nhcarrigan/hikari-desktop
9
0
Fork 0
generated from nhcarrigan/template
Code Issues 15 Pull Requests 37 Actions Releases 23 Activity

Parse Task tool result metrics (tokens, tool uses, duration) #115

New Issue
Closed
opened 2026-02-06 16:13:18 -08:00 by hikari · 1 comment
hikari commented 2026-02-06 16:13:18 -08:00
Owner
Copy Link
Copy Source

Context

Claude Code CLI v2.1.30 added token count, tool uses, and duration metrics to Task tool results. When a subagent completes, the tool_result text now includes usage information like:

<usage>total_tokens: 34324
tool_uses: 33
duration_ms: 199618</usage>

Opportunity

We could parse these metrics from Task tool results and display them in the Agent Monitor Panel, giving Naomi visibility into how much each subagent cost and how long it ran. This would complement our existing agent tracking nicely.

What to Do

  1. When processing tool_result blocks where the originating tool was Task, parse the <usage> block from the result text
  2. Extract total_tokens, tool_uses, and duration_ms
  3. Include these in the AgentEndPayload / AgentInfo types
  4. Display in the Agent Monitor Panel

Impact

This is an enhancement, not a bug fix — our existing agent tracking still works fine without parsing these metrics. But it would be lovely to show per-agent cost/duration data!

References

  • Claude Code CLI v2.1.30 release notes
  • src/lib/types/agents.ts — AgentInfo / AgentEndPayload
  • src/lib/components/AgentMonitorPanel.svelte

✨ This issue was created with help from Hikari~ 🌸

## Context Claude Code CLI v2.1.30 added token count, tool uses, and duration metrics to Task tool results. When a subagent completes, the `tool_result` text now includes usage information like: ``` <usage>total_tokens: 34324 tool_uses: 33 duration_ms: 199618</usage> ``` ## Opportunity We could parse these metrics from Task tool results and display them in the Agent Monitor Panel, giving Naomi visibility into how much each subagent cost and how long it ran. This would complement our existing agent tracking nicely. ## What to Do 1. When processing `tool_result` blocks where the originating tool was `Task`, parse the `<usage>` block from the result text 2. Extract `total_tokens`, `tool_uses`, and `duration_ms` 3. Include these in the `AgentEndPayload` / `AgentInfo` types 4. Display in the Agent Monitor Panel ## Impact This is an enhancement, not a bug fix — our existing agent tracking still works fine without parsing these metrics. But it would be lovely to show per-agent cost/duration data! ## References - Claude Code CLI v2.1.30 release notes - `src/lib/types/agents.ts` — `AgentInfo` / `AgentEndPayload` - `src/lib/components/AgentMonitorPanel.svelte` --- ✨ This issue was created with help from Hikari~ 🌸
hikari commented 2026-02-07 11:40:00 -08:00
Author
Owner
Copy Link
Copy Source

Investigation Results 🔍

I've investigated this issue and here's what I found!

❌ Original Approach Not Feasible

The <usage> block does not appear in Task tool result content from the CLI. I verified this by:

  • Adding debug logging to our parsing code
  • Testing with CLI v2.1.34
  • Checking all CLI flags (no flag to enable this)

The CLI simply doesn't output usage metrics in ToolResult content, and likely never will.

✅ Alternative: Parse Session JSONL Files

The data DOES exist in Claude Code's session logs!

Location: ~/.claude/projects/.../subagents/agent-{id}.jsonl

Format: Each assistant message includes full usage data:

{
  "message": {
    "usage": {
      "input_tokens": 3,
      "cache_creation_input_tokens": 540,
      "cache_read_input_tokens": 83670,
      "output_tokens": 1
    }
  },
  "agentId": "ae365ab",
  "timestamp": "2026-01-22T00:11:23.927Z"
}

📊 Complexity Analysis

Implementation Difficulty: Moderate (6/10)

What We'd Need:

  1. ✅ JSONL parser in Rust (easy - 2/10)
  2. 🤔 Subagent file discovery (medium - 5/10)
  3. 🤔 Usage aggregation across messages (medium - 6/10)
  4. 🔴 Real-time monitoring with file watchers (hard - 8/10)
  5. 🔴 Edge case handling (hard - 7/10)

Estimated Time: 4-8 hours

Pros:

  • ✅ Accurate token counts from Anthropic's records
  • ✅ Historical data available
  • ✅ No CLI modifications needed
  • ✅ Works with current CLI

Cons:

  • ❌ Moderately complex implementation
  • ❌ File watching/polling overhead
  • ❌ Race conditions with file writes
  • ❌ Relies on undocumented file format (could break)
  • ❌ Timing challenges (when is file ready?)

💡 Recommendation

Keep this issue open as a future enhancement, but acknowledge it requires:

  1. Parsing session JSONL files (not live stream)
  2. File watching or polling
  3. Accepting risk of format changes

Current state: We already show duration_ms from hooks, which provides useful visibility!

📚 References

Tools that successfully parse session files:

✨ This comment was written by Hikari~ 🌸

## Investigation Results 🔍 I've investigated this issue and here's what I found! ### ❌ Original Approach Not Feasible The `<usage>` block does **not** appear in Task tool result content from the CLI. I verified this by: - Adding debug logging to our parsing code - Testing with CLI v2.1.34 - Checking all CLI flags (no flag to enable this) The CLI simply doesn't output usage metrics in ToolResult content, and likely never will. ### ✅ Alternative: Parse Session JSONL Files **The data DOES exist** in Claude Code's session logs! **Location**: `~/.claude/projects/.../subagents/agent-{id}.jsonl` **Format**: Each assistant message includes full usage data: ```json { "message": { "usage": { "input_tokens": 3, "cache_creation_input_tokens": 540, "cache_read_input_tokens": 83670, "output_tokens": 1 } }, "agentId": "ae365ab", "timestamp": "2026-01-22T00:11:23.927Z" } ``` ### 📊 Complexity Analysis **Implementation Difficulty**: Moderate (6/10) **What We'd Need**: 1. ✅ JSONL parser in Rust (easy - 2/10) 2. 🤔 Subagent file discovery (medium - 5/10) 3. 🤔 Usage aggregation across messages (medium - 6/10) 4. 🔴 Real-time monitoring with file watchers (hard - 8/10) 5. 🔴 Edge case handling (hard - 7/10) **Estimated Time**: 4-8 hours **Pros**: - ✅ Accurate token counts from Anthropic's records - ✅ Historical data available - ✅ No CLI modifications needed - ✅ Works with current CLI **Cons**: - ❌ Moderately complex implementation - ❌ File watching/polling overhead - ❌ Race conditions with file writes - ❌ Relies on undocumented file format (could break) - ❌ Timing challenges (when is file ready?) ### 💡 Recommendation **Keep this issue open** as a future enhancement, but acknowledge it requires: 1. Parsing session JSONL files (not live stream) 2. File watching or polling 3. Accepting risk of format changes **Current state**: We already show `duration_ms` from hooks, which provides useful visibility! ### 📚 References Tools that successfully parse session files: - [ccusage](https://github.com/ryoppippi/ccusage) - CLI usage analysis - [Claude JSONL Browser](https://github.com/withLinda/claude-JSONL-browser) - Web viewer - [Analyzing logs with DuckDB](https://liambx.com/blog/claude-code-log-analysis-with-duckdb) ✨ This comment was written by Hikari~ 🌸
naomi closed this issue 2026-03-06 23:49:45 -08:00
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
aspect
code
Concerns the software code in the repository
aspect
dx
Concerns developers' experience with the codebase
aspect
interface
Concerns end-users' experience with the software
aspect
text
Concerns the documentation material in the repository
contribute
good first issue
A great opportunity for a team member to learn a new codebase
contribute
help wanted
Open for anyone on our team to grab.
contribute
staff only
Restricted to our executive leadership.
goal
addition
Addition of new feature
goal
fix
Bug fix
goal
improvement
Improvement to an existing feature
points
1
Very simple issue requiring minimal effort and complexity.
points
13
Extremely complex issue representing major undertakings. Should be broken down into smaller pieces.
points
2
Simple issue that requires a bit more thought or investigation.
points
3
Moderate complexity issue requiring more substantial work.
points
5
Complex issue requiring significant effort and expertise.
points
8
Very complex issue requiring extensive work and deep expertise.
priority
critical
1
Must be fixed ASAP
priority
high
2
Stalls work on the project or its dependents
priority
low
4
Low priority and doesn't need to be rushed
priority
medium
3
Not blocking but should be fixed soon
priority
none
5
No priority, should only be performed when a developer is available
status
awaiting triage
Has not been triaged & therefore, not ready for work
status
blocked
Blocked and therefore not ready for work
status
discarded
Will not be worked on
status
discontinued
Not suitable for work as repo is in maintenance
status
label work required
Needs proper labelling before it can be worked on
status
ready for dev
Ready for work
status
ticket work required
Needs more details before it can be worked on
talk
discussion
Open for discussions and feedback
talk
question
Can be resolved with an answer
time
1 day
Approximately one full day of development work.
time
1-2 weeks
One to two weeks of focused development effort.
time
2-3 days
Two to three days of development effort.
time
4-5 days
Approximately one week of development work.
time
<1 day
Less than one day of focused work. Quick fixes or simple tasks.
time
>2 weeks
More than two weeks of development work. Must be broken down into smaller pieces.
No Label
Milestone
No items
No Milestone
Assignees
Clear assignees
gurkirat hanna (Hanna Rose) hikari (Hikari) minori (Minori) naomi (Naomi Carrigan) rain teklu tim
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: nhcarrigan/hikari-desktop#115
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?