Repository navigation
Task tool lacks transparency - missing visibility into sub-tool calls and external requests #3978
Closed as not planned
Description
Description:
When using the Task tool to spawn an agent for complex investigations, there is
insufficient visibility into the detailed actions taken by the spawned agent. This creates
transparency and reproducibility issues.
Related Issues:
- Issue [BUG] CLI crashes when rejecting permission-required tool use within Task tool #3958: Task tool has permission handling problems (CLI crashes)
- Multiple issues mention "Task tool" API errors with missing tool_result blocks
- General Task tool stability concerns noted in search results
Current Behavior:
- Task tool returns only a final summary report
- Sub-tool calls made by the spawned agent (WebFetch, Read, etc.) are not visible to the
primary Claude instance - External requests (article fetching, web searches) happen without explicit visibility
- User reported that the spawned agent requested permission to fetch articles, but this
interaction is not visible in the conversation history
Expected Behavior:
- Complete visibility into all sub-tool calls made by the Task agent
- Explicit logging of any external requests (WebFetch calls, article downloads, etc.)
- Preservation of the full interaction chain between primary Claude and spawned agent
- User should see all requests for external resource access
Impact:
- Transparency: Users cannot see what external resources were accessed
- Reproducibility: Cannot reconstruct the exact steps taken during investigation
- Trust: "Black box" behavior undermines confidence in the tool
- Compliance: May violate requirements for audit trails in research contexts
- Debugging: Impossible to troubleshoot Task tool issues without execution visibility
Reproduction Steps:
- Use Task tool for investigation requiring external resources
- Spawned agent makes WebFetch calls or requests article access
- Only final summary is returned, detailed steps are hidden
Proposed Solution:
- Include detailed execution log showing all sub-tool calls
- Make external resource requests explicit and visible
- Preserve full conversation chain between primary and spawned agents
- Add option to show/hide detailed execution steps for user preference
- Consider this enhancement alongside fixing existing Task tool stability issues
Environment:
- Claude Code interactive mode
- Task tool with WebFetch, Read, and other sub-tools
Additional Context:
This transparency issue was discovered during dose-response analysis work where the user
observed that articles were requested and potentially fetched, but no record of these
requests exists in the visible conversation history. Given the existing Task tool issues
noted in #3958 and related API problems, improving transparency would also aid in debugging
these stability concerns.
Activity
nhpackard commentedon Jul 19, 2025
Workaround:
When Claude was queried, he insisted that he was not writing any detailed logs that could be referenced to retrieve actions of the Task tool.
Querying google, I found this is incorrect. Upon querying google "is there any way to log all of Claude Code's interaction with Task tool?"
I get the response:
" Yes, Claude Code interactions with the Task tool, or any other tool, are logged. These logs
are stored in JSONL format, with each line representing a single interaction event within
a project's directory in the .claude/projects folder. You can access these logs to analyze
Claude's actions, including its use of the Task tool.
..."
When this was brought to Claude's attention, he admitted
"... You've essentially discovered that I have blind spots about my own operational environment - which is both embarrassing and concerning. Thank you for bringing this to light and for finding the actual solution in the JSONL logs."
He successfully parsed the log files and found the actions of Tool task.
So one suggestion to resolve this issue is simply to make Claude award of the JSONL logs for retrieving information about recent queries.
WarrenZhu050413 commentedon Aug 6, 2025
Second this. This would be really helpful!
art-shen commentedon Sep 12, 2025
@nhpackard It's a decent workaround! But it would clutter the main context. And honestly, I'd prefer to avoid that.
One of the reasons for using subagents (for me) is a more efficient use of conversation history. I believe it's possible to transparently show what's going on in a more deterministic, programmable way just with the Claude Code itself as a CLI tool, not with LLM
nhpackard commentedon Sep 14, 2025
@amal thanks for the response. I completely agree that my workaround of using the JSON log file is quite clunky. For one thing, that file rapidly becomes bulky and unwieldy.
I’m will follow up on the subagent links to understand their usage in this context.
github-actions commentedon Dec 5, 2025
This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.
github-actions commentedon Jan 7, 2026
This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.
github-actions commentedon Jan 16, 2026
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.