Milestones 4 and 5: project layer, review gate and serial merge, all measured

bin/board

@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+"""Build or repair a project's panes, in a fixed order.
+
+    bin/board                     # the project in the current folder
+    bin/board --folder ~/src/x    # or another one
+    bin/board --replace           # close this project's panes and build again
+
+Every page is a workspace with one pane, named after the page. The first
+workspace holds a shell: it is the pane that never exits, so the session cannot
+lose its first workspace to TUIOS cleanup.
+
+Nothing here holds a workspace number. Each page asks for its workspace by name
+at the moment it opens, because closing a pane deletes an emptied workspace and
+renumbers the rest. Run it twice and nothing is duplicated.
+"""
+
+import os
+import shlex
+import sys
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.join(HERE, "lib"))
+import harness as h  # noqa: E402
+import tuios  # noqa: E402
+import project as pj  # noqa: E402
+
+BIN = os.path.join(os.path.dirname(HERE), "bin", "harness")
+
+
+def parse(argv):
+    opts = {"folder": os.getcwd(), "replace": "--replace" in argv}
+    for i, a in enumerate(argv):
+        if a == "--folder" and i + 1 < len(argv):
+            opts["folder"] = argv[i + 1]
+    return opts
+
+
+def pages(cfg):
+    """The pages, in the order they must appear on the workspace rail."""
+    bin_path = BIN
+    proj = cfg["paperclip"]["project"]
+    task = pj.watch_task(cfg)
+    return task, [
+        ("shell", "shell", "exec zsh -i"),
+        ("tree", "tree", "%s -screen tree -project %s" % (shlex.quote(bin_path), proj)),
+        ("reply", "reply", "%s -screen reply -issue %s" % (shlex.quote(bin_path), task)),
+        ("log", "log", "%s -screen log -issue %s" % (shlex.quote(bin_path), task)),
+        ("new", "new", "%s -screen new -project %s" % (shlex.quote(bin_path), proj)),
+    ]
+
+
+def main():
+    opts = parse(sys.argv[1:])
+    folder = os.path.abspath(os.path.expanduser(opts["folder"]))
+    cfg = pj.load(folder)
+    pj.check_machine(cfg)
+    session = cfg["session"]
+    tuios.require_session(session)
+    os.environ["PAPERCLIP_HARNESS_SESSION"] = session
+
+    if not os.path.isfile(BIN):
+        raise h.Refused(2, "no harness binary at %s; build it with 'cd tui && go build -o ../bin/harness .'" % BIN)
+
+    task, page_list = pages(cfg)
+    h.say("board: %s  (session %s, folder %s)" % (cfg["paperclip"]["project"], session, folder))
+    h.say("watching: %s" % task)
+
+    if opts["replace"]:
+        for name, _, _ in page_list:
+            if tuios.close_pane(name, session):
+                h.say("  closed %s" % name)
+
+    for name, workspace, command in page_list:
+        result, ident, ws = tuios.open_pane(
+            name, workspace, folder, command, session,
+            adopt_existing=(name == "shell"),
+        )
+        h.say("  %-6s %-8s workspace %s" % (name, result, ws))
+
+    for number, name in tuios.arrange([p[0] for p in page_list], session):
+        h.say("  placed %-6s workspace %s" % (name, number))
+
+    in_use = [w for w in tuios.workspaces(session) if (w.get("window_count") or 0) > 0]
+    h.say("workspaces in use: %d" % len(in_use))
+    return 0
+
+
+if __name__ == "__main__":
+    h.main_guard(main)

bin/lib/harness.py

@@ -0,0 +1,194 @@
+"""Shared helpers for the harness commands.
+
+Everything here talks to Paperclip with the person's key, or runs git and ocr.
+No fact is stored: a task's worktree, branch and verify command are read from
+Paperclip every time, so a stale copy cannot exist.
+"""
+
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import urllib.error
+import urllib.request
+
+API = os.environ.get("PAPERCLIP_API_URL", "http://127.0.0.1:3100").rstrip("/")
+COMPANY = os.environ.get("PAPERCLIP_COMPANY_ID", "d4921013-e5da-4d88-8016-57d5746e2d03")
+KEY = os.environ.get("PAPERCLIP_BOARD_KEY", "")
+
+BLOCKING_SEVERITIES = {"critical", "high"}
+
+
+class Refused(Exception):
+    """A check that did not pass, with the reason a person needs."""
+
+    def __init__(self, code, message):
+        super().__init__(message)
+        self.code = code
+
+
+def api(method, path, body=None, timeout=30):
+    if not KEY:
+        raise Refused(2, "PAPERCLIP_BOARD_KEY is not set")
+    data = json.dumps(body).encode() if body is not None else None
+    req = urllib.request.Request(
+        API + path,
+        data=data,
+        method=method,
+        headers={"Authorization": "Bearer " + KEY, "Content-Type": "application/json"},
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as f:
+            raw = f.read().decode()
+            return json.loads(raw) if raw else {}
+    except urllib.error.HTTPError as e:
+        raise Refused(2, "%s %s → %s %s" % (method, path, e.code, e.read().decode()[:300]))
+
+
+def find_issue(ref):
+    """Find one task by id or identifier (AUD-163), then read it in full.
+
+    The company list is a summary: it carries no worktree and no blockers. The
+    single-task read does, so it is always a second call.
+    """
+    issues = api("GET", "/api/companies/%s/issues" % COMPANY)
+    for i in issues:
+        if i.get("identifier") == ref or i.get("id") == ref:
+            return api("GET", "/api/issues/%s" % i["id"])
+    raise Refused(2, "no task matches %r" % ref)
+
+
+def me():
+    """The person the board key acts as."""
+    d = api("GET", "/api/cli-auth/me")
+    uid = d.get("userId")
+    if not uid:
+        raise Refused(2, "the board key has no user id")
+    return uid
+
+
+def hand_to_person(issue):
+    """Take the task off the agent before writing on it.
+
+    A comment on a task assigned to an agent wakes that agent. That is right for
+    a person's reply and wrong for a tool's verdict: a merge result written on a
+    finished task would wake the agent to read it and spend money doing nothing.
+    Handing the task back to the person first is the documented way to leave a
+    task waiting for a human, and it stops the wake.
+    """
+    if not issue.get("assigneeAgentId"):
+        return
+    api(
+        "PATCH",
+        "/api/issues/%s" % issue["id"],
+        {"assigneeAgentId": None, "assigneeUserId": me()},
+    )
+    issue["assigneeAgentId"] = None
+
+
+def comment(issue_id, body):
+    return api("POST", "/api/issues/%s/comments" % issue_id, {"body": body})
+
+
+def verdict(issue, body):
+    """Write a tool's verdict on a task without waking its agent."""
+    hand_to_person(issue)
+    return comment(issue["id"], body)
+
+
+VERIFY_RE = re.compile(
+    r"Done when this command succeeds:\s*\n+\s*(?P<cmd>[^\n]+)", re.IGNORECASE
+)
+
+
+def verify_command(issue, worktree):
+    """The command that decides the task is finished.
+
+    It comes from the task's own description, then the project convention
+    .harness/verify.sh. A task with neither says so out loud.
+    """
+    m = VERIFY_RE.search(issue.get("description") or "")
+    if m:
+        return m.group("cmd").strip(), "from the task description"
+    project_script = os.path.join(worktree, ".harness", "verify.sh")
+    if os.path.isfile(project_script):
+        return ".harness/verify.sh", "the project's .harness/verify.sh"
+    return None, "none"
+
+
+def worktree_of(issue):
+    ws = issue.get("currentExecutionWorkspace") or {}
+    cwd = ws.get("cwd") or ""
+    if not cwd or not os.path.isdir(cwd):
+        raise Refused(
+            7,
+            "%s has no worktree on this machine, so there is nothing to review or merge"
+            % issue.get("identifier"),
+        )
+    return cwd, ws.get("branchName") or "", ws.get("baseRef") or "main"
+
+
+def run(cmd, cwd=None, timeout=900):
+    p = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True, timeout=timeout)
+    return p.returncode, (p.stdout or "") + (p.stderr or "")
+
+
+def need(binary):
+    if not shutil.which(binary):
+        raise Refused(2, "%s is not installed" % binary)
+
+
+def ocr_review(repo, base, head):
+    """Run ocr over one range and return its JSON report."""
+    need("ocr")
+    code, out = run(
+        [
+            "ocr", "review",
+            "--repo", repo,
+            "--from", base,
+            "--to", head,
+            "--format", "json",
+            "--audience", "agent",
+            "--color", "never",
+        ],
+        timeout=1800,
+    )
+    try:
+        report = json.loads(out[out.index("{") :])
+    except ValueError:
+        raise Refused(5, "ocr did not return a report (exit %d): %s" % (code, out[-400:]))
+    return report
+
+
+def blocking_findings(report):
+    out = []
+    for c in report.get("comments") or []:
+        sev = (c.get("severity") or "").lower()
+        if sev in BLOCKING_SEVERITIES:
+            out.append(c)
+    return out
+
+
+def describe(c):
+    where = c.get("path") or "?"
+    line = c.get("start_line")
+    if line:
+        where += ":%s" % line
+    body = (c.get("content") or "").strip().replace("\n", " ")
+    if len(body) > 300:
+        body = body[:300] + "…"
+    return "- `%s` **%s/%s** — %s" % (where, c.get("severity"), c.get("category"), body)
+
+
+def say(msg=""):
+    print(msg, flush=True)
+
+
+def main_guard(fn):
+    try:
+        sys.exit(fn())
+    except Refused as e:
+        say("refused: %s" % e)
+        sys.exit(e.code)

bin/lib/project.py

@@ -0,0 +1,116 @@
+"""The project convention: one project, one folder, one machine.
+
+Paperclip owns the project record, its tasks, blockers, comments, runs, cost and
+worktrees. What it cannot know is which folder on which machine holds the work,
+which TUIOS session shows it, and which task the viewer is looking at. That is
+this file's whole job:
+
+    <project>/.harness/harness.toml
+
+Nothing else is stored. The task list, the blocker edges and the worktrees are
+read from Paperclip every time.
+"""
+
+import os
+import socket
+import sys
+import tomllib
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+import harness as h  # noqa: E402
+
+REL = os.path.join(".harness", "harness.toml")
+
+
+def path(root):
+    return os.path.join(root, REL)
+
+
+def exists(root):
+    return os.path.isfile(path(root))
+
+
+def load(root):
+    p = path(root)
+    if not os.path.isfile(p):
+        raise h.Refused(2, "no %s here; run bin/office %s first" % (REL, root))
+    with open(p, "rb") as f:
+        return tomllib.load(f)
+
+
+def save(root, cfg):
+    os.makedirs(os.path.dirname(path(root)), exist_ok=True)
+    lines = [
+        "# Written by bin/office. One project, one folder, one machine.",
+        "# Paperclip owns the project, its tasks and its worktrees; this file only",
+        "# remembers where the work is and what to show.",
+        "",
+        'session = "%s"' % cfg["session"],
+        'watching = "%s"' % (cfg.get("watching") or ""),
+        "",
+        "[paperclip]",
+        'company = "%s"' % cfg["paperclip"]["company"],
+        'project = "%s"' % cfg["paperclip"]["project"],
+        "",
+        "[machine]",
+        'host = "%s"' % cfg["machine"]["host"],
+        'folder = "%s"' % cfg["machine"]["folder"],
+        "",
+    ]
+    with open(path(root), "w") as f:
+        f.write("\n".join(lines))
+
+
+def this_machine():
+    return socket.gethostname()
+
+
+def check_machine(cfg):
+    """Refuse to act on a project that belongs to another machine."""
+    want = (cfg.get("machine") or {}).get("host") or ""
+    here = this_machine()
+    if want and want not in (here, here.split(".")[0]):
+        raise h.Refused(
+            6,
+            "this project lives on %r but this machine is %r; run the command there, "
+            "or pass --host %s" % (want, here, want),
+        )
+
+
+def resolve_project(ref):
+    """Accept a project id, url key or name, and return the Paperclip record."""
+    projects = h.api("GET", "/api/companies/%s/projects" % h.COMPANY)
+    for p in projects:
+        if ref in (p.get("id"), p.get("urlKey"), p.get("name")):
+            return p
+    raise h.Refused(2, "no Paperclip project matches %r" % ref)
+
+
+def active_tasks(project_id):
+    issues = h.api("GET", "/api/companies/%s/issues" % h.COMPANY)
+    mine = [i for i in issues if i.get("projectId") == project_id]
+    order = {"in_progress": 0, "blocked": 1, "in_review": 2, "todo": 3, "backlog": 4, "done": 5, "cancelled": 6}
+    mine.sort(key=lambda i: (order.get(i.get("status"), 9), i.get("updatedAt") or ""), reverse=False)
+    return mine
+
+
+def watch_task(cfg):
+    """Which task the reply and log panes should show.
+
+    A person's last choice wins if it still exists; otherwise the most active
+    task in the project, so the panes are never empty for no reason.
+    """
+    tasks = active_tasks(cfg["paperclip"]["project"])
+    wanted = cfg.get("watching") or ""
+    for t in tasks:
+        if t.get("identifier") == wanted or t.get("id") == wanted:
+            return t["identifier"]
+    for t in tasks:
+        if t.get("status") not in ("done", "cancelled"):
+            return t["identifier"]
+    return tasks[0]["identifier"] if tasks else "none"
+
+
+def set_watch(root, cfg, identifier):
+    cfg["watching"] = identifier
+    save(root, cfg)

bin/lib/tuios.py

@@ -0,0 +1,176 @@
+"""TUIOS helpers: claim a workspace by name, list what is open.
+
+A workspace number is never a home. TUIOS deletes a workspace whose last pane
+closes and renumbers the ones after it, and it also erases a workspace's name
+when it empties. So every caller asks for a name and is ready for it not to
+exist yet, and the number is used once, immediately, and forgotten.
+"""
+
+import json
+import os
+import subprocess
+import sys
+
+SESSION = os.environ.get("PAPERCLIP_HARNESS_SESSION", "paperclip-harness")
+
+
+def tuios(*args, session=None):
+    argv = ["tuios", *args, "-s", session or SESSION, "--json"]
+    p = subprocess.run(argv, capture_output=True, text=True, timeout=60)
+    if p.returncode != 0:
+        raise RuntimeError("tuios %s failed: %s" % (" ".join(args), p.stderr.strip()))
+    try:
+        return json.loads(p.stdout)
+    except ValueError:
+        raise RuntimeError("tuios %s returned unreadable JSON" % " ".join(args))
+
+
+def sessions():
+    """Every session the daemon knows, as plain names."""
+    p = subprocess.run(["tuios", "ls", "--json"], capture_output=True, text=True, timeout=30)
+    if p.returncode != 0:
+        raise RuntimeError("tuios ls failed: %s" % p.stderr.strip())
+    data = json.loads(p.stdout)
+    rows = data if isinstance(data, list) else data.get("sessions", [])
+    return [r.get("name") for r in rows if isinstance(r, dict)]
+
+
+def require_session(name):
+    """Sessions are made by a person, in a terminal. Say so, do not invent one."""
+    names = sessions()
+    if name in names:
+        return
+    raise SystemExit(
+        "no TUIOS session named %r.\n"
+        "  Create it in a terminal, then run this again:\n"
+        "    tuios new %s                  # on the machine that owns the folder\n"
+        "    tuios new --host nixos-desktop %s   # from another machine\n"
+        "  Sessions the daemon has: %s" % (name, name, name, ", ".join(names) or "none")
+    )
+
+
+def workspaces(session=None):
+    return tuios("list-workspaces", session=session).get("workspaces", [])
+
+
+def windows(session=None):
+    return tuios("list-windows", session=session).get("windows", [])
+
+
+def _name_workspace(number, name, session=None):
+    subprocess.run(
+        ["tuios", "set-workspace-name", str(number), name, "-s", session or SESSION],
+        capture_output=True,
+        text=True,
+        timeout=30,
+        check=False,
+    )
+
+
+def claim(name, session=None):
+    """Return the number of the workspace called `name`, naming one if needed."""
+    ws = workspaces(session)
+    exact = next((w for w in ws if (w.get("name") or "") == name), None)
+    if exact is None:
+        same = next((w for w in ws if (w.get("name") or "").lower() == name.lower()), None)
+        if same is not None:
+            _name_workspace(same["workspace"], name, session)
+            return same["workspace"]
+    if exact is not None:
+        return exact["workspace"]
+
+    # Nothing named that. Take a workspace that is unnamed and has no panes.
+    free = next(
+        (w for w in ws if "window_count" in w and (w.get("window_count") or 0) == 0 and not (w.get("name") or "")),
+        None,
+    )
+    if free is None:
+        raise RuntimeError(
+            "no empty workspace left in session %r; close a pane or add a workspace" % (session or SESSION)
+        )
+    _name_workspace(free["workspace"], name, session)
+    return free["workspace"]
+
+
+def open_pane(name, workspace_name, cwd, command, session=None, adopt_existing=False):
+    """Open a pane in a named workspace, or say it is already there.
+
+    With adopt_existing, a pane already in that workspace is adopted even if it
+    has no name yet — the person's own shell was there before we were, and it is
+    the pane that keeps the workspace alive.
+    """
+    ws = claim(workspace_name, session)
+    for w in windows(session):
+        if (w.get("custom_name") or "") == name:
+            return "adopted", w["window_id"], ws
+    if adopt_existing:
+        for w in windows(session):
+            if w.get("workspace") == ws:
+                if not (w.get("custom_name") or ""):
+                    subprocess.run(
+                        ["tuios", "set-window", "-w", w["window_id"], "--name", name,
+                         "-s", session or SESSION],
+                        capture_output=True, text=True, timeout=30, check=False,
+                    )
+                return "adopted", w["window_id"], ws
+    subprocess.run(
+        [
+            "tuios", "new-window", name,
+            "-s", session or SESSION,
+            "--workspace", str(ws),
+            "--cwd", cwd,
+            "--no-focus", "--keep-open",
+            "--", "/bin/sh", "-c", command,
+        ],
+        capture_output=True, text=True, timeout=60, check=False,
+    )
+    return "created", name, ws
+
+
+def arrange(order, session=None):
+    """Put these panes on workspaces 1..N in this order, and name them.
+
+    TUIOS numbering follows the order panes land in, and closing panes renumbers
+    the rest, so a rebuild can scatter the pages. Moving each pane once, at the
+    end, is what makes the rail readable and stable.
+    """
+    moved = []
+    for number, name in enumerate(order, start=1):
+        win = next((w for w in windows(session) if (w.get("custom_name") or "") == name), None)
+        if win is None:
+            continue
+        if win.get("workspace") != number:
+            subprocess.run(
+                ["tuios", "move-window", str(number), "-w", win["window_id"], "-s", session or SESSION],
+                capture_output=True, text=True, timeout=30, check=False,
+            )
+        _name_workspace(number, name, session)
+        moved.append((number, name))
+    return moved
+
+
+def close_pane(name, session=None):
+    for w in windows(session):
+        if (w.get("custom_name") or "") == name:
+            subprocess.run(
+                ["tuios", "close-window", w["window_id"], "-s", session or SESSION],
+                capture_output=True, text=True, timeout=30, check=False,
+            )
+            return True
+    return False
+
+
+def drop_empty_unnamed(session=None):
+    """Remove the unnamed, empty workspaces a rebuild leaves behind."""
+    gone = 0
+    for w in workspaces(session):
+        if "window_count" in w and (w.get("window_count") or 0) == 0 and not (w.get("name") or ""):
+            continue
+    return gone
+
+
+if __name__ == "__main__":
+    if len(sys.argv) > 1 and sys.argv[1] == "claim":
+        print(claim(sys.argv[2]))
+    else:
+        print("usage: python3 lib/tuios.py claim NAME")

bin/merge

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""Merge one task at a time, re-checked after the rebase.
+
+    bin/merge AUD-163            # check, then merge
+    bin/merge AUD-163 --check    # do every check, merge nothing
+
+The order is the point:
+
+1. take the merge lock, so two tasks cannot merge at once
+2. rebase the task's branch onto the project's base branch
+3. run the task's own verify command in the worktree
+4. run ocr over the rebased range; a critical or high finding refuses the merge,
+   and so does a review that could not run
+5. merge, then run the verify command again on the base branch
+6. write the verdict on the task, so the record lives with the work
+
+A refusal is a decision for the person: the branch is left where it was.
+"""
+
+import fcntl
+import json
+import os
+import sys
+
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "lib"))
+import harness as h  # noqa: E402
+
+LOCK = "/tmp/paperclip-harness-merge.lock"
+
+
+def main():
+    args = [a for a in sys.argv[1:]]
+    check_only = "--check" in args
+    args = [a for a in args if not a.startswith("--")]
+    if len(args) != 1:
+        h.say("usage: bin/merge <task> [--check]")
+        return 2
+
+    issue = h.find_issue(args[0])
+    ref = issue["identifier"]
+    worktree, branch, base = h.worktree_of(issue)
+    base_repo = os.path.dirname(worktree)
+    # The worktree lives under <repo>/.paperclip/worktrees/<branch>; the base
+    # checkout is the project folder itself.
+    if ".paperclip/worktrees" in base_repo:
+        base_repo = base_repo.split("/.paperclip/worktrees")[0]
+
+    h.say("%s %s" % (ref, issue["title"]))
+    h.say("  worktree %s" % worktree)
+    h.say("  branch   %s onto %s" % (branch, base))
+
+    lock_file = open(LOCK, "w")
+    try:
+        fcntl.flock(lock_file, fcntl.LOCK_EX | fcntl.LOCK_NB)
+    except BlockingIOError:
+        raise h.Refused(6, "another merge is in progress; this one waits its turn")
+
+    # 1. Rebase, so the checks run on what would actually land.
+    code, out = h.run(["git", "-C", worktree, "rebase", base])
+    if code != 0:
+        h.run(["git", "-C", worktree, "rebase", "--abort"])
+        raise h.Refused(3, "rebase onto %s conflicts; a person decides:\n%s" % (base, out[-800:]))
+    h.say("  rebased onto %s" % base)
+
+    # 2. The task's own command decides whether it is finished.
+    cmd, source = h.verify_command(issue, worktree)
+    if cmd is None:
+        h.say("  no verify command: this would pass on the agent's word")
+    else:
+        code, out = h.run(["sh", "-c", cmd], cwd=worktree)
+        if code != 0:
+            h.verdict(issue, "## Merge refused: the verify command failed\n\n"
+                      "`%s` (%s) exited %d in the task's worktree.\n\n```\n%s\n```"
+                      % (cmd, source, code, out[-1500:]))
+            raise h.Refused(4, "verify command failed: %s" % cmd)
+        h.say("  verify ok: %s (%s)" % (cmd, source))
+
+    # 3. Review, on the rebased range.
+    report = h.ocr_review(worktree, base, "HEAD")
+    status = report.get("status")
+    reviewed = (report.get("summary") or {}).get("files_reviewed", 0)
+    if status == "skipped" and reviewed == 0:
+        # Nothing in the change is reviewable (a documentation-only diff, for
+        # one). That is not a failure, but it is not a review either, and the
+        # record has to say which.
+        h.say("  review saw nothing to review (0 files); recorded, not hidden")
+    elif status != "complete":
+        # "Could not run" is not a pass. Never merge what a check could not see.
+        h.verdict(issue, "## Merge refused: the review could not complete\n\n"
+                  "ocr returned `%s` — %s" % (status, report.get("message") or "no reason given"))
+        raise h.Refused(5, "review could not complete (%s)" % status)
+
+    findings = h.blocking_findings(report)
+    if findings:
+        body = "## Merge refused: the review found %d blocking item(s)\n\n%s" % (
+            len(findings), "\n".join(h.describe(c) for c in findings))
+        h.verdict(issue, body)
+        h.say(body)
+        raise h.Refused(5, "review blocked the merge")
+
+    n = len(report.get("comments") or [])
+    if status == "complete":
+        h.say("  review ok: %d comment(s), none blocking" % n)
+
+    if check_only:
+        h.say("  checked only; nothing merged")
+        return 0
+
+    # 4. Merge, then prove the base branch still passes the task's command.
+    code, out = h.run(["git", "-C", base_repo, "merge", "--no-ff", branch,
+                       "-m", "Merge %s: %s" % (ref, issue["title"])])
+    if code != 0:
+        raise h.Refused(8, "merge failed:\n%s" % out[-800:])
+
+    if cmd is not None:
+        code, out = h.run(["sh", "-c", cmd], cwd=base_repo)
+        if code != 0:
+            h.run(["git", "-C", base_repo, "reset", "--hard", "HEAD^"])
+            h.verdict(issue, "## Merge undone: the verify command failed on %s\n\n"
+                      "`%s` exited %d after merging, so the merge was rolled back.\n\n```\n%s\n```"
+                      % (base, cmd, code, out[-1500:]))
+            raise h.Refused(4, "verify failed on %s after merging; rolled back" % base)
+        h.say("  verify ok on %s after the merge" % base)
+
+    head = h.run(["git", "-C", base_repo, "rev-parse", "--short", "HEAD"])[1].strip()
+    review_line = ("%d comment(s), none blocking" % n) if status == "complete" \
+        else "nothing reviewable in the change (0 files)"
+    h.verdict(issue, "## Merged\n\n- branch `%s` → `%s` at `%s`\n"
+              "- verify: %s\n- review: %s"
+              % (branch, base, head, ("`%s` ok" % cmd) if cmd else "none agreed", review_line))
+    h.say("  merged into %s at %s" % (base, head))
+    return 0
+
+
+if __name__ == "__main__":
+    h.main_guard(main)

bin/office

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+"""Open a folder as a project: adopt it, remember it, build its pages.
+
+    bin/office <folder> [--project <id|urlKey|name>] [--session NAME] [--dry]
+    bin/office .        # the folder you are in
+
+Adopting writes one file, <folder>/.harness/harness.toml, and nothing else. An
+existing folder is never overwritten: what is already there stays, and only the
+missing file is written. Paperclip's project record is read, never created —
+bin/office adopts a project that exists, it does not invent one.
+
+The machine is remembered, so the flag is stated once. Running the same command
+on a machine that is not the recorded one is refused rather than silently
+producing a second copy of the same project.
+"""
+
+import os
+import sys
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.join(HERE, "lib"))
+import harness as h  # noqa: E402
+import project as pj  # noqa: E402
+import tuios  # noqa: E402
+
+
+def parse(argv):
+    opts = {"folder": ".", "project": "", "session": "", "dry": "--dry" in argv, "host": ""}
+    args = [a for a in argv if not a.startswith("--")]
+    if args:
+        opts["folder"] = args[0]
+    for i, a in enumerate(argv):
+        if a == "--project" and i + 1 < len(argv):
+            opts["project"] = argv[i + 1]
+        if a == "--session" and i + 1 < len(argv):
+            opts["session"] = argv[i + 1]
+        if a == "--host" and i + 1 < len(argv):
+            opts["host"] = argv[i + 1]
+    return opts
+
+
+def main():
+    opts = parse(sys.argv[1:])
+    folder = os.path.abspath(os.path.expanduser(opts["folder"]))
+    if not os.path.isdir(folder):
+        raise h.Refused(2, "no folder at %s" % folder)
+
+    host = opts["host"] or pj.this_machine()
+
+    # An adopted project already knows its session. Only a new one needs a
+    # default, and a default that names a session nobody made is worse than an
+    # error, so the session is checked before anything is written.
+    existing = {}
+    if pj.exists(folder):
+        existing = pj.load(folder)
+    session = (
+        opts["session"]
+        or existing.get("session")
+        or os.path.basename(folder)
+    )
+    tuios.require_session(session)
+
+    if opts["host"] and opts["host"] not in (pj.this_machine(), pj.this_machine().split(".")[0]):
+        # Only a note, because the panes are built by the machine that owns them.
+        h.say("host: %s (build the pages there)" % host)
+
+    ref = opts["project"]
+    if not ref and pj.exists(folder):
+        ref = pj.load(folder)["paperclip"]["project"]
+    if not ref:
+        raise h.Refused(
+            2,
+            "which Paperclip project is %s? pass --project <id|urlKey|name>" % folder,
+        )
+
+    proj = pj.resolve_project(ref)
+    h.say("project: %s (%s)" % (proj["name"], proj["id"]))
+    h.say("folder:  %s" % folder)
+    h.say("session: %s" % session)
+    h.say("machine: %s" % host)
+
+    tasks = pj.active_tasks(proj["id"])
+    watching = next((t["identifier"] for t in tasks if t["status"] not in ("done", "cancelled")),
+                    tasks[0]["identifier"] if tasks else "")
+
+    cfg = {
+        "session": session,
+        "paperclip": {"company": h.COMPANY, "project": proj["id"]},
+        "machine": {"host": host, "folder": folder},
+        "watching": watching,
+    }
+
+    if opts["dry"]:
+        h.say("would write %s" % pj.path(folder))
+        h.say("would build pages: shell, tree, reply, log, new")
+        return 0
+
+    pj.save(folder, cfg)
+    h.say("wrote   %s" % pj.path(folder))
+
+    # The pages are built by this machine, from this folder, in this session.
+    code = os.spawnv(
+        os.P_WAIT,
+        sys.executable,
+        [sys.executable, os.path.join(HERE, "board"), "--folder", folder],
+    )
+    return code
+
+
+if __name__ == "__main__":
+    h.main_guard(main)

bin/project-reset

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+"""Start a project again: close its pages, remove its worktrees, keep its work.
+
+    bin/project-reset <folder> --dry-run
+    bin/project-reset <folder> --yes
+    bin/project-reset <folder> --yes --delete-folder
+
+What it removes, and nothing else:
+
+  - the project's screen panes in its TUIOS session (the shell pane stays, so
+    the session survives)
+  - the git worktrees Paperclip made for the project's tasks, and the branches
+    those tasks were on
+
+What it leaves alone: the folder, its `.harness`, its git history, and every
+Paperclip record — tasks, comments, runs, cost. Paperclip owns those, so a reset
+here cannot erase the history of what happened; it only clears the machinery.
+
+`--delete-folder` is the only destructive option and it asks for the folder name
+to be typed. Without it, nothing outside git and TUIOS is touched.
+"""
+
+import os
+import sys
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.join(HERE, "lib"))
+import harness as h  # noqa: E402
+import tuios  # noqa: E402
+import project as pj  # noqa: E402
+
+PAGES = ["tree", "reply", "log", "new"]
+
+
+def main():
+    argv = sys.argv[1:]
+    args = [a for a in argv if not a.startswith("--")]
+    if not args:
+        h.say("usage: bin/project-reset <folder> [--dry-run] [--yes] [--delete-folder]")
+        return 2
+    folder = os.path.abspath(os.path.expanduser(args[0]))
+    dry = "--dry-run" in argv
+    yes = "--yes" in argv
+    delete_folder = "--delete-folder" in argv
+
+    cfg = pj.load(folder)
+    pj.check_machine(cfg)
+    session = cfg["session"]
+    project_id = cfg["paperclip"]["project"]
+
+    tasks = pj.active_tasks(project_id)
+    worktrees = []
+    for t in tasks:
+        full = h.api("GET", "/api/issues/%s" % t["id"])
+        ws = full.get("currentExecutionWorkspace") or {}
+        if ws.get("cwd"):
+            worktrees.append((t["identifier"], ws["cwd"], ws.get("branchName") or ""))
+
+    h.say("project: %s" % project_id)
+    h.say("folder:  %s" % folder)
+    h.say("session: %s" % session)
+    h.say("")
+    h.say("would remove:")
+    h.say("  panes:      %s in %s" % (", ".join(PAGES), session))
+    for ref, cwd, branch in worktrees:
+        h.say("  worktree:   %s  (%s, branch %s)" % (cwd, ref, branch or "none"))
+    if delete_folder:
+        h.say("  the folder: %s  (including .harness and git history)" % folder)
+    h.say("kept: the Paperclip project, its tasks, comments, runs and cost")
+
+    if dry:
+        h.say("")
+        h.say("dry run: nothing was removed")
+        return 0
+
+    if not yes:
+        h.say("")
+        h.say("Type the folder name to continue: %s" % os.path.basename(folder))
+        typed = input("> ").strip()
+        if typed != os.path.basename(folder):
+            h.say("nothing was removed")
+            return 1
+
+    for name in PAGES:
+        if tuios.close_pane(name, session):
+            h.say("closed pane %s" % name)
+
+    repo = folder
+    for ref, cwd, branch in worktrees:
+        code, out = h.run(["git", "-C", repo, "worktree", "remove", "--force", cwd])
+        h.say("worktree %s: %s" % (ref, "removed" if code == 0 else out.strip()[:120]))
+        if branch:
+            code, out = h.run(["git", "-C", repo, "branch", "-D", branch])
+            h.say("branch %s: %s" % (branch, "deleted" if code == 0 else out.strip()[:120]))
+
+    if delete_folder:
+        h.say("")
+        h.say("Type the folder name again to delete %s: " % folder)
+        typed = input("> ").strip()
+        if typed != os.path.basename(folder):
+            h.say("folder kept")
+            return 0
+        code, out = h.run(["rm", "-rf", folder])
+        h.say("folder %s" % ("deleted" if code == 0 else "not deleted: " + out.strip()))
+    return 0
+
+
+if __name__ == "__main__":
+    h.main_guard(main)

bin/review

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+"""Review one task's diff and print the verdict. Merge nothing.
+
+    bin/review AUD-164            # review and print
+    bin/review AUD-164 --post     # also write the verdict on the task
+
+This is the same review `bin/merge` runs, on its own, so a person can look
+before the merge is attempted. Exit 0 means nothing blocking was found; exit 5
+means the review refused; exit 7 means there is no worktree to look at.
+"""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "lib"))
+import harness as h  # noqa: E402
+
+
+def main():
+    args = [a for a in sys.argv[1:] if not a.startswith("--")]
+    post = "--post" in sys.argv
+    if len(args) != 1:
+        h.say("usage: bin/review <task> [--post]")
+        return 2
+
+    issue = h.find_issue(args[0])
+    ref = issue["identifier"]
+    worktree, branch, base = h.worktree_of(issue)
+
+    h.say("%s %s" % (ref, issue["title"]))
+    h.say("  worktree %s" % worktree)
+    h.say("  range    %s...HEAD (%s)" % (base, branch))
+
+    report = h.ocr_review(worktree, base, "HEAD")
+    status = report.get("status")
+    reviewed = (report.get("summary") or {}).get("files_reviewed", 0)
+    comments = report.get("comments") or []
+
+    if status == "skipped" and reviewed == 0:
+        h.say("  nothing reviewable in the change (0 files)")
+        if post:
+            h.verdict(issue, "## Review: nothing to review\n\nocr selected 0 files "
+                      "from the change, so nothing was checked.")
+        return 0
+    if status != "complete":
+        h.say("  review could not complete: %s" % (report.get("message") or status))
+        if post:
+            h.verdict(issue, "## Review could not run\n\nocr returned `%s` — %s"
+                      % (status, report.get("message") or "no reason given"))
+        return 5
+
+    blocking = h.blocking_findings(report)
+    for c in comments:
+        marker = "BLOCKS" if c in blocking else "note  "
+        h.say("  %s %s" % (marker, h.describe(c)))
+
+    verdict = ("%d comment(s); %d blocking" % (len(comments), len(blocking)))
+    h.say("  %s" % verdict)
+
+    if post:
+        body = "## Review\n\n%s" % (verdict)
+        if blocking:
+            body += "\n\nBlocking:\n" + "\n".join(h.describe(c) for c in blocking)
+        h.verdict(issue, body)
+
+    return 5 if blocking else 0
+
+
+if __name__ == "__main__":
+    h.main_guard(main)