Replay your coding agent sessions as a map and a timeline
Python agent-observabilityai-agentsclaudeclaude-codecodexcoding-agentscursordeveloper-toolsdevtoolsgithub-copilotkirollmobservabilityopencodeprivacypythonsession-replaytypescriptvisualization
fix(gemini): only trust the CLI's exit verdict
shubhook1263207d6a37
diff --git a/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/parsers/gemini.py b/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/parsers/gemini.pyindex ceda4b0..07bcace 100644--- a/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/parsers/gemini.py+++ b/python/src/agentplay/_core/ingestion/coding_agent/_sources/local/parsers/gemini.py@@ -36,7 +36,11 @@ A tool call records `status` (`success` | `error` | `cancelled`). The shelltool reports `success` even when the command fails, and appends an`Exit Code: <n>` line to its output only when n is non-zero (and a `Signal:`line only when a signal ended the process). Those lines are the failure-signal for shell calls.+signal for shell calls. The CLI writes `Output: <program output>` first, then+its labels in a fixed order: `Error:`, `Exit Code:`, `Signal:`,+`Background PIDs:`, `Process Group PGID:`. Only that trailing block is read,+so a label line the program printed is not taken as the verdict unless it is+the last line of output with no CLI label after it to tell them apart.# Tokens@@ -76,10 +80,15 @@ _SESSION_CONTEXT_PREFIX = "<session_context>"_SHELL_TOOL = "run_shell_command"-# Full-line matches: the command's own output comes first in the result text,-# so the LAST match is the CLI's label, never a line the program printed.-_RE_EXIT_CODE = re.compile(r"^Exit Code: (-?\d+)\s*$", re.MULTILINE)-_RE_SIGNAL = re.compile(r"^Signal: (\S+)\s*$", re.MULTILINE)+# The labels the CLI may append after `Output: ...`, last first. The result is+# read from its last line upward, each label at most once and only in this+# order, so a label line inside the program's output ends the scan.+_RE_TRAILER = (+ ("pgid", re.compile(r"^Process Group PGID: \S+\s*$")),+ ("pids", re.compile(r"^Background PIDs: .*$")),+ ("signal", re.compile(r"^Signal: (\S+)\s*$")),+ ("exit", re.compile(r"^Exit Code: (-?\d+)\s*$")),+)def parse_gemini(@@ -391,11 +400,26 @@ def _failed(raw_name: str, response: dict) -> bool:output = response.get("output")if not isinstance(output, str):return False- codes = _RE_EXIT_CODE.findall(output)- if codes and codes[-1] != "0":+ verdict = _shell_trailer(output)+ if verdict.get("exit", "0") != "0":return True- signals = _RE_SIGNAL.findall(output)- return bool(signals and signals[-1] not in ("(none)", "0"))+ return verdict.get("signal", "(none)") not in ("(none)", "0")+++def _shell_trailer(output: str) -> dict[str, str]:+ """The CLI's own labels at the end of a shell result, keyed by label."""+ lines = output.split("\n")+ found: dict[str, str] = {}+ slot = 0+ # Line 0 is `Output: ...`, written by the CLI ahead of any label.+ while len(lines) > 1 and slot < len(_RE_TRAILER):+ key, pattern = _RE_TRAILER[slot]+ slot += 1+ m = pattern.match(lines[-1])+ if m:+ found[key] = m.group(1) if m.groups() else ""+ lines.pop()+ return found__all__ = ["parse_gemini"]diff --git a/python/tests/core/test_gemini_parser.py b/python/tests/core/test_gemini_parser.pyindex 5947e4a..b3827f3 100644--- a/python/tests/core/test_gemini_parser.py+++ b/python/tests/core/test_gemini_parser.py@@ -196,6 +196,47 @@ def test_exit_code_printed_by_the_program_is_not_the_verdict():assert tool.status_code == "OK"+def _shell_status(output: str) -> str:+ prompt = {"id": "u1", "timestamp": "2026-10-23T07:00:00Z", "type": "user", "content": [{"text": "go"}]}+ reply = {+ "id": "g1",+ "timestamp": "2026-10-23T07:00:01Z",+ "type": "gemini",+ "toolCalls": [_call("run_shell_command", "success", {"output": output}, command="cat log")],+ }+ (tool,) = _tools(parse_gemini(_session(prompt, reply), "t")[1])+ return tool.status_code+++ "output",+ [+ # The program prints a standalone label line, the command succeeds, and+ # the CLI appends no failure label of its own.+ "Output: Exit Code: 3\nExit Code: 3\ndone\nProcess Group PGID: 1",+ "Output: start\nSignal: SIGTERM\ndone\nProcess Group PGID: 1",+ "Output: Exit Code: 3\nSignal: SIGTERM\ndone",+ # The label sits before the CLI's own `Error:` line, so it is output.+ "Output: Exit Code: 3\nError: spawn failed\nProcess Group PGID: 1",+ ],+)+def test_label_lines_in_program_output_are_not_the_verdict(output):+ assert _shell_status(output) == "OK"+++ "output",+ [+ "Output: x\nExit Code: 1\nProcess Group PGID: 1",+ "Output: x\nError: boom\nExit Code: 127",+ "Output: (empty)\nSignal: SIGTERM\nBackground PIDs: 7, 8\nProcess Group PGID: 1",+ "Output: Exit Code: 0\nExit Code: 2\nSignal: 15\nProcess Group PGID: 1",+ ],+)+def test_the_clis_trailing_verdict_marks_failure(output):+ assert _shell_status(output) == "ERROR"++@pytest.mark.parametrize("data",[