@hivey/cli 0.9.40 → 0.9.43

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -6,19 +6,29 @@ Tasks live in <root>/.swarm/tasks.json (locked on every write) and are rendered
6
6
  swarm's timeline (addons such as the watcher may log there too). hivey's dashboard reads
7
7
  tasks.json, so keep its format.
8
8
 
9
- Statuses: open → in-progress → review (deliverable DONE, waiting on critic) → approved
9
+ Statuses: open → in-progress → review (deliverable submitted, fresh review running) → approved
10
10
  (+ blocked at any point)
11
11
  Rules: a task can be claimed only by its owner (or anyone if unowned) and only once all
12
12
  its deps are approved. If a task has a `check` command, `approved` requires it to pass.
13
- Only the critic or the coordinator can set a task to `approved`. Checks run outside the
14
- board lock, so a long test suite doesn't stall other agents.
13
+ Checks run outside the board lock, so a long test suite doesn't stall other agents.
14
+
15
+ Review: `submit` starts a fresh review in the background (the hivey skill's
16
+ fresh_review.py): N one-shot critics (config `<root>/.swarm/review.json`, default 2 × sonnet),
17
+ each in an empty context with no saved session, read the task spec, `<root>/.swarm/REQUEST.md`
18
+ (the user's request) and the artifact, run the check and exit. All PASS and the check exits 0
19
+ → the task is `approved` (by `fresh-review`); otherwise it goes back to `in-progress` with the
20
+ merged review, and the owner gets a message either way. After `rounds` failed rounds (default
21
+ 3) it's `blocked` and the coordinator is told. Besides that, only the coordinator approves.
22
+ With `critics: 0` there's no fresh review: `submit` sets `review` for the coordinator.
15
23
 
16
24
  Usage (root is found by walking up from cwd to the folder containing .swarm/):
17
- swarm_tasks.py add ID "title" [--owner A] [--deps T1,T2] [--path P] [--check "cmd"] [--as coordinator]
25
+ swarm_tasks.py add ID "title" [--owner A] [--deps T1,T2] [--path P] [--check "cmd"]
26
+ [--spec "done when …"] [--as coordinator]
18
27
  swarm_tasks.py list [--owner A] [--status S]
19
28
  swarm_tasks.py claim ID --as A
20
29
  swarm_tasks.py set ID STATUS --as A [--note TEXT]
21
30
  swarm_tasks.py verify ID --as A # run the task's check command
31
+ swarm_tasks.py submit ID --as A [--artifact P]... [--repo R --range main...swarm/A]
22
32
  swarm_tasks.py log TYPE --as A [--note TEXT] # append a free-form event
23
33
  """
24
34
  import argparse
@@ -32,7 +42,8 @@ from contextlib import contextmanager
32
42
  from pathlib import Path
33
43
 
34
44
  STATUSES = ["open", "in-progress", "blocked", "review", "approved"]
35
- APPROVERS = {"critic", "coordinator"}
45
+ APPROVERS = {"coordinator", "critic"} # "critic": swarms made before fresh review
46
+ REVIEWER = "fresh-review" # the actor of approvals that fresh_review.py earned
36
47
 
37
48
 
38
49
  def find_root(start=None):
@@ -97,18 +108,133 @@ def run_check(root, t):
97
108
  return False, "check timed out after 900s"
98
109
 
99
110
 
111
+
112
+ def fresh_review_script():
113
+ """$HIVEY_FRESH_REVIEW, else the installed hivey skill, else the repo copy next to this skill."""
114
+ candidates = [os.environ.get("HIVEY_FRESH_REVIEW")] + [
115
+ str(d / "fresh_review.py") for d in (Path.home() / ".claude/skills/hivey/scripts",
116
+ Path.home() / ".agents/skills/hivey/scripts",
117
+ Path(__file__).resolve().parents[2] / "hivey/scripts")]
118
+ for f in candidates:
119
+ if f and Path(f).is_file():
120
+ return Path(f)
121
+ sys.exit("fresh_review.py not found: install the hivey skill (hivey skill install)")
122
+
123
+
124
+ def review_config(root):
125
+ try:
126
+ cfg = json.loads((Path(root) / ".swarm" / "review.json").read_text())
127
+ except (OSError, ValueError):
128
+ cfg = {}
129
+ return {"critics": int(cfg.get("critics", 2)), "rounds": int(cfg.get("rounds", 3))}
130
+
131
+
132
+ def submit(root, args):
133
+ """Mark the task in review and start its fresh review in a detached process."""
134
+ cfg = review_config(root)
135
+ with board(root) as data:
136
+ t = get(data, args.id)
137
+ if t.get("owner") and t["owner"] != args.actor:
138
+ sys.exit(f"{args.id} is owned by {t['owner']}: only the owner submits it")
139
+ if t["status"] == "review":
140
+ sys.exit(f"{args.id} is already in review; wait for the verdict message")
141
+ if t["status"] == "approved":
142
+ sys.exit(f"{args.id} is already approved")
143
+ t["status"] = "review"
144
+ rnd = t.get("review_round", 0) + 1
145
+ if cfg["critics"] > 0:
146
+ t["review_round"] = rnd
147
+ log_event(root, "review_submitted", args.actor, task=args.id, round=rnd, critics=cfg["critics"])
148
+ task = dict(t)
149
+ if cfg["critics"] <= 0:
150
+ print(f"{args.id} is in review. Fresh review is off for this swarm (critics: 0): "
151
+ f"send DONE to the coordinator, who approves it.")
152
+ return
153
+ worker = [sys.executable, str(Path(__file__).resolve()), "--root", str(root), "_review", args.id,
154
+ "--as", args.actor, "--round", str(rnd)]
155
+ for a in args.artifact:
156
+ worker += ["--artifact", a]
157
+ if args.repo:
158
+ worker += ["--repo", args.repo, "--range", args.range or "HEAD~1..HEAD"]
159
+ reviews = Path(root) / ".swarm" / "reviews"
160
+ reviews.mkdir(parents=True, exist_ok=True)
161
+ log = open(reviews / f"{args.id}-r{rnd}.log", "w")
162
+ subprocess.Popen(worker, stdout=log, stderr=subprocess.STDOUT, start_new_session=True, cwd=str(root))
163
+ print(f"{args.id} round {rnd}: {cfg['critics']} fresh critics are reviewing it in the background "
164
+ f"(review: .swarm/reviews/{args.id}-r{rnd}.md). You'll get a message with the verdict; "
165
+ f"work on something else meanwhile.")
166
+
167
+
168
+ def review_worker(root, args):
169
+ """Detached: run fresh_review.py for one task, then record the verdict on the board."""
170
+ cfg = review_config(root)
171
+ t = get(load(root), args.id)
172
+ out = Path(root) / ".swarm" / "reviews" / f"{args.id}-r{args.round}.md"
173
+ spec = (f"Task {t['id']}: {t['title']}\nOwner: {t.get('owner') or '-'}\n"
174
+ f"Deliverable: {t.get('path') or '-'}\nDone when: {t.get('spec') or '(see the title and the check)'}")
175
+ cmd = [sys.executable, str(fresh_review_script()), "run", "--folder", str(root), "--label", t["id"],
176
+ "--round", str(args.round), "--spec", spec, "--out", str(out), "--check-cwd", str(root)]
177
+ request = Path(root) / ".swarm" / "REQUEST.md"
178
+ if request.exists():
179
+ cmd += ["--spec-file", str(request)]
180
+ if t.get("check"):
181
+ cmd += ["--check", t["check"]]
182
+ if t.get("last_review"):
183
+ cmd += ["--previous", t["last_review"]]
184
+ artifacts = list(args.artifact) + ([str(Path(root) / t["path"])] if t.get("path") else [])
185
+ for a in artifacts:
186
+ if Path(a).expanduser().exists():
187
+ cmd += ["--artifact", a]
188
+ if args.repo:
189
+ cmd += ["--repo", args.repo, "--range", args.range]
190
+ code = subprocess.run(cmd).returncode
191
+ verdict = {0: "APPROVED", 1: "CHANGES"}.get(code, "ERROR")
192
+ owner = t.get("owner") or args.actor
193
+ escalate = verdict == "CHANGES" and args.round >= cfg["rounds"]
194
+ with board(root) as data:
195
+ t = get(data, args.id)
196
+ t["last_review"] = str(out)
197
+ if verdict == "APPROVED":
198
+ t["status"], actor = "approved", REVIEWER
199
+ elif escalate:
200
+ t["status"], actor = "blocked", REVIEWER
201
+ else:
202
+ t["status"], actor = ("in-progress" if verdict == "CHANGES" else "review"), REVIEWER
203
+ t["notes"].append(f"{time.strftime('%Y-%m-%d %H:%M')} {REVIEWER}: r{args.round} {verdict} → {out}")
204
+ log_event(root, "review_" + verdict.lower(), actor, task=args.id, round=args.round, review=str(out))
205
+ rel = out.relative_to(root)
206
+ msgs = {"APPROVED": f"APPROVED {args.id} (fresh review r{args.round}, every critic PASS, check passed) → {rel}",
207
+ "CHANGES": f"CRITIQUE r{args.round} {args.id} → {rel}: fix every defect in its Defects list in one "
208
+ f"revision, then submit again",
209
+ "ERROR": f"Fresh review r{args.round} of {args.id} failed to run (see {rel} and its .log); "
210
+ f"tell the coordinator"}
211
+ if escalate:
212
+ msgs["CHANGES"] = (f"CRITIQUE r{args.round} {args.id} → {rel}: that was the last review round; "
213
+ f"the task is blocked and the coordinator decides")
214
+ send = ["hivey", "msg", "send"]
215
+ subprocess.run(send + [owner, msgs[verdict]], capture_output=True)
216
+ if escalate or verdict == "ERROR":
217
+ subprocess.run(send + ["coordinator", f"{args.id} needs you: fresh review r{args.round} "
218
+ f"{verdict} → {rel}"], capture_output=True)
219
+
100
220
  def main():
101
221
  p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
102
222
  p.add_argument("--root")
103
223
  sub = p.add_subparsers(dest="cmd", required=True)
104
224
  a = sub.add_parser("add"); a.add_argument("id"); a.add_argument("title")
105
225
  a.add_argument("--owner"); a.add_argument("--deps", default=""); a.add_argument("--path")
106
- a.add_argument("--check"); a.add_argument("--as", dest="actor", default="coordinator")
226
+ a.add_argument("--check"); a.add_argument("--spec", help="the deliverable's 'done when', for the critics")
227
+ a.add_argument("--as", dest="actor", default="coordinator")
107
228
  ls = sub.add_parser("list"); ls.add_argument("--owner"); ls.add_argument("--status")
108
229
  c = sub.add_parser("claim"); c.add_argument("id"); c.add_argument("--as", dest="actor", required=True)
109
230
  s = sub.add_parser("set"); s.add_argument("id"); s.add_argument("status", choices=STATUSES)
110
231
  s.add_argument("--as", dest="actor", required=True); s.add_argument("--note", default="")
111
232
  v = sub.add_parser("verify"); v.add_argument("id"); v.add_argument("--as", dest="actor", required=True)
233
+ sb = sub.add_parser("submit"); sb.add_argument("id"); sb.add_argument("--as", dest="actor", required=True)
234
+ sb.add_argument("--artifact", action="append", default=[]); sb.add_argument("--repo"); sb.add_argument("--range")
235
+ rw = sub.add_parser("_review"); rw.add_argument("id"); rw.add_argument("--as", dest="actor", required=True)
236
+ rw.add_argument("--round", type=int, required=True); rw.add_argument("--artifact", action="append", default=[])
237
+ rw.add_argument("--repo"); rw.add_argument("--range")
112
238
  lg = sub.add_parser("log"); lg.add_argument("type"); lg.add_argument("--as", dest="actor", required=True)
113
239
  lg.add_argument("--note", default="")
114
240
  args = p.parse_args()
@@ -119,12 +245,19 @@ def main():
119
245
  and (not args.status or t["status"] == args.status)]
120
246
  print(json.dumps(ts, indent=2))
121
247
  return
248
+ if args.cmd == "submit":
249
+ submit(root, args)
250
+ return
251
+ if args.cmd == "_review":
252
+ review_worker(root, args)
253
+ return
122
254
  if args.cmd == "log":
123
255
  log_event(root, args.type, args.actor, note=args.note)
124
256
  print(json.dumps({"ok": True}))
125
257
  return
126
258
  if args.cmd == "set" and args.status == "approved" and args.actor not in APPROVERS:
127
- sys.exit(f"only the critic or coordinator can approve {args.id} — set it to review and tell the critic")
259
+ sys.exit(f"only a fresh review or the coordinator can approve {args.id}: run "
260
+ f"`swarm_tasks.py submit {args.id} --as {args.actor}`")
128
261
  if args.cmd == "verify" or (args.cmd == "set" and args.status == "approved"):
129
262
  t = get(load(root), args.id)
130
263
  if t.get("check"): # run outside the board lock: a check can take minutes
@@ -146,7 +279,7 @@ def main():
146
279
  sys.exit(f"task {args.id} already exists")
147
280
  t = {"id": args.id, "title": args.title, "owner": args.owner, "status": "open",
148
281
  "deps": [d for d in args.deps.split(",") if d], "path": args.path, "check": args.check,
149
- "notes": []}
282
+ "spec": args.spec, "notes": []}
150
283
  data["tasks"].append(t)
151
284
  log_event(root, "task_added", args.actor, task=args.id, title=args.title, owner=args.owner)
152
285
  elif args.cmd == "claim":
@@ -1,35 +0,0 @@
1
- ---
2
- name: reviewer
3
- description: Independent reviewer for this agent's work. Use it before declaring any run or deliverable done. Give it the paths of what you produced. It checks them against The ask, the Deliverables table and the sources they cite, and returns PASS or a numbered list of concrete defects. It never edits files.
4
- tools: Read, Grep, Glob, Bash, WebFetch
5
- model: sonnet
6
- ---
7
-
8
- You are the reviewer for the agent whose `CLAUDE.md` is in the current directory. You didn't do
9
- the work, so judge it cold, with no charity for intentions. Your only output is a verdict.
10
-
11
- ## What to do
12
- 1. Read `CLAUDE.md` in the current directory: **The ask** (the user's verbatim request), the
13
- **Deliverables** table (paths, "done when", check commands) and the **Method** (plus
14
- `METHOD.md` if it exists).
15
- 2. Open every artifact you were given; don't trust the agent's summary of it.
16
- 3. Run each relevant **Check** command from the Deliverables table and record its exit code.
17
- 4. Verify the content:
18
- - Does it answer The ask, or something nearby?
19
- - Spot-check at least 3 concrete claims (numbers, names, dates, quotes) against the source
20
- they came from (open the URL or file). Report each mismatch with the correct value and
21
- where you found it.
22
- - Is anything required missing: sections, links, files, formats?
23
- - Is anything stated that the sources don't support?
24
- 5. Don't modify any file. Don't fix things yourself: the agent fixes, you report.
25
-
26
- ## Output format
27
- ```
28
- VERDICT: PASS | FAIL
29
- Checks: <command> → exit <n> (one line each)
30
- Defects:
31
- 1. <file:line or section> — <what's wrong> — <evidence / correct value + source>
32
- Verified OK: <the claims you checked that held up>
33
- ```
34
- PASS only if every check exits 0 and you found no defect that changes meaning. Keep it short.
35
- Style nits aren't defects.