@m13v/s4l 1.6.197-rc.10

package/scripts/log_claude_session.py

@@ -0,0 +1,636 @@
+#!/usr/bin/env python3
+"""Log a Claude Code session's cost into the claude_sessions table.
+
+Reads the session transcript at ~/.claude/projects/<encoded-cwd>/<session_id>.jsonl,
+sums per-model token usage from each assistant turn, applies a local pricing
+table to compute total cost, and inserts one row.
+
+Usage:
+    python3 scripts/log_claude_session.py \\
+        --session-id <uuid> \\
+        --script run-linkedin \\
+        [--started-at ISO8601] [--ended-at ISO8601]
+
+Designed to be called by run_claude.sh after `claude -p --session-id $UUID` exits.
+Idempotent: ON CONFLICT DO NOTHING on session_id.
+"""
+
+import argparse
+import glob
+import json
+import os
+import sys
+from datetime import datetime
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+
+
+PROJECTS_ROOT = os.path.expanduser("~/.claude/projects")
+
+# Archive root for post-facto investigation. ~/.claude/projects/ is Claude
+# Code's own scratch — it survives normal runs but is not under our control
+# for retention or rotation, and the encoded-cwd subdirectory layout is
+# annoying to navigate after the fact. We hardlink each finished session's
+# transcript here so investigations of watchdog-killed phases are
+# `tail skill/logs/claude-sessions/<date>/<HHMMSS>_<script>_<sid>.jsonl`
+# instead of forensics across `~/.claude/projects/-/<sid>.jsonl` candidates.
+ARCHIVE_ROOT = os.path.expanduser("~/social-autoposter/skill/logs/claude-sessions")
+
+
+def find_transcript(session_id: str):
+    """Locate the transcript .jsonl for a session id.
+
+    Claude Code writes transcripts under `~/.claude/projects/<encoded-cwd>/<uuid>.jsonl`.
+    The encoded-cwd depends on the working directory at invocation time:
+    interactive runs land under `-Users-matthewdi-social-autoposter`, but
+    launchd-fired runs (cwd=/) land under `-`. Glob across all project dirs.
+    """
+    matches = glob.glob(os.path.join(PROJECTS_ROOT, "*", f"{session_id}.jsonl"))
+    return matches[0] if matches else None
+
+
+def archive_transcript(transcript_path, session_id: str, script: str, started_iso):
+    """Hardlink (or copy) the live transcript into ARCHIVE_ROOT.
+
+    Best-effort: any failure returns None and the caller proceeds. The
+    archive lives under <date>/<HHMMSS>_<script>_<session_id>.jsonl so
+    investigators can navigate by day. Hardlink first (free, atomic, and
+    keeps the archive in sync if claude appends final bytes between our
+    archive call and parse_transcript); fall back to copy across volumes.
+    Idempotent: returns the existing path if already archived.
+    """
+    if not transcript_path or not os.path.exists(transcript_path):
+        return None
+    try:
+        dt = None
+        if started_iso:
+            try:
+                dt = datetime.fromisoformat(started_iso.replace("Z", "+00:00"))
+            except (ValueError, AttributeError):
+                dt = None
+        if dt is None:
+            try:
+                dt = datetime.utcfromtimestamp(os.path.getmtime(transcript_path))
+            except OSError:
+                dt = datetime.utcnow()
+
+        date_subdir = dt.strftime("%Y-%m-%d")
+        time_part = dt.strftime("%H%M%S")
+        safe_script = "".join(
+            c if (c.isalnum() or c in ("-", "_")) else "_"
+            for c in (script or "unknown")
+        ) or "unknown"
+
+        archive_dir = os.path.join(ARCHIVE_ROOT, date_subdir)
+        os.makedirs(archive_dir, exist_ok=True)
+        archive_path = os.path.join(
+            archive_dir, f"{time_part}_{safe_script}_{session_id}.jsonl"
+        )
+        if os.path.exists(archive_path):
+            return archive_path
+        try:
+            os.link(transcript_path, archive_path)
+        except OSError:
+            import shutil
+            shutil.copy2(transcript_path, archive_path)
+        return archive_path
+    except Exception:
+        return None
+
+# USD per 1M tokens. Cache_5m / cache_1h are the WRITE rates (Anthropic charges
+# a premium for caching writes); cache_read is the discounted re-read rate.
+# Fallback (unknown model) uses Opus rates so we never underestimate.
+PRICING = {
+    "opus":   {"input": 15.0, "output": 75.0, "cache_5m": 18.75, "cache_1h": 30.0, "cache_read": 1.5},
+    "sonnet": {"input": 3.0,  "output": 15.0, "cache_5m": 3.75,  "cache_1h": 6.0,  "cache_read": 0.3},
+    "haiku":  {"input": 1.0,  "output": 5.0,  "cache_5m": 1.25,  "cache_1h": 2.0,  "cache_read": 0.1},
+}
+
+
+def price_for_model(model_id: str) -> dict:
+    m = (model_id or "").lower()
+    if "opus" in m:
+        return PRICING["opus"]
+    if "sonnet" in m:
+        return PRICING["sonnet"]
+    if "haiku" in m:
+        return PRICING["haiku"]
+    return PRICING["opus"]
+
+
+def cost_from_usage(model: str, usage: dict) -> float:
+    p = price_for_model(model)
+    inp = usage.get("input_tokens", 0) or 0
+    out = usage.get("output_tokens", 0) or 0
+    cache_read = usage.get("cache_read_input_tokens", 0) or 0
+    cache_5m = (usage.get("cache_creation") or {}).get("ephemeral_5m_input_tokens", 0) or 0
+    cache_1h = (usage.get("cache_creation") or {}).get("ephemeral_1h_input_tokens", 0) or 0
+    if not (cache_5m or cache_1h):
+        cache_5m = usage.get("cache_creation_input_tokens", 0) or 0
+    return (
+        inp * p["input"]
+        + out * p["output"]
+        + cache_read * p["cache_read"]
+        + cache_5m * p["cache_5m"]
+        + cache_1h * p["cache_1h"]
+    ) / 1_000_000
+
+
+def _parse_subagent_transcript(path: str, meta: dict):
+    """Parse a single agent-<id>.jsonl into a cost summary.
+
+    Subagent transcripts have the same per-turn shape as the orchestrator
+    (assistant turns with usage), but every event carries ``isSidechain:
+    true``, an ``agentId``, and references back to the parent via
+    ``sessionId`` (matches orchestrator's session id).
+    """
+    if not os.path.exists(path):
+        return None
+    by_model = {}
+    first_ts = None
+    last_ts = None
+    turns = 0
+    with open(path) as f:
+        for line in f:
+            try:
+                ev = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            ts = ev.get("timestamp")
+            if ts:
+                first_ts = first_ts or ts
+                last_ts = ts
+            if ev.get("type") != "assistant":
+                continue
+            msg = ev.get("message") or {}
+            usage = msg.get("usage") or {}
+            if not usage:
+                continue
+            model = msg.get("model") or "unknown"
+            entry = by_model.setdefault(model, {
+                "input_tokens": 0, "output_tokens": 0,
+                "cache_read_tokens": 0, "cache_creation_tokens": 0,
+                "cost_usd": 0.0,
+            })
+            inp = usage.get("input_tokens", 0) or 0
+            out = usage.get("output_tokens", 0) or 0
+            cr = usage.get("cache_read_input_tokens", 0) or 0
+            cc = usage.get("cache_creation_input_tokens", 0) or 0
+            entry["input_tokens"] += inp
+            entry["output_tokens"] += out
+            entry["cache_read_tokens"] += cr
+            entry["cache_creation_tokens"] += cc
+            entry["cost_usd"] += cost_from_usage(model, usage)
+            turns += 1
+    cost = sum(v["cost_usd"] for v in by_model.values())
+    primary = None
+    if by_model:
+        primary = max(
+            by_model.items(),
+            key=lambda kv: (kv[1].get("output_tokens", 0), kv[1].get("input_tokens", 0)),
+        )[0]
+    return {
+        "by_model": {
+            m: {**v, "cost_usd": round(v["cost_usd"], 6)}
+            for m, v in by_model.items()
+        },
+        "cost_usd": round(cost, 6),
+        "turns": turns,
+        "first_ts": first_ts,
+        "last_ts": last_ts,
+        "model": primary or "unknown",
+        "agent_type": (meta or {}).get("agentType"),
+        "description": (meta or {}).get("description"),
+    }
+
+
+def _find_subagent_dir(orchestrator_path: str) -> str:
+    """Given path .../<session_id>.jsonl, return .../<session_id>/subagents/."""
+    if not orchestrator_path:
+        return None
+    base = orchestrator_path[:-len(".jsonl")] if orchestrator_path.endswith(".jsonl") else orchestrator_path
+    candidate = os.path.join(base, "subagents")
+    return candidate if os.path.isdir(candidate) else None
+
+
+def parse_transcript(path: str):
+    """Parse a Claude Code transcript .jsonl into per-session cost.
+
+    Subagent (Agent tool) handling
+    ------------------------------
+    Claude Code SDK >= 2.1.x writes subagent transcripts to a SEPARATE
+    sibling directory, NOT inline in the orchestrator's .jsonl. Layout:
+
+        ~/.claude/projects/<encoded-cwd>/<orchestrator-session-id>.jsonl
+        ~/.claude/projects/<encoded-cwd>/<orchestrator-session-id>/
+                                                    subagents/
+                                                        agent-<short-id>.jsonl
+                                                        agent-<short-id>.meta.json
+
+    The orchestrator's .jsonl only records the ``tool_use`` block (name
+    "Agent", with ``subagent_type``/``description``/``prompt`` input) and
+    the consolidated ``tool_result`` carrying the agent's final reply. The
+    Agent's internal chain of assistant turns lives entirely in
+    ``agent-<id>.jsonl`` with ``isSidechain: true`` on every event.
+
+    This parser:
+      * Sums orchestrator-only token usage and cost into by_model/totals
+        (so total_cost_usd matches the parent thread only — subagent cost
+        is broken out separately so the user can see exactly what the
+        subagents added).
+      * Counts ``Agent`` tool_use blocks in orchestrator turns -> the
+        legacy ``task_call_count`` field (kept the name for back-compat
+        even though the tool name is "Agent", not "Task").
+      * Scans the sibling ``subagents/`` directory; for each
+        agent-<id>.jsonl, parses it as an independent transcript and adds
+        its cost to ``subagent_cost_usd``. Per-subagent details land in
+        ``subagent_breakdown`` keyed by short agent id, with the
+        agentType/description from the .meta.json sidecar.
+      * As a defensive fallback, also detects legacy ``isSidechain: true``
+        events inside the same .jsonl (older SDK versions may have used
+        that layout). Currently zero hits across 14k+ historical sessions,
+        but kept so we don't have to revisit when the SDK changes again.
+
+    Historical note (2026-05-10): the prior parser version looked for
+    tool_use name="Task" and inline isSidechain entries. Both miss the
+    actual SDK layout. The corpus has 2041 Agent invocations (mostly in
+    seo_generate_page sessions) whose cost was previously invisible.
+    """
+    if not os.path.exists(path):
+        return None
+
+    by_model = {}
+    totals = {"input": 0, "output": 0, "cache_read": 0, "cache_creation": 0}
+    first_ts = None
+    last_ts = None
+
+    # Subagent (sidechain) accounting. Keyed by chain-root uuid so chained
+    # sidechain turns under the same Task() invocation aggregate together.
+    # When parentUuid linkage is ambiguous we fall back to a single synthetic
+    # group ("unknown") so the cost still gets counted.
+    sidechain_groups = {}  # root_uuid -> {model: per-model dict, cost_usd, turns, first_ts, last_ts, description}
+    # parentUuid -> root_uuid map, built as we walk the transcript. The first
+    # sidechain turn we see introduces its uuid as a root candidate; later
+    # turns chain by parentUuid.
+    uuid_to_root = {}
+
+    task_call_count = 0
+
+    def _bump_model_bucket(bucket, model, usage):
+        entry = bucket.setdefault(model, {
+            "input_tokens": 0, "output_tokens": 0,
+            "cache_read_tokens": 0, "cache_creation_tokens": 0,
+            "cost_usd": 0.0,
+        })
+        inp = usage.get("input_tokens", 0) or 0
+        out = usage.get("output_tokens", 0) or 0
+        cr = usage.get("cache_read_input_tokens", 0) or 0
+        cc = usage.get("cache_creation_input_tokens", 0) or 0
+        entry["input_tokens"] += inp
+        entry["output_tokens"] += out
+        entry["cache_read_tokens"] += cr
+        entry["cache_creation_tokens"] += cc
+        entry["cost_usd"] += cost_from_usage(model, usage)
+        return inp, out, cr, cc
+
+    with open(path) as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                ev = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+
+            ts = ev.get("timestamp")
+            if ts:
+                first_ts = first_ts or ts
+                last_ts = ts
+
+            is_sidechain = bool(ev.get("isSidechain"))
+            ev_uuid = ev.get("uuid")
+            parent_uuid = ev.get("parentUuid")
+
+            # Count subagent tool_use blocks (both legacy "Task" name and
+            # current "Agent" name) in orchestrator turns. This gives us an
+            # authoritative subagent-invocation count, independent of whether
+            # the sibling subagents/ transcripts came through fully (a
+            # watchdog SIGTERM mid-subagent can leave the tool_use stamped
+            # but the sibling .jsonl never finished). 2026-05-10 the actual
+            # SDK tool name is "Agent"; "Task" kept for forward-compat if
+            # the SDK ever renames it again.
+            msg = ev.get("message") or {}
+            if not is_sidechain and ev.get("type") == "assistant":
+                content = msg.get("content")
+                if isinstance(content, list):
+                    for c in content:
+                        if (isinstance(c, dict)
+                                and c.get("type") == "tool_use"
+                                and c.get("name") in ("Task", "Agent")):
+                            task_call_count += 1
+
+            if ev.get("type") != "assistant":
+                continue
+            usage = msg.get("usage") or {}
+            model = msg.get("model") or "unknown"
+
+            if not is_sidechain:
+                # Orchestrator turn.
+                inp, out, cr, cc = _bump_model_bucket(by_model, model, usage)
+                totals["input"] += inp
+                totals["output"] += out
+                totals["cache_read"] += cr
+                totals["cache_creation"] += cc
+            else:
+                # Sidechain (subagent) turn. Resolve to a chain root: if
+                # parentUuid is already mapped to a root, attach there;
+                # otherwise this is a new root.
+                root = uuid_to_root.get(parent_uuid)
+                if root is None:
+                    root = ev_uuid or "unknown"
+                if ev_uuid:
+                    uuid_to_root[ev_uuid] = root
+
+                grp = sidechain_groups.setdefault(root, {
+                    "by_model": {},
+                    "cost_usd": 0.0,
+                    "turns": 0,
+                    "first_ts": None,
+                    "last_ts": None,
+                    "root_uuid": root,
+                })
+                _bump_model_bucket(grp["by_model"], model, usage)
+                grp["cost_usd"] += cost_from_usage(model, usage)
+                grp["turns"] += 1
+                if ts:
+                    grp["first_ts"] = grp["first_ts"] or ts
+                    grp["last_ts"] = ts
+
+    if not by_model and not sidechain_groups:
+        return None
+
+    total_cost = sum(m["cost_usd"] for m in by_model.values())
+
+    # Dominant model = the one that produced the most output tokens in this
+    # session. Claude Code's transcript emits `"model": "<synthetic>"` on
+    # interrupted/stopped events with zero usage; those shouldn't win just
+    # because they sort alphabetically when all real candidates tie.
+    real_models = {k: v for k, v in by_model.items() if not k.startswith("<")}
+    pool = real_models or by_model
+    if pool:
+        primary_model = max(
+            pool.items(),
+            key=lambda kv: (kv[1].get("output_tokens", 0), kv[1].get("input_tokens", 0)),
+        )[0]
+    else:
+        # Subagents-only session (no orchestrator turns logged — unusual but
+        # possible if the orchestrator was SIGTERMed before its first
+        # assistant turn yet a sidechain had already started). Fall back to
+        # the dominant model across sidechains.
+        all_models = {}
+        for grp in sidechain_groups.values():
+            for m, v in grp["by_model"].items():
+                e = all_models.setdefault(m, {"output_tokens": 0, "input_tokens": 0})
+                e["output_tokens"] += v.get("output_tokens", 0)
+                e["input_tokens"] += v.get("input_tokens", 0)
+        primary_model = max(
+            all_models.items(),
+            key=lambda kv: (kv[1].get("output_tokens", 0), kv[1].get("input_tokens", 0)),
+        )[0] if all_models else "unknown"
+
+    # Compact per-subagent breakdown for the subagent_breakdown jsonb column.
+    # Two sources feed this map:
+    #   1. Inline isSidechain entries (legacy SDK layout, ~0 hits today).
+    #      Keyed by chain-root uuid.
+    #   2. Sibling agent-<id>.jsonl files (current SDK layout, post-2.1.x).
+    #      Keyed by short agent id (e.g. "ab24e352623c7d99b").
+    subagent_breakdown = {}
+    for root, grp in sidechain_groups.items():
+        # Dominant model for this subagent group.
+        bm = grp["by_model"]
+        if bm:
+            sg_model = max(
+                bm.items(),
+                key=lambda kv: (kv[1].get("output_tokens", 0), kv[1].get("input_tokens", 0)),
+            )[0]
+        else:
+            sg_model = "unknown"
+        subagent_breakdown[root] = {
+            "source": "inline_sidechain",
+            "cost_usd": round(grp["cost_usd"], 6),
+            "turns": grp["turns"],
+            "first_ts": grp["first_ts"],
+            "last_ts": grp["last_ts"],
+            "model": sg_model,
+            "by_model": {
+                m: {
+                    "input_tokens": v["input_tokens"],
+                    "output_tokens": v["output_tokens"],
+                    "cache_read_tokens": v["cache_read_tokens"],
+                    "cache_creation_tokens": v["cache_creation_tokens"],
+                    "cost_usd": round(v["cost_usd"], 6),
+                }
+                for m, v in bm.items()
+            },
+        }
+    subagent_cost_usd = sum(grp["cost_usd"] for grp in sidechain_groups.values())
+
+    # ------- Scan sibling subagents/ directory for current-SDK transcripts -------
+    sub_dir = _find_subagent_dir(path)
+    if sub_dir:
+        for agent_file in sorted(os.listdir(sub_dir)):
+            if not agent_file.endswith(".jsonl"):
+                continue
+            short_id = agent_file[len("agent-"):-len(".jsonl")] if agent_file.startswith("agent-") else agent_file
+            meta_path = os.path.join(sub_dir, agent_file.replace(".jsonl", ".meta.json"))
+            meta = {}
+            try:
+                if os.path.exists(meta_path):
+                    with open(meta_path) as mf:
+                        meta = json.load(mf)
+            except Exception:
+                meta = {}
+            sub = _parse_subagent_transcript(os.path.join(sub_dir, agent_file), meta)
+            if not sub:
+                continue
+            subagent_breakdown[short_id] = {
+                "source": "sibling_dir",
+                "cost_usd": sub["cost_usd"],
+                "turns": sub["turns"],
+                "first_ts": sub["first_ts"],
+                "last_ts": sub["last_ts"],
+                "model": sub["model"],
+                "agent_type": sub["agent_type"],
+                "description": sub["description"],
+                "by_model": sub["by_model"],
+            }
+            subagent_cost_usd += sub["cost_usd"]
+    # Total distinct subagents (inline + sibling-dir) for the count column.
+    subagent_count = len(subagent_breakdown)
+
+    return {
+        "by_model": by_model,
+        "totals": totals,
+        "total_cost_usd": total_cost,
+        "primary_model": primary_model,
+        "first_ts": first_ts,
+        "last_ts": last_ts,
+        "task_call_count": task_call_count,
+        "subagent_count": subagent_count,
+        "subagent_cost_usd": round(subagent_cost_usd, 6),
+        "subagent_breakdown": subagent_breakdown,
+    }
+
+
+_BACKFILL_TABLES = (
+    "posts", "replies", "dms", "dm_messages",
+    "seo_escalations", "seo_keywords", "seo_page_improvements", "gsc_queries",
+)
+
+
+def _persist_via_api(args, parsed, started, ended, duration_ms, orch_cost, cycle_id):
+    """Upsert claude_sessions row + backfill model column via HTTP routes.
+
+    Two calls:
+      POST /api/v1/claude-sessions             -> upsert by session_id
+      POST /api/v1/claude-sessions/backfill-model -> stamp model on activity rows
+    """
+    from http_api import api_post
+    api_post(
+        "/api/v1/claude-sessions",
+        {
+            "session_id": args.session_id,
+            "script": args.script,
+            "started_at": started,
+            "ended_at": ended,
+            "duration_ms": duration_ms,
+            "total_cost_usd": round(parsed["total_cost_usd"], 6),
+            "orchestrator_cost_usd": orch_cost,
+            "input_tokens": parsed["totals"]["input"],
+            "output_tokens": parsed["totals"]["output"],
+            "cache_read_tokens": parsed["totals"]["cache_read"],
+            "cache_creation_tokens": parsed["totals"]["cache_creation"],
+            "model_breakdown": parsed["by_model"],
+            "model": parsed["primary_model"],
+            "cycle_id": cycle_id,
+            "task_call_count": parsed.get("task_call_count", 0),
+            "subagent_count": parsed.get("subagent_count", 0),
+            "subagent_cost_usd": parsed.get("subagent_cost_usd", 0.0),
+            "subagent_breakdown": parsed.get("subagent_breakdown") or None,
+        },
+    )
+
+    resp = api_post(
+        "/api/v1/claude-sessions/backfill-model",
+        {
+            "session_id": args.session_id,
+            "model": parsed["primary_model"],
+            "tables": list(_BACKFILL_TABLES),
+        },
+    )
+    data = (resp or {}).get("data") or {}
+    backfill_counts = data.get("backfilled") or {}
+    for t in _BACKFILL_TABLES:
+        backfill_counts.setdefault(t, 0)
+    return backfill_counts
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--session-id", required=True)
+    parser.add_argument("--script", required=True)
+    parser.add_argument("--started-at", default=None,
+                        help="ISO8601 timestamp; falls back to first transcript ts")
+    parser.add_argument("--ended-at", default=None,
+                        help="ISO8601 timestamp; falls back to last transcript ts")
+    parser.add_argument("--orchestrator-cost-usd", default=None, type=float,
+                        help="Native SDK cost (streamRes.total_cost_usd) for the "
+                             "orchestrator session, captured from claude -p stdout. "
+                             "Stored in claude_sessions.orchestrator_cost_usd. "
+                             "Authoritative (matches Anthropic billing for the "
+                             "orchestrator), but excludes Task subagent costs "
+                             "(see anthropics/claude-code #43945). When omitted, "
+                             "the column stays NULL and dashboards fall back to "
+                             "total_cost_usd (manual transcript-derived estimate).")
+    parser.add_argument("--cycle-id", default=None,
+                        help="Optional per-cycle batch identifier (e.g. "
+                             "'rdcycle-20260510-110005'). Lets get_run_cost.py / "
+                             "the dashboard scope cost to ONE pipeline cycle "
+                             "even when multiple cycles of the same script "
+                             "(double-forked run-reddit-search.sh / "
+                             "run-twitter-cycle.sh) overlap in wall-clock time. "
+                             "Falls back to env SA_CYCLE_ID; NULL if unset.")
+    args = parser.parse_args()
+
+    # Allow callers (run_claude.sh, post_reddit.py spawning a child claude) to
+    # propagate cycle_id via env without re-plumbing every call site. CLI flag
+    # takes precedence so explicit overrides still work.
+    cycle_id = args.cycle_id or os.environ.get("SA_CYCLE_ID") or None
+    if cycle_id == "":
+        cycle_id = None
+
+    transcript = find_transcript(args.session_id)
+    # Archive the transcript BEFORE parsing so even an empty/short session
+    # leaves a forensics trail. This is the only path that runs reliably on
+    # watchdog SIGTERM — once the wrapper's EXIT trap fires, log_claude_session
+    # is the last chance to capture what claude was doing before death.
+    archive_path = archive_transcript(
+        transcript, args.session_id, args.script, args.started_at
+    )
+    parsed = parse_transcript(transcript) if transcript else None
+
+    if parsed is None:
+        print(json.dumps({
+            "logged": False,
+            "reason": "no-transcript-or-empty",
+            "transcript": transcript,
+            "archive_path": archive_path,
+            "session_id": args.session_id,
+        }))
+        return
+
+    started = args.started_at or parsed["first_ts"]
+    ended = args.ended_at or parsed["last_ts"]
+    duration_ms = None
+    try:
+        if started and ended:
+            s = datetime.fromisoformat(started.replace("Z", "+00:00"))
+            e = datetime.fromisoformat(ended.replace("Z", "+00:00"))
+            duration_ms = int((e - s).total_seconds() * 1000)
+    except (ValueError, AttributeError):
+        pass
+
+    # Orchestrator cost: prefer the native SDK value passed via flag (from
+    # streamRes.total_cost_usd in the caller); fall back to NULL so the column
+    # only holds authoritative values. Manual transcript-derived estimate goes
+    # into total_cost_usd unchanged.
+    orch_cost = (
+        round(args.orchestrator_cost_usd, 6)
+        if args.orchestrator_cost_usd is not None
+        else None
+    )
+
+    backfill_counts = _persist_via_api(args, parsed, started, ended, duration_ms,
+                                       orch_cost, cycle_id)
+
+    print(json.dumps({
+        "logged": True,
+        "session_id": args.session_id,
+        "script": args.script,
+        "cycle_id": cycle_id,
+        "total_cost_usd": round(parsed["total_cost_usd"], 6),
+        "orchestrator_cost_usd": orch_cost,
+        "duration_ms": duration_ms,
+        "model": parsed["primary_model"],
+        "models": list(parsed["by_model"].keys()),
+        "task_call_count": parsed.get("task_call_count", 0),
+        "subagent_count": parsed.get("subagent_count", 0),
+        "subagent_cost_usd": parsed.get("subagent_cost_usd", 0.0),
+        "backfilled": backfill_counts,
+        "archive_path": archive_path,
+    }))
+
+
+if __name__ == "__main__":
+    main()