octarin-cli 0.2.0

package/assets/claude_code/hook.py

@@ -0,0 +1,573 @@
+#!/usr/bin/env python3
+"""Claude Code -> Octarin capture hook (pure stdlib, fail-open).
+
+Registered as a Claude Code ``Stop`` hook. On each turn-end Claude Code pipes a
+small JSON payload on stdin (``session_id``, ``transcript_path``, ``cwd``, ...).
+This hook:
+
+  1. reads that payload and locates the session transcript JSONL;
+  2. parses user/assistant turns, tool calls, token usage, and model;
+  3. builds a single canonical ``IngestEvent`` (full ``spans`` form) covering the
+     turns produced since the last run (tracked via a per-session offset file);
+  4. POSTs it to ``${OCTARIN_INGEST_URL:-$OCTARIN_API_BASE/v1/ingest}`` with
+     ``Authorization: Bearer $OCTARIN_API_KEY``.
+
+It is deliberately tiny and dependency-free (stdlib only). Every failure path
+exits 0 so the host tool is never blocked, and the network call has a hard
+timeout. The canonical shape is defined in ``backend/app/schema/canonical.py``.
+"""
+
+from __future__ import annotations
+
+import base64
+import getpass
+import hashlib
+import json
+import os
+import subprocess
+import sys
+import time
+import urllib.request
+import uuid
+from datetime import datetime, timezone
+from pathlib import Path
+
+SOURCE = "claude-code"
+STATE_DIR = Path.home() / ".octarin"
+STATE_FILE = STATE_DIR / "claude_code_state.json"
+MAX_TEXT = 20_000  # cap stored input/output text so payloads stay small
+HTTP_TIMEOUT_S = 5.0
+# Cap per-attachment base64 payload we ship inline. Larger items are recorded
+# metadata-only (no b64) so a giant paste never bloats the POST or the backend.
+MAX_ATTACHMENT_BYTES = 5 * 1024 * 1024  # ~5MB of raw bytes
+# Map common file extensions -> mime for file refs that lack one.
+_EXT_MIME = {
+    ".png": "image/png",
+    ".jpg": "image/jpeg",
+    ".jpeg": "image/jpeg",
+    ".gif": "image/gif",
+    ".webp": "image/webp",
+    ".svg": "image/svg+xml",
+    ".pdf": "application/pdf",
+    ".txt": "text/plain",
+    ".md": "text/markdown",
+    ".json": "application/json",
+    ".csv": "text/csv",
+}
+# Same UUID5 namespace as backend deterministic_trace_id so retries de-dupe.
+_TRACE_NAMESPACE = uuid.UUID("6f8d2c1e-9a3b-4f5e-8c7d-1a2b3c4d5e6f")
+
+
+def _now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _truncate(text: str) -> str:
+    if not text:
+        return ""
+    return text if len(text) <= MAX_TEXT else text[:MAX_TEXT]
+
+
+def read_payload() -> dict:
+    """Read and parse the hook JSON from stdin; ``{}`` on any problem."""
+    try:
+        raw = sys.stdin.read()
+        if not raw.strip():
+            return {}
+        parsed = json.loads(raw)
+        return parsed if isinstance(parsed, dict) else {}
+    except Exception:
+        return {}
+
+
+def locate_transcript(payload: dict) -> tuple[str | None, Path | None, str | None]:
+    """Pull ``(session_id, transcript_path, cwd)`` from the hook payload."""
+    session_id = (
+        payload.get("session_id")
+        or payload.get("sessionId")
+        or (payload.get("session") or {}).get("id")
+    )
+    raw_path = (
+        payload.get("transcript_path")
+        or payload.get("transcriptPath")
+        or (payload.get("transcript") or {}).get("path")
+    )
+    cwd = payload.get("cwd") or payload.get("workspace") or None
+    path: Path | None = None
+    if raw_path:
+        try:
+            path = Path(raw_path).expanduser()
+        except Exception:
+            path = None
+    return session_id, path, cwd
+
+
+# ── transcript helpers (mirror Claude Code's JSONL shape) ──
+
+
+def _msg(entry: dict) -> dict:
+    m = entry.get("message")
+    return m if isinstance(m, dict) else {}
+
+
+def _role(entry: dict) -> str | None:
+    t = entry.get("type")
+    if t in ("user", "assistant"):
+        return t
+    r = _msg(entry).get("role")
+    return r if r in ("user", "assistant") else None
+
+
+def _content(entry: dict):
+    m = _msg(entry)
+    return m.get("content") if "message" in entry else entry.get("content")
+
+
+def _text(content) -> str:
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = []
+        for x in content:
+            if isinstance(x, dict) and x.get("type") == "text":
+                parts.append(x.get("text", ""))
+            elif isinstance(x, str):
+                parts.append(x)
+        return "\n".join(p for p in parts if p)
+    return ""
+
+
+def _blocks(content, block_type: str) -> list[dict]:
+    if not isinstance(content, list):
+        return []
+    return [x for x in content if isinstance(x, dict) and x.get("type") == block_type]
+
+
+def _attachment_from_image_block(block: dict) -> dict | None:
+    """Build an attachment dict from a Claude ``image`` content block.
+
+    Claude carries pasted images as ``{"type":"image","source":{"type":"base64",
+    "media_type":"image/png","data":"..."}}``. We capture the base64 bytes inline
+    when within the size cap; larger images are recorded metadata-only (no b64).
+    Returns ``None`` if the block carries no usable image data.
+    """
+    src = block.get("source")
+    if not isinstance(src, dict):
+        return None
+    mime = src.get("media_type") or "image/png"
+    name = block.get("name") or block.get("filename") or "pasted-image"
+    if src.get("type") == "base64":
+        data = src.get("data")
+        if not isinstance(data, str) or not data:
+            return None
+        # Authoritative size: decode once (cheap vs. the network cost we save).
+        try:
+            raw = base64.b64decode(data, validate=False)
+        except Exception:
+            return None
+        nbytes = len(raw)
+        att = {"kind": "image", "mime": mime, "name": str(name), "bytes": nbytes}
+        att["b64"] = data if nbytes <= MAX_ATTACHMENT_BYTES else None
+        return att
+    # URL-backed image (rare in transcripts): record metadata only.
+    if src.get("type") == "url" and src.get("url"):
+        return {"kind": "image", "mime": mime, "name": str(name), "bytes": 0, "b64": None}
+    return None
+
+
+def _mime_for_name(name: str) -> str:
+    """Best-effort mime from a filename extension; generic when unknown."""
+    lower = name.lower()
+    for ext, mime in _EXT_MIME.items():
+        if lower.endswith(ext):
+            return mime
+    return "application/octet-stream"
+
+
+def _attachment_from_file_block(block: dict) -> dict | None:
+    """Build a metadata attachment from a ``document``/file-ref content block.
+
+    Claude can carry document blocks (``{"type":"document","source":{...}}``) and
+    tool results sometimes reference files. We capture base64 ``document`` bytes
+    when present (within the cap); otherwise record the file name as metadata so
+    the trace at least shows that a file was attached.
+    """
+    src = block.get("source")
+    name = block.get("name") or block.get("title") or block.get("filename") or "attached-file"
+    name = str(name)
+    if isinstance(src, dict) and src.get("type") == "base64" and isinstance(src.get("data"), str):
+        data = src["data"]
+        mime = src.get("media_type") or _mime_for_name(name)
+        try:
+            raw = base64.b64decode(data, validate=False)
+        except Exception:
+            return None
+        nbytes = len(raw)
+        return {
+            "kind": "file",
+            "mime": mime,
+            "name": name,
+            "bytes": nbytes,
+            "b64": data if nbytes <= MAX_ATTACHMENT_BYTES else None,
+        }
+    # Bare reference with a name/path but no inline bytes: metadata only.
+    if block.get("name") or block.get("title") or block.get("filename"):
+        return {"kind": "file", "mime": _mime_for_name(name), "name": name, "bytes": 0, "b64": None}
+    return None
+
+
+def _extract_attachments(content) -> list[dict]:
+    """Pull image/file attachments from a message/tool-result content list.
+
+    Walks ``image`` and ``document`` content blocks (Claude's pasted-binary
+    shapes). Pure + fail-open: any malformed block is skipped, never raised, so
+    attachment capture can NEVER break the hook's core span extraction.
+    """
+    out: list[dict] = []
+    if not isinstance(content, list):
+        return out
+    for block in content:
+        if not isinstance(block, dict):
+            continue
+        try:
+            btype = block.get("type")
+            if btype == "image":
+                att = _attachment_from_image_block(block)
+            elif btype in ("document", "file"):
+                att = _attachment_from_file_block(block)
+            else:
+                att = None
+            if att:
+                out.append(att)
+        except Exception:
+            continue
+    return out
+
+
+def _is_tool_result(entry: dict) -> bool:
+    return _role(entry) == "user" and bool(_blocks(_content(entry), "tool_result"))
+
+
+def _usage(entry: dict) -> dict:
+    u = _msg(entry).get("usage")
+    if not isinstance(u, dict):
+        return {}
+    return {
+        "input": int(u.get("input_tokens") or 0),
+        "output": int(u.get("output_tokens") or 0),
+        "cache_read": int(u.get("cache_read_input_tokens") or 0),
+        "cache_write": int(u.get("cache_creation_input_tokens") or 0),
+    }
+
+
+def _ts(entry: dict) -> str | None:
+    v = entry.get("timestamp")
+    return v if isinstance(v, str) and v else None
+
+
+def read_new_entries(path: Path, state: dict, key: str) -> list[dict]:
+    """Return transcript entries appended since the last processed byte offset."""
+    if not path.exists():
+        return []
+    sess = state.get(key) or {}
+    offset = int(sess.get("offset", 0))
+    try:
+        size = path.stat().st_size
+        if size < offset:  # transcript rotated/truncated -> reprocess from start
+            offset = 0
+        with open(path, "rb") as fh:
+            fh.seek(offset)
+            chunk = fh.read()
+            new_offset = fh.tell()
+    except Exception:
+        return []
+    sess["offset"] = new_offset
+    state[key] = sess
+    out: list[dict] = []
+    for line in chunk.decode("utf-8", errors="replace").splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            obj = json.loads(line)
+            if isinstance(obj, dict):
+                out.append(obj)
+        except Exception:
+            continue
+    return out
+
+
+def build_spans(entries: list[dict]) -> tuple[list[dict], dict, list[str], str | None]:
+    """Turn transcript entries into canonical spans + rolled-up totals.
+
+    Each assistant message becomes one ``llm`` span (model + token usage); each
+    ``tool_use`` inside it becomes a child ``tool`` span. Returns
+    ``(spans, totals, models, repo)``.
+    """
+    # Map tool_use_id -> tool_result text for output enrichment, and
+    # tool_use_id -> attachments for any images a tool returned.
+    results_by_id: dict[str, str] = {}
+    attachments_by_tool_id: dict[str, list[dict]] = {}
+    # tool_use_id -> whether the tool_result reported a failure (Claude Code sets
+    # ``is_error: true`` on a failed tool result). Drives the tool span's status so
+    # the dashboard's error counts/rates reflect real failures instead of always 0.
+    errors_by_id: dict[str, bool] = {}
+    # tool_use_id -> ts of the message that returned the result. Used to give tool
+    # spans a real (assistant_ts -> result_ts) duration instead of zero.
+    result_ts_by_id: dict[str, str] = {}
+    for entry in entries:
+        if _is_tool_result(entry):
+            entry_ts = _ts(entry)
+            for tr in _blocks(_content(entry), "tool_result"):
+                tid = tr.get("tool_use_id")
+                if tid:
+                    out = tr.get("content")
+                    results_by_id[str(tid)] = (
+                        out if isinstance(out, str) else json.dumps(out, ensure_ascii=False)
+                    )
+                    errors_by_id[str(tid)] = bool(tr.get("is_error"))
+                    atts = _extract_attachments(out)
+                    if atts:
+                        attachments_by_tool_id[str(tid)] = atts
+                    if entry_ts:
+                        result_ts_by_id[str(tid)] = entry_ts
+
+    spans: list[dict] = []
+    models: list[str] = []
+    totals = {
+        "input_tokens": 0,
+        "output_tokens": 0,
+        "total_tokens": 0,
+        "cache_read_tokens": 0,
+        "cost_usd": 0.0,
+        "span_count": 0,
+        "tool_call_count": 0,
+    }
+
+    pending_user_text = ""
+    pending_user_attachments: list[dict] = []
+    # ts of the previous transcript entry; the LLM call started when the user
+    # prompt / tool result landed, finished when the assistant message appears.
+    prev_ts: str | None = None
+    for entry in entries:
+        role = _role(entry)
+        if role == "user" and not _is_tool_result(entry):
+            pending_user_text = _truncate(_text(_content(entry)))
+            # Images/files the user pasted into this turn ride along to the
+            # assistant span they prompted (accumulate across consecutive user
+            # messages until the next assistant generation consumes them).
+            pending_user_attachments.extend(_extract_attachments(_content(entry)))
+            prev_ts = _ts(entry) or prev_ts
+            continue
+        if role != "assistant":
+            prev_ts = _ts(entry) or prev_ts
+            continue
+
+        content = _content(entry)
+        usage = _usage(entry)
+        model = _msg(entry).get("model")
+        if model and model not in models:
+            models.append(model)
+        ts = _ts(entry) or _now_iso()
+        span_id = _msg(entry).get("id") or uuid.uuid4().hex
+        out_text = _truncate(_text(content))
+
+        in_tok = usage.get("input", 0)
+        out_tok = usage.get("output", 0)
+        cache_r = usage.get("cache_read", 0)
+        cache_w = usage.get("cache_write", 0)
+        llm_span = {
+            "span_id": str(span_id),
+            "parent_span_id": None,
+            "name": f"Claude generation ({model})" if model else "Claude generation",
+            "span_type": "llm",
+            "start_time": prev_ts or ts,
+            "end_time": ts,
+            "model": model,
+            "provider": "anthropic",
+            "input": pending_user_text or None,
+            "output": out_text or None,
+            "input_tokens": in_tok,
+            "output_tokens": out_tok,
+            "total_tokens": in_tok + out_tok,
+            "cache_read_tokens": cache_r,
+            "cache_write_tokens": cache_w,
+            "status": "ok",
+            "attributes": {"turn_role": "assistant"},
+        }
+        if pending_user_attachments:
+            llm_span["attachments"] = pending_user_attachments
+        spans.append(llm_span)
+        pending_user_text = ""  # consumed by this generation
+        pending_user_attachments = []  # consumed by this generation
+
+        totals["input_tokens"] += in_tok
+        totals["output_tokens"] += out_tok
+        totals["cache_read_tokens"] += cache_r
+        totals["total_tokens"] += in_tok + out_tok
+
+        for tu in _blocks(content, "tool_use"):
+            tid = str(tu.get("id") or uuid.uuid4().hex)
+            tname = tu.get("name") or "unknown"
+            tu_input = tu.get("input")
+            input_str = (
+                tu_input if isinstance(tu_input, str) else json.dumps(tu_input, ensure_ascii=False)
+            )
+            tool_span = {
+                "span_id": tid,
+                "parent_span_id": str(span_id),
+                "name": f"Tool: {tname}",
+                "span_type": "tool",
+                "start_time": ts,
+                "end_time": result_ts_by_id.get(tid, ts),
+                "input": _truncate(input_str),
+                "output": _truncate(results_by_id.get(tid, "")) or None,
+                "status": "error" if errors_by_id.get(tid) else "ok",
+                "attributes": {"tool_name": tname, "tool_id": tid},
+            }
+            tool_atts = attachments_by_tool_id.get(tid)
+            if tool_atts:
+                tool_span["attachments"] = tool_atts
+            spans.append(tool_span)
+            totals["tool_call_count"] += 1
+
+        prev_ts = ts
+
+    totals["span_count"] = len(spans)
+    return spans, totals, models, None
+
+
+def user_ref() -> str:
+    """Resolve the engineer's real identity for attribution.
+
+    Priority: an explicit ``OCTARIN_USER`` override → the Claude Code account
+    email (``~/.claude.json`` ``oauthAccount.emailAddress`` — the signed-in user)
+    → the git ``user.email`` → the OS username. We attribute to a real person
+    (matching ``backfill.py`` and the per-user ingest key) rather than an opaque
+    per-machine hash, so the dashboard shows who actually did the work. When the
+    request carries a per-user key the server overrides this with the key owner
+    anyway; a real identity here is what ANONYMOUS (slug-only) sends rely on.
+    """
+    ref = (os.environ.get("OCTARIN_USER") or "").strip()
+    if ref:
+        return ref
+    try:
+        with open(Path.home() / ".claude.json", encoding="utf-8") as fh:
+            account = json.load(fh).get("oauthAccount") or {}
+        email = (account.get("emailAddress") or "").strip()
+        if email:
+            return email
+    except Exception:
+        pass
+    try:
+        out = subprocess.check_output(
+            ["git", "config", "user.email"],
+            cwd=os.environ.get("CLAUDE_PROJECT_DIR") or os.getcwd(),
+            stderr=subprocess.DEVNULL,
+        )
+        email = out.decode().strip()
+        if email:
+            return email
+    except Exception:
+        pass
+    try:
+        return getpass.getuser()
+    except Exception:
+        return "unknown"
+
+
+def post_event(event: dict) -> bool:
+    """POST the IngestEvent. Returns True on 2xx, False otherwise (fail-open)."""
+    url = os.environ.get("OCTARIN_INGEST_URL")
+    if not url:
+        base = (os.environ.get("OCTARIN_API_BASE") or "").rstrip("/")
+        if not base:
+            return False
+        url = f"{base}/v1/ingest"
+    api_key = os.environ.get("OCTARIN_API_KEY", "")
+    body = json.dumps(event).encode("utf-8")
+    req = urllib.request.Request(url, data=body, method="POST")
+    req.add_header("Content-Type", "application/json")
+    if api_key:
+        req.add_header("Authorization", f"Bearer {api_key}")
+    try:
+        with urllib.request.urlopen(req, timeout=HTTP_TIMEOUT_S) as resp:
+            return 200 <= resp.status < 300
+    except Exception:
+        return False
+
+
+def load_state() -> dict:
+    try:
+        return json.loads(STATE_FILE.read_text(encoding="utf-8")) if STATE_FILE.exists() else {}
+    except Exception:
+        return {}
+
+
+def save_state(state: dict) -> None:
+    try:
+        STATE_DIR.mkdir(parents=True, exist_ok=True)
+        tmp = STATE_FILE.with_suffix(".tmp")
+        tmp.write_text(json.dumps(state, sort_keys=True), encoding="utf-8")
+        os.replace(tmp, STATE_FILE)
+    except Exception:
+        pass
+
+
+def build_event(payload: dict) -> dict | None:
+    """Assemble the canonical IngestEvent from a hook payload (or None to skip)."""
+    session_id, path, cwd = locate_transcript(payload)
+    if not session_id or path is None:
+        return None
+
+    state = load_state()
+    key = hashlib.sha256(f"{session_id}::{path}".encode()).hexdigest()
+    entries = read_new_entries(path, state, key)
+    save_state(state)
+    if not entries:
+        return None
+
+    spans, totals, models, _ = build_spans(entries)
+    if not spans:
+        return None
+
+    repo = Path(cwd).name if cwd else None
+    src_trace = f"{session_id}:{int(time.time())}"
+    trace_id = str(uuid.uuid5(_TRACE_NAMESPACE, f"{SOURCE}:{src_trace}"))
+    times = [s["start_time"] for s in spans]
+
+    return {
+        "trace_id": trace_id,
+        "source": SOURCE,
+        "session_id": session_id,
+        "user_ref": user_ref(),
+        "repo": repo,
+        "model": models[0] if models else None,
+        "spans": spans,
+        "start_time": min(times),
+        "end_time": max(times),
+        "total_tokens": totals["total_tokens"],
+        "input_tokens": totals["input_tokens"],
+        "output_tokens": totals["output_tokens"],
+        "cache_read_tokens": totals["cache_read_tokens"],
+        # extra (extra="allow"): handy for the backend rollup/audit
+        "totals": totals,
+        "models": models,
+    }
+
+
+def main() -> int:
+    try:
+        payload = read_payload()
+        event = build_event(payload)
+        if event is None:
+            return 0
+        post_event(event)
+    except Exception:
+        # Absolutely never let the hook break the host tool.
+        return 0
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())