"""wf res — shared memory/CPU ledger for agent jobs (files, /proc, systemd, printing). Pure logic: wflib/res.py. Spec: docs/resource-ledger.md. run --mem 10G [--cpus N] --for 40m --title T [--queue|--force] [--by NAME] [--commit REV] -- CMD… start a job or get `busy` (exit 3) status [ID] [--json] · wait ID · release ID [--stop] · note --mem 3G --for 20m [--force] [--by NAME] TITLE game on [--for 4h] | off · clean [--yes] · timer on|off · tick · shell-init """ from __future__ import annotations import re import argparse import contextlib import datetime as dt import fcntl import json import os import shlex import shutil import stat import subprocess import sys import time import tomllib from dataclasses import dataclass from pathlib import Path from typing import Callable HERE = Path(__file__).resolve().parent sys.path.insert(0, str(HERE)) from wflib import res # noqa: E402 LEDGER, LOCK = "resources.json", "resources.lock" HISTORY = "resources-history.jsonl" # finished runs pruned from the ledger: reservation sizes (wf res hist) ACTIVE = {"active", "activating", "deactivating", "reloading", "refreshing"} @dataclass class Env: """Everything wf res touches outside its own state; tests swap the callables.""" state: Path config: Path units: Path scratch: Path run: Callable[[list[str]], subprocess.CompletedProcess] meminfo: Callable[[], str] nproc: int now: Callable[[], dt.datetime] pid_alive: Callable[[int], bool] owner: Callable[[], int] root: Callable[[], Path] cwd: Callable[[], str] procs: Callable[[], list[tuple[int, int, str, str]]] # (pid, ppid, comm, cgroup) sleep: Callable[[float], None] tmp_used_gb: Callable[[], float] stdin: Callable[[], str] = lambda: sys.stdin.read() live_sessions: Callable[[], set[str]] = lambda: _live_sessions() tmp_dirs: tuple = (Path("/tmp"), Path("/var/tmp")) # swept for litter (res.litter_victims) cgread: Callable[[str, str], str] = lambda cg, name: _read(f"/sys/fs/cgroup{cg}/{name}") # ControlGroup, file environ: Callable[[], dict] = lambda: dict(os.environ) # ---------------------------------------------------------------- real machine def _read(path: str) -> str: try: return Path(path).read_text() except OSError: return "" def _live_sessions() -> set[str]: """Session ids of running Claude processes (Claude Code's ~/.claude/sessions/.json); unreadable → empty.""" out = set() for f in (Path.home() / ".claude" / "sessions").glob("*.json"): try: stat = Path(f"/proc/{f.stem}/stat").read_text() sid = res.live_session_id(f.read_text(), stat.rsplit(")", 1)[1].split()[19]) except (OSError, IndexError): continue if sid: out.add(sid) return out def _run(argv): return subprocess.run(argv, capture_output=True, text=True) def _xdg(var: str, default: str) -> Path: return Path(os.environ.get(var) or Path.home() / default) def _argv0(d: Path) -> str: try: return (d / "cmdline").read_bytes().split(b"\0", 1)[0].decode(errors="replace") except OSError: return "" def _comm(pid: int) -> str: """Process name; `claude` for any Claude Code process (res.proc_name).""" d = Path(f"/proc/{pid}") try: comm = (d / "comm").read_text().strip() except OSError: return "" return res.proc_name(comm, _argv0(d)) def _ppid(pid: int) -> int: return int(Path(f"/proc/{pid}/stat").read_text().rsplit(")", 1)[1].split()[1]) def _owner() -> int: """Nearest ancestor named `claude` (the session outlives its shells), else wf's parent.""" parent = p = os.getppid() while p > 1: if _comm(p) == "claude": return p try: p = _ppid(p) except (OSError, ValueError, IndexError): break return parent def _root() -> Path: r = _run(["git", "rev-parse", "--show-toplevel"]) return Path(r.stdout.strip()) if r.returncode == 0 and r.stdout.strip() else Path.cwd() def _procs() -> list[tuple[int, int, str, str]]: out = [] for d in Path("/proc").iterdir(): if not d.name.isdigit(): continue try: stat = (d / "stat").read_text() comm = stat[stat.index("(") + 1:stat.rindex(")")] ppid = int(stat.rsplit(")", 1)[1].split()[1]) cgroup = (d / "cgroup").read_text().strip().split(":", 2)[2] except (OSError, ValueError, IndexError): continue out.append((int(d.name), ppid, res.proc_name(comm, _argv0(d)), cgroup)) return out def _tmp_used_gb() -> float: st = os.statvfs("/tmp") return (st.f_blocks - st.f_bfree) * st.f_frsize / res.GB def real_env() -> Env: runtime = Path(os.environ.get("XDG_RUNTIME_DIR") or f"/run/user/{os.getuid()}") if not (runtime / "systemd").is_dir(): raise res.ResError("wf res needs a systemd user session") return Env(state=_xdg("XDG_STATE_HOME", ".local/state") / "wf", config=_xdg("XDG_CONFIG_HOME", ".config") / "wf" / "resources.toml", units=_xdg("XDG_CONFIG_HOME", ".config") / "systemd" / "user", scratch=Path(f"/tmp/claude-{os.getuid()}"), run=_run, meminfo=lambda: Path("/proc/meminfo").read_text(), nproc=os.cpu_count() or 1, now=lambda: dt.datetime.now().astimezone(), pid_alive=lambda p: Path(f"/proc/{p}").exists(), owner=_owner, root=_root, cwd=os.getcwd, procs=_procs, sleep=time.sleep, tmp_used_gb=_tmp_used_gb) # ---------------------------------------------------------------- ledger file def warn(msg: str) -> None: print(f"wf: {msg}", file=sys.stderr) def write_atomic(path: Path, text: str) -> None: tmp = path.with_name(path.name + ".tmp") tmp.write_text(text) os.replace(tmp, path) @contextlib.contextmanager def locked(env: Env): """Exclusive flock; yields the ledger; writes it back atomically when the block ends without error.""" env.state.mkdir(parents=True, exist_ok=True) with open(env.state / LOCK, "w") as lock: fcntl.flock(lock, fcntl.LOCK_EX) path = env.state / LEDGER text = path.read_text() if path.exists() else "" try: led = res.loads(text) except res.ResError as e: bad = path.with_name(f"{LEDGER}.bad-{env.now():%Y%m%d-%H%M%S}") path.rename(bad) warn(f"{e}; moved to {bad.name}, starting empty") led = res.Ledger(next=res.salvage_next(text)) yield led write_atomic(path, res.dumps(led)) # ---------------------------------------------------------------- systemd def systemctl(env: Env, args: list[str]) -> str: r = env.run(["systemctl", "--user", *args]) if r.returncode != 0: first = (r.stderr or r.stdout).strip().splitlines() raise res.ResError(f"systemctl {args[0]}: " + (first[0] if first else f"exit {r.returncode}")) return r.stdout def ensure_slice(env: Env) -> None: f = env.units / res.SLICE if not f.exists(): env.units.mkdir(parents=True, exist_ok=True) f.write_text(res.slice_unit_text()) systemctl(env, ["daemon-reload"]) def _result(env: Env, e: res.Entry) -> tuple: """(rc, peak_gb, why); no .rc (wrapper killed with the job) → why from the journal.""" def read(name): try: return (env.state / "logs" / name).read_text().strip() except OSError: return "" rc, peak = read(f"{e.id}.rc"), read(f"{e.id}.peak") rc, peak = int(rc) if rc.lstrip("-").isdigit() else None, int(peak) / res.GB if peak.isdigit() else None if rc is not None: return rc, peak, "" argv = ["journalctl", "--user", "-u", e.unit, "-o", "cat", "--no-pager"] if e.started: argv += ["--since", f"@{int(e.started.timestamp())}"] why, jpeak = res.journal_reason(env.run(argv).stdout or "", e.mem_gb) return rc, peak if peak is not None else jpeak, why or f"no exit code; see journalctl --user -u {e.unit}" def bare_units(env: Env) -> list: names = [l.split()[0] for l in systemctl(env, ["list-units", "--type=service", "--state=running", "--no-legend", "--plain"]).splitlines() if l.strip()] if not names: return [] return res.bare_units(res.show_units(systemctl(env, ["show", "-p", "Id,Transient,Slice,MemoryCurrent", *names]))) def session_scopes(env: Env) -> dict[str, dict[str, str]]: """Running scopes in agents.slice (Claude sessions) → show props.""" names = [l.split()[0] for l in systemctl(env, ["list-units", "--type=scope", "--state=running", "--no-legend", "--plain"]).splitlines() if l.strip()] if not names: return {} blocks = res.show_units(systemctl(env, ["show", "-p", "Id,Slice,MemoryHigh,MemoryCurrent,ControlGroup", *names])) return {n: b for n, b in blocks.items() if b.get("Slice") == res.SLICE} def cap_sessions(env: Env, cfg: res.Config) -> None: for unit, value in res.session_caps(session_scopes(env), cfg.session_mem_gb): systemctl(env, ["set-property", "--runtime", unit, f"MemoryHigh={value}"]) def gather(env: Env, led: res.Ledger) -> res.Facts: total, available = res.meminfo(env.meminfo()) names = sorted({e.unit for e in led.entries if e.state == "running"}) blocks = res.show_units(systemctl(env, ["show", "-p", "Id,ActiveState,MemoryCurrent,ControlGroup", res.SLICE, *names])) units = {} for n in names: b = blocks.get(n, {}) active, cg = b.get("ActiveState") in ACTIVE, b.get("ControlGroup", "") units[n] = res.Unit(active, res.gb_or_none(b.get("MemoryCurrent", "")), res.psi_some_avg60(env.cgread(cg, "memory.pressure")) if active and cg else None) scg = blocks.get(res.SLICE, {}).get("ControlGroup", "") return res.Facts( now=env.now(), total_gb=total, available_gb=available, nproc=env.nproc, units=units, live={e.owner for e in led.entries if e.state == "note" and env.pid_alive(e.owner)}, results={e.id: _result(env, e) for e in led.entries if e.state == "running" and not units[e.unit].active}, slice_gb=res.gb_or_none(blocks.get(res.SLICE, {}).get("MemoryCurrent", "")) or 0.0, slice_cache_gb=res.cache_gb(env.cgread(scg, "memory.stat")) if scg else 0.0) def start(env: Env, e: res.Entry, now: dt.datetime) -> str | None: """Launch e as a unit; on success e is running. Returns an error line or None.""" logs = env.state / "logs" logs.mkdir(parents=True, exist_ok=True) e.unit, e.log = f"wf-{e.id}.service", str(logs / f"{e.id}.log") for stale in (f"{e.id}.rc", f"{e.id}.peak"): # a reused id must not inherit an old exit code (logs / stale).unlink(missing_ok=True) ensure_slice(env) r = env.run(res.run_argv(e, str(logs))) if r.returncode != 0: first = (r.stderr or r.stdout).strip().splitlines() return "systemd-run: " + (first[0] if first else f"exit {r.returncode}") e.state, e.started = "running", now e.expires = now + dt.timedelta(minutes=2 * e.est_min) return None def housekeep(env: Env, cfg: res.Config, led: res.Ledger, facts: res.Facts) -> list[str]: """Every command, under the lock: prune, start queued work.""" done = [e for e in led.entries if e.state == "done"] lines = res.prune(led, facts) kept = {id(e) for e in led.entries} save_history(env, [r for r in (res.history_record(e) for e in done if id(e) not in kept) if r]) for e in res.to_start(cfg, led, facts): err = start(env, e, facts.now) if err: e.state, e.ended, e.why = "done", facts.now, f"start failed: {err}" lines.append(f"{e.id} {e.why}") else: lines.append(f"{e.id} started (queued)") if res.gaming(led, facts.now) or "game off (expired)" in lines: apply_slice(env, cfg, led, facts) if not led.last_clean or facts.now - led.last_clean >= res.CLEAN_EVERY: lines.extend(clean_auto(env, cfg, led, facts.now)) return lines def read_history(env: Env) -> list[dict]: path = env.state / HISTORY out = [] for line in (path.read_text().splitlines() if path.exists() else []): try: r = json.loads(line) except ValueError: continue if isinstance(r, dict): out.append(r) return out def save_history(env: Env, recs: list[dict]) -> None: """Append runs leaving the ledger; keep the newest res.HIST_KEEP (call under the lock).""" if not recs: return path = env.state / HISTORY old = path.read_text().splitlines() if path.exists() else [] lines = old + [json.dumps(r) for r in recs] if len(lines) > res.HIST_KEEP: write_atomic(path, "".join(x + "\n" for x in lines[-res.HIST_KEEP:])) else: with open(path, "a") as f: f.write("".join(json.dumps(r) + "\n" for r in recs)) def suggestion(env: Env, led: res.Ledger, title: str) -> tuple[float, int, int] | None: """History's (mem GB, minutes, n) for this project's kind of title (res.suggest), else None.""" return res.suggest(res.all_runs(read_history(env), led), main_tree(env).name, res.title_kind(title)) def hint(env: Env, led: res.Ledger, title: str, mem: float, est: int) -> None: line = res.hint_line(mem, est, suggestion(env, led, title)) if line: print(line, file=sys.stderr) def session(env: Env, cfg: res.Config, led: res.Ledger) -> res.Facts: facts = gather(env, led) housekeep(env, cfg, led, facts) return facts # ---------------------------------------------------------------- commands def _git(env: Env, *args) -> str: r = env.run(["git", "-C", env.cwd(), *args]) return r.stdout.strip() if r.returncode == 0 else "" def main_tree(env: Env) -> Path: """Project's main tree: git common dir's parent (same for its lane worktrees), else env.root().""" common = _git(env, "rev-parse", "--path-format=absolute", "--git-common-dir") return Path(common).parent if common and Path(common).name == ".git" else env.root() def _session_records(env: Env) -> list[dict]: """wf lane session records (.wf/sessions/*.json) of this project: main tree + root.""" out = [] for top in dict.fromkeys([main_tree(env), env.root()]): for f in sorted((top / ".wf" / "sessions").glob("*.json")): try: out.append(json.loads(f.read_text())) except (OSError, ValueError): pass return out def owner_by(env: Env, name: str = "") -> dict: return res.owner_by(env.environ(), _git(env, "branch", "--show-current"), _session_records(env), name) def _new_entry(env, led, facts, args, mem, est, state) -> res.Entry: return res.Entry(id=led.new_id(), project=main_tree(env).name, owner=env.owner(), title=args.title, mem_gb=mem, cpus=args.cpus, est_min=est, state=state, cwd=env.cwd(), queued=facts.now, by=owner_by(env, getattr(args, "by", "") or "")) def _new_probe(env, args) -> res.Entry: """Entry stand-in (no id) for capacity checks that depend on who starts it (res.loan).""" return res.Entry(id="", project="", owner=0, title=args.title, mem_gb=0.0, cpus=0, est_min=0, state="queued", by=owner_by(env, getattr(args, "by", "") or "")) def job_commit(env: Env, args, lock: str) -> str: """Full sha the job tests: --commit REV, else (locked job) a hex word of the title that names a commit; else ''.""" def sha(rev: str) -> str: out = _git(env, "rev-parse", "--verify", "--quiet", f"{rev}^{{commit}}") return out if re.fullmatch(r"[0-9a-f]{40}", out) else "" rev = getattr(args, "commit", "") or "" if rev: return sha(rev) or _raise(res.ResError(f"--commit {rev}: not a commit here")) m = res.COMMIT_TITLE_RE.search(args.title) if lock else None return sha(m.group()) if m else "" def _raise(err: Exception): raise err def enqueue(env, led, facts, args, mem, est, cmd, caller_env, lock, commit) -> tuple[res.Entry, str]: """Queue a job; same-lock queued entries coalesce by commit (res.coalesce): (entry waited for, note).""" e = _new_entry(env, led, facts, args, mem, est, "queued") e.cmd, e.env, e.lock, e.commit = cmd, caller_env, lock, commit anc = lambda a, b: env.run(["git", "-C", env.cwd(), "merge-base", "--is-ancestor", a, b]).returncode == 0 cov, gone = res.coalesce(led, e, anc) if cov: led.next -= 1 # never used res.cover(cov, commit) return cov, f"; covers {commit[:7]} (queued on descendant {cov.commit[:7]}, nothing new queued): wf res wait {cov.id}" led.entries.append(e) res.supersede(e, gone, facts.now) return e, "".join(f"; supersedes {g.id} ({g.commit[:7]})" for g in gone) def force_busy(cfg, led, facts, mem, cpus, room) -> str: """--force refused: not even the really free memory (beyond the user reserve) holds it.""" free = f"{res.fmt_gb(max(0.0, room[0]))}" if mem > room[0] + res.EPS else f"{max(0, room[1])} cpus" return f"busy even with --force: only {free} really free beyond the reserve; retry later or work on something else" def cmd_run(env, cfg, args) -> int: mem, est = res.parse_size(args.mem), res.parse_duration(args.for_) cmd = args.cmd[1:] if args.cmd[:1] == ["--"] else args.cmd if not cmd: raise res.ResError("no command after --") fail = busy = out = None shown = env.run(["systemctl", "--user", "show-environment"]) caller_env = res.env_diff(env.environ(), res.parse_show_environment(shown.stdout if shown.returncode == 0 else "")) lock = res.lock_for(args.title, getattr(args, "lock", ""), str(main_tree(env))) commit = job_commit(env, args, lock) with locked(env) as led: facts = session(env, cfg, led) hint(env, led, args.title, mem, est) fail = res.never_fits(cfg, facts, mem, args.cpus) holder = res.lock_holder(led, lock) if not fail and holder and not args.queue: busy = res.lock_busy_line(holder, facts) elif not fail and holder: e, note = enqueue(env, led, facts, args, mem, est, cmd, caller_env, lock, commit) when = res.queue_estimate(cfg, led, facts, e) held = res.lock_holder(led, lock) held = f" (lock held by {held.id})" if held and held is not e else "" out = (f"{e.id} queued, position {res.position(led, e)}{held}; est. start " + (f"~{res.hhmm(when)}" if when else "unknown") + f"; cancel: wf res release {e.id}{note}") elif not fail: q_gb, q_cpus = res.queued_total(led) room = res.force_room(cfg, led, facts) if args.force else res.budget(cfg, led, facts) if args.force: q_gb, q_cpus = 0.0, 0 probe = _new_probe(env, args) if not args.force: # a batch's own job borrows the batch's claim (res.loan) l_gb, l_cpus = res.loan(led, probe) room = (room[0] + l_gb, room[1] + l_cpus) if res.fits(mem + q_gb, args.cpus + q_cpus, room): e = _new_entry(env, led, facts, args, mem, est, "queued") e.cmd, e.env, e.lock, e.commit = cmd, caller_env, lock, commit fail = start(env, e, facts.now) if not fail: led.entries.append(e) out = f"{e.id} started; log {e.log}; ETA ~{res.hhmm(e.eta())} (estimate: not killed when over)" elif args.queue: e, note = enqueue(env, led, facts, args, mem, est, cmd, caller_env, lock, commit) when = res.queue_estimate(cfg, led, facts, e) out = (f"{e.id} queued, position {res.position(led, e)}; est. start " + (f"~{res.hhmm(when)}" if when else "unknown") + f"; cancel: wf res release {e.id}{note}") elif args.force: busy = force_busy(cfg, led, facts, mem, args.cpus, room) else: busy = res.busy_line(cfg, led, facts, mem + q_gb, args.cpus + q_cpus, res.force_hint(cfg, led, facts, mem, args.cpus)) if fail: raise res.ResError(fail) if busy: raise res.Busy(busy) print(out) return 0 def cmd_status(env, cfg, args) -> int: with locked(env) as led: facts = session(env, cfg, led) if args.json: text = res.status_json(cfg, led, facts) elif args.id: text = "\n".join(res.detail_lines(led.get(args.id))) else: outside = len(res.adopt_groups(env.procs())) scopes = session_scopes(env) stall = {n: res.psi_some_avg60(env.cgread(b.get("ControlGroup", ""), "memory.pressure")) or 0.0 for n, b in scopes.items() if b.get("ControlGroup")} text = "\n".join(res.status_lines(cfg, led, facts) + res.throttle_lines(led, facts) + res.session_cap_lines(scopes, stall, cfg.session_mem_gb) + res.warning_lines(facts.total_gb, env.tmp_used_gb(), outside, bare_units(env))) print(text) return 0 def cmd_hist(env, cfg, args) -> int: with locked(env) as led: runs = res.all_runs(read_history(env), led) print("\n".join(res.hist_lines(runs, args.project))) return 0 def cmd_release(env, cfg, args) -> int: fail = out = None with locked(env) as led: facts = session(env, cfg, led) e = led.get(args.id) if e.state == "running" and not args.stop: fail = f"{e.id} is running; --stop to kill it" elif e.state == "running": systemctl(env, ["stop", e.unit]) e.rc, e.peak_gb, _ = _result(env, e) e.state, e.ended, e.why = "done", facts.now, "released" out = f"{e.id} stopped" else: led.entries.remove(e) out = f"{e.id} released" if fail: raise res.ResError(fail) print(out) return 0 POLL_SECONDS = 15 def cmd_wait(env, cfg, args) -> int: limit = res.parse_duration(args.timeout) deadline = env.now() + dt.timedelta(minutes=limit) warned = False while True: with locked(env) as led: facts = session(env, cfg, led) e = led.get(args.id) if e.superseded_by: # coalesced into a gate on a descendant commit: wait for that one print(f"{e.id} superseded by {e.superseded_by} (same lock, descendant commit); waiting for it") args.id = e.superseded_by continue line, state = (res.done_line(e) if e.state == "done" else None), e.state hot = res.throttle_lines(res.Ledger(entries=[e]), facts) if hot and not warned: # once per wait warn(hot[0]) warned = True if line: print(line) return 0 if env.now() >= deadline: raise res.ResError(f"{args.id} not done after {res.fmt_dur(limit)} ({state})") env.sleep(POLL_SECONDS) def cmd_note(env, cfg, args) -> int: mem, est = res.parse_size(args.mem), res.parse_duration(args.for_) fail = busy = out = None shown = env.run(["systemctl", "--user", "show-environment"]) caller_env = res.env_diff(env.environ(), res.parse_show_environment(shown.stdout if shown.returncode == 0 else "")) with locked(env) as led: facts = session(env, cfg, led) hint(env, led, args.title, mem, est) fail = res.never_fits(cfg, facts, mem, args.cpus) q_gb, q_cpus = (0.0, 0) if args.force else res.queued_total(led) room = res.force_room(cfg, led, facts) if args.force else res.budget(cfg, led, facts) if not fail and res.fits(mem + q_gb, args.cpus + q_cpus, room): e = _new_entry(env, led, facts, args, mem, est, "note") e.started, e.expires = facts.now, facts.now + dt.timedelta(minutes=est) led.entries.append(e) out = f"{e.id} noted {res.fmt_gb(mem)} until ~{res.hhmm(e.expires)} (then freed; nothing is killed)" elif not fail and args.force: busy = force_busy(cfg, led, facts, mem, args.cpus, room) elif not fail: busy = res.busy_line(cfg, led, facts, mem + q_gb, args.cpus + q_cpus, res.force_hint(cfg, led, facts, mem, args.cpus)) if fail: raise res.ResError(fail) if busy: raise res.Busy(busy) print(out) return 0 def apply_slice(env: Env, cfg: res.Config, led: res.Ledger, facts: res.Facts) -> None: ensure_slice(env) props = res.slice_props(cfg, led, facts) systemctl(env, ["set-property", "--runtime", res.SLICE, *(f"{k}={v}" for k, v in props.items())]) def cmd_game(env, cfg, args) -> int: with locked(env) as led: facts = session(env, cfg, led) if args.state == "on": minutes = res.parse_duration(args.for_) if args.for_ else int(cfg.game_hours * 60) led.game_until = max(led.game_until or facts.now, facts.now + dt.timedelta(minutes=minutes)) lines = res.game_on_lines(cfg, led, facts, minutes) else: led.game_until = None lines = ["game off"] apply_slice(env, cfg, led, facts) print("\n".join(lines)) return 0 def _lstat(path) -> os.stat_result | None: """None when the path vanished meanwhile (scratch is shared with running sessions).""" try: return os.lstat(path) except OSError: return None def _newest(p: Path) -> float: st = _lstat(p) newest = st.st_mtime if st else 0.0 if p.is_dir() and not p.is_symlink(): for dirpath, dirnames, filenames in os.walk(p, followlinks=False): for name in dirnames + filenames: if st := _lstat(os.path.join(dirpath, name)): newest = max(newest, st.st_mtime) return newest def _size(p: Path) -> int: if p.is_symlink() or not p.is_dir(): st = _lstat(p) return st.st_size if st else 0 return sum(st.st_size for d, _, fs in os.walk(p, followlinks=False) for f in fs if (st := _lstat(os.path.join(d, f)))) def _remove(p: Path) -> None: if p.is_dir() and not p.is_symlink(): shutil.rmtree(p, ignore_errors=True) else: p.unlink(missing_ok=True) def scan_scratch(env: Env) -> dict: tree = {} if not env.scratch.is_dir(): return tree for top in env.scratch.iterdir(): kids = {} if top.is_dir() and not top.is_symlink(): kids = {str(c): _newest(c) for c in top.iterdir()} tree[str(top)] = (_newest(top), kids) return tree def scan_litter(env: Env) -> list[tuple[str, str, float]]: uid, out = os.getuid(), [] for d in env.tmp_dirs: try: names = os.listdir(d) except OSError: continue for n in names: p = Path(d) / n st = _lstat(p) if st is None or st.st_uid != uid: continue if stat.S_ISDIR(st.st_mode): try: kind = "dir" if os.listdir(p) else "emptydir" except OSError: continue else: kind = "file" if stat.S_ISREG(st.st_mode) and st.st_size else "other" out.append((str(p), kind, st.st_mtime)) return out def sweep_litter(env: Env, cfg: res.Config, now: dt.datetime) -> int: n = 0 for path in res.litter_victims(scan_litter(env), now.timestamp(), cfg.scratch_hours, env.pid_alive): try: (os.rmdir if os.path.isdir(path) and not os.path.islink(path) else os.unlink)(path) n += 1 except OSError: pass # refilled or gone meanwhile return n def clean_auto(env: Env, cfg: res.Config, led: res.Ledger, now: dt.datetime) -> list[str]: """Scratch untouched for scratch_hours (D4) and failed wf units; no confirmation needed.""" lines = [] for path in res.scratch_victims(scan_scratch(env), now.timestamp(), cfg.scratch_hours, env.live_sessions()): p = Path(path) size = _size(p) _remove(p) lines.append(f"freed {res.fmt_bytes(size)}: {p}") swept = sweep_litter(env, cfg, now) if swept: lines.append(f"swept {swept} temp-dir litter entries (empty dirs, dead .NET pipes)") env.run(["systemctl", "--user", "reset-failed", "wf-r-*"]) led.last_clean = now return lines def project_rules(env: Env) -> tuple[Path, list[tuple[str, int | None]]]: root = env.root() toml = root / "workflow.toml" if not toml.exists(): return root, [] try: patterns = tomllib.loads(toml.read_text()).get("cleanup", []) except tomllib.TOMLDecodeError as e: raise res.ResError(f"workflow.toml: {e}") from None return root, [res.cleanup_rule(p) for p in patterns] def project_matches(env: Env, now: dt.datetime) -> list[Path]: root, rules = project_rules(env) out = [] for pattern, days in rules: for p in sorted(root.glob(pattern)): if p.is_symlink() or not p.resolve().is_relative_to(root.resolve()): continue if days is not None and now.timestamp() - _newest(p) < days * 86400: continue out.append(p) return out def cmd_clean(env, cfg, args) -> int: with locked(env) as led: facts = gather(env, led) lines = [l for l in housekeep(env, cfg, led, facts) if l.startswith("freed ")] lines += clean_auto(env, cfg, led, facts.now) root = env.root() for p in project_matches(env, facts.now): size, rel = res.fmt_bytes(_size(p)), p.relative_to(root) if args.yes: _remove(p) lines.append(f"deleted {size}: {rel}") else: lines.append(f"would delete {size}: {rel} (wf res clean --yes)") print("\n".join(lines) if lines else "nothing to clean") return 0 def cmd_timer(env, cfg, args) -> int: service, timer = env.units / "wf-res.service", env.units / "wf-res.timer" if args.state == "on": env.units.mkdir(parents=True, exist_ok=True) service.write_text(res.service_text(sys.executable, str(HERE / "wf.py"))) timer.write_text(res.timer_text()) systemctl(env, ["daemon-reload"]) systemctl(env, ["enable", "--now", "wf-res.timer"]) print("timer on: wf-res.timer every 1 min") else: if timer.exists(): systemctl(env, ["disable", "--now", "wf-res.timer"]) service.unlink(missing_ok=True) timer.unlink(missing_ok=True) systemctl(env, ["daemon-reload"]) print("timer off") return 0 def adopt(env: Env) -> list[str]: """Move running Claude sessions (and their descendants) into agents.slice; one line per session.""" lines = [] groups = res.adopt_groups(env.procs()) if groups: ensure_slice(env) for root, pids in groups.items(): r = env.run(res.adopt_argv(root, pids)) if r.returncode == 0: lines.append(f"adopted claude {root} ({len(pids)} processes)") else: first = (r.stderr or r.stdout).strip().splitlines() lines.append(f"claude {root}: " + (first[0] if first else f"exit {r.returncode}")) return lines def cmd_adopt(env, cfg, args) -> int: print("\n".join(adopt(env)) or "all claude sessions already in agents.slice") return 0 def cmd_tick(env, cfg, args) -> int: with locked(env) as led: session(env, cfg, led) adopt(env) # silent; a session that vanished mid-move is retried next minute try: cap_sessions(env, cfg) except res.ResError: pass # a session that ended between list and set-property; retried next minute return 0 def cmd_hook(env, cfg, args) -> int: """PreToolUse hook; read-only (no lock, no systemctl) and silent on any error: it runs before every command.""" try: payload = json.loads(env.stdin()) led = res.loads((env.state / LEDGER).read_text()) out = res.hook_output(payload, led.game_until, env.now()) except (OSError, ValueError, AttributeError, res.ResError): return 0 if out: print(json.dumps(out)) return 0 def cmd_shell_init(env, cfg, args) -> int: print(res.shell_init_line()) return 0 # ---------------------------------------------------------------- main def parser() -> argparse.ArgumentParser: ap = argparse.ArgumentParser(prog="wf res", description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) sub = ap.add_subparsers(dest="command", metavar="COMMAND", required=True) def cmd(name, func, help): sp = sub.add_parser(name, help=help, description=help) sp.set_defaults(func=func) return sp sp = cmd("run", cmd_run, "reserve and start a job as a systemd user unit; busy → exit 3") sp.add_argument("--mem", required=True, metavar="SIZE", help="e.g. 10G, 512M") sp.add_argument("--cpus", type=int, default=1, choices=range(1, 1025), metavar="N") sp.add_argument("--for", dest="for_", required=True, metavar="DURATION", help="estimate, e.g. 40m, 2h (ETA and queue; the job is never killed at it)") sp.add_argument("--title", required=True) sp.add_argument("--queue", action="store_true", help="wait in the FIFO instead of exit 3") sp.add_argument("--force", action="store_true", help=FORCE_HELP) sp.add_argument("--by", default="", metavar="NAME", help="your session name (ListAgents), shown as the owner; default: WF_SESSION_NAME, else the wf lane session record") sp.add_argument("--lock", default="", metavar="KEY", help="one queued/running job per KEY in this project (main tree " "and its lane worktrees) at a time, e.g. a shared checkout; a title starting with 'gate' locks " "'gate' unasked; held → exit 3, --queue waits for it, --force does not override it") sp.add_argument("--commit", default="", metavar="REV", help="commit the job tests (default for a locked job: a " "hex word of the title, e.g. 'gate 1a2b3c4'); a queued job on a descendant commit replaces queued " "same-lock jobs on its ancestors (wf res wait follows), one on an ancestor is covered, not queued") sp.add_argument("cmd", nargs=argparse.REMAINDER, metavar="-- CMD …") sp = cmd("status", cmd_status, "reservations, budget, game mode, warnings") sp.add_argument("id", nargs="?") sp.add_argument("--json", action="store_true") sp = cmd("wait", cmd_wait, "poll every 15 s until the entry is done; prints rc and peak") sp.add_argument("id") sp.add_argument("--timeout", default="2h", metavar="DURATION") sp = cmd("hist", cmd_hist, "reservation sizes from history, per project and title kind (title minus trailing " "sha/hex words and (…)): runs n, request, peak p50/p95, estimate, duration p50/p90, suggestion " "(≥ 3 runs: mem = p95 peak ×1.15 ≥ 0.2 GB, for = p90 duration ×1.5 ≥ 5 min); run/note print a hint " "when asking > 2× it") sp.add_argument("--project", metavar="P", help="only this project (dir name of its main tree)") sp = cmd("note", cmd_note, "reserve for foreground work; freed when this Claude session exits or at --for") sp.add_argument("--mem", required=True, metavar="SIZE") sp.add_argument("--cpus", type=int, default=1, choices=range(1, 1025), metavar="N") sp.add_argument("--for", dest="for_", required=True, metavar="DURATION", help="reservation freed after this (nothing is killed)") sp.add_argument("--force", action="store_true", help=FORCE_HELP) sp.add_argument("--by", default="", metavar="NAME", help="your session name (ListAgents), shown as the owner; default: WF_SESSION_NAME, else the wf lane session record") sp.add_argument("title") sp = cmd("game", cmd_game, "gaming reserve on (auto-off, default 4h) or off; jobs slow, never stop") sp.add_argument("state", choices=["on", "off"]) sp.add_argument("--for", dest="for_", metavar="DURATION") sp = cmd("clean", cmd_clean, "delete stale /tmp/claude-* scratch (auto) and list project cleanup patterns") sp.add_argument("--yes", action="store_true", help="also delete the project's workflow.toml cleanup matches") cmd("adopt", cmd_adopt, "move running Claude sessions into agents.slice (no restart); the timer does this every minute") sp = cmd("timer", cmd_timer, "install/remove the 1-minute wf-res.timer (queue, game expiry, clean)") sp.add_argument("state", choices=["on", "off"]) cmd("tick", cmd_tick, "housekeeping once (run by the timer); silent") cmd("hook", cmd_hook, "Claude Code PreToolUse hook (Bash): no display for agent commands while game is on") cmd("shell-init", cmd_shell_init, "print the alias that starts Claude inside agents.slice (add to ~/.bashrc)") sp = cmd("release", cmd_release, "drop an entry (cancels a queued job, frees a note); --stop kills a running job") sp.add_argument("id") sp.add_argument("--stop", action="store_true") return ap # ---------------------------------------------------------------- wf batch FORCE_HELP = ("memory is really free though the ledger says busy (unused reservations, stale notes): fit against " "MemAvailable minus the user reserve and headroom only, ignoring claims and the queue; still recorded; " "the busy line names --force when it would fit") CEILING_VAR = "CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS" BATCH_TITLE = "wf-batch" BATCH_FOR = "6h" # default --for and ceiling (no history) PREP_MEM = "1G" STOP_FILE = "out/wf-batch.stop" def cmd_batch(env, cfg, args) -> int: if args.stop: f = env.root() / STOP_FILE f.parent.mkdir(exist_ok=True) f.touch() print(f"stop requested: {STOP_FILE} (a batch or loop stops before its next spawn; running workers finish)") return 0 if args.status: return batch_status(env, cfg) if args.time_left: return batch_time_left(env, args.time_left) if args.n is None: batch_parser().print_usage(sys.stderr) print("wf batch: error: N is required (or --status)", file=sys.stderr) return 2 if args.left: if args.prep: raise res.ResError("--left is for work batches, not --prep (use --for)") left = res.parse_duration(args.left) p90, runs = _task_p90(env) k = res.batch_fit(args.n, left, p90) line = f"fit: {k} of {args.n} tasks (left {res.fmt_dur(left)}, task p90 {_p90_text(p90, runs)})" if not k: print(f"{line}; nothing started") return 0 print(line) args.n = k args.for_ = args.for_ or args.left sug = None if not args.prep and (args.mem is None or args.for_ is None): # prep jobs share the title, not the size with locked(env) as led: sug = suggestion(env, led, BATCH_TITLE) minutes = res.parse_duration(args.for_ or BATCH_FOR) # the ceiling: never shortened by history if args.for_ is None: args.for_ = f"{sug[1]}m" if sug else BATCH_FOR if args.mem is None: # prep: one claude -p, subagents in-process, no builds → small args.mem = PREP_MEM if args.prep else f"{sug[0]:g}G" if sug else f"{min(8.0, max(4.0, 1.5 * args.n)):g}G" claude = shutil.which("claude", path=env.environ().get("PATH")) if not claude: raise res.ResError("claude not found on PATH") out = f"out/wf-batch-{env.now().strftime('%Y-%m-%d-%H%M')}.md" names = [x.strip() for x in (args.lanes or "").split(",") if x.strip()] stamp = env.now().strftime('%Y-%m-%d %H:%M') if args.prep: ids, tasks_rel = prep_ids(env.root(), set(names) if names else None, args.n) if not ids: print(f"prep: no pending task without Done (lanes: {', '.join(names) or 'all'}); nothing started") return 0 prompt = (HERE / "templates" / "prep-prompt.md").read_text().strip().format( here=HERE, root=env.root(), ids=" ".join(ids), out=out, tasks=tasks_rel) header = f"prep K={args.n}: {' '.join(ids)}" else: lanes = ", ".join(names) if names else \ "every lane with ready tasks, see wf lanes" prompt = (HERE / "templates" / "batch-prompt.md").read_text().strip().format( here=HERE, n=args.n, lanes=lanes, out=out, stop=env.root() / STOP_FILE, deadline=(env.now() + dt.timedelta(minutes=minutes)).isoformat(timespec="minutes")) header = f"N={args.n}, lanes: {lanes}" summary = env.root() / out if not args.dry_run: # dry-run touches nothing; an existing summary (same minute) is appended to, never clobbered summary.parent.mkdir(exist_ok=True) with summary.open("a") as f: f.write(f"# wf-batch {env.root().name} {stamp} ({header})\n") cmd = [claude, "-p", "--model", args.model, "--permission-mode", "auto", "--permission-prompts", "none", prompt] if not args.prep and _cloud_project(env.root()): cmd = [sys.executable, str(HERE / "wf_res.py"), SIDECAR, "--root", str(env.root()), "--summary", out, "--every", str(SIDECAR_EVERY), "--", *cmd] ceiling = str(minutes * 60_000) if args.dry_run: print(f"{CEILING_VAR}={ceiling} wf res run --mem {args.mem} --for {args.for_} --title {BATCH_TITLE} -- " + " ".join(shlex.quote(a) for a in cmd)) return 0 stop = env.root() / STOP_FILE if stop.exists(): stop.unlink() print(f"warning: stale {STOP_FILE} cleared; starting batch", file=sys.stderr) base = env.environ env.environ = lambda: {**base(), CEILING_VAR: ceiling} try: code = cmd_run(env, cfg, argparse.Namespace(mem=args.mem, cpus=1, for_=args.for_, title=BATCH_TITLE, queue=False, force=args.force, cmd=cmd)) finally: env.environ = base if stop.exists(): stop.unlink() print(f"summary: {out} · status: wf batch --status") return code SIDECAR = "batch-sidecar" SIDECAR_EVERY = 600 # s between the sidecar's wf cloud pull --all runs PULLED = re.compile(r"^(\S+): (\w+) \(") TAIL_CAP = 24 * 3600 # s: the tail after the orchestrator exits pulls at most this long TAIL_MEM_GB, TAIL_CPUS = 0.2, 1 # the job's reservation during the tail def _cloud_project(root: Path) -> bool: from wflib import config try: return config.load(config.find_root(root)).cloud except (config.ConfigError, OSError): return False def sidecar(child: list[str], root: Path, summary: Path, every: float, wf_cmd: list[str], cap: float = TAIL_CAP, shrink: Callable[[], str] | None = None) -> int: """Run the batch orchestrator `child`; while it lives, every `every` s (until the stop file appears): wf cloud pull --all, each ended ': (…' line → wf orch post cloud --result [--commit ] --no-pick + one summary line. After it exits with .wf/cloud records left: the tail — shrink() the job's reservation, keep pulling every `every` s (no picks, no sends) until no record is left, the stop file appears or `cap` s passed; one summary line says why it ended. Returns the child's exit code.""" proc = subprocess.Popen(child) while True: try: rc = proc.wait(timeout=every) break except subprocess.TimeoutExpired: pass if (root / STOP_FILE).exists(): print("sidecar: stop file, no more pulls", flush=True) return proc.wait() _pull_post(root, summary, wf_cmd) if not _cloud_records(root): return rc print(f"sidecar: orchestrator exited ({rc}), cloud records left: {' '.join(_cloud_records(root))}; " f"pulling every {every:g}s (no picks) until none, stop file or {cap / 3600:g}h", flush=True) if shrink: print(shrink(), flush=True) t0, pulls = time.monotonic(), 0 while True: time.sleep(every) if (root / STOP_FILE).exists(): why = "stop file" elif time.monotonic() - t0 >= cap: why = f"{cap / 3600:g}h cap" else: _pull_post(root, summary, wf_cmd) pulls += 1 if _cloud_records(root): continue why = "no cloud records left" break left = " ".join(_cloud_records(root)) or "none" line = f"cloud tail ended: {why}; {pulls} pulls; left: {left}" print(f"sidecar: {line}", flush=True) with summary.open("a") as f: f.write(f"{time.strftime('%H:%M')} {line} (sidecar)\n") return rc def _cloud_records(root: Path) -> list[str]: return sorted(p.stem for p in (root / ".wf" / "cloud").glob("*.json")) def _pull_post(root: Path, summary: Path, wf_cmd: list[str]) -> None: pull = subprocess.run([*wf_cmd, "cloud", "pull", "--all", "--project", str(root)], cwd=root, capture_output=True, text=True) print(pull.stdout + pull.stderr, end="", flush=True) for line in pull.stdout.splitlines(): m = PULLED.match(line) if not m or m[2] == "running" or line.startswith("archive "): continue sha = re.search(r"merged ([0-9a-f]{7,40})\b", line) post = [*wf_cmd, "orch", "post", m[1], "cloud", "--result", m[2], "--no-pick"] post += ["--commit", sha[1]] if sha else [] r = subprocess.run(post, cwd=root, capture_output=True, text=True) print(r.stdout + r.stderr, end="", flush=True) with summary.open("a") as f: f.write(f"{time.strftime('%H:%M')} cloud {m[1]} {m[2]} {sha[1] if sha else '-'} (sidecar pull)\n") def shrink_reservation(env: Env, res_id: str, mem_gb: float = TAIL_MEM_GB, cpus: int = TAIL_CPUS) -> str: """The batch job's ledger entry down to the tail's size (ledger only: the unit's MemoryMax stays, a pull must never be OOM-killed).""" with locked(env) as led: e = next((x for x in led.entries if x.id == res_id), None) if not e or e.state != "running": return f"sidecar: {res_id} not running in the ledger; reservation kept" old = e.mem_gb e.mem_gb, e.cpus = min(e.mem_gb, mem_gb), min(e.cpus, cpus) return (f"sidecar: {res_id} reservation {res.fmt_gb(old)} -> {res.fmt_gb(e.mem_gb)}, {e.cpus} cpu " "for the cloud tail") def sidecar_main(argv: list[str]) -> int: ap = argparse.ArgumentParser(prog=f"wf_res.py {SIDECAR}", description="wf batch job wrapper of cloud projects " "(started by wf batch, not by hand): runs the orchestrator and pulls cloud tasks") ap.add_argument("--root", type=Path, required=True) ap.add_argument("--summary", required=True, help="summary file, relative to --root") ap.add_argument("--every", type=float, default=SIDECAR_EVERY) ap.add_argument("--wf", default=f"{sys.executable} {HERE / 'wf.py'}", help="wf command (tests: a fake)") ap.add_argument("--cap", type=float, default=TAIL_CAP, help="s the tail (cloud records left after the " "orchestrator exits) pulls at most") ap.add_argument("cmd", nargs=argparse.REMAINDER) a = ap.parse_args(argv) cmd = a.cmd[1:] if a.cmd[:1] == ["--"] else a.cmd if not cmd: ap.error("no command after --") rid = os.environ.get(res.RES_ID_VAR, "") def shrink() -> str: try: return shrink_reservation(real_env(), rid) except (res.ResError, OSError) as e: return f"sidecar: reservation of {rid} kept: {e}" return sidecar(cmd, a.root, a.root / a.summary, a.every, shlex.split(a.wf), a.cap, shrink if rid else None) def prep_ids(root: Path, only: set[str] | None, k: int) -> tuple[list[str], str]: """Up to k ids a prep batch works (wflib.lanes.prep_targets) + the task file path relative to root.""" from wflib import config, lanes, tasks try: cfg = config.load(config.find_root(root)) doc = tasks.parse(cfg.tasks.read_text(encoding="utf-8")) archived = tasks.archive_ids(cfg.archive.read_text(encoding="utf-8")) except (config.ConfigError, OSError) as e: raise res.ResError(f"batch --prep: {e}") from None items = lanes.prep_targets(doc, archived, cfg.lanes, cfg.slice_above, only) return [i.id for i in items[:k]], cfg.rel(cfg.tasks) def _orch_progress(root: Path, stem: str) -> list[str]: """out/wf-orch.log lines (start 'YYYY-MM-DD HH:MM' or 'T') at/after the batch start (from the summary name).""" m = re.fullmatch(r"wf-batch-(\d{4}-\d\d-\d\d)-(\d\d)(\d\d)", stem) log = root / "out" / "wf-orch.log" if not m or not log.is_file(): return [] start = f"{m[1]} {m[2]}:{m[3]}" done = [ln for ln in log.read_text().splitlines() if re.match(r"\d{4}-\d\d-\d\d[ T]\d\d:\d\d", ln) and ln[:16].replace("T", " ") >= start] return ["finished since batch start (out/wf-orch.log):"] + done[-30:] if done else [] def _task_p90(env) -> tuple[int, int]: log = env.root() / "out" / "wf-orch.log" return res.task_p90(log.read_text(errors="replace") if log.exists() else "") def _p90_text(p90: int, runs: int) -> str: return f"{res.fmt_dur(p90)} from {runs} runs" if runs >= res.TASK_P90_MIN_RUNS else \ f"{res.fmt_dur(p90)} default ({runs} runs)" def batch_time_left(env, deadline: str) -> int: """Per-round check of a running batch: spawn while the time to the deadline is ≥ the task p90.""" try: end = dt.datetime.fromisoformat(deadline) except ValueError: raise res.ResError(f"--time-left '{deadline}' (want an ISO time, e.g. 2026-10-01T20:00+02:00)") now = env.now() if end.tzinfo is None: end = end.replace(tzinfo=now.tzinfo) left = max(0, int((end - now).total_seconds() // 60)) p90, runs = _task_p90(env) word = "spawn" if left >= p90 else "stop" print(f"time left {res.fmt_dur(left)}, task p90 {res.fmt_dur(p90)} ({runs} runs): {word}") return 0 def batch_status(env, cfg) -> int: root = env.root() with locked(env) as led: facts = session(env, cfg, led) mine = [e for e in led.entries if e.title == BATCH_TITLE and e.project == main_tree(env).name] lines = [res.entry_line(mine[-1], led, facts), f" log {mine[-1].log}"] if mine else [] summaries = sorted((root / "out").glob("wf-batch-*.md")) if summaries: rel = summaries[-1].relative_to(root) lines += [f"{rel}:"] + summaries[-1].read_text().rstrip("\n").splitlines()[-30:] lines += _orch_progress(root, summaries[-1].stem) if (root / STOP_FILE).exists(): lines.append(f"stop requested: {STOP_FILE}") print("\n".join(lines) or "no batch in this project") return 0 def _count(text: str) -> int: if not text.isdigit() or int(text) < 1: raise argparse.ArgumentTypeError("N must be ≥ 1") return int(text) def batch_parser() -> argparse.ArgumentParser: ap = argparse.ArgumentParser( prog="wf batch", description="Start an unattended batch orchestrator: claude -p (auto mode, no prompts) " "as a wf res job, prompt templates/batch-prompt.md; it works up to N tasks and writes out/wf-batch-*.md. " "Owner allow rule (once): Bash(python3 /projects/public/workflow/wf.py batch:*).") ap.add_argument("n", nargs="?", type=_count, metavar="N", help="at most N tasks") ap.add_argument("--lanes", metavar="L,…", help="lanes to work (default: every lane with ready tasks)") ap.add_argument("--model", default="opus", help="the batch orchestrator's model (default opus)") ap.add_argument("--for", dest="for_", default=None, metavar="DURATION", help=f"estimate and {CEILING_VAR} (default: estimate from wf res hist for wf-batch in this " f"project when ≥ 3 runs (not --prep), else {BATCH_FOR}; the ceiling stays {BATCH_FOR} unless given)") ap.add_argument("--mem", default=None, metavar="SIZE", help="reservation (default: wf res hist suggestion for wf-batch in this project when ≥ 3 runs, " "else max(4G, 1.5G x N), cap 8G; --prep: 1G; explicit wins; a request > 2x history prints a hint)") ap.add_argument("--dry-run", action="store_true", help="print the command, start nothing") ap.add_argument("--force", action="store_true", help=FORCE_HELP) ap.add_argument("--prep", action="store_true", help="prep batch: for up to N pending tasks without Done (pickable first, then priority) sonnet " "workers write Done (+Model) from the task text, unclear ones → awaiting; prompt " "templates/prep-prompt.md; summary lines 'id → Done: …' | 'id → awaiting a-…'. " "None to prep → prints so, starts nothing") ap.add_argument("--left", metavar="DURATION", help="time left to the pilot's deadline: N becomes min(N, floor(left / task p90)) (p90 of done/" f"handback durations in out/wf-orch.log, < {res.TASK_P90_MIN_RUNS} runs → {res.TASK_P90_DEFAULT}m), " "prints 'fit: K of N tasks (…)'; K = 0 → nothing started (rc 0); --for defaults to DURATION") ap.add_argument("--time-left", metavar="DEADLINE", help="(run by the batch orchestrator before each round) print 'time left X, task p90 Y (n runs): " "spawn|stop' for the ISO DEADLINE its prompt names (= start + ceiling); stop = left < p90") ap.add_argument("--status", action="store_true", help="newest batch job of this project + its summary file + tasks finished since start (out/wf-orch.log)") ap.add_argument("--stop", action="store_true", help=f"graceful stop: create {STOP_FILE}; batch and loop check it before every spawn (latency ≤ the running " "tasks), stop with 'stopped: stop file' and removes it") ap.set_defaults(func=cmd_batch) return ap def batch_main(argv: list[str], env: Env | None = None) -> int: return main(argv, env, batch_parser()) PREP_ALL = 1000 # "wf prep all": more than any task list def prep_main(argv: list[str], env: Env | None = None) -> int: """wf prep [N|all] [batch opts] = wf batch --prep N (all = every not-ready task).""" if argv and argv[0] == "all": argv = [str(PREP_ALL), *argv[1:]] return batch_main([*argv, "--prep"], env) def load_cfg(env: Env) -> res.Config: return res.load_config(env.config.read_text() if env.config.exists() else "") def main(argv: list[str], env: Env | None = None, ap: argparse.ArgumentParser | None = None) -> int: try: args = (ap or parser()).parse_args(argv) except SystemExit as e: # usage error → 2, -h → 0; callers (and tests) get a code, not an exception return e.code if isinstance(e.code, int) else 2 try: env = env or real_env() return args.func(env, load_cfg(env), args) except res.Busy as e: print(e) return 3 except res.ResError as e: warn(str(e)) return 1 except BrokenPipeError: return 0 except OSError as e: warn(str(e)) return 1 if __name__ == "__main__": sys.exit(sidecar_main(sys.argv[2:]) if sys.argv[1:2] == [SIDECAR] else main(sys.argv[1:]))