6da0cba96d
Some checks failed
Tests / Declarative: Post Actions testing.tests.test_repo_structure.test_knowledge_service_mirror_matches_source failed
The Hermes stage proved only that a pod was running, when the substantive evidence is the JSON Hermes returned; it now reads that from the audit trail, so classification, confidence and requested action are on screen. The policy gates stage never showed the verdict it was describing; authorized and authorize_reason now appear with it. The chart paths were paraphrases. They are now the node labels from mermaid/TestAutomation.mmd verbatim - Response check, Scoped repair guard, Action authorizer, Action registry, Scoped ConfigMap repair, Action result - so a line on screen can be found on the diagram. The checklist carries them too. The Optional source proposal branch was silently absent rather than explained. The route stage now states it was not taken and why: a fixture repair is an operational action, and source fixes are not actions at all. Also widened the log window and its timeout; the fetch occasionally exceeded 25s and dropped a poll. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
332 lines
13 KiB
Python
Executable File
332 lines
13 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""Narrate the triage flow live, showing every command it runs.
|
|
|
|
Run in a second terminal beside the demo. When a stage of
|
|
`mermaid/TestAutomation.mmd` is reached this prints, in order:
|
|
|
|
the stage banner, naming the chart subgraph
|
|
the service UI to look at, if one changes at that stage
|
|
each command, echoed before it runs, then its output
|
|
|
|
Echoing the commands is the point. An audience watching a dashboard has to
|
|
take the result on trust; watching `kubectl` run against the cluster and
|
|
reading the raw answer is the difference between a demonstration and an
|
|
assertion.
|
|
|
|
Two honest limits. This reports at subgraph granularity, not per node: the
|
|
chart draws the collector, response check, marker check and authorizer
|
|
separately, and a refusal here names the gate rather than walking all twelve.
|
|
And it follows one incident on one job, whereas the chart describes the whole
|
|
system.
|
|
|
|
Read-only. Every command below is a read; nothing here changes the cluster.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import os
|
|
import subprocess
|
|
import sys
|
|
import time
|
|
from datetime import datetime, timezone
|
|
|
|
JOB = os.environ.get("MONITOR_JOB", "hermes-triage-demo")
|
|
NS_DEMO = "hermes-triage-demo"
|
|
NS_ARIADNE = "maintenance"
|
|
JENKINS = os.environ.get("JENKINS_URL", "https://ci.bstein.dev")
|
|
GITEA = os.environ.get("GITEA_URL", "https://scm.bstein.dev")
|
|
GRAFANA = os.environ.get("GRAFANA_URL", "https://metrics.bstein.dev")
|
|
HERMES_UI = os.environ.get("HERMES_URL", "https://agent.bstein.dev")
|
|
POLL_SECONDS = 6
|
|
|
|
BOLD, DIM, GREEN, CYAN, YELLOW, RED, RESET = (
|
|
"\033[1m", "\033[2m", "\033[32m", "\033[36m", "\033[33m", "\033[31m", "\033[0m"
|
|
)
|
|
|
|
# stage key -> (chart subgraph, what it means, chart path, UI worth showing)
|
|
STAGES: dict[str, tuple[str, str, str, str]] = {
|
|
"detect": (
|
|
"Detect and gather",
|
|
"A failed build becomes one incident",
|
|
"Jenkins detector -> Evidence sources",
|
|
f"{JENKINS}/job/{JOB}/ — the red build",
|
|
),
|
|
"evidence": (
|
|
"Detect and gather",
|
|
"Ariadne assembles the bounded evidence bundle",
|
|
"Console reader -> Failure ranker -> Context filter -> Incident bundle",
|
|
"",
|
|
),
|
|
"hermes": (
|
|
"Hermes analysis",
|
|
"Hermes returns a recommendation it cannot act on",
|
|
"Triage skills -> Structured recommendation",
|
|
f"{HERMES_UI} — the agent run appears here",
|
|
),
|
|
"gates": (
|
|
"Ariadne policy gates",
|
|
"Ariadne decides on its own reading of the evidence",
|
|
"Response check -> Scoped repair guard -> Action authorizer",
|
|
"",
|
|
),
|
|
"route": (
|
|
"Ariadne response",
|
|
"The policy result, and which branch it opens",
|
|
"Policy result -> Action registry",
|
|
"",
|
|
),
|
|
"response": (
|
|
"Ariadne response",
|
|
"Ariadne executes, proposes, or escalates",
|
|
"Action registry -> Scoped ConfigMap repair -> Action result",
|
|
"",
|
|
),
|
|
"verify": (
|
|
"Ariadne response",
|
|
"One rebuild decides whether the repair held",
|
|
"Action result -> validation build",
|
|
f"{JENKINS}/job/{JOB}/ — a new build starts on its own",
|
|
),
|
|
"outputs": (
|
|
"Inspectable outputs",
|
|
"Incident closed; the artifacts remain",
|
|
"Ariadne records every outcome -> Audit events, Triage metrics",
|
|
f"{GRAFANA}/d/atlas-testing — triage panels",
|
|
),
|
|
}
|
|
ORDER = ["detect", "evidence", "hermes", "gates", "route", "response", "verify", "outputs"]
|
|
|
|
|
|
def stamp() -> str:
|
|
return datetime.now(timezone.utc).strftime("%H:%M:%S")
|
|
|
|
|
|
def run(cmd: list[str], *, show: bool = True, limit: int = 12) -> str:
|
|
"""Echo a command, run it, print its output, and return stdout."""
|
|
|
|
if show:
|
|
print(f" {DIM}${RESET} {CYAN}{' '.join(cmd)}{RESET}")
|
|
try:
|
|
# Ariadne emits ~200 lines per tick, so a ten-minute window is a
|
|
# few thousand lines; the fetch needs room to finish.
|
|
done = subprocess.run(cmd, capture_output=True, text=True, timeout=60)
|
|
out = done.stdout.strip()
|
|
except Exception as exc: # a failed read must never stop the narration
|
|
print(f" {RED}(command failed: {exc}){RESET}")
|
|
return ""
|
|
if show:
|
|
for line in (out.split("\n")[:limit] or ["(no output)"]):
|
|
print(f" {line[:150]}")
|
|
print()
|
|
return out
|
|
|
|
|
|
def quiet(cmd: list[str]) -> str:
|
|
return run(cmd, show=False)
|
|
|
|
|
|
def banner(key: str) -> None:
|
|
subgraph, meaning, path, ui = STAGES[key]
|
|
print(f"\n{GREEN}{'─' * 70}{RESET}")
|
|
print(f"{GREEN}{BOLD} {stamp()} {subgraph}{RESET}{GREEN} — {meaning}{RESET}")
|
|
print(f"{DIM} chart: {path}{RESET}")
|
|
if ui:
|
|
print(f"{YELLOW} look at: {ui}{RESET}")
|
|
print(f"{GREEN}{'─' * 70}{RESET}")
|
|
|
|
|
|
def ariadne_records(since: str = "10m") -> list[dict]:
|
|
"""Return recent Ariadne log records that parse as JSON.
|
|
|
|
Windowed by time, not line count. Ariadne emits roughly two hundred lines
|
|
per tick once it is polling every job, so a --tail window wide enough to
|
|
be useful is impossible to guess and a narrow one silently drops ticks.
|
|
"""
|
|
|
|
raw = quiet(
|
|
["kubectl", "-n", NS_ARIADNE, "logs", "deploy/ariadne", "-c", "ariadne", f"--since={since}"]
|
|
)
|
|
records = []
|
|
for line in raw.split("\n"):
|
|
line = line.strip()
|
|
if line.startswith("{"):
|
|
try:
|
|
records.append(json.loads(line))
|
|
except ValueError:
|
|
continue
|
|
return records
|
|
|
|
|
|
def job_states(records: list[dict]) -> list[dict]:
|
|
"""Return every tick result for this job in the window, oldest first.
|
|
|
|
The stages are reconstructed from all of them rather than from the newest
|
|
alone. A tick reporting the repair is replaced by the next tick within
|
|
seconds, so reading only the latest state means a poll landing at the
|
|
wrong moment loses that step permanently.
|
|
"""
|
|
|
|
states = []
|
|
for record in records:
|
|
if record.get("event") == "hermes_autotriage" and record.get("jobs"):
|
|
try:
|
|
state = json.loads(record["jobs"]).get(JOB)
|
|
except ValueError:
|
|
continue
|
|
if state:
|
|
states.append(state)
|
|
return states
|
|
|
|
|
|
_DIAG_QUERY = (
|
|
". /vault/secrets/ariadne-env.sh >/dev/null 2>&1; python3 -c \""
|
|
"import json,os,psycopg;"
|
|
"conn=psycopg.connect(os.environ['ARIADNE_DATABASE_URL']);"
|
|
"cur=conn.cursor();"
|
|
"cur.execute(\\\"select detail from ariadne_events where event_type='hermes_autotriage_diagnosis'"
|
|
" order by id desc limit 1\\\");"
|
|
"r=cur.fetchone()[0];"
|
|
"d=r if isinstance(r,dict) else json.loads(r);"
|
|
"print(json.dumps({'incident':d.get('incident_id'),'authorized':d.get('authorized'),"
|
|
"'authorize_reason':d.get('authorize_reason'),'evidence_marker':d.get('evidence_marker'),"
|
|
"'outcome':d.get('outcome')},indent=1))\""
|
|
)
|
|
|
|
|
|
def diagnosis_event() -> None:
|
|
"""Show the diagnosis Ariadne stored: what Hermes said, and the verdict."""
|
|
|
|
run(["kubectl", "-n", NS_ARIADNE, "exec", "deploy/ariadne", "-c", "ariadne", "--",
|
|
"sh", "-c", _DIAG_QUERY], limit=16)
|
|
|
|
|
|
def evidence_for(key: str) -> None:
|
|
"""Run the reads that show this stage actually happened."""
|
|
|
|
if key == "detect":
|
|
run(["kubectl", "-n", NS_DEMO, "get", "cm", "hermes-triage-demo-fixture",
|
|
"-o", "jsonpath={.data.state}"])
|
|
elif key == "evidence":
|
|
print(f" {DIM}the bundle is what Ariadne sends Hermes; the audit trail keeps it{RESET}\n")
|
|
run(["kubectl", "-n", NS_ARIADNE, "exec", "deploy/ariadne", "-c", "ariadne", "--",
|
|
"printenv", "ARIADNE_HERMES_AUTOTRIAGE_JOB_ALLOWLIST"], limit=2)
|
|
elif key == "hermes":
|
|
print(f" {DIM}what Hermes actually returned, as Ariadne stored it:{RESET}")
|
|
diagnosis_event()
|
|
print(f" {DIM}Hermes holds no Git or Kubernetes write access; this JSON is its"
|
|
f" entire output{RESET}\n")
|
|
elif key == "gates":
|
|
print(f" {DIM}authorized/authorize_reason above is the policy result: the verdict"
|
|
f" Ariadne reached on its own reading{RESET}")
|
|
run(["kubectl", "-n", NS_ARIADNE, "exec", "deploy/ariadne", "-c", "ariadne", "--",
|
|
"printenv", "ARIADNE_HERMES_ALLOWED_ACTIONS"], limit=2)
|
|
run(["kubectl", "-n", NS_ARIADNE, "exec", "deploy/ariadne", "-c", "ariadne", "--",
|
|
"printenv", "ARIADNE_HERMES_MIN_CONFIDENCE"], limit=2)
|
|
print(f" {DIM}only these two ids may ever be executed. Source fixes are not actions:"
|
|
f" they become pull requests and are never executed by Ariadne.{RESET}\n")
|
|
elif key == "route":
|
|
print(f" {DIM}the fixture repair is an operational action, so the Optional source"
|
|
f" proposal branch is not taken for this incident{RESET}\n")
|
|
elif key in {"response", "outputs"}:
|
|
run(["kubectl", "-n", NS_DEMO, "get", "cm", "hermes-triage-demo-fixture",
|
|
"-o", "jsonpath={.data.state}"])
|
|
if key == "response":
|
|
print(f" {DIM}Scoped ConfigMap repair on the chart: one merge patch to one"
|
|
f" ConfigMap, in process, no pod created{RESET}\n")
|
|
if key == "outputs":
|
|
print(f" {YELLOW}issues filed by triage: {GITEA}/bstein/ariadne/issues{RESET}\n")
|
|
elif key == "verify":
|
|
print(f" {DIM}exactly one rebuild is triggered; it never retries in a loop{RESET}\n")
|
|
|
|
|
|
class Monitor:
|
|
"""Track which flow-chart stage the current incident has reached."""
|
|
|
|
def __init__(self) -> None:
|
|
self.done: set[str] = set()
|
|
self.incident = ""
|
|
self.summarised = False
|
|
self.last_tick = ""
|
|
|
|
def mark(self, key: str, evidence: str) -> None:
|
|
if key in self.done:
|
|
return
|
|
self.done.add(key)
|
|
banner(key)
|
|
print(f" {BOLD}what happened:{RESET} {evidence}\n")
|
|
evidence_for(key)
|
|
|
|
def new_incident(self, incident: str) -> None:
|
|
if incident and incident != self.incident:
|
|
self.incident = incident
|
|
self.done.clear()
|
|
self.summarised = False
|
|
print(f"\n{BOLD}{YELLOW}══ incident {incident} ══{RESET}")
|
|
|
|
def checklist(self) -> None:
|
|
print(f"\n{BOLD} flow chart progress{RESET}")
|
|
for key in ORDER:
|
|
tick = f"{GREEN}✓{RESET}" if key in self.done else f"{DIM}·{RESET}"
|
|
print(f" {tick} {key:<9} {STAGES[key][0]:<22} {DIM}{STAGES[key][2]}{RESET}")
|
|
print()
|
|
|
|
|
|
def main() -> None:
|
|
print(f"{BOLD}Hermes triage monitor — following mermaid/TestAutomation.mmd{RESET}")
|
|
print(f"{DIM}Every command is echoed before it runs. Read-only. Ctrl-C to stop.{RESET}")
|
|
print(f"{YELLOW}Jenkins: {JENKINS}/job/{JOB}/{RESET}")
|
|
print(f"{YELLOW}Gitea: {GITEA}/bstein{RESET}")
|
|
print(f"{YELLOW}Grafana: {GRAFANA}/d/atlas-testing{RESET}\n")
|
|
monitor = Monitor()
|
|
while True:
|
|
states = job_states(ariadne_records())
|
|
# Replay every tick in the window, so a step that lasted one tick is
|
|
# never lost to poll timing. mark() is idempotent.
|
|
for state in states:
|
|
status = str(state.get("status") or "")
|
|
monitor.new_incident(str(state.get("incident_id") or ""))
|
|
if status in {"human_required", "awaiting_rebuild", "deduped", "failed"}:
|
|
monitor.mark("detect", f"incident {monitor.incident} opened from a terminal failure")
|
|
if status == "awaiting_rebuild":
|
|
monitor.mark("evidence", "bundle collected: console regions, tests, logs")
|
|
monitor.mark("hermes", "diagnosis returned and parsed against the frozen schema")
|
|
monitor.mark("gates", "every gate passed; the predefined action was authorized")
|
|
monitor.mark(
|
|
"route",
|
|
"policy result: authorized -> Action registry (Optional source proposal not taken)",
|
|
)
|
|
monitor.mark(
|
|
"response",
|
|
f"Scoped ConfigMap repair: {state.get('repair', 'action')} on {state.get('target', '')}",
|
|
)
|
|
monitor.mark("verify", "one rebuild triggered with seeding disabled")
|
|
if status == "human_required":
|
|
monitor.mark("evidence", "bundle collected")
|
|
monitor.mark("hermes", "diagnosis returned")
|
|
monitor.mark("gates", f"Ariadne refused: {state.get('reason', '')} — nothing ran")
|
|
monitor.mark(
|
|
"route",
|
|
"policy result: refused -> Diagnosis and next checks -> Ariadne opens an issue",
|
|
)
|
|
monitor.mark("response", "escalated; issue filed in the service repository")
|
|
if status == "healthy" and state.get("resolved") and not monitor.summarised:
|
|
monitor.mark("verify", "rebuild finished green")
|
|
monitor.mark("outputs", f"resolved: {', '.join(state['resolved'])}")
|
|
monitor.checklist()
|
|
monitor.summarised = True
|
|
status = str(states[-1].get("status") or "") if states else ""
|
|
|
|
if status and status != monitor.last_tick:
|
|
monitor.last_tick = status
|
|
print(f"{DIM}{stamp()} tick: {status}{RESET}")
|
|
sys.stdout.flush()
|
|
time.sleep(POLL_SECONDS)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
try:
|
|
main()
|
|
except KeyboardInterrupt:
|
|
print("\nstopped")
|