cardinal-ai 0.2.13 → 0.2.14

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 2d68fd5849d2147217f98894631cb725af6c4dd786bf4e2c5438f374101f0e34
-  data.tar.gz: 905b5c3af9bd9fdad3aa2c5f19b69968b495ff2e2788a9ca04567c0e2be15e6a
+  metadata.gz: b30a673a110becb4019a1d537443a14ec37ebeb326c29ff310c41edd3dc59dee
+  data.tar.gz: ca0b8933f66d185cf62b30e9b32fc6b9e6217045654bc2dc61f884be82ff212c
 SHA512:
-  metadata.gz: 184cbda0e27ce7e49f5b921e39b8e532a77ec27e80105f186161f2d01197ad36a5da1fdf549c2636f31b09b318459dcd9f3b19f4a74ffadd17834c54fbeb04f2
-  data.tar.gz: 82e3aeb820bbb9252304147f37fd57bb50cb65d40e3ae709b7a9c8b911bb5b51930a19233a528eda0caec09e86181c4d95391349bf70f03bac8791aefa320ce4
+  metadata.gz: 585edbeff0fc41cef7ad16fcb76bb61c5bd041fc44a51f0ce3ca71d7dfcfa22144951638f0f9c3f400821fda0a5ffb84eaff2bde32d2e29ecf7448783c689672
+  data.tar.gz: d00bca1a2862b301a4204545ff0d79dfd14855f36a8a45e538e595a7eb7e1e81983e6aeeda3c0806fce789d7fea69f0fbb2690b8ef6e97e59dfc6774dd20036d

data/app/assets/stylesheets/cardinal.css

@@ -108,6 +108,8 @@ a { color: var(--blue); text-decoration: none; }
 .col-search.open { display: block; }
 .filter-hidden { display: none !important; }
 
+.assistant-spend { margin-left: auto; font-size: 11px; color: var(--text-dim); cursor: default; }
+
 .pull-form { display: inline; }
 #repo-pull-status { font-size: 0.85rem; }
 #repo-pull-status .pull-ok  { color: var(--green); }

data/app/jobs/ai_task_job.rb

@@ -17,7 +17,8 @@ class AiTaskJob < ApplicationJob
     text = ClaudeCli.prompt(
       prompt,
       system: "You are a maintenance agent on a Cardinal board performing one bounded task on card ##{card.number}. Be concise; your output is posted directly to the card's timeline.",
-      model: model.presence || AssistantReplyJob::FALLBACK_MODEL
+      model: model.presence || AssistantReplyJob::FALLBACK_MODEL,
+      ledger: { kind: "ai_task", card: card }
     )
     card.log!("assistant_message", actor: "assistant", text: text) if text.present?
   rescue ClaudeCli::Error => e

data/app/jobs/assistant_reply_job.rb

@@ -57,6 +57,7 @@ class AssistantReplyJob < ApplicationJob
         # affirming reply ("yes, that approach") must not escalate the
         # planner into implementing.
         return ClaudeCli.prompt(unheard_messages(card), resume: card.assistant_session_id,
+                                 ledger: { kind: "assistant", card: card },
                                 system: ROLE_REMINDER, **common)
       rescue ClaudeCli::Error
         card.update!(assistant_session_id: nil) # stale/expired — start fresh below
@@ -64,7 +65,7 @@ class AssistantReplyJob < ApplicationJob
     end
 
     opener = kickoff ? kickoff_prompt(card) : transcript_prompt(card)
-    ClaudeCli.prompt(opener, system: system_prompt(card), **common)
+    ClaudeCli.prompt(opener, system: system_prompt(card), ledger: { kind: "assistant", card: card }, **common)
   end
 
   def kickoff_prompt(card)

data/app/jobs/compact_job.rb

@@ -28,7 +28,8 @@ class CompactJob < ApplicationJob
     return clear_working(card) unless ClaudeCli.available?
 
     model = card.board.columns.find_by(archetype: "planning")&.model.presence || FALLBACK_MODEL
-    compact = ClaudeCli.prompt(build_prompt(card), system: SYSTEM, model: model, max_turns: 1)
+    compact = ClaudeCli.prompt(build_prompt(card), system: SYSTEM, model: model, max_turns: 1,
+                               ledger: { kind: "compact", card: card })
 
     card.update!(compact: compact.to_s.strip, compact_generated_at: Time.current, compact_status: nil)
     card.broadcast_replace_to card, target: "card_compact",

data/app/jobs/deep_dive_job.rb

@@ -44,7 +44,8 @@ class DeepDiveJob < ApplicationJob
     sha = board.head_sha
     model = board.columns.find_by(archetype: "planning")&.model.presence || FALLBACK_MODEL
 
-    brief = ClaudeCli.prompt(PROMPT, model:, tools: "Read,Glob,Grep", cwd: repo, max_turns: MAX_TURNS)
+    brief = ClaudeCli.prompt(PROMPT, model:, tools: "Read,Glob,Grep", cwd: repo, max_turns: MAX_TURNS,
+                             ledger: { kind: "deep_dive" })
 
     File.write(board.brief_path, brief.to_s)
     board.update!(brief_sha: sha, brief_generated_at: Time.current,

data/app/jobs/summary_job.rb

@@ -22,7 +22,8 @@ class SummaryJob < ApplicationJob
     return clear_working(card) unless ClaudeCli.available?
 
     model = card.board.columns.find_by(archetype: "planning")&.model.presence || FALLBACK_MODEL
-    summary = ClaudeCli.prompt(build_prompt(card), system: SYSTEM, model: model, max_turns: 1)
+    summary = ClaudeCli.prompt(build_prompt(card), system: SYSTEM, model: model, max_turns: 1,
+                               ledger: { kind: "summary", card: card })
 
     card.update!(summary: summary.to_s.strip, summary_generated_at: Time.current, summary_status: nil)
     card.broadcast_replace_to card, target: "card_summary",

data/app/models/ai_call.rb

@@ -0,0 +1,10 @@
+# One row per one-shot AI call (§ money honesty): what it was for, what it
+# cost. Worker runs keep their usage on Run; everything that goes through
+# ClaudeCli lands here — including board-level calls with no card (deep dive).
+class AiCall < ApplicationRecord
+  KINDS = %w[assistant ai_task deep_dive summary compact rules_compile].freeze
+
+  belongs_to :card, optional: true
+
+  validates :kind, inclusion: { in: KINDS }
+end

data/app/models/card.rb

@@ -23,6 +23,7 @@ class Card < ApplicationRecord
                       dependent: :nullify, inverse_of: :parent
   has_many :events, -> { order(:created_at, :id) }, dependent: :destroy
   has_many :agent_sessions, dependent: :destroy
+  has_many :ai_calls, dependent: :delete_all
   has_many :runs, through: :agent_sessions
 
   # Card-face status glyphs. Keyed on status, except `ready_for_approval?`
@@ -101,8 +102,12 @@ class Card < ApplicationRecord
 
   # Running tally across every run on the card — the closed-card cost footer
   # (card #20). Sums stopped/restarted segments so the total reflects real spend.
-  def total_cost = runs.sum(:cost)
-  def total_output_tokens = runs.sum(:output_tokens)
+  # Honest money: worker runs PLUS every one-shot call made on this card's
+  # behalf (planning assistant, ai_task, summary/compact) — see AiCall.
+  def total_cost = runs.sum(:cost) + ai_calls.sum(:cost)
+  def total_output_tokens = runs.sum(:output_tokens) + ai_calls.sum(:output_tokens)
+
+  def assistant_cost = ai_calls.where(kind: "assistant").sum(:cost)
 
   # Is the planning assistant expected to post next? True right after entering
   # a planning column (kickoff inspection pending) or after a user message.

data/app/models/column.rb

@@ -134,9 +134,10 @@ class Column < ApplicationRecord
   def footer_value(compute)
     case compute.to_s
     when "sum_cost"
-      "$%.2f" % column_runs.sum(:cost)
+      "$%.2f" % (column_runs.sum(:cost) + column_ai_calls.sum(:cost))
     when "sum_tokens"
-      ActiveSupport::NumberHelper.number_to_delimited(column_runs.sum("input_tokens + output_tokens"))
+      ActiveSupport::NumberHelper.number_to_delimited(
+        column_runs.sum("input_tokens + output_tokens") + column_ai_calls.sum("input_tokens + output_tokens"))
     when "count_cards"
       cards.count.to_s
     when "model"
@@ -196,4 +197,9 @@ class Column < ApplicationRecord
   def column_runs
     Run.joins(agent_session: :card).where(cards: { column_id: id })
   end
+
+  # One-shot AI spend (assistant/ai_task/summary/…) of this column's cards.
+  def column_ai_calls
+    AiCall.where(card_id: cards.select(:id))
+  end
 end

data/app/services/claude_cli.rb

@@ -34,11 +34,15 @@ module ClaudeCli
   # resume: continue an existing claude session (context carries over).
   # with_session: return [text, session_id] instead of just text, so callers
   # can keep a continuing conversation (the planning assistant does).
+  # ledger: { kind:, card: } — record this call's tokens/cost as an AiCall
+  # (§ money honesty: planning conversations and maintenance calls spend real
+  # dollars; only worker runs used to be counted).
   def self.prompt(text, system: nil, model: nil, tools: nil, cwd: nil, max_turns: 1,
-                  resume: nil, with_session: false)
+                  resume: nil, with_session: false, ledger: nil)
     raise Error.new("claude CLI not found on PATH") unless available?
 
     json = invoke(text, system:, model:, tools:, cwd:, max_turns:, resume:)
+    record_usage!(json, ledger, model)
     if success?(json)
       return with_session ? [json["result"].to_s, json["session_id"]] : json["result"].to_s
     end
@@ -47,6 +51,7 @@ module ClaudeCli
     # force an answer from the context it already gathered.
     if json["subtype"] == "error_max_turns" && json["session_id"].present?
       wrapped = invoke(WRAP_UP, model:, cwd:, tools: "", max_turns: 2, resume: json["session_id"])
+      record_usage!(wrapped, ledger, model)
       if success?(wrapped)
         return with_session ? [wrapped["result"].to_s, wrapped["session_id"] || json["session_id"]] : wrapped["result"].to_s
       end
@@ -57,6 +62,23 @@ module ClaudeCli
     raise Error.new(friendly_failure(json), detail: json.to_json)
   end
 
+  # Best-effort by design: a ledger failure must never break the AI call that
+  # already succeeded. Failed calls are recorded too — they cost money.
+  def self.record_usage!(json, ledger, model)
+    return unless ledger.is_a?(Hash) && ledger[:kind].present?
+    usage = json["usage"] || {}
+    AiCall.create!(
+      card: ledger[:card],
+      kind: ledger[:kind].to_s,
+      model: json["model"] || model,
+      input_tokens: usage["input_tokens"].to_i,
+      output_tokens: usage["output_tokens"].to_i,
+      cost: json["total_cost_usd"].to_f
+    )
+  rescue StandardError => e
+    Rails.logger.warn("AiCall ledger write failed: #{e.class}: #{e.message}")
+  end
+
   def self.success?(json)
     json["subtype"] == "success" && !json["is_error"]
   end

data/app/services/rules/compiler.rb

@@ -22,6 +22,7 @@ module Rules
 
       raw = ClaudeCli.prompt(
         text,
+        ledger: { kind: "rules_compile" },
         model: AssistantReplyJob::FALLBACK_MODEL,
         system: <<~SYS
           You compile plain-English descriptions of Kanban column automation into JSON rule

data/app/services/run_sweeper.rb

@@ -14,8 +14,7 @@ module RunSweeper
   def self.fail_dead_runs
     Run.where(status: %w[queued running]).find_each do |run|
       next if alive?(run)
-      next if run.heartbeat_at && run.heartbeat_at > HEARTBEAT_GRACE.ago
-      next if run.heartbeat_at.nil? && run.created_at > HEARTBEAT_GRACE.ago
+      next if recently_active?(run)
 
       run.update!(status: "failed", finished_at: Time.current,
                   result_summary: "Runner died without finishing (swept)")
@@ -32,7 +31,11 @@ module RunSweeper
   def self.repair_stuck_cards
     Card.where(status: "working").find_each do |card|
       next unless card.column.ai? # non-AI columns: "working" means a human is
-      next if card.runs.where(status: %w[queued running needs_input]).any? { |r| r.needs_input? || alive?(r) }
+      # Same grace as fail_dead_runs: a freshly started run has no pid until
+      # AFTER workspace provisioning (clone/fetch) — recency is its proof of
+      # life, or every just-dragged card risks a bogus "stuck" verdict.
+      next if card.runs.where(status: %w[queued running needs_input])
+                  .any? { |r| r.needs_input? || alive?(r) || recently_active?(r) }
       card.update!(status: "failed")
       card.log!("error", text: "Card was stuck working with no live run; marked failed.")
     end
@@ -42,6 +45,12 @@ module RunSweeper
     Column.where(archetype: "execution").find_each(&:kick_queue)
   end
 
+  # Between state writes (provisioning, spawn) a live run has no pid yet —
+  # a recent heartbeat or recent birth counts as alive.
+  def self.recently_active?(run)
+    (run.heartbeat_at || run.created_at) > HEARTBEAT_GRACE.ago
+  end
+
   def self.alive?(run)
     pid = run.agent_session&.config&.dig("pid")
     return false if pid.blank?

data/app/views/boards/brief.html.erb

@@ -18,6 +18,8 @@
             Generated <%= @board.brief_generated_at&.strftime("%b %-d, %H:%M") %>
             from <code><%= @board.brief_sha&.first(7) %></code>
             with <%= @board.brief_model %>
+            <% dive_cost = AiCall.where(kind: "deep_dive").order(:id).last&.cost.to_f %>
+            <% if dive_cost.positive? %> for $<%= sprintf("%.2f", dive_cost) %><% end %>
             <% behind = @board.commits_behind_brief %>
             <% if behind&.positive? %>
               · <span class="brief-behind"><%= pluralize(behind, "commit") %> behind HEAD</span>

data/app/views/cards/_detail.html.erb

@@ -35,6 +35,9 @@
               <%= link_to zoom.capitalize, card_path(@card, zoom: zoom),
                           class: ("active" if @zoom == zoom) %>
             <% end %>
+            <% if (spend = @card.assistant_cost).positive? %>
+              <span class="assistant-spend" title="What this card's planning conversation has cost so far (assistant replies — separate from agent runs)">🪶 $<%= sprintf("%.2f", spend) %></span>
+            <% end %>
           </nav>
 
           <% if @zoom == "summary" %>

data/db/migrate/20260705120000_create_ai_calls.rb

@@ -0,0 +1,17 @@
+# Usage ledger for every one-shot AI call (card #-less deep dives included).
+# Worker runs track their own usage on Run; this covers the ClaudeCli tier:
+# planning assistant, ai_task rules, deep dive, summary/compact, compiler.
+class CreateAiCalls < ActiveRecord::Migration[8.1]
+  def change
+    create_table :ai_calls do |t|
+      t.references :card, foreign_key: true, null: true
+      t.string :kind, null: false
+      t.string :model
+      t.integer :input_tokens, default: 0, null: false
+      t.integer :output_tokens, default: 0, null: false
+      t.decimal :cost, precision: 10, scale: 6, default: 0, null: false
+      t.datetime :created_at, null: false
+    end
+    add_index :ai_calls, :kind
+  end
+end

data/lib/cardinal/version.rb

@@ -1,3 +1,3 @@
 module Cardinal
-  VERSION = "0.2.13"
+  VERSION = "0.2.14"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: cardinal-ai
 version: !ruby/object:Gem::Version
-  version: 0.2.13
+  version: 0.2.14
 platform: ruby
 authors:
 - Jason Ellis
@@ -196,6 +196,7 @@ files:
 - app/jobs/summary_job.rb
 - app/mailers/application_mailer.rb
 - app/models/agent_session.rb
+- app/models/ai_call.rb
 - app/models/application_record.rb
 - app/models/artifact.rb
 - app/models/board.rb
@@ -267,6 +268,7 @@ files:
 - db/migrate/20260704130000_add_summary_to_cards.rb
 - db/migrate/20260704140000_add_compact_to_cards.rb
 - db/migrate/20260704231436_add_model_and_effort_to_cards.rb
+- db/migrate/20260705120000_create_ai_calls.rb
 - db/queue_schema.rb
 - db/seeds.rb
 - docker/agent/Dockerfile