Skip to content

Replay your coding agent sessions as a map and a timeline

fix(gemini): bound and harden the discovery scan

shubhook6d7ce90e2b4e
diff --git a/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/source.py b/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/source.py
index d17a336..9a7d8ae 100644
--- a/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/source.py
+++ b/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/source.py
@@ -24,6 +24,7 @@ from agentplay._core.ingestion.coding_agent._sources._base import (
MAX_SESSION_BYTES,
OversizedSessionError,
)
+from agentplay.discover import MAX_SOURCE_BYTES
from agentplay._core.ingestion.coding_agent._sources.local.parsers import (
AgentType,
detect_agent,
@@ -37,6 +38,9 @@ from agentplay._core.ingestion.coding_agent._sources.local.parsers import (
parse_kiro,
parse_opencode,
)
+from agentplay._core.ingestion.coding_agent._sources.local.parsers.gemini import (
+ _user_prompt as _gemini_user_prompt,
+)
from agentplay._core.ingestion.coding_agent.spans import DiscoveredSession, SpanData
logger = logging.getLogger(__name__)
@@ -1130,52 +1134,35 @@ def _gemini_header(head: bytes) -> dict | None:
return None
-# Injected as the first user message of every Gemini CLI session. Not a prompt.
-# The same prefix `parsers.gemini._user_prompt` ignores.
-_GEMINI_SESSION_CONTEXT_PREFIX = "<session_context>"
-
-
-def _gemini_user_counts(content: object) -> bool:
- """True when a `user` record carries prompt text, plus tool-result records.
-
- Empty text and the CLI's `<session_context>` block are ignored, the same
- filter as `parsers.gemini._user_prompt` (`text` must be non-empty and not
- that prefix). A tool-result record has no text parts and still counts.
- """
- if isinstance(content, str):
- texts = [content]
- elif isinstance(content, list):
- texts = [
- part["text"]
- for part in content
- if isinstance(part, dict) and isinstance(part.get("text"), str)
- ]
- else:
- return True
- if not texts:
- return True
- return any(text and not text.lstrip().startswith(_GEMINI_SESSION_CONTEXT_PREFIX) for text in texts)
-
-
def _gemini_has_messages(path: Path) -> bool:
"""True when the file holds a real prompt or a `gemini` message.
Streams one line at a time and stops at the first qualifying record, so a
- prompt that sits past `_read_head` (after a multi-MB `<session_context>`
- line) is still found. A file of only the header, that context block, or
- empty user text is skipped at EOF, same as `_kiro_event_count == 0`.
- The whole transcript is never held at once.
+ prompt that sits past `_read_head` is still found. Reads at most
+ ``MAX_SOURCE_BYTES``. A file of only the header or user content
+ ``_user_prompt`` rejects is skipped at EOF. A ``RecursionError`` from a
+ deeply nested record skips that line instead of aborting the index.
"""
+ remaining = MAX_SOURCE_BYTES
with path.open("rb") as handle:
- for line in handle:
+ while remaining > 0:
+ line = handle.readline(remaining)
+ if not line:
+ break
+ remaining -= len(line)
+ # `readline(n)` cuts a line at n bytes. A partial record is not JSON.
+ if remaining == 0 and not line.endswith(b"\n"):
+ break
try:
obj = json.loads(line)
- except ValueError:
+ except (ValueError, RecursionError):
continue
if not isinstance(obj, dict):
continue
kind = obj.get("type")
- if kind == "gemini" or (kind == "user" and _gemini_user_counts(obj.get("content"))):
+ if kind == "gemini" or (
+ kind == "user" and _gemini_user_prompt(obj.get("content")) is not None
+ ):
return True
return False
@@ -1946,6 +1933,12 @@ class LocalSource:
if size == 0:
self._record_discovery("gemini-cli", "empty_or_unreadable")
return None
+ # `build_map` refuses the same cap with SkipSession("too_large"). Skip
+ # here, before any read, so discovery does not allocate the file.
+ if size > MAX_SOURCE_BYTES:
+ logger.debug("gemini-cli: %s exceeds the source byte cap; session skipped", jsonl_file)
+ self._record_discovery("gemini-cli", "too_large")
+ return None
try:
head = _read_head(jsonl_file, stop_when=_agent_detected)
except OSError:
diff --git a/python/tests/core/test_gemini_parser.py b/python/tests/core/test_gemini_parser.py
index 2465922..df2cc08 100644
--- a/python/tests/core/test_gemini_parser.py
+++ b/python/tests/core/test_gemini_parser.py
@@ -21,6 +21,7 @@ import pytest
from agentplay import builder, discover
from agentplay._core.code_authorship import extract_edit_evidence
+from agentplay._core.ingestion.coding_agent._sources.local import source as local_source
from agentplay._core.ingestion.coding_agent._sources.local.parsers import (
AgentType,
detect_agent,
@@ -495,6 +496,94 @@ def test_prompt_past_the_head_is_indexed_and_empty_sessions_are_skipped(tmp_path
assert source.discovery_stats()["gemini-cli"]["skipped"]["empty_events"] == 2
+def _chats(tmp_path: Path) -> Path:
+ home = tmp_path / ".gemini"
+ chats = home / "tmp" / "project" / "chats"
+ chats.mkdir(parents=True)
+ (chats.parent / ".project_root").write_text(FIXTURE_ROOT, encoding="utf-8")
+ return chats
+
+
+def _user_record(msg_id: str, content: object) -> dict:
+ return {"id": msg_id, "timestamp": "2026-10-23T07:02:00Z", "type": "user", "content": content}
+
+
+def test_non_prompt_user_content_is_skipped_as_empty_events(tmp_path):
+ # Same shapes `_user_prompt` refuses: a non-text value, empty text, and
+ # the CLI's context block. Each file is only that record.
+ chats = _chats(tmp_path)
+ context = "<session_context>\nCONTEXT-PLACEHOLDER\n</session_context>"
+ cases = {
+ "number": 5,
+ "blank": "",
+ "empty-part": [{"text": ""}],
+ "context": [{"text": context}],
+ }
+ for session_id, content in cases.items():
+ (chats / f"session-{session_id}.jsonl").write_bytes(
+ _session(_user_record("u0", content), session_id=session_id)
+ )
+ source = LocalSource(claude_code_homes=[], codex_homes=[], gemini_homes=[chats.parents[2]])
+ index = source._build_index()
+ assert set(index) == set()
+ assert source.discovery_stats()["gemini-cli"]["skipped"]["empty_events"] == len(cases)
+
+
+def test_deeply_nested_json_line_does_not_abort_discovery(tmp_path, monkeypatch):
+ nested = b'{"a":' * 1200 + b"0" + b"}" * 1200
+ try:
+ json.loads(nested)
+ except RecursionError:
+ pass
+ else:
+ # Python 3.15 parses this depth. Older interpreters raise here, which
+ # is the failure the scan has to survive, so reproduce it.
+ real_loads = local_source.json.loads
+
+ def loads(raw, *args, **kwargs):
+ text = raw.decode("utf-8", "replace") if isinstance(raw, (bytes, bytearray)) else raw
+ if isinstance(text, str) and text.startswith('{"a":' * 40):
+ raise RecursionError("maximum recursion depth exceeded")
+ return real_loads(raw, *args, **kwargs)
+
+ monkeypatch.setattr(local_source.json, "loads", loads)
+
+ chats = _chats(tmp_path)
+ prompt = _user_record("u1", [{"text": "PROMPT-PLACEHOLDER-1"}])
+ prompt_line = json.dumps(prompt).encode() + b"\n"
+ header = json.dumps(
+ {"sessionId": "nested-then-prompt", "projectHash": "h", "startTime": "2026-10-23T07:00:00Z", "kind": "main"}
+ ).encode() + b"\n"
+ (chats / "session-nested-then-prompt.jsonl").write_bytes(header + nested + b"\n" + prompt_line)
+ only = header.replace(b"nested-then-prompt", b"nested-only")
+ (chats / "session-nested-only.jsonl").write_bytes(only + nested + b"\n")
+
+ source = LocalSource(claude_code_homes=[], codex_homes=[], gemini_homes=[chats.parents[2]])
+ index = source._build_index()
+ assert "local:gemini-cli:nested-then-prompt" in index
+ assert "local:gemini-cli:nested-only" not in index
+ assert source.discovery_stats()["gemini-cli"]["skipped"]["empty_events"] == 1
+
+
+def test_oversized_transcript_is_skipped_as_too_large(tmp_path, monkeypatch):
+ # The real cap is 512MB. Lower it so the test file is over the cap without
+ # writing one. The prompt sits at the start, so a scan that ignored the
+ # size and stopped at the cap would still index this session.
+ monkeypatch.setattr(local_source, "MAX_SOURCE_BYTES", 800)
+ chats = _chats(tmp_path)
+ prompt = _user_record("u1", [{"text": "PROMPT-PLACEHOLDER-1"}])
+ (chats / "session-small.jsonl").write_bytes(_session(prompt, session_id="small"))
+ huge = _session(prompt, {"id": "pad", "type": "info", "content": "x" * 2000}, session_id="huge")
+ assert len(huge) > 800
+ (chats / "session-huge.jsonl").write_bytes(huge)
+
+ source = LocalSource(claude_code_homes=[], codex_homes=[], gemini_homes=[chats.parents[2]])
+ index = source._build_index()
+ assert "local:gemini-cli:small" in index
+ assert "local:gemini-cli:huge" not in index
+ assert source.discovery_stats()["gemini-cli"]["skipped"]["too_large"] == 1
+
+
def test_local_source_indexes_gemini_sessions(tmp_path):
source = LocalSource(claude_code_homes=[], codex_homes=[], gemini_homes=[_gemini_home(tmp_path)])
index = source._build_index()