Show child process cpu usage in dtop

[aclauer]

Problem

dtop only shows cpu usage for Python workers spawned by DimOS. Any native modules spawned by that worker don't show up in the cpu statistics. Also add --log flag to log dtop statistics and dtop-plot to generate plots of cpu usage.

Closes DIM-XXX

Solution

Read the pids of any processes spawned and include their cpu usage in a drop down of the main worker.

Breaking Changes

None

How to Test

dimos --dtop --replay --replay-db=go2_bigoffice run unitree-go2

and

dtop

When dimos spawns the viewer, it will show up as a subprocess of the rerun bridge worker.

Contributor License Agreement

• I have read and approved the CLA.

[greptile-apps]

Greptile Summary

This PR extends dtop to collect and display CPU usage of child processes spawned by each worker, adds a --log flag for JSONL output, and introduces a new dtop-plot CLI for generating matplotlib plots from those logs. The core logic—collecting children via proc.children(), caching them through _get_process, and folding their CPU into the parent's total—is sound and well-structured.

Confidence Score: 5/5

Safe to merge; all findings are P2 memory-growth issues that don't affect correctness.

Both new comments are P2 (unbounded growth in _proc_cache and _child_cpu_history for terminated child PIDs). These don't cause crashes or incorrect data in typical usage and are not regressions — they only matter for very long-running sessions with high child-process churn. No P0 or P1 issues found.

dimos/core/resource_monitor/stats.py and dimos/utils/cli/dtop.py for the cache-cleanup gaps.

Important Files Changed

Filename	Overview
dimos/core/resource_monitor/stats.py	Adds ChildProcessStats dataclass and collect_children_stats() to gather per-child CPU usage; extends WorkerStats with a children field. Child PIDs are cached via _get_process but dead child entries are never evicted from _proc_cache.
dimos/utils/cli/dtop.py	Adds child-process rows beneath each worker in the TUI, extracts CPU rendering into _cpu_metric, adds --log flag for JSONL output. _child_cpu_history grows without bound as children come and go.
dimos/utils/cli/dtop_plot.py	New dtop-plot CLI tool that reads a JSONL log and produces matplotlib plots. Works correctly for well-formed logs; previously-flagged NaN/KeyError edge cases remain unaddressed.
dimos/core/resource_monitor/monitor.py	Calls the new collect_children_stats after collect_process_stats and folds children's CPU into the parent's cpu_percent; passes children list to WorkerStats. Logic is straightforward and correct.
pyproject.toml	Registers dtop-plot as a console script entry point pointing to the new dtop_plot:main.

Sequence Diagram

sequenceDiagram
    participant M as StatsMonitor
    participant S as stats.py
    participant LC as LCM
    participant D as dtop TUI

    loop every poll interval
        M->>S: collect_process_stats(worker_pid)
        S-->>M: ProcessStats (parent CPU, mem, …)
        M->>S: collect_children_stats(worker_pid)
        S->>S: proc.children(recursive=False)
        loop each child pid
            S->>S: _get_process(child.pid) → cached Process
            S->>S: cpu_percent(interval=None)
        end
        S-->>M: list[ChildProcessStats]
        M->>M: ps_dict[cpu_percent] += Σ child cpu
        M->>M: WorkerStats(…, children=[…])
        M->>LC: publish ResourceStatsMessage
    end

    LC-->>D: _on_msg(msg)
    D->>D: write JSONL line (if --log)
    D->>D: _refresh() → render child rows under each worker

Loading

_{Reviews (6): Last reviewed commit: "Add log flush" | Re-trigger Greptile}