From 454ed0e357d35af4f4c10d5aa5689c7fa426386c Mon Sep 17 00:00:00 2001
From: itscooleric <itscooleric@users.noreply.github.com>
Date: Mon, 16 Mar 2026 22:50:14 +0000
Subject: [PATCH] feat: token/cost tracking per agent session (#43)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Parses Claude Code conversation JSONL for token usage data and adds
it to the session_end event in events.jsonl:
- input_tokens, output_tokens, total_tokens
- estimated_cost_usd (per-model pricing: Opus, Sonnet, Haiku)
- turns count

New script: scripts/token-cost.py — standalone parser, can also be
run manually on any conversation.jsonl file.

Closes #43

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 Dockerfile                |   3 +-
 scripts/session-logger.sh |  26 +++++++++
 scripts/token-cost.py     | 120 ++++++++++++++++++++++++++++++++++++++
 3 files changed, 148 insertions(+), 1 deletion(-)
 create mode 100755 scripts/token-cost.py

diff --git a/Dockerfile b/Dockerfile
index 68b2228..77dc5ee 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -90,7 +90,8 @@ COPY claude-entrypoint.sh /usr/local/bin/claude-entrypoint.sh
 COPY firewall.sh /usr/local/bin/firewall.sh
 COPY scripts/session-logger.sh /usr/local/bin/session-logger.sh
 COPY scripts/notify.sh /usr/local/bin/notify.sh
-RUN chmod +x /usr/local/bin/entrypoint.sh /usr/local/bin/claude-entrypoint.sh /usr/local/bin/firewall.sh /usr/local/bin/session-logger.sh /usr/local/bin/notify.sh
+COPY scripts/token-cost.py /usr/local/bin/token-cost.py
+RUN chmod +x /usr/local/bin/entrypoint.sh /usr/local/bin/claude-entrypoint.sh /usr/local/bin/firewall.sh /usr/local/bin/session-logger.sh /usr/local/bin/notify.sh /usr/local/bin/token-cost.py
 
 # Default CLAUDE.md template — seeded into /workspace on first run if none exists
 COPY CLAUDE.md.template /usr/local/share/clide/CLAUDE.md.template
diff --git a/scripts/session-logger.sh b/scripts/session-logger.sh
index 1b3204c..0cf6035 100755
--- a/scripts/session-logger.sh
+++ b/scripts/session-logger.sh
@@ -273,6 +273,32 @@ except: pass
   fi
   if [[ -f "${SESSION_DIR}/conversation.jsonl" ]]; then
     _end_args+=("has_conversation=true")
+
+    # ── Token / cost tracking ──────────────────────────────────────
+    # Parse the conversation log for token counts and estimated cost.
+    # Uses the Python script for reliable JSON parsing + pricing math.
+    COST_SCRIPT="$(dirname "$0")/token-cost.py"
+    if [[ ! -x "$COST_SCRIPT" ]]; then
+      COST_SCRIPT="/usr/local/bin/token-cost.py"
+    fi
+    if [[ -x "$COST_SCRIPT" || -f "$COST_SCRIPT" ]]; then
+      _cost_json=$(python3 "$COST_SCRIPT" "${SESSION_DIR}/conversation.jsonl" 2>/dev/null || echo '{}')
+      if [[ -n "$_cost_json" && "$_cost_json" != "{}" ]]; then
+        # Extract fields and add to session_end args
+        _input_tokens=$(echo "$_cost_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('input_tokens',0))" 2>/dev/null || echo 0)
+        _output_tokens=$(echo "$_cost_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('output_tokens',0))" 2>/dev/null || echo 0)
+        _total_tokens=$(echo "$_cost_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('total_tokens',0))" 2>/dev/null || echo 0)
+        _cost_usd=$(echo "$_cost_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('estimated_cost_usd',0))" 2>/dev/null || echo 0)
+        _turns=$(echo "$_cost_json" | python3 -c "import json,sys; print(json.load(sys.stdin).get('turns',0))" 2>/dev/null || echo 0)
+
+        _end_args+=("input_tokens=${_input_tokens}")
+        _end_args+=("output_tokens=${_output_tokens}")
+        _end_args+=("total_tokens=${_total_tokens}")
+        _end_args+=("estimated_cost_usd=${_cost_usd}")
+        _end_args+=("turns=${_turns}")
+        echo "[session-logger] Tokens: ${_input_tokens} in / ${_output_tokens} out (${_total_tokens} total) — \$${_cost_usd}"
+      fi
+    fi
   fi
   emit_event "session_end" "${_end_args[@]}"
 
diff --git a/scripts/token-cost.py b/scripts/token-cost.py
new file mode 100755
index 0000000..7641eb7
--- /dev/null
+++ b/scripts/token-cost.py
@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+"""Parse a Claude Code conversation JSONL and output token/cost summary.
+
+Usage:
+    token-cost.py <conversation.jsonl>
+
+Outputs a JSON object with:
+    input_tokens, output_tokens, cache_creation_tokens, cache_read_tokens,
+    total_tokens, estimated_cost_usd, model, turns
+
+Pricing (per million tokens, as of 2026-03):
+    Claude Sonnet 4:   input $3, output $15, cache_write $3.75, cache_read $0.30
+    Claude Opus 4:     input $15, output $75, cache_write $18.75, cache_read $1.50
+    Claude Haiku 3.5:  input $0.80, output $4, cache_write $1, cache_read $0.08
+"""
+
+import json
+import sys
+from pathlib import Path
+
+# Pricing per million tokens
+PRICING = {
+    "claude-sonnet-4-20250514": {
+        "input": 3.0, "output": 15.0,
+        "cache_write": 3.75, "cache_read": 0.30,
+    },
+    "claude-opus-4-20250514": {
+        "input": 15.0, "output": 75.0,
+        "cache_write": 18.75, "cache_read": 1.50,
+    },
+    "claude-haiku-3-5-20241022": {
+        "input": 0.80, "output": 4.0,
+        "cache_write": 1.0, "cache_read": 0.08,
+    },
+}
+
+# Default to sonnet pricing if model unknown
+DEFAULT_PRICING = PRICING["claude-sonnet-4-20250514"]
+
+
+def parse_conversation(path: str) -> dict:
+    """Parse conversation JSONL and return token/cost summary."""
+    input_tokens = 0
+    output_tokens = 0
+    cache_creation = 0
+    cache_read = 0
+    model = ""
+    turns = 0
+
+    with open(path) as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                obj = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+
+            # Extract model from summary or first assistant message
+            if not model:
+                if obj.get("type") == "summary":
+                    model = obj.get("model", "")
+                elif obj.get("type") == "assistant":
+                    model = obj.get("message", {}).get("model", "")
+
+            # Count turns
+            if obj.get("type") in ("user", "human"):
+                turns += 1
+
+            # Extract usage from assistant messages
+            if obj.get("type") == "assistant":
+                usage = obj.get("message", {}).get("usage", {})
+                if usage:
+                    input_tokens += usage.get("input_tokens", 0)
+                    output_tokens += usage.get("output_tokens", 0)
+                    cache_creation += usage.get("cache_creation_input_tokens", 0)
+                    cache_read += usage.get("cache_read_input_tokens", 0)
+
+    # Calculate cost — match model family by keyword
+    pricing = DEFAULT_PRICING
+    model_lower = model.lower()
+    if "opus" in model_lower:
+        pricing = PRICING["claude-opus-4-20250514"]
+    elif "haiku" in model_lower:
+        pricing = PRICING["claude-haiku-3-5-20241022"]
+    elif "sonnet" in model_lower:
+        pricing = PRICING["claude-sonnet-4-20250514"]
+
+    cost = (
+        (input_tokens / 1_000_000) * pricing["input"]
+        + (output_tokens / 1_000_000) * pricing["output"]
+        + (cache_creation / 1_000_000) * pricing["cache_write"]
+        + (cache_read / 1_000_000) * pricing["cache_read"]
+    )
+
+    return {
+        "input_tokens": input_tokens,
+        "output_tokens": output_tokens,
+        "cache_creation_tokens": cache_creation,
+        "cache_read_tokens": cache_read,
+        "total_tokens": input_tokens + output_tokens + cache_creation + cache_read,
+        "estimated_cost_usd": round(cost, 6),
+        "model": model or "unknown",
+        "turns": turns,
+    }
+
+
+if __name__ == "__main__":
+    if len(sys.argv) != 2:
+        print(f"Usage: {sys.argv[0]} <conversation.jsonl>", file=sys.stderr)
+        sys.exit(1)
+
+    path = sys.argv[1]
+    if not Path(path).exists():
+        print(json.dumps({"error": f"File not found: {path}"}))
+        sys.exit(1)
+
+    result = parse_conversation(path)
+    print(json.dumps(result))