fix: publish Hermes cost line immediately in post_llm_call (0.4.5)

The classic CLI's GatewayAgent._pump returns at message.complete and
_drain_stale discards later frames, so the previous settlement-ordered
publish (wait for the execution stamp to clear) was silently dropped
client-side even though the daemon logged it as published. Publish
immediately instead; the CLI queues it as a notice flushed after
run_conversation, still below the response.
This commit is contained in:
nvaldez
2026-11-03 17:07:55 +01:00
parent dd89d1b699
commit a5f92d4bbb
6 changed files with 42 additions and 58 deletions
+1 -1
View File
@@ -498,7 +498,7 @@ class TurnCompletionHandler(FileSystemEventHandler):
self._check_journal(event)
def _check_journal(self, event):
if not event.is_directory and "journal" in str(event.src_path) and event.src_path.endswith(".jsonl"):
if not event.is_directory and ("journal" in str(event.src_path) or "messages" in str(event.src_path)) and event.src_path.endswith(".jsonl"):
session_dir = Path(event.src_path).parent.parent
self._run_hooks(session_dir)
+15 -17
View File
@@ -317,6 +317,10 @@ class AgentCostEstimatorGUI:
def _run_tkinter(self, tk, ttk):
"""Run the GUI using Tkinter."""
# Store for use in other methods
self.tk = tk
self.ttk = ttk
root = tk.Tk()
root.title("Agent Cost Estimator")
root.geometry("900x600")
@@ -396,7 +400,6 @@ class AgentCostEstimatorGUI:
# Store root for refresh
self.root = root
self.tk = tk
# Initial load
self._refresh_sessions()
@@ -424,7 +427,7 @@ class AgentCostEstimatorGUI:
# Still add harness node even if no sessions
self.tree.insert(
"",
tk.END,
self.tk.END,
text=harness_name,
values=("", "", ""),
tags=("harness",)
@@ -433,7 +436,7 @@ class AgentCostEstimatorGUI:
# Add harness as parent
harness_node = self.tree.insert(
"",
tk.END,
self.tk.END,
text=harness_name,
values=("", "", ""),
tags=("harness",)
@@ -443,7 +446,7 @@ class AgentCostEstimatorGUI:
for session in harness_info.sessions:
self.tree.insert(
harness_node,
tk.END,
self.tk.END,
text=f"{session.session_id[:8]}..." if len(session.session_id) > 8
else session.session_id,
values=(
@@ -470,20 +473,18 @@ class AgentCostEstimatorGUI:
def _show_harness_dialog(self):
"""Show dialog to select default harness."""
import tkinter as tk
available = list(SessionDiscovery.HARNESS_ADAPTERS.keys())
current = self.settings.default_harness
harness_var = tk.StringVar(value=current)
harness_var = self.tk.StringVar(value=current)
dialog = tk.Toplevel(self.root)
dialog = self.tk.Toplevel(self.root)
dialog.title("Select Default Harness")
dialog.transient(self.root)
dialog.grab_set()
tk.Label(dialog, text="Default Harness:").pack(pady=5)
self.tk.Label(dialog, text="Default Harness:").pack(pady=5)
combo = ttk.Combobox(
combo = self.ttk.Combobox(
dialog,
textvariable=harness_var,
values=available,
@@ -502,8 +503,8 @@ class AgentCostEstimatorGUI:
def cancel():
dialog.destroy()
tk.Button(dialog, text="Save", command=save).pack(side=tk.LEFT, padx=5, pady=5)
tk.Button(dialog, text="Cancel", command=cancel).pack(side=tk.RIGHT, padx=5, pady=5)
self.tk.Button(dialog, text="Save", command=save).pack(side=self.tk.LEFT, padx=5, pady=5)
self.tk.Button(dialog, text="Cancel", command=cancel).pack(side=self.tk.RIGHT, padx=5, pady=5)
dialog.wait_window()
@@ -530,16 +531,13 @@ class AgentCostEstimatorGUI:
def _show_session_details(self, harness_name: str, session_id: str):
"""Show detailed information for a session."""
import tkinter as tk
from tkinter import messagebox
# Find the session
if self.session_discovery:
harnesses = self.session_discovery.discover_sessions()
if harness_name in harnesses:
for session in harnesses[harness_name].sessions:
if session.session_id == session_id:
messagebox.showinfo(
self.tk.messagebox.showinfo(
f"Session: {session.session_id}",
f"Model: {session.model}\n"
f"Session Price: ${session.session_price:.2f}\n"
@@ -548,7 +546,7 @@ class AgentCostEstimatorGUI:
)
return
messagebox.showwarning("Not Found", f"Session {session_id} not found")
self.tk.messagebox.showwarning("Not Found", f"Session {session_id} not found")
def _auto_refresh(self):
"""Auto-refresh sessions."""
@@ -52,10 +52,6 @@ logger = logging.getLogger("agent_cost_estimator.hermes_plugin")
DB_NAME = "state.db"
LEDGER_DIR = Path("plugin-data") / "agent-cost-estimator" / "usage"
LINE_PREFIX = "-- "
# How long the ordering wait gives the turn's message.complete to
# be published before the cost line publishes anyway (late beats
# absent; matches the bg-review hold pattern's spirit).
COMPLETION_WAIT_S = 60.0
# Interactive terminal surfaces that get the cost line. In current
# Hermes builds the CLI chat runs through the gateway's local
# adapter (Platform.LOCAL), so its platform value is "local";
@@ -464,16 +460,14 @@ def _in_daemon() -> bool:
def _publish_notification(session_id: str, line: str) -> bool:
"""Publish a notification.show event after the turn's response.
Ordering: post_llm_call fires in the agent's worker thread
BEFORE the gateway settles the turn and publishes
message.complete, so an immediate publish lands ABOVE the
response. Wait (bounded) until the session's execution stamp
is cleared - the settlement clears it strictly after
message.complete is published - then publish, making the cost
line the last line of the turn. This is the same hold pattern
the gateway's background-review delivery uses (run
detached; post_llm_call is timeout-bounded and must return
fast).
Publish immediately: post_llm_call fires inside the turn, before
settlement, so the frame reaches the viewer while its pump is
still dispatching. The classic CLI queues it as an AgentNotice
and flushes it after run_conversation (still below the response);
waiting for the turn's execution stamp to clear would land the
frame AFTER message.complete, which the classic CLI's pump has
already returned from and _drain_stale discards - silently
dropping the line client-side (verified 2026-11-03).
Returns True when the event was published (or the session is
gone and the daemon owns display either way), False when this
@@ -505,15 +499,14 @@ def _publish_notification(session_id: str, line: str) -> bool:
"[cost] session %s has no event stream; printing",
session_id)
return False
execution = getattr(stream, "execution", None)
if isinstance(execution, dict):
generation = execution.get("execution_generation")
if generation is not None:
deadline = time.monotonic() + COMPLETION_WAIT_S
while (stream.execution.get(
"execution_generation") == generation
and time.monotonic() < deadline):
time.sleep(0.05)
# Publish immediately. The classic CLI's pump returns at
# message.complete and discards later frames (_drain_stale),
# so waiting for the turn's execution stamp to clear (the
# old ordering wait) made the line land after the frame the
# CLI stops at and get dropped. post_llm_call fires inside
# the turn, so the frame is dispatched before message.complete
# and the classic CLI queues it as a notice, flushed after
# run_conversation — still printed below the response.
stream.publish(
session_id,
{"text": line, "level": "info", "kind": "sticky",
@@ -1,5 +1,5 @@
name: agent-cost-estimator
version: 0.4.4
version: 0.4.5
description: "Print an estimated cost line (provider:model | session:$X | turn:$Y) after every assistant turn, recorded from Hermes session usage."
author: "nvaldez"
provides_hooks:
+1 -1
View File
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
[project]
name = "agent-cost-estimator"
version = "0.4.4"
version = "0.4.5"
description = "Report token usage and estimated USD cost of AI agent sessions from local harness logs"
readme = "README.md"
requires-python = ">=3.9"
+8 -15
View File
@@ -951,15 +951,16 @@ def test_plugin_publishes_notification_in_daemon(tmp_path, monkeypatch,
assert capsys.readouterr().out == ""
def test_plugin_waits_for_completion_before_publish(monkeypatch):
"""While the turn's execution stamp is set, the publish waits;
once cleared it publishes (ordering: line lands AFTER the
response)."""
import threading
def test_plugin_publishes_immediately_without_waiting(monkeypatch):
"""The publish must NOT wait for the turn's execution stamp to
clear: the classic CLI's pump returns at message.complete and
drops later frames, so a settlement-ordered publish is lost
client-side. The frame must go out while the stamp is still set
(post_llm_call fires inside the turn)."""
import types
class _Stream:
execution = {"execution_generation": 7}
execution = {"execution_generation": 7} # still running
published_at = None
def publish(self, session_id, payload, *, event_type=None):
@@ -980,20 +981,12 @@ def test_plugin_waits_for_completion_before_publish(monkeypatch):
fake_gate.session_cron = fake_cron # type: ignore[attr-defined]
monkeypatch.setitem(sys.modules, "gateway", fake_gate)
monkeypatch.setitem(sys.modules, "gateway.session_cron", fake_cron)
monkeypatch.setattr(hp, "COMPLETION_WAIT_S", 5.0)
# Clear the stamp shortly after the publish attempt starts.
def clear_later():
time.sleep(0.3)
stream.execution = {}
clearer = threading.Thread(target=clear_later, daemon=True)
clearer.start()
started = time.monotonic()
ok = hp._publish_notification("s1", "line")
assert ok is True
assert stream.published_at is not None
assert stream.published_at >= started + 0.3 # waited for the clear
assert stream.published_at < started + 0.3 # did NOT wait for the stamp
def test_plugin_label_uses_billing_route(tmp_path, monkeypatch, capsys):