atlas-iac/services/hermes/scripts/cli_lane_runner.py
2026-08-16 09:19:23 -03:00

984 lines
36 KiB
Python

#!/usr/bin/env python3
"""Run durable Codex and Claude workers from Hermes' authoritative Kanban."""
from __future__ import annotations
import concurrent.futures
import json
import os
import re
import selectors
import signal
import subprocess
import sys
import threading
import time
import uuid
import urllib.error
import urllib.request
from dataclasses import asdict, dataclass
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Callable
DATA_ROOT = Path(os.environ.get("HERMES_HOME", "/opt/data"))
SWITCHYARD_URL = os.environ.get(
"HERMES_SWITCHYARD_URL",
"http://hermes-switchyard.hermes.svc.cluster.local:9005/v1/chat/completions",
)
STATE_ROOT = DATA_ROOT / "cli-lanes"
CODEX_BIN = DATA_ROOT / "tools/bin/codex"
CLAUDE_BIN = DATA_ROOT / "tools/bin/claude"
CLAUDE_SETTINGS = DATA_ROOT / "home/.claude/settings.json"
RESULT_SCHEMA_PATH = STATE_ROOT / "worker-result.schema.json"
EFFORTS = ("low", "medium", "high", "xhigh")
EXTERNAL_PREFIX = "cli-"
DEFAULT_CLAIM_TTL = 7 * 24 * 60 * 60
DEFAULT_MAX_RUNTIME = 12 * 60 * 60
HEARTBEAT_SECONDS = 20
PROVIDER_HEALTH_MAX_AGE_SECONDS = 5 * 60
PROVIDER_HEALTH_PATHS = {
"codex": DATA_ROOT / "provider-health/codex.json",
"claude": DATA_ROOT / "provider-health/claude.json",
}
WORKTREE_LOCK = threading.Lock()
BOARD_CORRUPTION_ERRORS: dict[str, str] = {}
CAPACITY_PATTERN = re.compile(
r"(?:rate.?limit|capacity|overload|usage.?limit|quota|credit|exhaust|429|529|authentication|oauth|token.*expired)",
re.I,
)
NO_CLAUDE_SESSION = "No conversation found with session ID:"
CLAUDE_SESSION_COLLISION = "Session ID already in use"
NO_CODEX_THREAD = "no rollout found for thread id"
RESULT_SCHEMA: dict[str, Any] = {
"type": "object",
"additionalProperties": False,
"required": ["status", "summary", "changed_files", "tests_run", "artifacts", "blockers"],
"properties": {
"status": {"type": "string", "enum": ["completed", "blocked"]},
"summary": {"type": "string"},
"changed_files": {"type": "array", "items": {"type": "string"}},
"tests_run": {"type": "array", "items": {"type": "string"}},
"artifacts": {"type": "array", "items": {"type": "string"}},
"blockers": {"type": "array", "items": {"type": "string"}},
},
}
@dataclass(frozen=True)
class Route:
"""One provider/model/effort decision for an observable worker boundary."""
provider: str
model: str
effort: str
profile: str
classifier: str
reason: str
latency_ms: int
fallback_chain: tuple[str, ...]
@dataclass
class ProcessResult:
"""Captured outcome from one provider CLI invocation."""
returncode: int
output: str
structured: dict[str, Any] | None
capacity_failure: bool
def utc_now() -> str:
return datetime.now(timezone.utc).isoformat()
def atomic_json(path: Path, value: dict[str, Any], mode: int = 0o600) -> None:
"""Durably replace a small non-secret state document."""
path.parent.mkdir(parents=True, exist_ok=True)
temporary = path.with_name(f".{path.name}.{os.getpid()}.tmp")
temporary.write_text(json.dumps(value, indent=2, sort_keys=True) + "\n", encoding="utf-8")
temporary.chmod(mode)
os.replace(temporary, path)
def load_json(path: Path) -> dict[str, Any]:
try:
value = json.loads(path.read_text(encoding="utf-8"))
except (OSError, json.JSONDecodeError):
return {}
return value if isinstance(value, dict) else {}
def parse_assignee(assignee: str) -> tuple[str | None, str | None]:
"""Parse external lane overrides while leaving cli-auto fully automatic."""
value = str(assignee or "").strip().lower()
if value == "cli-auto":
return None, None
match = re.fullmatch(r"cli-(codex|claude)-(low|medium|high|xhigh)", value)
if not match:
raise ValueError(f"unsupported external lane assignee: {assignee}")
return match.group(1), match.group(2)
def fresh_unavailable_provider(now: float | None = None) -> str | None:
"""Return one recently proven-down provider for automatic route exclusion."""
current = time.time() if now is None else now
unavailable: list[str] = []
for provider, path in PROVIDER_HEALTH_PATHS.items():
health = load_json(path)
try:
age = current - path.stat().st_mtime
except OSError:
continue
if (
0 <= age <= PROVIDER_HEALTH_MAX_AGE_SECONDS
and health.get("state") == "unavailable"
):
unavailable.append(provider)
# If both providers are down, keep the normal attempt/fallback path so the
# card records authoritative current errors instead of trusting snapshots.
return unavailable[0] if len(unavailable) == 1 else None
def _decode_worker_target(value: str) -> tuple[str, str, str]:
"""Decode the selected model header emitted by a worker decision route."""
parts = value.split("/", 3)
if len(parts) != 4 or parts[0] != "worker":
raise RuntimeError(f"invalid Switchyard worker target: {value}")
provider, model, effort = parts[1:]
if provider not in {"codex", "claude"} or effort not in EFFORTS:
raise RuntimeError(f"unsupported Switchyard worker target: {value}")
return provider, model, effort
def select_route(
prompt: str,
assignee: str,
*,
exclude_provider: str | None = None,
exclude_reason: str | None = None,
switchyard_url: str = SWITCHYARD_URL,
open_request: Callable[..., Any] = urllib.request.urlopen,
) -> Route:
"""Ask Switchyard to select one native CLI worker at this boundary."""
started = time.monotonic()
manual_provider, manual_effort = parse_assignee(assignee)
if manual_provider and manual_effort:
route_id = f"atlas/worker/manual/{manual_provider}/{manual_effort}"
source = "switchyard-manual"
else:
route_id = "atlas/worker/auto/maximum"
source = "switchyard-classifier"
context = prompt
if exclude_provider:
reason = exclude_reason or "failed or exhausted capacity at this boundary"
context += (
f"\n\nRouting constraint: the {exclude_provider} provider {reason}. "
"Do not select it."
)
payload = json.dumps(
{
"model": route_id,
"messages": [{"role": "user", "content": context}],
"stream": False,
"max_tokens": 1,
}
).encode("utf-8")
request = urllib.request.Request(
switchyard_url,
data=payload,
headers={"Content-Type": "application/json"},
method="POST",
)
try:
with open_request(request, timeout=60) as response:
selected = str(response.headers.get("x-model-router-selected-model") or "")
rationale = str(response.headers.get("x-model-router-rationale") or "")
response_body = response.read()
except (OSError, urllib.error.URLError) as exc:
raise RuntimeError(f"Switchyard worker routing failed: {exc}") from exc
provider, model, effort = _decode_worker_target(selected)
# Switchyard preserves the stable tier target in the selection header and
# top-level response model. The worker broker's assistant content contains
# the steward-resolved provider model required by the native CLI.
try:
response_document = json.loads(response_body)
resolved_target = str(response_document["choices"][0]["message"]["content"] or "")
resolved_provider, resolved_model, resolved_effort = _decode_worker_target(resolved_target)
required_prefix = "gpt-" if provider == "codex" else "claude-"
if (
resolved_provider == provider
and resolved_effort == effort
and resolved_model.startswith(required_prefix)
):
model = resolved_model
except (AttributeError, IndexError, KeyError, RuntimeError, TypeError, ValueError, json.JSONDecodeError):
pass
if exclude_provider and provider == exclude_provider:
alternate = "claude" if provider == "codex" else "codex"
guarded = select_route(
prompt,
f"cli-{alternate}-{effort}",
switchyard_url=switchyard_url,
open_request=open_request,
)
return Route(
provider=guarded.provider,
model=guarded.model,
effort=guarded.effort,
profile=guarded.profile,
classifier=f"{source}-health-guard",
reason=(
f"Switchyard selected excluded {exclude_provider}; preserved "
f"its {effort} effort on healthy-provider route. {guarded.reason}"
),
latency_ms=int((time.monotonic() - started) * 1000),
fallback_chain=(),
)
return Route(
provider=provider,
model=model,
effort=effort,
profile=f"{provider}-{effort}",
classifier=source,
reason=rationale or f"Switchyard selected {selected}",
latency_ms=int((time.monotonic() - started) * 1000),
fallback_chain=(),
)
def state_path(board: str, task_id: str) -> Path:
safe_board = re.sub(r"[^a-zA-Z0-9_.-]+", "-", board)
safe_task = re.sub(r"[^a-zA-Z0-9_.-]+", "-", task_id)
return STATE_ROOT / safe_board / f"{safe_task}.json"
def build_prompt(context: str, workspace: Path, handoff: str = "") -> str:
"""Create a bounded worker contract with an explicit machine-readable result."""
return f"""You are a durable coding worker managed by Hermes Kanban.
Work only on this objective and its acceptance criteria:
{context}
Workspace: {workspace}
{handoff}
Operate autonomously inside the workspace. Inspect before editing, preserve unrelated user changes, run proportionate tests, and do not claim completion without evidence. You have owner-level Kubernetes access in every namespace. Prefer Flux-tracked manifests for durable changes, but use kubectl, Flux, exec, port-forwarding, rollout operations, and existing Vault workflows when the objective or incident requires them. Persist any desired-state mutation back to Git. Do not force-push, hard-reset, clean untracked files, or expose credentials.
Return a final JSON object matching the supplied schema. Use status=blocked only for a concrete unresolved blocker. List changed files, tests run, durable artifact paths, and blockers explicitly.
"""
def git_handoff(workspace: Path, prior_output: str) -> str:
"""Build an explicit cross-provider handoff without transferring hidden state."""
def read_git(*args: str) -> str:
completed = subprocess.run(
["git", "-C", str(workspace), *args],
text=True,
capture_output=True,
timeout=30,
check=False,
)
return (completed.stdout or completed.stderr).strip()[-12000:]
return (
"\nCross-provider handoff from a failed or exhausted worker:\n"
f"Git status:\n{read_git('status', '--short', '--branch')}\n\n"
f"Current diff summary:\n{read_git('diff', '--stat')}\n\n"
f"Prior worker tail:\n{prior_output[-10000:]}\n"
"Reinspect the workspace and verify all inherited claims before continuing."
)
def workspace_artifacts(workspace: Path, values: Any) -> list[str]:
"""Return existing regular artifacts contained by the task worktree."""
if not isinstance(values, list):
return []
root = workspace.resolve()
artifacts: list[str] = []
for value in values:
if not isinstance(value, str) or not value.strip():
continue
candidate = Path(value.strip()).expanduser()
if not candidate.is_absolute():
candidate = root / candidate
try:
resolved = candidate.resolve(strict=True)
resolved.relative_to(root)
except (OSError, RuntimeError, ValueError):
continue
if resolved.is_file():
artifacts.append(str(resolved))
return list(dict.fromkeys(artifacts))
def _extract_json(value: Any) -> dict[str, Any] | None:
if isinstance(value, dict) and value.get("status") in {"completed", "blocked"}:
return value
if not isinstance(value, str):
return None
candidates = [value]
match = re.search(r"\{.*\}", value, re.DOTALL)
if match:
candidates.append(match.group(0))
for candidate in candidates:
try:
parsed = json.loads(candidate)
except json.JSONDecodeError:
continue
if isinstance(parsed, dict) and parsed.get("status") in {"completed", "blocked"}:
return parsed
return None
def _event_payload(provider: str, line: str, state: dict[str, Any], state_file: Path) -> dict[str, Any] | None:
"""Persist provider session identifiers before interpreting final output."""
try:
event = json.loads(line)
except json.JSONDecodeError:
return None
if not isinstance(event, dict):
return None
if provider == "codex" and event.get("type") == "thread.started":
thread_id = event.get("thread_id") or event.get("thread", {}).get("id")
if thread_id:
state["codex_thread_id"] = str(thread_id)
atomic_json(state_file, state)
if provider == "claude" and event.get("session_id"):
state["claude_session_id"] = str(event["session_id"])
atomic_json(state_file, state)
for key in ("structured_output", "result", "output", "text"):
parsed = _extract_json(event.get(key))
if parsed:
return parsed
item = event.get("item")
if isinstance(item, dict):
for key in ("text", "content"):
parsed = _extract_json(item.get(key))
if parsed:
return parsed
return None
def stream_process(
command: list[str],
*,
provider: str,
cwd: Path,
env: dict[str, str],
log_path: Path,
state: dict[str, Any],
state_file: Path,
heartbeat: Callable[[str], bool],
max_runtime: int,
) -> ProcessResult:
"""Stream JSONL to Kanban logs while maintaining the authoritative lease."""
process = subprocess.Popen(
command,
cwd=cwd,
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
bufsize=1,
start_new_session=True,
)
assert process.stdout is not None
selector = selectors.DefaultSelector()
selector.register(process.stdout, selectors.EVENT_READ)
started = time.monotonic()
last_heartbeat = 0.0
lines: list[str] = []
structured: dict[str, Any] | None = None
log_path.parent.mkdir(parents=True, exist_ok=True)
forced_failure = ""
with log_path.open("a", encoding="utf-8") as log:
log.write(f"\n[{utc_now()}] starting {provider} worker\n")
log.flush()
while process.poll() is None:
now = time.monotonic()
if now - started > max_runtime:
_terminate_worker_process(process)
forced_failure = "worker exceeded its maximum runtime"
lines.append(forced_failure + "\n")
break
if now - last_heartbeat >= HEARTBEAT_SECONDS:
if not heartbeat(f"{provider} worker active for {round(now - started)}s"):
_terminate_worker_process(process)
forced_failure = "Kanban lease was lost; provider process terminated"
lines.append(forced_failure + "\n")
break
last_heartbeat = now
for key, _ in selector.select(timeout=1.0):
line = key.fileobj.readline()
if not line:
continue
lines.append(line)
if len(lines) > 4000:
lines = lines[-4000:]
log.write(line)
log.flush()
parsed = _event_payload(provider, line, state, state_file)
structured = parsed or structured
_terminate_worker_process(process)
remainder = process.stdout.read()
if remainder:
lines.append(remainder)
log.write(remainder)
for line in remainder.splitlines():
parsed = _event_payload(provider, line, state, state_file)
structured = parsed or structured
if forced_failure:
log.write(forced_failure + "\n")
log.write(f"\n[{utc_now()}] {provider} exit={process.returncode}\n")
selector.close()
output = "".join(lines)[-100000:]
returncode = int(process.returncode if process.returncode is not None else 1)
return ProcessResult(
returncode=returncode,
output=output,
structured=structured,
capacity_failure=returncode != 0 and bool(CAPACITY_PATTERN.search(output)),
)
def _terminate_worker_process(process: subprocess.Popen[str]) -> None:
"""Reap a worker and every subprocess in its isolated process group."""
if process.poll() is None:
try:
os.killpg(process.pid, signal.SIGTERM)
except ProcessLookupError:
pass
try:
process.wait(timeout=10)
except subprocess.TimeoutExpired:
try:
os.killpg(process.pid, signal.SIGKILL)
except ProcessLookupError:
pass
process.wait(timeout=10)
# A provider can exit before one of its terminal or language-server children.
# The process group remains addressable after its leader exits, so reap those
# children as well before a retry is allowed to own the same worktree.
try:
os.killpg(process.pid, signal.SIGKILL)
except ProcessLookupError:
pass
def _base_env() -> dict[str, str]:
"""Build a worker environment while preserving Vault-backed CLI homes."""
env = os.environ.copy()
env.update(
{
"HOME": str(DATA_ROOT / "home"),
"CODEX_HOME": os.environ.get("CODEX_HOME", "/runtime-access/codex"),
"CLAUDE_CONFIG_DIR": os.environ.get(
"CLAUDE_CONFIG_DIR", "/runtime-access/claude"
),
"GIT_TERMINAL_PROMPT": "0",
}
)
env["PATH"] = f"{DATA_ROOT / 'tools/bin'}:/opt/coordinator:" + env.get("PATH", "")
return env
def _codex_command(route: Route, prompt: str, workspace: Path, state: dict[str, Any], result_file: Path) -> list[str]:
base = [
str(CODEX_BIN),
"--dangerously-bypass-approvals-and-sandbox",
"--dangerously-bypass-hook-trust",
"exec",
]
thread_id = str(state.get("codex_thread_id") or "")
common = [
"--json",
"-m",
route.model,
"-c",
f'model_reasoning_effort="{route.effort}"',
"--output-schema",
str(RESULT_SCHEMA_PATH),
"-o",
str(result_file),
]
if thread_id:
return [*base, "resume", *common, thread_id, prompt]
return [
*base,
*common,
"-C",
str(workspace),
prompt,
]
def _claude_command(route: Route, prompt: str, state: dict[str, Any], resume: bool) -> list[str]:
session_id = str(state["claude_session_id"])
session = ["--resume", session_id] if resume else ["--session-id", session_id]
denied = [
"Bash(git push --force *)",
"Bash(git reset --hard *)",
"Bash(git clean -f *)",
]
return [
str(CLAUDE_BIN),
"--dangerously-skip-permissions",
"--autocompact",
"auto",
"--settings",
str(CLAUDE_SETTINGS),
"--disallowedTools",
*denied,
"--model",
route.model,
"--effort",
route.effort,
"--output-format",
"stream-json",
"--verbose",
"--json-schema",
json.dumps(RESULT_SCHEMA, separators=(",", ":")),
*session,
"-p",
prompt,
]
def run_provider(
route: Route,
prompt: str,
workspace: Path,
state: dict[str, Any],
state_file: Path,
log_path: Path,
heartbeat: Callable[[str], bool],
max_runtime: int,
) -> ProcessResult:
"""Run or resume one provider session using its pinned structured CLI."""
state.setdefault("attempts", []).append({"route": asdict(route), "started_at": utc_now()})
if route.provider == "claude" and not state.get("claude_session_id"):
# The reservation is durable before Claude starts, closing the crash gap.
state["claude_session_id"] = str(uuid.uuid4())
state["current_route"] = asdict(route)
state["updated_at"] = utc_now()
atomic_json(state_file, state)
env = _base_env()
result_file = state_file.with_suffix(".result.json")
if route.provider == "codex":
result_file.unlink(missing_ok=True)
command = _codex_command(route, prompt, workspace, state, result_file)
result = stream_process(
command,
provider="codex",
cwd=workspace,
env=env,
log_path=log_path,
state=state,
state_file=state_file,
heartbeat=heartbeat,
max_runtime=max_runtime,
)
if (
state.get("codex_thread_id")
and result.returncode != 0
and NO_CODEX_THREAD in result.output.lower()
):
state.pop("codex_thread_id", None)
atomic_json(state_file, state)
result_file.unlink(missing_ok=True)
result = stream_process(
_codex_command(route, prompt, workspace, state, result_file),
provider="codex",
cwd=workspace,
env=env,
log_path=log_path,
state=state,
state_file=state_file,
heartbeat=heartbeat,
max_runtime=max_runtime,
)
file_result = load_json(result_file)
if file_result.get("status") in {"completed", "blocked"}:
result.structured = file_result
return result
resume = bool(state.get("claude_started"))
result = stream_process(
_claude_command(route, prompt, state, resume),
provider="claude",
cwd=workspace,
env=env,
log_path=log_path,
state=state,
state_file=state_file,
heartbeat=heartbeat,
max_runtime=max_runtime,
)
if resume and result.returncode == 1 and NO_CLAUDE_SESSION in result.output:
result = stream_process(
_claude_command(route, prompt, state, False),
provider="claude",
cwd=workspace,
env=env,
log_path=log_path,
state=state,
state_file=state_file,
heartbeat=heartbeat,
max_runtime=max_runtime,
)
elif not resume and result.returncode == 1 and CLAUDE_SESSION_COLLISION in result.output:
result = stream_process(
_claude_command(route, prompt, state, True),
provider="claude",
cwd=workspace,
env=env,
log_path=log_path,
state=state,
state_file=state_file,
heartbeat=heartbeat,
max_runtime=max_runtime,
)
if result.returncode == 0 or NO_CLAUDE_SESSION not in result.output:
state["claude_started"] = True
atomic_json(state_file, state)
return result
def _task_value(task: Any, name: str, default: Any = None) -> Any:
return getattr(task, name, default)
def _task_context(kanban_db: Any, conn: Any, task_id: str) -> str:
value = kanban_db.build_worker_context(conn, task_id)
if isinstance(value, str):
return value
return json.dumps(value, indent=2, default=str)
def _resolve_workspace(kanban_db: Any, conn: Any, task: Any, board: str) -> Path:
# External coding lanes always receive their own linked worktree. A task
# that does not resolve to a Git repository is blocked rather than sharing
# a mutable checkout with another unattended worker.
# Git worktree creation updates common-repository metadata. Serialize that
# short materialization step while allowing the provider workers themselves
# to run concurrently in independent worktrees.
with WORKTREE_LOCK:
value, branch_name = kanban_db._resolve_worktree_workspace(task, board=board)
kanban_db.set_branch_name(conn, str(_task_value(task, "id")), branch_name)
return Path(value).resolve()
def execute_claim(board: str, task_id: str) -> None:
"""Execute one already-claimed task and commit its outcome to Kanban."""
from hermes_cli import kanban_db
with kanban_db.scoped_current_board(board):
conn = kanban_db.connect(board=board)
task = kanban_db.get_task(conn, task_id)
if task is None:
return
run_id = _task_value(task, "current_run_id", None)
assignee = str(_task_value(task, "assignee", "cli-auto") or "cli-auto")
try:
state_file = state_path(board, task_id)
state = load_json(state_file)
state.update(
{
"board": board,
"task_id": task_id,
"run_id": run_id,
"assignee": assignee,
}
)
atomic_json(state_file, state)
log_path = Path(kanban_db.worker_log_path(task_id, board=board))
workspace = _resolve_workspace(kanban_db, conn, task, board)
kanban_db.set_workspace_path(conn, task_id, str(workspace))
context = _task_context(kanban_db, conn, task_id)
except Exception as error:
kanban_db.block_task(
conn,
task_id,
reason=f"Direct CLI lane preparation failed: {type(error).__name__}: {error}",
kind="capability",
expected_run_id=run_id,
)
conn.close()
return
def heartbeat(note: str) -> bool:
return bool(
kanban_db.heartbeat_worker(
conn,
task_id,
note=note,
expected_run_id=run_id,
)
)
try:
previous_route = state.get("current_route")
excluded_provider = (
fresh_unavailable_provider() if assignee == "cli-auto" else None
)
route = select_route(
context,
assignee,
exclude_provider=excluded_provider,
exclude_reason="is unavailable according to fresh native health"
if excluded_provider
else None,
)
if excluded_provider:
kanban_db.add_comment(
conn,
task_id,
"cli-lane-runner",
f"Provider health guard excluded {excluded_provider} before automatic routing.",
)
kanban_db.add_comment(
conn,
task_id,
"cli-lane-runner",
f"CLI route: {route.provider}/{route.model} at {route.effort}; classifier={route.classifier}; {route.reason}",
)
resume_handoff = ""
if (
isinstance(previous_route, dict)
and previous_route.get("provider")
and previous_route.get("provider") != route.provider
):
try:
previous_output = log_path.read_text(encoding="utf-8")[-10000:]
except OSError:
previous_output = "Previous provider log was unavailable after restart."
resume_handoff = git_handoff(workspace, previous_output)
kanban_db.add_comment(
conn,
task_id,
"cli-lane-runner",
f"Restart-time provider change: {previous_route.get('provider')} -> {route.provider}; explicit workspace handoff attached.",
)
prompt = build_prompt(context, workspace, resume_handoff)
max_runtime = int(
_task_value(task, "max_runtime_seconds", 0) or DEFAULT_MAX_RUNTIME
)
result = run_provider(route, prompt, workspace, state, state_file, log_path, heartbeat, max_runtime)
if result.capacity_failure:
retry_context = (
context
+ "\n\nRouting boundary: the first provider failed from capacity/authentication. "
+ "Select the alternate hosted provider at an appropriate effort."
)
alternate = "claude" if route.provider == "codex" else "codex"
fallback = select_route(
retry_context,
f"cli-{alternate}-{route.effort}",
)
kanban_db.add_comment(
conn,
task_id,
"cli-lane-runner",
f"Provider fallback: {route.provider} -> {fallback.provider}; Jetson reclassified the retry boundary.",
)
result = run_provider(
fallback,
build_prompt(context, workspace, git_handoff(workspace, result.output)),
workspace,
state,
state_file,
log_path,
heartbeat,
max_runtime,
)
route = fallback
structured = result.structured
if structured:
structured = dict(structured)
structured["artifacts"] = workspace_artifacts(
workspace,
structured.get("artifacts"),
)
metadata = {
"executor": "direct-cli-lane",
"provider": route.provider,
"model": route.model,
"effort": route.effort,
"classifier": route.classifier,
"state_file": str(state_file),
"codex_thread_id": state.get("codex_thread_id"),
"claude_session_id": state.get("claude_session_id"),
}
if structured:
for key in ("changed_files", "tests_run", "artifacts", "blockers"):
value = structured.get(key)
metadata[key] = value if isinstance(value, list) else []
if structured and structured.get("status") == "completed" and result.returncode == 0:
kanban_db.complete_task(
conn,
task_id,
result=json.dumps(structured, sort_keys=True),
summary=str(structured.get("summary") or "Completed"),
metadata=metadata,
expected_run_id=run_id,
)
else:
reason = (
"; ".join(str(item) for item in (structured or {}).get("blockers", []))
if structured
else result.output[-4000:]
)
kanban_db.block_task(
conn,
task_id,
reason=reason or f"{route.provider} worker failed with exit {result.returncode}",
kind="transient" if result.capacity_failure else "capability",
expected_run_id=run_id,
)
except Exception as error:
kanban_db.block_task(
conn,
task_id,
reason=f"Direct CLI lane failed: {type(error).__name__}: {error}",
kind="capability",
expected_run_id=run_id,
)
finally:
conn.close()
def _board_slug(board: Any) -> str:
if isinstance(board, dict):
return str(board.get("slug") or board.get("id") or "")
return str(getattr(board, "slug", None) or getattr(board, "id", None) or board)
def _external(task: Any) -> bool:
return str(_task_value(task, "assignee", "") or "").startswith(EXTERNAL_PREFIX)
def _connect_healthy_board(kanban_db: Any, board: str) -> Any | None:
"""Open one board without letting localized corruption stop other lanes."""
try:
conn = kanban_db.connect(board=board)
except getattr(kanban_db, "KanbanDbCorruptError", ()) as error:
detail = str(error)
if BOARD_CORRUPTION_ERRORS.get(board) != detail:
print(
f"quarantining corrupt Kanban board {board!r}: {detail}",
file=sys.stderr,
flush=True,
)
BOARD_CORRUPTION_ERRORS[board] = detail
return None
BOARD_CORRUPTION_ERRORS.pop(board, None)
return conn
def recover_orphans() -> None:
"""Return external running tasks to ready after a runner/pod restart."""
from hermes_cli import kanban_db
for raw_board in kanban_db.list_boards(include_archived=False):
board = _board_slug(raw_board)
if not board:
continue
with kanban_db.scoped_current_board(board):
conn = _connect_healthy_board(kanban_db, board)
if conn is None:
continue
try:
for task in kanban_db.list_tasks(conn):
if _external(task) and str(_task_value(task, "status", "")) == "running":
kanban_db.reclaim_task(
conn,
str(_task_value(task, "id")),
reason="direct CLI lane restarted; provider session will resume",
)
finally:
conn.close()
def claim_ready(active: set[tuple[str, str]], limit: int) -> list[tuple[str, str]]:
"""Atomically claim external ready tasks across all non-archived boards."""
from hermes_cli import kanban_db
claimed: list[tuple[str, str]] = []
if limit <= 0:
return claimed
for raw_board in kanban_db.list_boards(include_archived=False):
board = _board_slug(raw_board)
if not board:
continue
with kanban_db.scoped_current_board(board):
conn = _connect_healthy_board(kanban_db, board)
if conn is None:
continue
try:
kanban_db.recompute_ready(conn)
for task in kanban_db.list_tasks(conn):
task_id = str(_task_value(task, "id", ""))
assignee = str(_task_value(task, "assignee", "") or "")
if task_id and not assignee and str(_task_value(task, "status", "")) == "ready":
if kanban_db.assign_task(conn, task_id, "cli-auto"):
task = kanban_db.get_task(conn, task_id)
assignee = "cli-auto"
if (
not task_id
or (board, task_id) in active
or not assignee.startswith(EXTERNAL_PREFIX)
or str(_task_value(task, "status", "")) != "ready"
):
continue
try:
result = kanban_db.claim_task(
conn,
task_id,
ttl_seconds=DEFAULT_CLAIM_TTL,
claimer="direct-cli-lane",
)
except Exception:
continue
if result is not None:
claimed.append((board, task_id))
if len(claimed) >= limit:
return claimed
finally:
conn.close()
return claimed
def main() -> int:
"""Continuously bridge external Kanban lanes to provider CLIs."""
RESULT_SCHEMA_PATH.parent.mkdir(parents=True, exist_ok=True)
atomic_json(RESULT_SCHEMA_PATH, RESULT_SCHEMA, 0o644)
recover_orphans()
workers = max(1, min(int(os.environ.get("HERMES_CLI_LANE_CONCURRENCY", "4")), 8))
futures: dict[concurrent.futures.Future[None], tuple[str, str]] = {}
with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as pool:
while True:
for future in list(futures):
if future.done():
try:
future.result()
except Exception as error:
print(f"worker future failed: {error}", file=sys.stderr, flush=True)
del futures[future]
active = set(futures.values())
for board, task_id in claim_ready(active, workers - len(futures)):
future = pool.submit(execute_claim, board, task_id)
futures[future] = (board, task_id)
time.sleep(5)
return 0
if __name__ == "__main__":
raise SystemExit(main())