Class: McpLogs::Call

Inherits:
ApplicationRecord show all
Defined in:
app/models/mcp_logs/call.rb

Overview

One JSON-RPC request served by an MCP server. Inserted as "running" the moment the request starts, updated when it finishes — so a long tool call is visible while it runs rather than only after it returns.

Constant Summary collapse

STATUSES = 
%w[running ok error].freeze
ABANDONED =

error_type for a row whose finish never landed. Its own value rather than an exception class name, because no exception was raised here — the process that would have raised one is gone.

"abandoned".freeze
ABANDONED_MESSAGE = 
"No completion was ever recorded for this call. Whatever was " \
"serving it stopped first — a deploy, a crash, or a failed log " \
"write — so its result and duration are unknown.".freeze
FILTER_OPTIONS_LIMIT =

How many distinct values a filter dropdown will offer. Bounded because tool_name is raw client input (see Recorder::STRING_LIMIT): the number of distinct values is capped by requests served, not by tools that exist, so an unbounded SELECT DISTINCT would let one looping client put an 

200

Class Method Summary collapse

Instance Method Summary collapse

Class Method Details

.filter_values(column) ⇒ Object 



37
38
39
 
# File 'app/models/mcp_logs/call.rb', line 37

def self.filter_values(column)
  where.not(column => nil).distinct.order(column).limit(FILTER_OPTIONS_LIMIT).pluck(column)
end

.purge!(retention_days = McpLogs.retention_days) ⇒ Object 



45
46
47
 
# File 'app/models/mcp_logs/call.rb', line 45

def self.purge!(retention_days = McpLogs.retention_days)
  where(started_at: ...retention_days.to_i.days.ago).delete_all
end

.sweep_abandoned!(hours = McpLogs.abandoned_after_hours) ⇒ Object

A row is inserted "running" and updated when the request finishes. If that update never lands — the container restarted mid-call, or the write itself failed — nothing retries it, so the row would claim to be running for the entire retention window: a permanent false positive on the page an operator reads to answer "what is the agent doing right now", and an inflated count in tool_counts.

completed_at and duration_ms stay nil on purpose. The sweep knows the call did not finish; it does not know when it stopped, and stamping "now" would invent a fact. Returns how many rows it settled.



59
60
61
62
63
64
 
# File 'app/models/mcp_logs/call.rb', line 59

def self.sweep_abandoned!(hours = McpLogs.abandoned_after_hours)
  where(status: "running")
    .where(started_at: ...hours.to_i.hours.ago)
    .update_all(status: "error", error_type: ABANDONED, error_message: ABANDONED_MESSAGE,
                updated_at: Time.current)
end

.tool_counts ⇒ Object 



41
42
43
 
# File 'app/models/mcp_logs/call.rb', line 41

def self.tool_counts
  where.not(tool_name: nil).group(:tool_name).count
end

Instance Method Details

#duration_seconds ⇒ Object 



68
 
# File 'app/models/mcp_logs/call.rb', line 68

def duration_seconds = duration_ms && duration_ms / 1000.0

#running? ⇒ Boolean

Returns:

  • (Boolean) 


66
 
# File 'app/models/mcp_logs/call.rb', line 66

def running? = status == "running"