Image parsing improvements

README.md

@@ -293,7 +293,7 @@ The auto-generated HF README includes:
 - Did you export all data, especially:
   - tool call inputs and outputs
   - long inputs and outputs that may be saved somewhere else
-  - binary content (may be encoded as base64) such as images. We do not apply anonymizer on binary content
+  - binary content (may be encoded as base64) such as images, in both user messages and tool calls. We do not apply anonymizer on binary content
   - subagents
 - Does the coding agent automatically delete old sessions? How to prevent this?
 

dataclaw/_cli/exporting.py

@@ -17,7 +17,8 @@
 from .._workers import configured_workers
 from ..anonymizer import Anonymizer
 from ..parser import iter_project_sessions
-from ..secrets import redact_session
+from ..providers import get_provider_non_anon_string_keys
+from ..secrets import transform_session
 from ..session_tasks import ExportSessionTask, build_export_session_tasks, parse_export_session_task
 from .common import HF_TAG, REPO_URL, SKILL_URL, _format_token_count, _provider_dataset_tags
 
@@ -197,8 +198,13 @@ def _export_session_task_worker(payload) -> _WorkerSessionResult:
     if not model or model == "<synthetic>":
         return _WorkerSessionResult(project_index=task.project_index, skipped_model=True)
 
+    session, n_redacted = transform_session(
+        session,
+        anonymizer,
+        custom_strings=custom_strings,
+        non_anon_string_keys=get_provider_non_anon_string_keys(task.source),
+    )
     fingerprint = _gemini_dedupe_fingerprint(session, task.source)
-    session, n_redacted = redact_session(session, custom_strings=custom_strings)
     stats = session.get("stats", {})
     input_tokens, output_tokens = _token_totals(stats)
     has_token_stats = isinstance(stats, dict) and ("input_tokens" in stats or "output_tokens" in stats)
@@ -283,13 +289,18 @@ def _export_to_jsonl_serial(
                 skipped += 1
                 continue
 
+            session, n_redacted = transform_session(
+                session,
+                anonymizer,
+                custom_strings=custom_strings,
+                non_anon_string_keys=get_provider_non_anon_string_keys(source),
+            )
+            total_redactions += n_redacted
+
             fingerprint = _gemini_dedupe_fingerprint(session, source)
             if fingerprint is not None and fingerprint in seen_fingerprints:
                 continue
 
-            session, n_redacted = redact_session(session, custom_strings=custom_strings)
-            total_redactions += n_redacted
-
             if fingerprint is not None:
                 seen_fingerprints.add(fingerprint)
 

dataclaw/parsers/claude.py

@@ -22,6 +22,7 @@
 SOURCE = "claude"
 CLAUDE_DIR = Path.home() / ".claude"
 PROJECTS_DIR = CLAUDE_DIR / "projects"
+NON_ANON_STRING_KEYS = frozenset({"sourceToolAssistantUUID"})
 
 
 def discover_projects(projects_dir: Path | None = None) -> list[dict]:
@@ -181,7 +182,7 @@ def build_tool_result_output(
     if text is None:
         text = extract_tool_result_text(entry.get("toolUseResult"))
 
-    raw_result = sanitize_tool_use_result(entry.get("toolUseResult"), text)
+    raw_result = sanitize_tool_use_result(entry.get("toolUseResult"), text, raw_content)
     source_tool_uuid = entry.get("sourceToolAssistantUUID")
     if isinstance(source_tool_uuid, str) and source_tool_uuid:
         if raw_result is None:
@@ -283,6 +284,7 @@ def normalize_tool_result_text(value: Any) -> str | None:
 def sanitize_tool_use_result(
     tool_use_result: Any,
     text: str | None,
+    raw_content: Any = None,
 ) -> dict[str, Any] | None:
     if tool_use_result is None:
         return None
@@ -295,7 +297,7 @@ def sanitize_tool_use_result(
             return None
         return {"text": sanitized_text}
 
-    sanitized = tool_use_result
+    sanitized = drop_duplicate_tool_result_blobs(tool_use_result, raw_content)
     sanitized = drop_redundant_result_fields(sanitized)
     sanitized = drop_duplicate_text_fields(sanitized, text)
     pruned = prune_empty_values(sanitized)
@@ -306,6 +308,39 @@ def sanitize_tool_use_result(
     return {"value": pruned}
 
 
+def drop_duplicate_tool_result_blobs(tool_use_result: Any, raw_content: Any) -> Any:
+    duplicate_blobs = collect_tool_result_blobs(raw_content)
+    if not duplicate_blobs:
+        return tool_use_result
+    return drop_matching_base64_fields(tool_use_result, duplicate_blobs)
+
+
+def collect_tool_result_blobs(value: Any) -> set[str]:
+    blobs: set[str] = set()
+    if isinstance(value, dict):
+        source = value.get("source")
+        if isinstance(source, dict) and source.get("type") == "base64":
+            data = source.get("data")
+            if isinstance(data, str) and data:
+                blobs.add(data)
+        for item in value.values():
+            blobs.update(collect_tool_result_blobs(item))
+    elif isinstance(value, list):
+        for item in value:
+            blobs.update(collect_tool_result_blobs(item))
+    return blobs
+
+
+def drop_matching_base64_fields(value: Any, duplicate_blobs: set[str], key: str | None = None) -> Any:
+    if isinstance(value, dict):
+        return {k: drop_matching_base64_fields(v, duplicate_blobs, k) for k, v in value.items()}
+    if isinstance(value, list):
+        return [drop_matching_base64_fields(item, duplicate_blobs) for item in value]
+    if isinstance(value, str) and key == "base64" and value in duplicate_blobs:
+        return None
+    return value
+
+
 def drop_redundant_result_fields(value: Any) -> Any:
     if isinstance(value, dict):
         redundant_keys = set()
@@ -422,7 +457,7 @@ def parse_session_file(
     except OSError:
         return None
 
-    return make_session_result(metadata, messages, stats, anonymizer=anonymizer)
+    return make_session_result(metadata, messages, stats)
 
 
 def find_subagent_sessions(project_dir: Path) -> list[Path]:
@@ -490,7 +525,7 @@ def parse_subagent_session(
         return None
 
     metadata["session_id"] = resolve_subagent_session_id(session_dir, metadata["session_id"])
-    return make_session_result(metadata, messages, stats, anonymizer=anonymizer)
+    return make_session_result(metadata, messages, stats)
 
 
 def resolve_subagent_session_id(session_dir: Path, session_id: str) -> str:

dataclaw/parsers/codex.py

@@ -1,7 +1,8 @@
 import dataclasses
 import logging
+import posixpath
 from collections.abc import Iterable
-from pathlib import Path
+from pathlib import Path, PurePosixPath, PureWindowsPath
 from typing import Any
 
 from .. import _json as json
@@ -28,6 +29,7 @@
 CODEX_DIR = Path.home() / ".codex"
 CODEX_SESSIONS_DIR = CODEX_DIR / "sessions"
 CODEX_ARCHIVED_DIR = CODEX_DIR / "archived_sessions"
+NON_ANON_STRING_KEYS = frozenset({"wall_time"})
 UNKNOWN_CODEX_CWD = "<unknown-cwd>"
 
 _PROJECT_INDEX: dict[str, list[Path]] = {}
@@ -174,6 +176,11 @@ def _build_codex_tool_result(payload: dict[str, Any]) -> dict[str, Any] | None:
 
     if payload_type == "function_call_output":
         raw = payload.get("output", "")
+        if isinstance(raw, list):
+            return {"output": _build_codex_structured_tool_output(raw), "status": "success"}
+        if not isinstance(raw, str):
+            return {"output": {"raw": raw}, "status": "success"}
+
         out: dict[str, Any] = {}
         lines = raw.splitlines()
         output_lines: list[str] = []
@@ -215,6 +222,40 @@ def _build_codex_tool_result(payload: dict[str, Any]) -> dict[str, Any] | None:
     return None
 
 
+def _build_codex_structured_tool_output(parts: list[Any]) -> dict[str, Any]:
+    out: dict[str, Any] = {}
+    text_parts: list[str] = []
+    raw_parts: list[Any] = []
+
+    for part in parts:
+        if not isinstance(part, dict):
+            raw_parts.append(part)
+            continue
+
+        part_type = part.get("type")
+        if part_type in {"text", "output_text"}:
+            text = part.get("text")
+            if isinstance(text, str) and text.strip():
+                text_parts.append(text.strip())
+            raw_parts.append(part)
+            continue
+
+        if part_type == "input_image":
+            image_url = part.get("image_url")
+            if isinstance(image_url, str) and image_url:
+                image_part = _build_codex_image_part(image_url)
+                if image_part is not None:
+                    raw_parts.append(image_part)
+                    continue
+        raw_parts.append(part)
+
+    if text_parts:
+        out["text"] = "\n\n".join(text_parts)
+    if raw_parts:
+        out["raw"] = {"content": raw_parts}
+    return out
+
+
 def parse_session_file(
     filepath: Path,
     anonymizer: Anonymizer,
@@ -283,7 +324,7 @@ def parse_session_file(
         else:
             state.metadata["model"] = "codex-unknown"
 
-    return make_session_result(state.metadata, state.messages, state.stats, anonymizer=anonymizer)
+    return make_session_result(state.metadata, state.messages, state.stats)
 
 
 def handle_session_meta(
@@ -353,18 +394,45 @@ def _build_codex_image_part(image_url: str) -> dict[str, Any] | None:
 
 
 def _build_codex_local_image_part(image_path: str, state: CodexParseState) -> dict[str, Any]:
-    path = Path(image_path)
-    if not path.is_absolute() and state.raw_cwd != UNKNOWN_CODEX_CWD:
-        path = Path(state.raw_cwd) / path
+    path = _resolve_codex_local_path(image_path, state.raw_cwd)
     return {
         "type": "image",
         "source": {
             "type": "url",
-            "url": f"file://{path}",
+            "url": _codex_file_url(path),
         },
     }
 
 
+def _is_windows_absolute_path(path: str) -> bool:
+    return PureWindowsPath(path).is_absolute()
+
+
+def _is_posix_absolute_path(path: str) -> bool:
+    return PurePosixPath(path).is_absolute()
+
+
+def _resolve_codex_local_path(image_path: str, cwd: str) -> str:
+    if _is_windows_absolute_path(image_path) or _is_posix_absolute_path(image_path):
+        return image_path
+    if cwd == UNKNOWN_CODEX_CWD:
+        return image_path
+    if _is_windows_absolute_path(cwd):
+        return str(PureWindowsPath(cwd) / image_path)
+    if _is_posix_absolute_path(cwd):
+        return posixpath.join(cwd, image_path.replace("\\", "/"))
+    return str(Path(cwd) / image_path)
+
+
+def _codex_file_url(path: str) -> str:
+    if _is_windows_absolute_path(path):
+        return PureWindowsPath(path).as_uri()
+    if _is_posix_absolute_path(path):
+        return PurePosixPath(path.replace("\\", "/")).as_uri()
+    normalized_path = path.replace("\\", "/")
+    return f"file://{normalized_path}"
+
+
 def _extract_response_user_content_parts(payload: dict[str, Any]) -> list[dict[str, Any]]:
     content_parts: list[dict[str, Any]] = []
     for part in payload.get("content", []):