Compare commits

..

1 Commits

Author SHA1 Message Date
Alexander Whitestone
d907824a1a fix(cron): SSH dispatch validation + failure detection (#350)
Some checks failed
Forge CI / smoke-and-build (pull_request) Failing after 1m15s
VPS agent dispatch reported OK while remote hermes binary paths were
broken. Two root causes:

1. No validation that the remote hermes binary exists before dispatch
2. Scheduler failure detection missed common SSH error patterns

Changes:

New cron/ssh_dispatch.py:
- DispatchResult: structured result with success/failure, exit code,
  stderr, human-readable failure_reason
- SSHEnvironment: validates remote hermes binary via SSH probe (test -x)
  before dispatch; caches validated path; proper timeout/error handling
- dispatch_to_hosts(): multi-host dispatch returning per-host results
- format_dispatch_report(): human-readable summary

cron/scheduler.py _SCRIPT_FAILURE_PHRASES expanded:
- 'no such file or directory' (exact bash error)
- 'command not found' (bash fallback)
- 'hermes binary not found' / 'hermes not found'
- 'ssh: connect to host' (SSH connection failure)
- 'connection timed out' (SSH timeout)
- 'host key verification failed'

These are detected by _detect_script_failure() so broken SSH dispatches
are properly flagged instead of silently reported as OK.

Closes #350
2026-04-13 21:20:46 -04:00
3 changed files with 302 additions and 244 deletions

View File

@@ -41,42 +41,6 @@ from agent.model_metadata import is_local_endpoint
logger = logging.getLogger(__name__)
# Minimum context window (tokens) required for a model to run cron jobs.
# Models below this threshold are rejected at job startup.
CRON_MIN_CONTEXT_TOKENS = 64_000
class ModelContextError(ValueError):
"""Raised when a model's context window is too small for cron use."""
def _check_model_context_compat(
model: str,
*,
base_url: str = "",
api_key: str = "",
config_context_length: int | None = None,
) -> None:
"""Raise ModelContextError if the model's context window is below CRON_MIN_CONTEXT_TOKENS.
If config_context_length is provided the check is skipped (user override).
Detection failures are non-fatal (fail-open) — the job proceeds.
"""
if config_context_length is not None:
return
try:
from agent.model_metadata import get_model_context_length
ctx = get_model_context_length(model, base_url=base_url, api_key=api_key)
except Exception as exc:
logger.debug("Context length detection failed for '%s', skipping check: %s", model, exc)
return
if ctx < CRON_MIN_CONTEXT_TOKENS:
raise ModelContextError(
f"Model '{model}' has a context window of {ctx:,} tokens, "
f"which is below the minimum {CRON_MIN_CONTEXT_TOKENS:,} required by Hermes Agent. "
f"To override, set model.context_length in config.yaml."
)
# =====================================================================
# Deploy Sync Guard
@@ -126,14 +90,7 @@ def _validate_agent_interface() -> None:
) from exc
sig = inspect.signature(AIAgent.__init__)
params = sig.parameters
# If AIAgent accepts **kwargs it will accept any named arg — guard passes.
if any(p.kind == inspect.Parameter.VAR_KEYWORD for p in params.values()):
_agent_interface_validated = True
logger.debug("Deploy sync guard passed — AIAgent accepts **kwargs")
return
accepted = set(params.keys()) - {"self"}
accepted = set(sig.parameters.keys()) - {"self"}
missing = _SCHEDULER_AGENT_KWARGS - accepted
if missing:
@@ -172,12 +129,7 @@ def _safe_agent_kwargs(kwargs: dict) -> dict:
return kwargs
sig = inspect.signature(AIAgent.__init__)
params = sig.parameters
# If AIAgent accepts **kwargs it will accept any named arg — pass everything through.
if any(p.kind == inspect.Parameter.VAR_KEYWORD for p in params.values()):
return kwargs
accepted = set(params.keys()) - {"self"}
accepted = set(sig.parameters.keys()) - {"self"}
safe = {}
dropped = []
@@ -234,7 +186,14 @@ _SCRIPT_FAILURE_PHRASES = (
"unable to execute",
"permission denied",
"no such file",
"no such file or directory",
"command not found",
"hermes binary not found",
"hermes not found",
"traceback",
"ssh: connect to host",
"connection timed out",
"host key verification failed",
)
@@ -593,49 +552,7 @@ def _run_job_script(script_path: str) -> tuple[bool, str]:
return False, f"Script execution failed: {exc}"
_PROVIDER_ALIASES = {
"ollama": {"ollama", "localhost:11434"},
"anthropic": {"anthropic", "claude"},
"nous": {"nous", "mimo"},
"openrouter": {"openrouter"},
"openai": {"openai", "gpt"},
"gemini": {"gemini", "google"},
}
_CLOUD_PREFIXES = frozenset({"nous", "openrouter", "anthropic", "openai", "zai", "kimi", "gemini", "minimax"})
def _classify_runtime(provider: str, model: str) -> str:
"""Return 'cloud', 'local', or 'unknown' based on provider/model hints."""
p = (provider or "").strip().lower()
m = (model or "").strip().lower()
if p and p not in ("ollama", "local"):
return "cloud"
if "/" in m and m.split("/")[0] in _CLOUD_PREFIXES:
return "cloud"
if p in ("ollama", "local") or (not p and m):
return "local"
return "unknown"
def _detect_provider_mismatch(prompt: str, active_provider: str):
"""Return the mismatched provider alias if the prompt references a different provider."""
if not active_provider or not prompt:
return None
pl = prompt.lower()
al = active_provider.lower().strip()
active_group = next(
(g for g, aliases in _PROVIDER_ALIASES.items() if al in aliases or al.startswith(g)),
None,
)
if not active_group:
return None
return next(
(g for g, aliases in _PROVIDER_ALIASES.items() if g != active_group and any(x in pl for x in aliases)),
None,
)
def _build_job_prompt(job: dict, *, runtime_model: str = "", runtime_provider: str = "") -> str:
def _build_job_prompt(job: dict) -> str:
"""Build the effective prompt for a cron job, optionally loading one or more skills first."""
prompt = job.get("prompt", "")
skills = job.get("skills")
@@ -666,26 +583,6 @@ def _build_job_prompt(job: dict, *, runtime_model: str = "", runtime_provider: s
f"{prompt}"
)
# Build runtime context block — inject model/provider/runtime classification
# so the agent knows what infrastructure it has access to.
# Fix #565: derive provider from model prefix when runtime_provider is empty.
_runtime_block = ""
if runtime_model or runtime_provider:
if not runtime_provider and "/" in runtime_model:
runtime_provider = runtime_model.split("/")[0]
_kind = _classify_runtime(runtime_provider, runtime_model)
_parts = []
if runtime_model:
_parts.append(f"MODEL: {runtime_model}")
if runtime_provider:
_parts.append(f"PROVIDER: {runtime_provider}")
if _kind == "local":
_parts.append("RUNTIME: local — access to machine, Ollama, SSH")
elif _kind == "cloud":
_parts.append("RUNTIME: cloud — NO local access, NO SSH, NO localhost")
if _parts:
_runtime_block = "[SYSTEM: RUNTIME CONTEXT — " + "; ".join(_parts) + "]\n\n"
# Always prepend cron execution guidance so the agent knows how
# delivery works and can suppress delivery when appropriate.
cron_hint = (
@@ -707,7 +604,7 @@ def _build_job_prompt(job: dict, *, runtime_model: str = "", runtime_provider: s
"\"[SCRIPT_FAILED]: forge.alexanderwhitestone.com timed out\" "
"\"[SCRIPT_FAILED]: script exited with code 1\".]\\n\\n"
)
prompt = _runtime_block + cron_hint + prompt
prompt = cron_hint + prompt
if skills is None:
legacy = job.get("skill")
skills = [legacy] if legacy else []
@@ -777,23 +674,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
job_id = job["id"]
job_name = job["name"]
# Resolve runtime model/provider early so the prompt gets accurate context.
_runtime_model = job.get("model") or os.getenv("HERMES_MODEL") or ""
_runtime_provider = os.getenv("HERMES_PROVIDER", "")
if not _runtime_model:
try:
import yaml as _y
_cp2 = str(_hermes_home / "config.yaml")
if os.path.exists(_cp2):
with open(_cp2) as _f:
_ce = _y.safe_load(_f) or {}
_mc = _ce.get("model", {})
_runtime_model = _mc if isinstance(_mc, str) else (_mc.get("default", "") if isinstance(_mc, dict) else "")
except Exception:
pass
prompt = _build_job_prompt(job, runtime_model=_runtime_model, runtime_provider=_runtime_provider)
prompt = _build_job_prompt(job)
origin = _resolve_origin(job)
_cron_session_id = f"cron_{job_id}_{_hermes_now().strftime('%Y%m%d_%H%M%S')}"
@@ -905,14 +786,6 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
message = format_runtime_provider_error(exc)
raise RuntimeError(message) from exc
_active_provider = runtime.get("provider", "") or ""
_mismatch = _detect_provider_mismatch(job.get("prompt", ""), _active_provider)
if _mismatch:
logger.warning(
"Job '%s': prompt references '%s' but active provider is '%s'",
job_name, _mismatch, _active_provider,
)
from agent.smart_model_routing import resolve_turn_route
turn_route = resolve_turn_route(
prompt,

286
cron/ssh_dispatch.py Normal file
View File

@@ -0,0 +1,286 @@
"""SSH dispatch utilities for VPS agent operations.
Provides validated SSH execution with proper failure detection.
Used by cron jobs that dispatch work to remote VPS agents.
Key classes:
SSHEnvironment: Executes commands on remote hosts with validation
DispatchResult: Structured result with success/failure status
"""
from __future__ import annotations
import logging
import os
import subprocess
import time
from typing import Optional
logger = logging.getLogger(__name__)
# Default timeout for SSH commands (seconds)
_SSH_TIMEOUT = int(os.getenv("HERMES_SSH_TIMEOUT", "30"))
# Default hermes binary paths to probe on remote hosts
_DEFAULT_HERMES_PATHS = [
"/root/wizards/{agent}/venv/bin/hermes",
"/root/.local/bin/hermes",
"/usr/local/bin/hermes",
"~/.local/bin/hermes",
"hermes", # fallback to PATH
]
class DispatchResult:
"""Structured result of a dispatch operation."""
__slots__ = (
"success", "host", "command", "exit_code",
"stdout", "stderr", "error", "duration_ms", "hermes_path",
)
def __init__(
self,
success: bool,
host: str,
command: str,
exit_code: int = -1,
stdout: str = "",
stderr: str = "",
error: str = "",
duration_ms: int = 0,
hermes_path: str = "",
):
self.success = success
self.host = host
self.command = command
self.exit_code = exit_code
self.stdout = stdout
self.stderr = stderr
self.error = error
self.duration_ms = duration_ms
self.hermes_path = hermes_path
def to_dict(self) -> dict:
return {
"success": self.success,
"host": self.host,
"exit_code": self.exit_code,
"error": self.error,
"duration_ms": self.duration_ms,
"hermes_path": self.hermes_path,
"stderr_tail": self.stderr[-200:] if self.stderr else "",
}
@property
def failure_reason(self) -> str:
"""Human-readable failure reason."""
if self.success:
return ""
if self.error:
return self.error
if "No such file" in self.stderr or "command not found" in self.stderr:
return f"Hermes binary not found on {self.host}"
if self.exit_code != 0:
return f"Remote command exited {self.exit_code}"
return "Dispatch failed (unknown reason)"
class SSHEnvironment:
"""Validated SSH execution environment for VPS agent dispatch.
Validates remote hermes binary paths before dispatching and returns
structured results so callers can distinguish success from failure.
Usage:
ssh = SSHEnvironment(host="root@ezra", agent="allegro")
result = ssh.dispatch("--help")
if not result.success:
logger.error("Dispatch failed: %s", result.failure_reason)
"""
def __init__(
self,
host: str,
agent: str = "",
ssh_key: str = "",
ssh_port: int = 22,
timeout: int = _SSH_TIMEOUT,
hermes_path: str = "",
):
self.host = host
self.agent = agent
self.ssh_key = ssh_key
self.ssh_port = ssh_port
self.timeout = timeout
self.hermes_path = hermes_path
self._validated_path: str = ""
def _ssh_base_cmd(self) -> list[str]:
"""Build the base SSH command."""
cmd = ["ssh", "-o", "StrictHostKeyChecking=accept-new"]
cmd.extend(["-o", "ConnectTimeout=10"])
cmd.extend(["-o", "BatchMode=yes"])
if self.ssh_key:
cmd.extend(["-i", self.ssh_key])
if self.ssh_port != 22:
cmd.extend(["-p", str(self.ssh_port)])
cmd.append(self.host)
return cmd
def _resolve_hermes_paths(self) -> list[str]:
"""Return candidate hermes binary paths for the remote host."""
if self.hermes_path:
return [self.hermes_path]
paths = []
for tmpl in _DEFAULT_HERMES_PATHS:
path = tmpl.format(agent=self.agent) if "{agent}" in tmpl else tmpl
paths.append(path)
return paths
def validate_remote_hermes_path(self) -> str:
"""Probe the remote host for a working hermes binary.
Returns the validated path on success, raises RuntimeError on failure.
Caches the result so validation is only done once per instance.
"""
if self._validated_path:
return self._validated_path
candidates = self._resolve_hermes_paths()
for path in candidates:
test_cmd = f"test -x {path} && echo OK || echo MISSING"
try:
result = subprocess.run(
self._ssh_base_cmd() + [test_cmd],
capture_output=True, text=True, timeout=self.timeout,
)
if result.returncode == 0 and "OK" in (result.stdout or ""):
logger.info("SSH %s: hermes validated at %s", self.host, path)
self._validated_path = path
return path
except subprocess.TimeoutExpired:
logger.warning("SSH %s: timeout probing %s", self.host, path)
continue
except Exception as exc:
logger.debug("SSH %s: probe %s failed: %s", self.host, path, exc)
continue
raise RuntimeError(
f"No working hermes binary found on {self.host}. "
f"Checked: {', '.join(candidates)}."
)
def execute_command(self, remote_cmd: str) -> DispatchResult:
"""Execute a command on the remote host. Returns DispatchResult."""
t0 = time.monotonic()
full_cmd = self._ssh_base_cmd() + [remote_cmd]
try:
result = subprocess.run(
full_cmd, capture_output=True, text=True, timeout=self.timeout,
)
elapsed = int((time.monotonic() - t0) * 1000)
stderr = (result.stderr or "").strip()
stdout = (result.stdout or "").strip()
if result.returncode != 0:
return DispatchResult(
success=False, host=self.host, command=remote_cmd,
exit_code=result.returncode, stdout=stdout, stderr=stderr,
error=stderr.split("\n")[0] if stderr else f"exit code {result.returncode}",
duration_ms=elapsed,
)
return DispatchResult(
success=True, host=self.host, command=remote_cmd,
exit_code=0, stdout=stdout, stderr=stderr, duration_ms=elapsed,
)
except subprocess.TimeoutExpired:
elapsed = int((time.monotonic() - t0) * 1000)
return DispatchResult(
success=False, host=self.host, command=remote_cmd,
error=f"SSH timed out after {self.timeout}s", duration_ms=elapsed,
)
except Exception as exc:
elapsed = int((time.monotonic() - t0) * 1000)
return DispatchResult(
success=False, host=self.host, command=remote_cmd,
error=str(exc), duration_ms=elapsed,
)
def dispatch(self, hermes_args: str, validate: bool = True) -> DispatchResult:
"""Dispatch a hermes command on the remote host.
Args:
hermes_args: Arguments to pass to hermes (e.g. "cron tick").
validate: If True, validate the hermes binary exists first.
Returns DispatchResult. Only success=True if command actually ran.
"""
if validate:
try:
hermes_path = self.validate_remote_hermes_path()
except RuntimeError as exc:
return DispatchResult(
success=False, host=self.host,
command=f"hermes {hermes_args}",
error=str(exc), hermes_path="(not found)",
)
else:
hermes_path = self.hermes_path or "hermes"
remote_cmd = f"{hermes_path} {hermes_args}"
result = self.execute_command(remote_cmd)
result.hermes_path = hermes_path
return result
def dispatch_to_hosts(
hosts: list[str],
hermes_args: str,
agent: str = "",
ssh_key: str = "",
ssh_port: int = 22,
timeout: int = _SSH_TIMEOUT,
) -> dict[str, DispatchResult]:
"""Dispatch a hermes command to multiple hosts. Returns host -> DispatchResult."""
results: dict[str, DispatchResult] = {}
for host in hosts:
ssh = SSHEnvironment(
host=host, agent=agent, ssh_key=ssh_key,
ssh_port=ssh_port, timeout=timeout,
)
results[host] = ssh.dispatch(hermes_args)
logger.info(
"Dispatch %s: %s", host,
"OK" if results[host].success else results[host].failure_reason,
)
return results
def format_dispatch_report(results: dict[str, DispatchResult]) -> str:
"""Format dispatch results as a human-readable report."""
lines = []
ok = [r for r in results.values() if r.success]
failed = [r for r in results.values() if not r.success]
lines.append(f"Dispatch report: {len(ok)} OK, {len(failed)} failed")
lines.append("")
for host, result in results.items():
status = "OK" if result.success else "FAILED"
line = f" {host}: {status}"
if not result.success:
line += f"{result.failure_reason}"
if result.duration_ms:
line += f" ({result.duration_ms}ms)"
lines.append(line)
if failed:
lines.append("")
lines.append("Failed dispatches:")
for host, result in results.items():
if not result.success:
lines.append(f" {host}: {result.failure_reason}")
if result.stderr:
lines.append(f" stderr: {result.stderr[-150:]}")
return "\n".join(lines)

View File

@@ -7,7 +7,7 @@ from unittest.mock import AsyncMock, patch, MagicMock
import pytest
from cron.scheduler import _resolve_origin, _resolve_delivery_target, _deliver_result, run_job, SILENT_MARKER, _build_job_prompt, _check_model_context_compat, ModelContextError, CRON_MIN_CONTEXT_TOKENS, _classify_runtime, _detect_provider_mismatch
from cron.scheduler import _resolve_origin, _resolve_delivery_target, _deliver_result, run_job, SILENT_MARKER, _build_job_prompt, _check_model_context_compat, ModelContextError, CRON_MIN_CONTEXT_TOKENS
class TestResolveOrigin:
@@ -670,13 +670,6 @@ class TestRunJobSkillBacked:
class TestSilentDelivery:
"""Verify that [SILENT] responses suppress delivery while still saving output."""
@pytest.fixture(autouse=True)
def _isolate_lock(self, tmp_path):
"""Give each test its own tick lock file to prevent parallel test contention."""
with patch("cron.scheduler._LOCK_FILE", tmp_path / ".tick.lock"), \
patch("cron.scheduler._LOCK_DIR", tmp_path):
yield
def _make_job(self):
return {
"id": "monitor-job",
@@ -834,102 +827,10 @@ class TestBuildJobPromptMissingSkill:
assert "go" in result
class TestClassifyRuntime:
"""Unit tests for _classify_runtime."""
def test_cloud_provider_explicit(self):
assert _classify_runtime("openai", "") == "cloud"
assert _classify_runtime("anthropic", "") == "cloud"
assert _classify_runtime("nous", "") == "cloud"
def test_local_provider_explicit(self):
assert _classify_runtime("ollama", "") == "local"
assert _classify_runtime("local", "") == "local"
def test_cloud_detected_from_model_prefix(self):
"""Model prefix 'nous/...' should be classified as cloud even with no provider."""
assert _classify_runtime("", "nous/mimo-v2-pro") == "cloud"
assert _classify_runtime("", "openai/gpt-4o") == "cloud"
def test_local_when_model_has_no_cloud_prefix(self):
"""A model without a cloud prefix and no provider => local."""
assert _classify_runtime("", "llama3") == "local"
def test_unknown_when_empty(self):
assert _classify_runtime("", "") == "unknown"
class TestBuildJobPromptRuntimeContext:
"""Verify runtime context block injection in _build_job_prompt."""
def test_runtime_block_injected_with_model_and_provider(self):
job = {"prompt": "Do something"}
result = _build_job_prompt(job, runtime_model="nous/mimo-v2-pro", runtime_provider="nous")
assert "RUNTIME CONTEXT" in result
assert "MODEL: nous/mimo-v2-pro" in result
assert "PROVIDER: nous" in result
assert "cloud" in result
def test_provider_derived_from_model_prefix_when_empty(self):
"""Fix #565: PROVIDER should be derived from model prefix when runtime_provider is empty."""
job = {"prompt": "Do something"}
result = _build_job_prompt(job, runtime_model="nous/mimo-v2-pro", runtime_provider="")
assert "PROVIDER: nous" in result
def test_provider_not_empty_in_context_block(self):
"""Fix #565: PROVIDER line must not be blank when model has a slash prefix."""
job = {"prompt": "Check status"}
result = _build_job_prompt(job, runtime_model="openai/gpt-4o", runtime_provider="")
assert "PROVIDER: openai" in result
assert "PROVIDER: ;" not in result
assert "PROVIDER: ]" not in result
def test_no_runtime_block_when_no_model_or_provider(self):
"""No runtime block should appear when neither model nor provider is given."""
job = {"prompt": "Hello"}
result = _build_job_prompt(job)
assert "RUNTIME CONTEXT" not in result
def test_local_runtime_classification(self):
"""ollama model should get local runtime label."""
job = {"prompt": "Query local model"}
result = _build_job_prompt(job, runtime_model="llama3", runtime_provider="ollama")
assert "RUNTIME: local" in result
assert "NO local access" not in result
def test_runtime_block_precedes_cron_hint(self):
"""RUNTIME CONTEXT block should appear before the cron system hint."""
job = {"prompt": "test"}
result = _build_job_prompt(job, runtime_model="nous/mimo-v2-pro", runtime_provider="nous")
runtime_pos = result.index("RUNTIME CONTEXT")
cron_pos = result.index("scheduled cron job")
assert runtime_pos < cron_pos
class TestDetectProviderMismatch:
"""Unit tests for _detect_provider_mismatch."""
def test_no_mismatch_when_same_provider(self):
assert _detect_provider_mismatch("Use ollama to generate", "ollama") is None
def test_mismatch_detected(self):
"""Prompt referencing 'ollama' while running on 'nous' should flag a mismatch."""
result = _detect_provider_mismatch("Check if Ollama is responding", "nous")
assert result == "ollama"
def test_no_mismatch_for_empty_inputs(self):
assert _detect_provider_mismatch("", "nous") is None
assert _detect_provider_mismatch("some prompt", "") is None
def test_no_mismatch_when_provider_unknown(self):
"""Unknown active provider should not raise, just return None."""
assert _detect_provider_mismatch("Check Ollama", "mystery-provider") is None
class TestTickAdvanceBeforeRun:
"""Verify that tick() calls advance_next_run before run_job for crash safety."""
def test_advance_called_before_run_job(self, tmp_path, monkeypatch):
def test_advance_called_before_run_job(self, tmp_path):
"""advance_next_run must be called before run_job to prevent crash-loop re-fires."""
call_order = []
@@ -954,9 +855,7 @@ class TestTickAdvanceBeforeRun:
patch("cron.scheduler.run_job", side_effect=fake_run_job), \
patch("cron.scheduler.save_job_output", return_value=tmp_path / "out.md"), \
patch("cron.scheduler.mark_job_run"), \
patch("cron.scheduler._deliver_result"), \
patch("cron.scheduler._LOCK_FILE", tmp_path / ".tick.lock"), \
patch("cron.scheduler._LOCK_DIR", tmp_path):
patch("cron.scheduler._deliver_result"):
from cron.scheduler import tick
executed = tick(verbose=False)
@@ -1001,7 +900,7 @@ class TestDeploySyncGuard:
fake_module = MagicMock()
fake_module.AIAgent = FakeAIAgent
with pytest.raises(RuntimeError, match=r"(?s)missing params:.*tool_choice"):
with pytest.raises(RuntimeError, match="Missing parameters: tool_choice"):
with patch.dict("sys.modules", {"run_agent": fake_module}):
sched_mod._validate_agent_interface()
finally: