@opengsd/gsd-path 1.0.0 → 1.2.0

package/skills/gsd-path-define/scripts/loop_run.py

@@ -0,0 +1,839 @@
+#!/usr/bin/env python3
+# gsd-path project runtime
+"""Deterministic gates for a GSD Path LOOP.md loop spec.
+
+The loop skill (a model) reads the prose sections of a LOOP.md; this helper
+reads only its fixed machine fields — exact `key: value` lines, one per
+line, no nesting, read only from the header above the first `## `
+section. `check` decides run/skip from status, cooldown, and period
+budgets without running skip_when; `claim` runs skip_when before
+admitting a pass; `verify` re-runs the spec's verifier
+commands independently of any worker; `claim` atomically admits one pass and
+`finish` closes it append-only from helper-owned timing and verify events;
+`status` aggregates completed outcomes and exposes unfinished recovery.
+Relative paths (spec `log:`, skip_when, verify commands) resolve against the
+caller's working directory — invoke from the repository root.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+import re
+import subprocess
+import sys
+sys.dont_write_bytecode = True
+import time
+import uuid
+from contextlib import contextmanager
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from typing import Iterator, Optional, Sequence
+
+if sys.platform == "win32":
+    import msvcrt
+else:
+    import fcntl
+
+
+class LoopError(RuntimeError):
+    pass
+
+
+MANDATORY_FIELDS = ("loop", "status", "trigger", "verify", "max_iterations", "wall_clock", "log")
+OPTIONAL_FIELDS = ("cooldown", "skip_when", "period", "period_budget")
+KNOWN_FIELDS = frozenset((*MANDATORY_FIELDS, *OPTIONAL_FIELDS))
+REPEATABLE_FIELDS = frozenset({"verify"})
+STATUSES = ("active", "paused", "done")
+RESULTS = ("pass", "fail", "blocked", "skipped")
+CLAIM_EVENT = "claim"
+VERIFY_EVENT = "verify"
+FINISH_EVENT = "finish"
+FIELD_LINE = re.compile(r"^(?P<key>[a-z_]+):(?P<rest>[ \t].*)?$")
+TRAILING_COMMENT = re.compile(r"[ \t]*<!--.*?-->[ \t]*$")
+DURATION = re.compile(r"^(?P<amount>\d+)(?P<unit>s|m|h|d)?$")
+UNIT_SECONDS = {"s": 1, "m": 60, "h": 3600, "d": 86400, None: 1}
+OUTPUT_TAIL_LINES = 20
+
+
+def parse_duration(raw: str, field: str) -> int:
+    match = DURATION.fullmatch(raw.strip())
+    if not match:
+        raise LoopError(f"{field} must be a duration like 300s, 15m, 2h, or 1d")
+    return int(match.group("amount")) * UNIT_SECONDS[match.group("unit")]
+
+
+def parse_spec(path: Path) -> dict:
+    if path.is_symlink() or not path.is_file():
+        raise LoopError(f"spec must be a real file: {path}")
+    fields: dict[str, object] = {"verify": []}
+    for number, line in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
+        if line.startswith("## "):
+            break  # machine fields live in the header; prose sections follow
+        match = FIELD_LINE.match(line)
+        if not match or match.group("key") not in KNOWN_FIELDS:
+            continue
+        key = match.group("key")
+        rest = match.group("rest")
+        value = TRAILING_COMMENT.sub("", rest).strip() if rest else ""
+        if not value:
+            raise LoopError(f"{path}:{number}: field {key} has no value")
+        if key in REPEATABLE_FIELDS:
+            fields[key].append(value)
+        elif key in fields:
+            raise LoopError(f"{path}:{number}: duplicate field {key}")
+        else:
+            fields[key] = value
+    missing = [key for key in MANDATORY_FIELDS if not fields.get(key)]
+    if missing:
+        raise LoopError(f"{path}: missing mandatory fields: {', '.join(missing)}")
+    if fields["status"] not in STATUSES:
+        raise LoopError(f"{path}: status must be one of {', '.join(STATUSES)}")
+    fields["max_iterations"] = parse_positive_int(fields["max_iterations"], "max_iterations")
+    fields["wall_clock"] = parse_duration(fields["wall_clock"], "wall_clock")
+    for key in ("cooldown", "period", "period_budget"):
+        if key in fields:
+            fields[key] = parse_duration(fields[key], key)
+    if "period_budget" in fields and "period" not in fields:
+        raise LoopError(f"{path}: period_budget requires period")
+    return fields
+
+
+def parse_positive_int(raw: str, field: str) -> int:
+    try:
+        value = int(raw)
+    except ValueError as error:
+        raise LoopError(f"{field} must be a positive integer") from error
+    if value < 1:
+        raise LoopError(f"{field} must be a positive integer")
+    return value
+
+
+def log_path(fields: dict) -> Path:
+    return Path(fields["log"])
+
+
+@contextmanager
+def log_lock(fields: dict) -> Iterator[None]:
+    path = log_path(fields)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    lock_path = path.with_name(f".{path.name}.lock")
+    with lock_path.open("a+b") as handle:
+        if sys.platform == "win32":
+            handle.seek(0, os.SEEK_END)
+            if handle.tell() == 0:
+                handle.write(b"\0")
+                handle.flush()
+            handle.seek(0)
+            msvcrt.locking(handle.fileno(), msvcrt.LK_LOCK, 1)
+        else:
+            fcntl.flock(handle.fileno(), fcntl.LOCK_EX)
+        try:
+            yield
+        finally:
+            if sys.platform == "win32":
+                handle.seek(0)
+                msvcrt.locking(handle.fileno(), msvcrt.LK_UNLCK, 1)
+            else:
+                fcntl.flock(handle.fileno(), fcntl.LOCK_UN)
+
+
+def load_log(fields: dict) -> list[dict]:
+    path = log_path(fields)
+    if not path.exists():
+        return []
+    records = []
+    for number, line in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
+        if not line.strip():
+            continue
+        try:
+            record = json.loads(line)
+        except json.JSONDecodeError as error:
+            raise LoopError(f"{path}:{number}: log line is not valid JSON") from error
+        if not isinstance(record, dict):
+            raise LoopError(f"{path}:{number}: log line must be a JSON object")
+        records.append(record)
+    return records
+
+
+def append_log(fields: dict, record: dict) -> None:
+    path = log_path(fields)
+    payload = (json.dumps(record, sort_keys=True) + "\n").encode()
+    with path.open("ab", buffering=0) as handle:
+        written = handle.write(payload)
+        if written != len(payload):
+            raise LoopError(f"could not append a complete record to {path}")
+        os.fsync(handle.fileno())
+
+
+def parse_timestamp(record: dict, path: Path, number: int) -> datetime:
+    raw = record.get("timestamp")
+    if not isinstance(raw, str):
+        raise LoopError(f"{path}:{number}: log record lacks a timestamp")
+    try:
+        parsed = datetime.fromisoformat(raw)
+    except ValueError as error:
+        raise LoopError(f"{path}:{number}: log record timestamp is invalid") from error
+    if parsed.tzinfo is None:
+        raise LoopError(f"{path}:{number}: log record timestamp must be timezone-aware")
+    return parsed
+
+
+def claim_id(record: dict, path: Path, number: int) -> str:
+    value = record.get("claim")
+    if not isinstance(value, str) or not value:
+        raise LoopError(f"{path}:{number}: claim event lacks a claim id")
+    return value
+
+
+def require_record_fields(
+    record: dict, expected: set[str], path: Path, number: int, event: str
+) -> None:
+    if set(record) != expected:
+        raise LoopError(f"{path}:{number}: {event} record has unsupported fields")
+
+
+def validate_text_field(record: dict, field: str, path: Path, number: int) -> str:
+    value = record.get(field)
+    if not isinstance(value, str):
+        raise LoopError(f"{path}:{number}: {field} must be text")
+    return value
+
+
+def log_state(
+    fields: dict, records: list[dict]
+) -> tuple[list[dict], dict, dict, dict, Optional[dict]]:
+    path = log_path(fields)
+    outcomes = []
+    claims = {}
+    verifications: dict[str, list[dict]] = {}
+    finishes = {}
+    for number, record in enumerate(records, 1):
+        event = record.get("event")
+        if event is None:
+            require_record_fields(
+                record,
+                {
+                    "timestamp",
+                    "loop",
+                    "trigger",
+                    "iterations",
+                    "result",
+                    "wall_clock_used",
+                    "rescue",
+                    "notes",
+                },
+                path,
+                number,
+                "legacy outcome",
+            )
+            parse_timestamp(record, path, number)
+            if record.get("loop") != fields["loop"] or record.get("trigger") != fields["trigger"]:
+                raise LoopError(f"{path}:{number}: outcome belongs to another loop")
+            if record.get("result") not in RESULTS:
+                raise LoopError(f"{path}:{number}: outcome has an invalid result")
+            if integer_field(record, "iterations", path, number) > fields["max_iterations"]:
+                raise LoopError(f"{path}:{number}: outcome exceeds max_iterations")
+            integer_field(record, "wall_clock_used", path, number)
+            if not isinstance(record.get("rescue"), bool):
+                raise LoopError(f"{path}:{number}: rescue must be boolean")
+            validate_text_field(record, "notes", path, number)
+            outcomes.append(record)
+            continue
+        if event not in {CLAIM_EVENT, VERIFY_EVENT, FINISH_EVENT}:
+            raise LoopError(f"{path}:{number}: unknown log event {event}")
+        identifier = claim_id(record, path, number)
+        if event == CLAIM_EVENT:
+            require_record_fields(
+                record,
+                {
+                    "event",
+                    "claim",
+                    "timestamp",
+                    "loop",
+                    "trigger",
+                    "max_iterations",
+                    "wall_clock_limit",
+                },
+                path,
+                number,
+                "claim",
+            )
+            if identifier in claims:
+                raise LoopError(f"{path}:{number}: duplicate claim {identifier}")
+            parse_timestamp(record, path, number)
+            if record.get("loop") != fields["loop"] or record.get("trigger") != fields["trigger"]:
+                raise LoopError(f"{path}:{number}: claim belongs to another loop")
+            if integer_field(record, "max_iterations", path, number) < 1:
+                raise LoopError(f"{path}:{number}: claim max_iterations must be positive")
+            if integer_field(record, "wall_clock_limit", path, number) < 1:
+                raise LoopError(f"{path}:{number}: claim wall_clock_limit must be positive")
+            claims[identifier] = record
+            continue
+        if event == VERIFY_EVENT:
+            require_record_fields(
+                record,
+                {
+                    "event",
+                    "claim",
+                    "timestamp",
+                    "sequence",
+                    "iteration",
+                    "result",
+                    "commands",
+                },
+                path,
+                number,
+                "verify",
+            )
+            if identifier not in claims:
+                raise LoopError(
+                    f"{path}:{number}: verify references unknown claim {identifier}"
+                )
+            if identifier in finishes:
+                raise LoopError(
+                    f"{path}:{number}: verify follows finish for claim {identifier}"
+                )
+            if record.get("result") not in {"pass", "fail"}:
+                raise LoopError(
+                    f"{path}:{number}: verify result must be pass or fail"
+                )
+            verified_at = parse_timestamp(record, path, number)
+            claim_record = claims[identifier]
+            claim_number = records.index(claim_record) + 1
+            if verified_at < parse_timestamp(claim_record, path, claim_number):
+                raise LoopError(f"{path}:{number}: verify predates its claim")
+            sequence = integer_field(record, "sequence", path, number)
+            expected = len(verifications.get(identifier, [])) + 1
+            if sequence != expected:
+                raise LoopError(
+                    f"{path}:{number}: verify sequence must be {expected}"
+                )
+            iteration = integer_field(record, "iteration", path, number)
+            if iteration != max(0, sequence - 1):
+                raise LoopError(
+                    f"{path}:{number}: verify iteration does not match sequence"
+                )
+            max_iterations = integer_field(
+                claim_record, "max_iterations", path, claim_number
+            )
+            if sequence > max_iterations + 1:
+                raise LoopError(f"{path}:{number}: verify exceeds max_iterations")
+            if integer_field(record, "commands", path, number) != len(fields["verify"]):
+                raise LoopError(f"{path}:{number}: verify command count does not match spec")
+            verifications.setdefault(identifier, []).append(record)
+            continue
+        if identifier not in claims:
+            raise LoopError(f"{path}:{number}: finish references unknown claim {identifier}")
+        if identifier in finishes:
+            raise LoopError(f"{path}:{number}: duplicate finish for claim {identifier}")
+        require_record_fields(
+            record,
+            {
+                "event",
+                "claim",
+                "timestamp",
+                "loop",
+                "trigger",
+                "iterations",
+                "result",
+                "wall_clock_used",
+                "rescue",
+                "notes",
+            },
+            path,
+            number,
+            "finish",
+        )
+        claim_record = claims[identifier]
+        claim_number = records.index(claim_record) + 1
+        started = parse_timestamp(claim_record, path, claim_number)
+        finished_at = parse_timestamp(record, path, number)
+        if finished_at < started:
+            raise LoopError(f"{path}:{number}: finish predates its claim")
+        if record.get("loop") != claim_record.get("loop") or record.get("trigger") != claim_record.get("trigger"):
+            raise LoopError(f"{path}:{number}: finish does not match its claim")
+        verification_records = verifications.get(identifier, [])
+        iterations = integer_field(record, "iterations", path, number)
+        if iterations != max(0, len(verification_records) - 1):
+            raise LoopError(f"{path}:{number}: finish iterations do not match verify events")
+        elapsed = max(0, math.ceil((finished_at - started).total_seconds()))
+        if integer_field(record, "wall_clock_used", path, number) != elapsed:
+            raise LoopError(f"{path}:{number}: finish wall clock does not match claim timestamps")
+        result = record.get("result")
+        if result not in RESULTS:
+            raise LoopError(f"{path}:{number}: finish has an invalid result")
+        if not isinstance(record.get("rescue"), bool):
+            raise LoopError(f"{path}:{number}: finish rescue must be boolean")
+        validate_text_field(record, "notes", path, number)
+        latest = verification_records[-1]["result"] if verification_records else None
+        if result == "pass" and latest != "pass":
+            raise LoopError(f"{path}:{number}: passing finish lacks a passing verify")
+        if latest == "pass" and result != "pass":
+            raise LoopError(f"{path}:{number}: passing verify must finish as pass")
+        if result == "fail":
+            max_iterations = integer_field(
+                claim_record, "max_iterations", path, claim_number
+            )
+            wall_clock_limit = integer_field(
+                claim_record, "wall_clock_limit", path, claim_number
+            )
+            if latest != "fail" or (
+                iterations < max_iterations and elapsed < wall_clock_limit
+            ):
+                raise LoopError(f"{path}:{number}: failing finish has no exhausted limit")
+        finishes[identifier] = record
+        outcomes.append(record)
+    active = [record for identifier, record in claims.items() if identifier not in finishes]
+    if len(active) > 1:
+        raise LoopError(f"{path}: multiple unfinished claims require recovery")
+    return outcomes, claims, verifications, finishes, active[0] if active else None
+
+
+def integer_field(record: dict, field: str, path: Path, number: int) -> int:
+    value = record.get(field, 0)
+    if isinstance(value, bool):
+        raise LoopError(f"{path}:{number}: {field} must be an integer")
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError) as error:
+        raise LoopError(f"{path}:{number}: {field} must be an integer") from error
+    if parsed < 0:
+        raise LoopError(f"{path}:{number}: {field} must not be negative")
+    return parsed
+
+
+def remaining_period_budget(fields: dict, outcomes: list[dict], now: datetime) -> Optional[int]:
+    if "period_budget" not in fields:
+        return None
+    path = log_path(fields)
+    window_start = now - timedelta(seconds=fields["period"])
+    consumed = 0
+    for number, record in enumerate(outcomes, 1):
+        if parse_timestamp(record, path, number) >= window_start:
+            consumed += integer_field(record, "wall_clock_used", path, number)
+    return max(0, fields["period_budget"] - consumed)
+
+
+def precheck(fields: dict) -> Optional[dict]:
+    if fields["status"] != "active":
+        return {"decision": "skip", "reason": f"status {fields['status']}"}
+    return None
+
+
+def skip_when_matched(fields: dict) -> Optional[dict]:
+    if "skip_when" in fields:
+        exit_code, _ = run_shell(fields["skip_when"], fields["wall_clock"])
+        if exit_code == 0:
+            return {"decision": "skip", "reason": "skip_when matched"}
+    return None
+
+
+def admission(fields: dict, records: list[dict], now: datetime) -> dict:
+    outcomes, _claims, _verifications, _finishes, active = log_state(fields, records)
+    if active is not None:
+        line = records.index(active) + 1
+        started = parse_timestamp(active, log_path(fields), line)
+        wall_clock_limit = integer_field(
+            active, "wall_clock_limit", log_path(fields), line
+        )
+        elapsed = max(0, int((now - started).total_seconds()))
+        expired = elapsed >= wall_clock_limit
+        result = {
+            "decision": "skip",
+            "reason": (
+                f"claim {active['claim']} is expired"
+                if expired
+                else "another loop pass is still active"
+            ),
+            "recovery": {
+                "elapsed_seconds": min(elapsed, wall_clock_limit),
+                "wall_clock_limit": wall_clock_limit,
+                "recoverable": expired,
+            },
+        }
+        if expired:
+            result["claim"] = active["claim"]
+        return result
+    path = log_path(fields)
+    timestamps = [
+        parse_timestamp(record, path, number)
+        for number, record in enumerate(outcomes, 1)
+    ]
+    remaining_cooldown = 0
+    if "cooldown" in fields and timestamps:
+        elapsed = (now - max(timestamps)).total_seconds()
+        remaining_cooldown = max(0, int(fields["cooldown"] - elapsed))
+        if remaining_cooldown > 0:
+            return {
+                "decision": "skip",
+                "reason": f"cooldown active ({remaining_cooldown}s remaining)",
+            }
+    remaining_budget = remaining_period_budget(fields, outcomes, now)
+    if remaining_budget == 0:
+        return {"decision": "skip", "reason": "period budget exhausted"}
+    return {
+        "decision": "run",
+        "reason": "gates green",
+        "remaining": {
+            "cooldown_seconds": remaining_cooldown,
+            "period_budget_seconds": remaining_budget,
+            "wall_clock_seconds": fields["wall_clock"],
+            "max_iterations": fields["max_iterations"],
+        },
+    }
+
+
+def command_check(fields: dict) -> dict:
+    skipped = precheck(fields)
+    if skipped is not None:
+        return skipped
+    decision = admission(fields, load_log(fields), datetime.now(timezone.utc))
+    if "skip_when" in fields:
+        decision["skip_when"] = "not run by check; claim runs it"
+    return decision
+
+
+def command_claim(fields: dict) -> dict:
+    skipped = precheck(fields) or skip_when_matched(fields)
+    if skipped is not None:
+        return skipped
+    with log_lock(fields):
+        now = datetime.now(timezone.utc)
+        decision = admission(fields, load_log(fields), now)
+        if decision["decision"] != "run":
+            return decision
+        remaining_budget = decision["remaining"]["period_budget_seconds"]
+        wall_clock_limit = fields["wall_clock"]
+        if remaining_budget is not None:
+            wall_clock_limit = min(wall_clock_limit, remaining_budget)
+        identifier = uuid.uuid4().hex
+        append_log(
+            fields,
+            {
+                "event": CLAIM_EVENT,
+                "claim": identifier,
+                "timestamp": now.isoformat(),
+                "loop": fields["loop"],
+                "trigger": fields["trigger"],
+                "max_iterations": fields["max_iterations"],
+                "wall_clock_limit": wall_clock_limit,
+            },
+        )
+        decision["claim"] = identifier
+        decision["remaining"]["wall_clock_seconds"] = wall_clock_limit
+        return decision
+
+
+def run_shell(command: str, timeout: float) -> tuple[Optional[int], str]:
+    """Run one spec command; a hung command counts as failed (exit None)."""
+    try:
+        completed = subprocess.run(
+            command, shell=True, capture_output=True, text=True, check=False, timeout=timeout
+        )
+    except subprocess.TimeoutExpired as expired:
+        output = (expired.stdout or b"").decode(errors="replace") + (expired.stderr or b"").decode(errors="replace")
+        return None, f"{output}\ntimed out at the admitted wall-clock deadline"
+    return completed.returncode, completed.stdout + completed.stderr
+
+
+def command_verify(fields: dict, claim_identifier: Optional[str]) -> dict:
+    if not claim_identifier:
+        raise LoopError("verify requires --claim")
+    with log_lock(fields):
+        records = load_log(fields)
+        _outcomes, claims, verifications, _finishes, active = log_state(
+            fields, records
+        )
+        claim = claims.get(claim_identifier)
+        if claim is None:
+            raise LoopError(f"unknown claim {claim_identifier}")
+        if active is None or active.get("claim") != claim_identifier:
+            raise LoopError(f"claim {claim_identifier} is not the unfinished claim")
+        claim_line = records.index(claim) + 1
+        prior = verifications.get(claim_identifier, [])
+        max_iterations = integer_field(
+            claim, "max_iterations", log_path(fields), claim_line
+        )
+        if len(prior) > max_iterations:
+            raise LoopError(f"claim {claim_identifier} has exhausted max_iterations")
+        if prior and prior[-1]["result"] == "pass":
+            raise LoopError(f"claim {claim_identifier} already verified pass")
+        started = parse_timestamp(claim, log_path(fields), claim_line)
+        wall_clock_limit = integer_field(
+            claim, "wall_clock_limit", log_path(fields), claim_line
+        )
+        remaining = (
+            started
+            + timedelta(seconds=wall_clock_limit)
+            - datetime.now(timezone.utc)
+        ).total_seconds()
+        if remaining <= 0:
+            raise LoopError(f"claim {claim_identifier} has exhausted its wall clock")
+        deadline = time.monotonic() + remaining
+        failures = []
+        for command in fields["verify"]:
+            command_budget = deadline - time.monotonic()
+            if command_budget <= 0:
+                failures.append(
+                    {
+                        "command": command,
+                        "exit_code": None,
+                        "output_tail": "admitted wall-clock deadline exhausted before command",
+                    }
+                )
+                break
+            exit_code, output = run_shell(command, command_budget)
+            if exit_code != 0:
+                lines = output.splitlines()
+                failures.append(
+                    {
+                        "command": command,
+                        "exit_code": exit_code,
+                        "output_tail": "\n".join(lines[-OUTPUT_TAIL_LINES:]),
+                    }
+                )
+        result = "fail" if failures else "pass"
+        sequence = len(prior) + 1
+        iteration = max(0, sequence - 1)
+        append_log(
+            fields,
+            {
+                "event": VERIFY_EVENT,
+                "claim": claim_identifier,
+                "timestamp": datetime.now(timezone.utc).isoformat(),
+                "sequence": sequence,
+                "iteration": iteration,
+                "result": result,
+                "commands": len(fields["verify"]),
+            },
+        )
+    payload = {
+        "result": result,
+        "claim": claim_identifier,
+        "commands": len(fields["verify"]),
+        "iteration": iteration,
+    }
+    if failures:
+        payload["failures"] = failures
+    return payload
+
+
+def command_finish(fields: dict, arguments: argparse.Namespace) -> dict:
+    if not arguments.claim:
+        raise LoopError("finish requires --claim")
+    with log_lock(fields):
+        records = load_log(fields)
+        _outcomes, claims, verifications, finishes, active = log_state(fields, records)
+        claim = claims.get(arguments.claim)
+        if claim is None:
+            raise LoopError(f"unknown claim {arguments.claim}")
+        completed = finishes.get(arguments.claim)
+        if completed is not None:
+            candidate = {
+                "result": arguments.result,
+                "iterations": arguments.iterations,
+                "rescue": arguments.rescue,
+                "notes": arguments.notes,
+            }
+            if any(completed.get(key) != value for key, value in candidate.items()):
+                raise LoopError(f"claim {arguments.claim} is already finished differently")
+            return {
+                "recorded": completed["result"],
+                "claim": arguments.claim,
+                "log": str(log_path(fields)),
+                "reason": "already finished",
+            }
+        if active is None or active.get("claim") != arguments.claim:
+            raise LoopError(f"claim {arguments.claim} is not the unfinished claim")
+        claim_line = records.index(claim) + 1
+        started = parse_timestamp(claim, log_path(fields), claim_line)
+        max_iterations = integer_field(
+            claim, "max_iterations", log_path(fields), claim_line
+        )
+        wall_clock_limit = integer_field(
+            claim, "wall_clock_limit", log_path(fields), claim_line
+        )
+        verification_records = verifications.get(arguments.claim, [])
+        iterations = max(0, len(verification_records) - 1)
+        if arguments.iterations != iterations:
+            raise LoopError(
+                f"--iterations must match helper-observed fix cycles "
+                f"({arguments.iterations} != {iterations})"
+            )
+        now = datetime.now(timezone.utc)
+        elapsed = max(0, math.ceil((now - started).total_seconds()))
+        if elapsed > wall_clock_limit and arguments.result not in {"fail", "blocked"}:
+            raise LoopError(
+                f"claim {arguments.claim} exhausted its wall clock; "
+                "finish with fail or blocked"
+            )
+        latest_result = verification_records[-1]["result"] if verification_records else None
+        if arguments.result == "pass" and latest_result != "pass":
+            raise LoopError("pass requires a recorded passing verify")
+        if latest_result == "pass" and arguments.result != "pass":
+            raise LoopError("a recorded passing verify must finish as pass")
+        if arguments.result == "fail" and (
+            latest_result != "fail"
+            or (iterations < max_iterations and elapsed < wall_clock_limit)
+        ):
+            raise LoopError("fail requires a failed verify and an exhausted limit")
+        record = {
+            "event": FINISH_EVENT,
+            "claim": arguments.claim,
+            "timestamp": now.isoformat(),
+            "loop": fields["loop"],
+            "trigger": fields["trigger"],
+            "iterations": iterations,
+            "result": arguments.result,
+            "wall_clock_used": elapsed,
+            "rescue": arguments.rescue,
+            "notes": arguments.notes,
+        }
+        append_log(fields, record)
+    return {
+        "recorded": record["result"],
+        "claim": arguments.claim,
+        "log": str(log_path(fields)),
+    }
+
+
+def command_recover(fields: dict, claim_identifier: Optional[str]) -> dict:
+    """Close one expired unfinished claim without caller-reported telemetry."""
+
+    if not claim_identifier:
+        raise LoopError("recover requires --claim")
+    with log_lock(fields):
+        records = load_log(fields)
+        _outcomes, claims, verifications, finishes, active = log_state(fields, records)
+        claim = claims.get(claim_identifier)
+        if claim is None:
+            raise LoopError(f"unknown claim {claim_identifier}")
+        completed = finishes.get(claim_identifier)
+        verification_records = verifications.get(claim_identifier, [])
+        latest = verification_records[-1]["result"] if verification_records else None
+        recovered_result = latest if latest in {"pass", "fail"} else "blocked"
+        recovery_notes = f"recovered unfinished claim as {recovered_result}"
+        if completed is not None:
+            if (
+                completed.get("result") != recovered_result
+                or completed.get("notes") != recovery_notes
+            ):
+                raise LoopError(f"claim {claim_identifier} is already finished differently")
+            return {
+                "recorded": recovered_result,
+                "claim": claim_identifier,
+                "log": str(log_path(fields)),
+                "reason": "already recovered",
+            }
+        if active is None or active.get("claim") != claim_identifier:
+            raise LoopError(f"claim {claim_identifier} is not the unfinished claim")
+        claim_line = records.index(claim) + 1
+        started = parse_timestamp(claim, log_path(fields), claim_line)
+        wall_clock_limit = integer_field(
+            claim, "wall_clock_limit", log_path(fields), claim_line
+        )
+        now = datetime.now(timezone.utc)
+        elapsed_raw = max(0.0, (now - started).total_seconds())
+        if elapsed_raw < wall_clock_limit:
+            raise LoopError(f"claim {claim_identifier} has not expired")
+        iterations = max(0, len(verification_records) - 1)
+        record = {
+            "event": FINISH_EVENT,
+            "claim": claim_identifier,
+            "timestamp": now.isoformat(),
+            "loop": fields["loop"],
+            "trigger": fields["trigger"],
+            "iterations": iterations,
+            "result": recovered_result,
+            "wall_clock_used": math.ceil(elapsed_raw),
+            "rescue": False,
+            "notes": recovery_notes,
+        }
+        append_log(fields, record)
+    return {
+        "recorded": recovered_result,
+        "claim": claim_identifier,
+        "iterations": iterations,
+        "log": str(log_path(fields)),
+    }
+
+
+def command_status(fields: dict) -> dict:
+    records = load_log(fields)
+    outcomes, _claims, _verifications, _finishes, active = log_state(fields, records)
+    counts = {result: 0 for result in RESULTS}
+    wall_clock_used = 0
+    human_rescues = 0
+    path = log_path(fields)
+    for number, record in enumerate(outcomes, 1):
+        result = record.get("result")
+        if result in counts:
+            counts[result] += 1
+        wall_clock_used += integer_field(record, "wall_clock_used", path, number)
+        if record.get("rescue"):
+            human_rescues += 1
+    accepted = counts["pass"]
+    evaluated = accepted + counts["fail"]
+    return {
+        "loop": fields["loop"],
+        "runs": len(outcomes),
+        "accepted": accepted,
+        "rejected": counts["fail"],
+        "blocked": counts["blocked"],
+        "skipped": counts["skipped"],
+        "human_rescues": human_rescues,
+        "wall_clock_used_seconds": wall_clock_used,
+        "acceptance_rate": round(accepted / evaluated, 4) if evaluated else None,
+        "wall_clock_per_accept_seconds": (
+            round(wall_clock_used / accepted, 2) if accepted else None
+        ),
+        "active_claim": active["claim"] if active is not None else None,
+    }
+
+
+def parser() -> argparse.ArgumentParser:
+    result = argparse.ArgumentParser(description=__doc__)
+    result.add_argument(
+        "command", choices=("check", "claim", "verify", "finish", "recover", "status")
+    )
+    result.add_argument("--spec", type=Path, required=True)
+    result.add_argument("--claim")
+    result.add_argument("--result", choices=RESULTS)
+    result.add_argument("--iterations", type=int, default=0)
+    result.add_argument("--notes", default="")
+    result.add_argument("--rescue", action="store_true")
+    return result
+
+
+def main(argv: Optional[Sequence[str]] = None) -> int:
+    arguments = parser().parse_args(argv)
+    try:
+        fields = parse_spec(arguments.spec)
+        if arguments.command == "check":
+            payload = command_check(fields)
+        elif arguments.command == "claim":
+            payload = command_claim(fields)
+        elif arguments.command == "verify":
+            payload = command_verify(fields, arguments.claim)
+        elif arguments.command == "recover":
+            payload = command_recover(fields, arguments.claim)
+        elif arguments.command == "finish":
+            if arguments.result is None:
+                raise LoopError(f"{arguments.command} requires --result")
+            payload = command_finish(fields, arguments)
+        else:
+            payload = command_status(fields)
+    except LoopError as error:
+        print(str(error), file=sys.stderr)
+        return 2
+    print(json.dumps(payload, sort_keys=True))
+    if arguments.command == "verify" and payload["result"] == "fail":
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())