[BREAKING][core] refactor: centralize distributed session logging

[yyDing1]

What does this PR do?

This PR centralizes logging ownership at execution entry points and makes per-session logging reliable across inline and Ray-dispatched runners.

Major Changes

• Introduces explicit LogContext propagation across process boundaries.
• Writes Task/Agent/Tool logs to task.log.
• Writes Framework logs to framework.log for Ray-dispatched sessions.
• Stores logs and trajectory artifacts under:
  • <log_dir>/step_<global_step>/<session_id>/
  • <log_dir>/<session_id>/ when global_steps is None
• Adds descriptive session IDs such as session-sample-0-rollout-1-<uuid>.
• Redacts API keys, bearer tokens, base URLs, and endpoint URLs from file and console logs.
• Removes implicit Task-owned logging, TaskConfig.log_dir, and Task.episode_logging().
• Adds per-rollout logging to API inference and verification entry points.
• Adds --log-dir and global --concurrency support to inference scripts.
• Removes the legacy dashboard implementation.
• Improves Claude Code execution:
  • Falls back to the native installer when npm is unavailable.
  • Forwards external API credentials through ANTHROPIC_AUTH_TOKEN.
• Accepts JSON-encoded view_range values in str_replace_editor.
• Updates Quickstart and Concepts documentation with API and verl inference recipes.

API and Usage Example

from uni_agent.logging import LogContext, sample_logging

log_context = LogContext(
    log_id="sample-0-rollout-0",
    log_path="/mnt/shared/logs/sample-0-rollout-0/task.log",
)

async with sample_logging.from_context(log_context):
    await task.run()

Framework logging is configured through:

actor_rollout_ref:
  rollout:
    custom:
      agent_framework:
        log_dir: /mnt/shared/uni_agent_logs

Set log_dir to an empty value to disable file logging.

Breaking Changes

• run.log is renamed to task.log.
• The following logging APIs are removed:
  • current_run_id
  • get_logger
  • add_file_handler
  • cleanup_handlers
  • install_console_sink
• TaskConfig.log_dir and Task.episode_logging() are removed.
• Consumers of the deleted dashboard or old log paths must migrate.

Test

PYTHONPATH=.:verl python -m pytest -q tests/uni_agent/framework

PYTHONPATH=. python -m pytest -q \
  tests/uni_agent/agents/test_claude_code_agent.py \
  tests/uni_agent/logging/test_redaction.py \
  tests/uni_agent/tasks/test_parallel_infer_api.py \
  tests/uni_agent/tools/test_edit_file_arguments.py \
  tests/uni_agent/tools/test_toolbox_call_errors.py

mkdocs build --strict

[gemini-code-assist]

Code Review

This pull request removes the lightweight streaming dashboard and refactors the logging system to use a centralized LogContext and log IDs instead of run IDs. It also introduces a log redaction module to hide sensitive credentials, removes log_dir from TaskConfig, updates the Claude Code agent to support native installation when npm is missing, and enhances argument parsing for the file editor tool. The review feedback highlights a few critical issues: a class-level Semaphore in parallel_verify_swe.py that could raise a RuntimeError outside of an active event loop, a regex pattern in the redaction module that fails to handle escaped quotes, and unhandled command failures in the SWE-bench task execution when applying gold patches.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Declaring _semaphore = asyncio.Semaphore(...) as a class-level attribute causes it to be instantiated at import time. In Python 3.10+, this will raise a RuntimeError: no running event loop if the module is imported outside of an active event loop (which is typical for CLI entry points).

To fix this, remove the class-level _semaphore attribute and initialize it inside __init__ where the event loop is guaranteed to be running.

Suggested change

	def __init__(self, log_dir: str | None):
	self.log_dir = log_dir
	def __init__(self, log_dir: str | None):
	self.log_dir = log_dir
	self._semaphore = asyncio.Semaphore(max(1, GLOBAL_CONCURRENCY // NUM_WORKERS))

[gemini-code-assist]

The non-greedy pattern (?P<value>.*?) followed by (?P=quote) will stop matching at the first occurrence of the quote character, even if it is escaped (e.g., \" inside a double-quoted string). This can result in incomplete redaction of sensitive values that contain escaped quotes.

To handle escaped quotes correctly, use a pattern that explicitly matches escaped characters, such as:
rf"(?P<quote>[\"'])(?P<value>(?:(?!(?P=quote))[^\]|\.)*)(?P=quote)"

Suggested change

	rf"(?P<quote>[\"'])(?P<value>.*?)(?P=quote)",
	rf"(?P<quote>[\"'])(?P<value>(?:(?!(?P=quote))[^\\]|\\.)*)(?P=quote)",

[gemini-code-assist]

sandbox.exec returns an ExecResult and does not raise an exception if the command fails (returns a non-zero exit code). If git apply fails to apply the gold patch, the task will silently proceed to evaluate the unpatched codebase, leading to incorrect evaluation results.

Consider checking the exit code of the git apply command and raising an error if it fails.

Suggested change

	if cfg.run_gold_patch:
	logger.info("applying gold patch to /testbed")
	await sandbox.write_file("/tmp/gold_patch.patch", sample["patch"])
	await sandbox.exec(["git", "apply", "--whitespace=fix", "/tmp/gold_patch.patch"], workdir="/testbed")
	if cfg.run_gold_patch:
	logger.info("applying gold patch to /testbed")
	await sandbox.write_file("/tmp/gold_patch.patch", sample["patch"])
	res = await sandbox.exec(["git", "apply", "--whitespace=fix", "/tmp/gold_patch.patch"], workdir="/testbed")
	if res.exit_code != 0:
	raise RuntimeError(f"Failed to apply gold patch: {res.stderr}")

[gemini-code-assist]

sandbox.exec returns an ExecResult and does not raise an exception if the command fails. If git apply fails to apply the gold patch, the task will silently proceed to evaluate the unpatched codebase, leading to incorrect evaluation results.

Consider checking the exit code of the git apply command and raising an error if it fails.

Suggested change

	if cfg.run_gold_patch:
	logger.info("applying gold patch to /testbed")
	await sandbox.write_file("/tmp/gold_patch.patch", sample["patch"])
	await sandbox.exec(["git", "apply", "--whitespace=fix", "/tmp/gold_patch.patch"], workdir="/testbed")
	if cfg.run_gold_patch:
	logger.info("applying gold patch to /testbed")
	await sandbox.write_file("/tmp/gold_patch.patch", sample["patch"])
	res = await sandbox.exec(["git", "apply", "--whitespace=fix", "/tmp/gold_patch.patch"], workdir="/testbed")
	if res.exit_code != 0:
	raise RuntimeError(f"Failed to apply gold patch: {res.stderr}")