Files
meshtastic_firmware/.claude/commands/leakhunt.md
T
2026-07-01 19:01:27 -05:00

104 lines
8.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
---
description: Hunt for memory leaks (and other slow degradations) by reading the persistent recorder's heap timeline + log slice over a window
argument-hint: [window=1h] [field=free_heap] [variant=local]
---
<!-- markdownlint-disable MD029 -->
# `/leakhunt` - read the recorder, classify a memory leak
Use the always-on recorder (`mcp-server/.mtlog/`) to read a heap timeline plus the matching log slice and produce a one-page verdict: **steady / slow leak / fragmentation / OOM-imminent**. No firmware changes, no special build flags - the LocalStats telemetry packet that the firmware already broadcasts every ~60 s carries `heap_free_bytes` and `heap_total_bytes`.
## Two signal paths - pick the right one
| Path | Build flag | Cadence | Per-thread attribution | Cost |
| --------------------- | ---------------- | -------------- | ---------------------- | ------------------------- |
| LocalStats packet | (default) | ~60 s | No | Free - always on |
| `[heap N]` log prefix | `-DDEBUG_HEAP=1` | every log line | Yes (Thread X leaked) | Bigger flash + log volume |
Both feed the same `telemetry_timeline(field="free_heap")` query - when DEBUG_HEAP is on, the recorder synthesizes telemetry rows from log prefixes (tagged `source: debug_heap`), so a single timeline call gets whichever signal is available. **For a slow leak diagnosis, the default path is plenty** (60 s cadence over 6 h = 360 points; linear regression over that nails sub-100-byte/min slopes). **DEBUG_HEAP is for attribution** - when the slope is real and you need to know which thread is leaking.
## What to do
1. **Parse `$ARGUMENTS`**: optional `window` (default `1h`, accepts `30m`/`6h`/`-3d`/etc.), optional `field` (default `free_heap`; alternates: `total_heap`, `battery_level`, anything in the LocalStats variant), optional `variant` (default `local`; alternates: `device`, `environment`, `power`, `airQuality`, `health`).
2. **Verify the recorder is alive** - call `mcp__meshtastic__recorder_status`. Check:
- `running == True`
- `files.telemetry.lines > 0` (at least one telemetry packet recorded - if zero, the device hasn't broadcast LocalStats yet OR `set_debug_log_api` has never been on; tell the operator to run `mcp__meshtastic__set_debug_log_api(enabled=True)` and wait one device-update interval)
- `files.telemetry.last_ts` within the last 5 minutes (if older, the device is silent - log that, not "leak detected")
3. **Detect whether DEBUG_HEAP is active** - `mcp__meshtastic__logs_window(start="-2m", grep=r"\\[heap \\d+\\]", max_lines=3)`. If any line matches, the firmware has the prefix → DEBUG_HEAP is on, expect higher-cadence data and `heap_event` rows. If zero matches over the last 2 minutes, you're on the LocalStats-only path.
4. **Pull the timeline** - `mcp__meshtastic__telemetry_timeline(window=$window, variant=$variant, field=$field, max_points=200)`. Read:
- `samples` - how many raw points contributed
- `min`, `max` - total swing
- `slope_per_min` - units per minute (linear regression over the whole window)
5. **Pull the log context for the same window** - `mcp__meshtastic__logs_window(start="-${window}", grep="Heap status|leaked heap|freed heap|out of memory|Alloc an err|panic|abort", max_lines=200)`. These are the strings the firmware emits when something memory-related happens (`DEBUG_HEAP` builds emit `"Heap status:"` and `"leaked heap"` lines; production builds emit `"Alloc an err"` on failure and `"out of memory"` on OOM).
6. **Pull marker events** so we know if the operator labeled phases - `mcp__meshtastic__events_window(start="-${window}", kind="mark|connection_lost|connection_established")`. If a `connection_lost` overlaps a sharp drop, that's not a leak; that's a reboot.
6a. **(DEBUG_HEAP only) Per-thread attribution** - `mcp__meshtastic__logs_window(start="-${window}", grep="leaked heap", max_lines=200)`. Each row has a structured `heap_event` field with `{kind, thread, before, after, delta}`. Aggregate by thread: sum the `delta` over the window per thread name. The thread with the largest cumulative negative delta is your suspect. Note the count too - a thread with 50× small leaks is different from 1× big leak.
7. **Classify** based on what the data says, NOT on what you wish it said. Use these rules in order:
- **Insufficient data** (< 5 samples): say so. Suggest a longer window or longer wait. Stop.
- **Reboot mid-window**: if any `connection_lost` event is present AND `free_heap` jumped UP at that timestamp, the device rebooted. Note it; pre-reboot trend may be a leak but you only have part of the curve.
- **OOM-imminent**: any `Alloc an err=` or `out of memory` line in the log slice. This trumps everything; flag urgently.
- **Slow leak**: `slope_per_min < -50` AND `max - min > 1000` AND no reboot. The heap is monotonically (or near-monotonically) declining. Estimate time-to-zero: `min / -slope_per_min` minutes. Surface it.
- **Fragmentation suspect**: `slope_per_min` close to zero (|x| < 50) BUT min trends down across the window AND the log slice shows `Alloc an err` warnings WITHOUT total OOM. Means free total is OK but largest contiguous block is shrinking. Recommend a `DEBUG_HEAP` build to confirm.
- **Steady**: |slope_per_min| < 50, no error lines. Heap is fine.
- **Recovery curve**: slope is POSITIVE - heap recovered. Either a workload completed or GC fired. Note it; not a leak.
8. **Report**:
```text
/leakhunt window=6h field=free_heap variant=local
────────────────────────────────────────────────────
recorder : running, telem last_ts 8s ago
build : DEBUG_HEAP=ON (per-line prefix detected)
samples : 14,200 over 6h (cadence ~1.5s, log-line synth)
free_heap : min 92,344 / max 124,008 / range 31,664
slope : -82 bytes/min (negative - heap declining)
reboots : none in window
OOM events : none
error lines : 3× "Alloc an err=ESP_ERR_NO_MEM" at +4h12m, +5h08m, +5h44m
thread leaks : (DEBUG_HEAP) MeshPacket -3,124 B over 18 events
Router -1,408 B over 4 events
others -240 B
verdict : SLOW LEAK - primary suspect MeshPacket thread
est. time-to-OOM: ~1,127 min (~18.8 h) at current slope
evidence : (3 log line citations with uptimes)
```
Then: **what to do next.**
- SLOW LEAK, **DEBUG_HEAP off** → recommend rebuilding with the flag and re-running this skill. Concrete one-liner the operator can copy:
```text
mcp__meshtastic__build(env="<env>", build_flags={"DEBUG_HEAP": 1})
mcp__meshtastic__pio_flash(env="<env>", port="<port>", confirm=True)
```
After flash, set debug_log_api back on and wait one window; re-run `/leakhunt`.
- SLOW LEAK, **DEBUG_HEAP on** → cite the top-leaking thread name from step 6a. Point at the corresponding source file (`grep -rn "ThreadName(\"<name>\")" src/`); the operator decides what to fix.
- FRAGMENTATION SUSPECT → propose pre-allocating any per-packet buffers; or rebuilding with `CONFIG_HEAP_TASK_TRACKING=y` on ESP32 to see who's holding the largest blocks.
- OOM-IMMINENT → flag for immediate attention; don't wait for the next telemetry interval.
- STEADY → say so; stop. Don't invent problems.
## What NOT to do
- Don't assume a leak from a single dip. LocalStats fires every ~60 s and the firmware naturally allocates+frees on each broadcast cycle; one packet sees the trough. Look at the slope, not the deltas.
- Don't recommend code changes. This skill diagnoses; the operator decides what to fix.
- Don't enable `set_debug_log_api` automatically - if it's off, telemetry isn't reaching pubsub anyway, and the recorder will be empty. Tell the operator to flip it on and wait, then re-run.
- Don't run heavy workloads to "trigger the leak." The recorder is passive; we read what's there.
## Companion: `mark_event` for stress runs
If the operator wants to test under stimulus (e.g. blast 50 broadcasts and see what the heap does), they can frame the experiment with markers:
```text
mark_event("burst-start")
… run the workload …
mark_event("burst-end")
/leakhunt window=15m
```
The markers land in both `events.jsonl` and `logs.jsonl`, so the report can show "free_heap dipped 8 KB during the burst window, recovered to baseline within 2 LocalStats cycles" → not a leak.