 1)
|
||||
self.assertEqual(self.calls[0][0], "nvidia-smi")
|
||||
self.assertIn(f"--id={AWAKE}", self.calls[0])
|
||||
self.assertNotIn(ASLEEP, " ".join(self.calls[0]))
|
||||
|
||||
def test_nvidia_smi_never_runs_while_every_nvidia_gpu_sleeps_or_lacks_a_runtime_status(self):
|
||||
device(self.root, ASLEEP, "0x10de", "0x030000", "suspended")
|
||||
device(self.root, "0000:04:00.0", "0x10de", "0x030000", None)
|
||||
self.assertEqual(sampler.gpus(self.root, self.run_smi), [{"name": ASLEEP, "asleep": True}, {"name": "0000:04:00.0", "asleep": True}])
|
||||
self.assertEqual(self.calls, [])
|
||||
|
||||
def test_an_awake_gpu_nvidia_smi_gives_no_reading_for_is_recorded_unknown(self):
|
||||
unknown = {"name": AWAKE, "util": None, "vram_used": None, "vram_total": None}
|
||||
self.assertEqual(sampler.gpus(host_sys(self.root), lambda argv: None), [unknown, {"name": ASLEEP, "asleep": True}])
|
||||
self.assertEqual(sampler.gpus(self.root, lambda argv: NVIDIA_SMI.splitlines()[1]), [unknown, {"name": ASLEEP, "asleep": True}])
|
||||
self.assertEqual(sampler.mean_row(NOW, [row(NOW + 60, 1.0, 1, 0, unknown), row(NOW + 120, 1.0, 1, 0, unknown)])["gpus"], [unknown])
|
||||
|
||||
def test_no_nvidia_gpu_means_no_gpu_row_and_no_call(self):
|
||||
device(self.root, "0000:00:02.0", "0x8086", "0x030000", "active")
|
||||
self.assertEqual(sampler.gpus(self.root, self.run_smi), [])
|
||||
self.assertEqual(sampler.gpus(self.root / "missing", self.run_smi), [])
|
||||
self.assertEqual(self.calls, [])
|
||||
|
||||
|
||||
class TestRetention(unittest.TestCase):
|
||||
def test_rows_older_than_seven_days_fold_into_fifteen_minute_means_and_means_past_ninety_days_drop(self):
|
||||
bucket = NOW - 8 * DAY
|
||||
asleep, awake = {"name": AWAKE, "asleep": True}, {"name": AWAKE, "util": 50.0, "vram_used": 300, "vram_total": 1000}
|
||||
mean = {**row(NOW - 30 * DAY, 12.0, 100, 0, asleep)}
|
||||
rows = [
|
||||
row(NOW - 91 * DAY, 5.0, 1, 0, asleep),
|
||||
mean,
|
||||
row(bucket, 10.0, 100, 0, asleep),
|
||||
row(bucket + 60, 20.0, 200, 1, asleep),
|
||||
{**row(bucket + 120, 60.0, 300, 2, awake), "disks": [{"mount": "/", "used": 300, "total": 1000}, {"mount": "/mnt", "used": 10, "total": 20}]},
|
||||
row(bucket + 1000, 40.0, 400, 0, awake),
|
||||
row(NOW - 60, 70.0, 700, 3, awake),
|
||||
]
|
||||
kept = sampler.retain(rows, NOW)
|
||||
self.assertEqual([entry["t"] for entry in kept], [NOW - 30 * DAY, bucket, bucket + 900, NOW - 60])
|
||||
self.assertIs(kept[0], mean)
|
||||
self.assertEqual(kept[1], {
|
||||
"t": bucket, "disks": [{"mount": "/", "used": 200, "total": 1000}, {"mount": "/mnt", "used": 10, "total": 20}], "cpu": 30.0, "ram": 15.0, "swap": 0.0, "load": 3.0,
|
||||
"gpus": [{"name": AWAKE, "util": 50.0, "vram_used": 300, "vram_total": 1000}], "queue": {"waiting": 1.0, "active": 1.0},
|
||||
})
|
||||
self.assertEqual((kept[2]["cpu"], kept[2]["disks"]), (40.0, [{"mount": "/", "used": 400, "total": 1000}]))
|
||||
self.assertEqual(kept[3], rows[-1])
|
||||
self.assertEqual(sampler.retain(kept, NOW), kept)
|
||||
|
||||
def test_a_bucket_still_reaching_into_the_last_seven_days_stays_raw_and_asleep_buckets_stay_asleep(self):
|
||||
now = NOW + 300
|
||||
asleep = {"name": ASLEEP, "asleep": True}
|
||||
rows = [row(NOW - 7 * DAY - 890, 10.0, 1, 0, asleep), row(NOW - 7 * DAY - 830, 30.0, 3, 0, asleep), row(NOW - 7 * DAY + 10, 50.0, 5, 0, asleep)]
|
||||
kept = sampler.retain(rows, now)
|
||||
self.assertEqual([entry["t"] for entry in kept], [NOW - 7 * DAY - 900, NOW - 7 * DAY + 10])
|
||||
self.assertEqual((kept[0]["cpu"], kept[0]["gpus"]), (20.0, [asleep]))
|
||||
self.assertEqual(kept[1], rows[2])
|
||||
|
||||
def test_record_appends_and_folds_only_on_the_first_minute_of_a_bucket(self):
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
state = Path(tmp)
|
||||
asleep = {"name": AWAKE, "asleep": True}
|
||||
old = [row(NOW - 8 * DAY, 10.0, 1, 0, asleep), row(NOW - 8 * DAY + 60, 30.0, 1, 0, asleep)]
|
||||
(state / "samples.jsonl").write_text("".join(json.dumps(entry) + "\n" for entry in old), encoding="utf-8")
|
||||
later = row(NOW + 400, 1.0, 1, 0, asleep)
|
||||
sampler.record(state, later, NOW + 400)
|
||||
lines = [json.loads(line) for line in (state / "samples.jsonl").read_text(encoding="utf-8").splitlines()]
|
||||
self.assertEqual(lines, [*old, later])
|
||||
first = row(NOW + 930, 2.0, 1, 0, asleep)
|
||||
sampler.record(state, first, NOW + 930)
|
||||
lines = [json.loads(line) for line in (state / "samples.jsonl").read_text(encoding="utf-8").splitlines()]
|
||||
self.assertEqual([entry["t"] for entry in lines], [NOW - 8 * DAY, NOW + 400, NOW + 930])
|
||||
self.assertEqual(lines[0]["cpu"], 20.0)
|
||||
|
||||
|
||||
class TestSamplerInScratchHome(unittest.TestCase):
|
||||
def setUp(self):
|
||||
tmp = tempfile.TemporaryDirectory()
|
||||
self.addCleanup(tmp.cleanup)
|
||||
base = Path(tmp.name)
|
||||
self.home, self.council, stubs, self.proc = base / "home", base / "council", base / "bin", base / "proc"
|
||||
for folder in (self.home, self.council, stubs, self.proc):
|
||||
folder.mkdir()
|
||||
for name, text in (("stat", STAT_BEFORE), ("meminfo", MEMINFO), ("loadavg", LOADAVG)):
|
||||
(self.proc / name).write_text(text, encoding="utf-8")
|
||||
for tool in ("df", "nvidia-smi"):
|
||||
(stubs / tool).write_text(STUB, encoding="utf-8")
|
||||
(stubs / tool).chmod(0o755)
|
||||
self.sys = host_sys(base / "sys")
|
||||
self.log = base / "calls.jsonl"
|
||||
(self.council / "queue.json").write_text(json.dumps({"entries": [{"status": "waiting"}, {"status": "active"}, {"status": "active"}]}), encoding="utf-8")
|
||||
self.env = {
|
||||
**os.environ, "HOME": str(self.home), "PATH": f"{stubs}:/usr/bin:/bin",
|
||||
"STUB_LOG": str(self.log), "STUB_DF": DF, "STUB_NVIDIA_SMI": NVIDIA_SMI.splitlines()[0] + "\n",
|
||||
}
|
||||
self.resources = self.home / ".local" / "state" / "agents-council" / "resources"
|
||||
|
||||
def sample(self) -> None:
|
||||
argv = [sys.executable, str(SCRIPT), "sample", "--state", str(self.resources), "--council", str(self.council), "--proc", str(self.proc), "--sys", str(self.sys)]
|
||||
done = subprocess.run(argv, env=self.env, capture_output=True, text=True, timeout=60, check=False)
|
||||
self.assertEqual(done.returncode, 0, done.stderr)
|
||||
|
||||
def test_the_sampler_appends_one_row_per_run_under_the_scratch_home(self):
|
||||
self.sample()
|
||||
(self.proc / "stat").write_text(STAT_AFTER, encoding="utf-8")
|
||||
self.sample()
|
||||
samples = self.resources / "samples.jsonl"
|
||||
rows = [json.loads(line) for line in samples.read_text(encoding="utf-8").splitlines()]
|
||||
self.assertEqual(len(rows), 2)
|
||||
for found in rows:
|
||||
with self.subTest(t=found["t"]):
|
||||
self.assertEqual(sorted(found), ["cpu", "disks", "gpus", "load", "queue", "ram", "swap", "t"])
|
||||
self.assertEqual(found["disks"], DISKS)
|
||||
self.assertEqual(found["gpus"], [{"name": AWAKE, "util": 37.0, "vram_used": 2048 * MIB, "vram_total": 8188 * MIB}, {"name": ASLEEP, "asleep": True}])
|
||||
self.assertEqual(found["queue"], {"waiting": 1, "active": 2})
|
||||
self.assertEqual((found["ram"], found["swap"], found["load"]), (25.0, 25.0, 1.25))
|
||||
self.assertEqual([found["cpu"] for found in rows], [None, 61.5])
|
||||
self.assertLessEqual(rows[0]["t"], rows[1]["t"])
|
||||
calls = [json.loads(line) for line in self.log.read_text(encoding="utf-8").splitlines()]
|
||||
self.assertEqual([call[0] for call in calls], ["df", "nvidia-smi"] * 2)
|
||||
self.assertTrue(all(ASLEEP not in " ".join(call) for call in calls), calls)
|
||||
self.assertEqual(sorted(path.name for path in samples.parent.iterdir()), ["cpu.json", "samples.jsonl"])
|
||||
|
||||
|
||||
class TestTimer(unittest.TestCase):
|
||||
def test_a_user_timer_runs_sample_and_tokens_every_minute_without_a_page_condition(self):
|
||||
timer = (UNITS / "kvb-orchestrator-resources.timer").read_text(encoding="utf-8").splitlines()
|
||||
service = (UNITS / "kvb-orchestrator-resources.service").read_text(encoding="utf-8").splitlines()
|
||||
self.assertIn("OnCalendar=*-*-* *:*:00", timer)
|
||||
self.assertIn("WantedBy=timers.target", timer)
|
||||
script = "/usr/bin/env python3 %h/.local/share/kvb-orchestrator/client/resources.py"
|
||||
shared = "--state ${KVB_ORCHESTRATOR_KEPT}/resources --council ${KVB_ORCHESTRATOR_COUNCIL}"
|
||||
self.assertEqual([line for line in service if line.startswith("ExecStart=")], [
|
||||
f"ExecStart={script} sample {shared} --proc /proc --sys /sys",
|
||||
f"ExecStart={script} tokens {shared} --transcripts ${{KVB_ORCHESTRATOR_PROJECTS}}",
|
||||
])
|
||||
self.assertIn("EnvironmentFile=%h/.local/share/kvb-orchestrator/.env", service)
|
||||
self.assertIn("Type=oneshot", service)
|
||||
self.assertFalse(any(line.startswith("ExecCondition=") for line in service))
|
||||
|
||||
def test_orchestrator_install_enables_the_resources_timer(self):
|
||||
script = (hub_support.REPO_ROOT / "scripts" / "orchestrator.sh").read_text(encoding="utf-8")
|
||||
self.assertIn('RESOURCES_UNIT="${NAME}-resources"', script)
|
||||
self.assertIn('systemctl --user enable --now "${RESOURCES_UNIT}.timer"', script)
|
||||
self.assertIn('"${HUB}/client/resources.py"', script)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,212 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import ast
|
||||
import datetime
|
||||
import json
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
from unittest import mock
|
||||
|
||||
import hub_support
|
||||
import resources as sampler
|
||||
|
||||
SCRIPT = hub_support.HUB / "client" / "resources.py"
|
||||
NOW = 1_800_000_000
|
||||
HOUR = 3600
|
||||
SESSION, OTHER = "5d88e9aa-0000-4000-8000-000000000001", "f45bf5bb-0000-4000-8000-000000000002"
|
||||
OPUS, HAIKU = "claude-opus-5-5", "claude-haiku-4-5"
|
||||
|
||||
|
||||
def iso(ts: float) -> str:
|
||||
return datetime.datetime.fromtimestamp(ts, datetime.UTC).isoformat().replace("+00:00", "Z")
|
||||
|
||||
|
||||
def assistant(ts: float, session: str, model: str, ident: str, counts: tuple[int, int, int, int], request: str = "") -> dict:
|
||||
usage = dict(zip(("input_tokens", "output_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"), counts, strict=True))
|
||||
message = {"id": ident, "type": "message", "role": "assistant", "model": model, "content": [{"type": "text", "text": "ok"}], "usage": {**usage, "service_tier": "standard"}}
|
||||
return {"type": "assistant", "sessionId": session, "timestamp": iso(ts), "message": message, **({"requestId": request} if request else {})}
|
||||
|
||||
|
||||
def lines(rows: list[dict]) -> str:
|
||||
return "".join(json.dumps(entry) + "\n" for entry in rows)
|
||||
|
||||
|
||||
class TokensCase(unittest.TestCase):
|
||||
def setUp(self):
|
||||
tmp = tempfile.TemporaryDirectory()
|
||||
self.addCleanup(tmp.cleanup)
|
||||
base = Path(tmp.name)
|
||||
self.projects, self.council, self.state = base / "projects", base / "council", base / "state"
|
||||
for folder in (self.projects / "-repo", self.council / "work" / "s1", self.state):
|
||||
folder.mkdir(parents=True)
|
||||
(self.council / "work" / "s1" / "claim.json").write_text(json.dumps({"office": "work/s1", "session": SESSION, "name": " Skills ", "beat": NOW}), encoding="utf-8")
|
||||
self.transcript = self.projects / "-repo" / f"{SESSION}.jsonl"
|
||||
|
||||
def count(self, now: float = NOW + 120, budget: int = sampler.BUDGET) -> list[dict]:
|
||||
return sampler.count_tokens(self.projects, self.council, self.state, now, budget)
|
||||
|
||||
def offsets(self) -> dict:
|
||||
return json.loads((self.state / "offsets.json").read_text(encoding="utf-8"))
|
||||
|
||||
def written(self) -> list[dict]:
|
||||
path = self.state / "tokens.jsonl"
|
||||
return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines()] if path.exists() else []
|
||||
|
||||
def totals(self) -> dict[tuple[int, str, str], tuple[int, ...]]:
|
||||
found: dict[tuple[int, str, str], tuple[int, ...]] = {}
|
||||
for entry in self.written():
|
||||
key = (entry["t"], entry["agent"], entry["model"])
|
||||
before = found.get(key, (0, 0, 0, 0))
|
||||
found[key] = tuple(a + b for a, b in zip(before, (entry["input"], entry["output"], entry["cache_read"], entry["cache_write"]), strict=True))
|
||||
return found
|
||||
|
||||
|
||||
class TestTokens(TokensCase):
|
||||
def test_usage_fields_add_up_per_hour_agent_and_model_and_a_message_split_over_lines_counts_once(self):
|
||||
hour = NOW
|
||||
rows = [
|
||||
{"type": "user", "sessionId": SESSION, "timestamp": iso(NOW + 5), "message": {"role": "user", "content": "x", "usage": {"input_tokens": 999}}},
|
||||
assistant(NOW + 10, SESSION, OPUS, "msg_1", (10, 20, 300, 40)),
|
||||
assistant(NOW + 10, SESSION, OPUS, "msg_1", (10, 20, 300, 40)),
|
||||
assistant(NOW + 20, SESSION, HAIKU, "msg_2", (1, 2, 3, 4)),
|
||||
assistant(NOW + 30, SESSION, OPUS, "msg_3", (5, 6, 7, 8)),
|
||||
assistant(NOW + 40, SESSION, "<synthetic>", "msg_4", (0, 0, 0, 0)),
|
||||
assistant(NOW - HOUR + 10, OTHER, OPUS, "msg_5", (100, 200, 0, 0)),
|
||||
]
|
||||
self.transcript.write_text(lines(rows) + "{not json\n", encoding="utf-8")
|
||||
added = self.count()
|
||||
self.assertEqual(sorted((entry["t"], entry["agent"], entry["model"]) for entry in added), sorted(self.totals()))
|
||||
self.assertEqual(self.totals(), {
|
||||
(hour, "Skills", OPUS): (15, 26, 307, 48),
|
||||
(hour, "Skills", HAIKU): (1, 2, 3, 4),
|
||||
(hour - HOUR, OTHER[:8], OPUS): (100, 200, 0, 0),
|
||||
})
|
||||
self.assertEqual(sorted(self.written()[0]), ["agent", "cache_read", "cache_write", "input", "model", "output", "t"])
|
||||
|
||||
def test_a_second_run_adds_only_the_bytes_after_the_last_offset(self):
|
||||
hour = NOW
|
||||
self.transcript.write_text(lines([assistant(NOW + 10, SESSION, OPUS, "msg_1", (10, 20, 30, 40))]), encoding="utf-8")
|
||||
self.count()
|
||||
self.assertEqual(self.offsets()["files"], {f"-repo/{SESSION}.jsonl": self.transcript.stat().st_size})
|
||||
self.assertEqual(self.count(), [])
|
||||
last = json.dumps(assistant(NOW + 50, SESSION, OPUS, "msg_3", (7, 7, 7, 7)))
|
||||
more = [assistant(NOW + 10, SESSION, OPUS, "msg_1", (10, 20, 30, 40)), assistant(NOW + 40, SESSION, OPUS, "msg_2", (1, 1, 1, 1))]
|
||||
with self.transcript.open("a", encoding="utf-8") as handle:
|
||||
handle.write(lines(more) + last[:40])
|
||||
self.count()
|
||||
self.assertEqual(self.totals(), {(hour, "Skills", OPUS): (11, 21, 31, 41)})
|
||||
with self.transcript.open("a", encoding="utf-8") as handle:
|
||||
handle.write(last[40:] + "\n")
|
||||
self.count()
|
||||
self.assertEqual(self.totals(), {(hour, "Skills", OPUS): (18, 28, 38, 48)})
|
||||
self.assertEqual(self.offsets()["files"][f"-repo/{SESSION}.jsonl"], self.transcript.stat().st_size)
|
||||
|
||||
def test_a_byte_budget_spreads_a_long_backlog_over_runs_without_losing_tokens(self):
|
||||
rows = [assistant(NOW + index, SESSION, OPUS, f"msg_{index}", (1, 0, 0, 0)) for index in range(40)]
|
||||
self.transcript.write_text(lines(rows), encoding="utf-8")
|
||||
line = len(lines(rows[:1]))
|
||||
self.count(budget=10 * line)
|
||||
self.assertEqual(sum(entry["input"] for entry in self.written()), 10)
|
||||
for _ in range(3):
|
||||
self.count(budget=10 * line)
|
||||
self.assertEqual(sum(entry["input"] for entry in self.written()), 40)
|
||||
|
||||
def test_a_shrunk_file_restarts_at_zero_a_vanished_one_leaves_the_offsets_and_an_old_one_starts_at_its_end(self):
|
||||
self.transcript.write_text(lines([assistant(NOW - 40, SESSION, OPUS, "msg_1", (5, 0, 0, 0)), assistant(NOW - 30, SESSION, OPUS, "msg_2", (5, 0, 0, 0))]), encoding="utf-8")
|
||||
old = self.projects / "-repo" / f"{OTHER}.jsonl"
|
||||
old.write_text(lines([assistant(NOW - 100 * 86400, OTHER, OPUS, "msg_9", (9, 0, 0, 0))]), encoding="utf-8")
|
||||
os.utime(old, (NOW - 100 * 86400, NOW - 100 * 86400))
|
||||
self.count()
|
||||
self.assertEqual(sum(entry["input"] for entry in self.written()), 10)
|
||||
self.transcript.write_text(lines([assistant(NOW - 20, SESSION, OPUS, "msg_7", (1, 0, 0, 0))]), encoding="utf-8")
|
||||
old.unlink()
|
||||
self.count()
|
||||
self.assertEqual(sum(entry["input"] for entry in self.written()), 11)
|
||||
self.assertEqual(sorted(self.offsets()["files"]), [f"-repo/{SESSION}.jsonl"])
|
||||
|
||||
def test_a_message_that_fork_transcripts_copy_counts_once_across_files_and_runs(self):
|
||||
shared = assistant(NOW + 10, SESSION, OPUS, "msg_1", (10, 20, 30, 40), "req_1")
|
||||
subagents = self.projects / "-repo" / SESSION / "subagents"
|
||||
subagents.mkdir(parents=True)
|
||||
self.transcript.write_text(lines([shared]), encoding="utf-8")
|
||||
(subagents / "agent-a.jsonl").write_text(lines([{**shared, "timestamp": iso(NOW + 40)}, assistant(NOW + 50, SESSION, OPUS, "msg_2", (1, 1, 1, 1), "req_2")]), encoding="utf-8")
|
||||
self.count()
|
||||
self.assertEqual(self.totals(), {(NOW, "Skills", OPUS): (11, 21, 31, 41)})
|
||||
(subagents / "agent-b.jsonl").write_text(lines([shared, assistant(NOW + 60, SESSION, OPUS, "msg_1", (5, 5, 5, 5), "req_9")]), encoding="utf-8")
|
||||
self.count(now=NOW + 180)
|
||||
self.assertEqual(self.totals(), {(NOW, "Skills", OPUS): (16, 26, 36, 46)})
|
||||
|
||||
def test_the_seen_messages_keep_the_hours_of_this_run_and_the_last_two_hours(self):
|
||||
self.transcript.write_text(lines([assistant(NOW - 5 * HOUR, SESSION, OPUS, "msg_1", (1, 0, 0, 0), "req_1"), assistant(NOW + 10, SESSION, OPUS, "msg_2", (1, 0, 0, 0), "req_2")]), encoding="utf-8")
|
||||
self.count()
|
||||
self.assertEqual({hour: len(keys) for hour, keys in self.offsets()["seen"].items()}, {str(NOW - 5 * HOUR): 1, str(NOW): 1})
|
||||
self.count(now=NOW + HOUR + 120)
|
||||
self.assertEqual(sorted(self.offsets()["seen"]), [str(NOW)])
|
||||
self.count(now=NOW + 2 * HOUR + 120)
|
||||
self.assertEqual(self.offsets()["seen"], {})
|
||||
|
||||
def test_a_transcript_that_fails_to_read_keeps_its_offset_and_counts_nothing_twice(self):
|
||||
old = NOW - 5 * HOUR
|
||||
self.transcript.write_text(lines([assistant(old + 10, SESSION, OPUS, "msg_1", (5, 0, 0, 0), "req_1")]), encoding="utf-8")
|
||||
self.count()
|
||||
end = self.transcript.stat().st_size
|
||||
with self.transcript.open("a", encoding="utf-8") as handle:
|
||||
handle.write(lines([assistant(old + 20, SESSION, OPUS, "msg_2", (1, 0, 0, 0), "req_2")]))
|
||||
with mock.patch.object(sampler, "scan", side_effect=PermissionError("denied")):
|
||||
self.assertEqual(self.count(now=NOW + HOUR + 120), [])
|
||||
self.assertEqual(self.offsets()["files"], {f"-repo/{SESSION}.jsonl": end})
|
||||
self.count(now=NOW + 2 * HOUR + 120)
|
||||
self.assertEqual(self.totals(), {(old, "Skills", OPUS): (6, 0, 0, 0)})
|
||||
|
||||
def test_a_run_without_new_bytes_leaves_offsets_json_unwritten(self):
|
||||
self.transcript.write_text(lines([assistant(NOW + 10, SESSION, OPUS, "msg_1", (1, 0, 0, 0), "req_1")]), encoding="utf-8")
|
||||
self.count()
|
||||
stored = self.offsets()
|
||||
with mock.patch.object(sampler, "replace", side_effect=AssertionError("offsets.json rewritten")):
|
||||
self.assertEqual(self.count(now=NOW + 180), [])
|
||||
self.assertEqual(self.offsets(), stored)
|
||||
|
||||
def test_the_first_minute_of_an_hour_merges_rows_and_drops_those_older_than_ninety_days(self):
|
||||
hour = NOW
|
||||
stale = {"t": hour - 91 * 86400, "agent": "Old", "model": OPUS, "input": 1, "output": 1, "cache_read": 1, "cache_write": 1}
|
||||
parts = [{"t": hour - HOUR, "agent": "Skills", "model": OPUS, "input": value, "output": 0, "cache_read": 0, "cache_write": 0} for value in (1, 2, 3)]
|
||||
(self.state / "tokens.jsonl").write_text(lines([stale, *parts]), encoding="utf-8")
|
||||
self.transcript.write_text(lines([assistant(NOW + 10, SESSION, OPUS, "msg_1", (4, 0, 0, 0))]), encoding="utf-8")
|
||||
self.count(now=hour + 1800)
|
||||
self.assertEqual(len(self.written()), 5)
|
||||
self.count(now=hour + 30)
|
||||
self.assertEqual(self.written(), [{**parts[0], "input": 6}, {"t": hour, "agent": "Skills", "model": OPUS, "input": 4, "output": 0, "cache_read": 0, "cache_write": 0}])
|
||||
|
||||
|
||||
class TestNoLanguageModel(unittest.TestCase):
|
||||
def test_the_sampler_and_the_parser_import_no_network_client(self):
|
||||
tree = ast.parse(SCRIPT.read_text(encoding="utf-8"))
|
||||
imported = {alias.name.split(".")[0] for node in ast.walk(tree) if isinstance(node, ast.Import) for alias in node.names}
|
||||
imported |= {node.module.split(".")[0] for node in ast.walk(tree) if isinstance(node, ast.ImportFrom) and node.module}
|
||||
self.assertEqual(imported & {"urllib", "http", "socket", "ssl", "requests", "hub_api", "anthropic", "openai"}, set())
|
||||
|
||||
|
||||
class TestTokensInScratchHome(unittest.TestCase):
|
||||
def test_the_tokens_command_reads_the_scratch_home_transcripts_and_writes_beside_the_samples(self):
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
home, council = Path(tmp) / "home", Path(tmp) / "council"
|
||||
folder = home / ".claude" / "projects" / "-repo"
|
||||
folder.mkdir(parents=True)
|
||||
council.mkdir()
|
||||
ts = datetime.datetime.now(datetime.UTC).timestamp() - 60
|
||||
(folder / f"{SESSION}.jsonl").write_text(lines([assistant(ts, SESSION, OPUS, "msg_1", (3, 4, 5, 6))]), encoding="utf-8")
|
||||
state = home / ".local" / "state" / "agents-council" / "resources"
|
||||
argv = [sys.executable, str(SCRIPT), "tokens", "--state", str(state), "--council", str(council), "--transcripts", str(home / ".claude" / "projects")]
|
||||
done = subprocess.run(argv, env={**os.environ, "HOME": str(home), "PATH": "/usr/bin:/bin"}, capture_output=True, text=True, timeout=60, check=False)
|
||||
self.assertEqual(done.returncode, 0, done.stderr)
|
||||
self.assertEqual(sorted(path.name for path in state.iterdir()), ["offsets.json", "tokens.jsonl"])
|
||||
written = [json.loads(line) for line in (state / "tokens.jsonl").read_text(encoding="utf-8").splitlines()]
|
||||
self.assertEqual(written, [{"t": int(ts - ts % HOUR), "agent": SESSION[:8], "model": OPUS, "input": 3, "output": 4, "cache_read": 5, "cache_write": 6}])
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in new issue
Block a user