feat: crash reports sent from the game to a bot (with consent) #81

New issue

Open

opened 2026-11-13 02:23:14 +00:00 by HumanBean · 1 comment

HumanBean commented

2026-11-13 02:23:14 +00:00

Owner

Why

When Longshot crashes, the player has to find the log, open an issue or a Discord message, and paste it by hand. Most
don't, so we only hear about a fraction of crashes, often without the version, the GPU or the backtrace. A bot that
receives a crash report straight from the game would tell us how often each crash happens and on which hardware.

What we already have

On a crash, the handler in host/mkw_glue.c (report_crash) writes to the log:

CRASH: exception <code> (<read|write> <addr>, guest <addr>) at <module>+<offset>
CRASH: host backtrace: <module>+<offset> ... (x64 unwind data on Windows, backtrace() on Linux)

and log_crash_filter (host/hos_thread.c) drains the log queue before the process dies. So the summary of a crash
is already there; it never leaves the PC.

Proposal

In the game, after a crash: on the next start (not in the crashing process), offer to send the report:
"Longshot closed unexpectedly last time. Send a crash report?" with Send, Don't send and Always send.
Nothing is sent without that consent, and it can be turned off in the settings.
What a report holds, shown to the player before sending:
- the CRASH: lines and the last ~200 lines of the log before them
- Longshot's version and commit, the OS, the GPU, its driver and VRAM, the settings that differ from the defaults
- an optional comment ("what were you doing?") and an optional contact (Discord name)
What it never holds: game files or anything made from them (CONTRIBUTING.md), save data, keys, or personal
paths: user names in paths become <user>.
Where it goes: a small endpoint on the Magic Bean server (next to the relay; MagicBeanDev/Infra), which:
- rate-limits by IP and caps the report size
- groups reports by a signature: the exception code plus the top frames of the backtrace
- posts the first report of each signature to a Discord #crash-reports channel, and counts the rest on the same
  message (versions, GPUs and OSes seen)
- opens a Forgejo issue (label crash) once a signature reaches a few reports, linking the Discord message, and
  comments on it when a new version still crashes the same way
Symbols: the backtrace is module+offset. Each release keeps its symbol map so the endpoint (or a script) can
turn the top frames into function names for the post.

Open questions

Minidumps: send one too (much more useful for debugging, but larger and may hold memory from the game), or the log
only?
Should nightlies report as well (most crashes would come from them), and should they be told apart in the channel?
Sign the request with a per-build key, or accept anonymous reports and rely on rate limits?

Done when

A crash in a test build, then a restart and Send, shows up in #crash-reports with its version, hardware and
backtrace, and nothing is sent without consent.
A second report of the same crash updates the count instead of posting again.
tests/ has a check that the report builder leaves out paths and game data (no game needed).

## Why When Longshot crashes, the player has to find the log, open an issue or a Discord message, and paste it by hand. Most don't, so we only hear about a fraction of crashes, often without the version, the GPU or the backtrace. A bot that receives a crash report straight from the game would tell us how often each crash happens and on which hardware. ## What we already have On a crash, the handler in `host/mkw_glue.c` (`report_crash`) writes to the log: - `CRASH: exception <code> (<read|write> <addr>, guest <addr>) at <module>+<offset>` - `CRASH: host backtrace: <module>+<offset> ...` (x64 unwind data on Windows, `backtrace()` on Linux) and `log_crash_filter` (`host/hos_thread.c`) drains the log queue before the process dies. So the summary of a crash is already there; it never leaves the PC. ## Proposal 1. **In the game, after a crash:** on the next start (not in the crashing process), offer to send the report: "Longshot closed unexpectedly last time. Send a crash report?" with *Send*, *Don't send* and *Always send*. Nothing is sent without that consent, and it can be turned off in the settings. 2. **What a report holds**, shown to the player before sending: - the `CRASH:` lines and the last ~200 lines of the log before them - Longshot's version and commit, the OS, the GPU, its driver and VRAM, the settings that differ from the defaults - an optional comment ("what were you doing?") and an optional contact (Discord name) 3. **What it never holds:** game files or anything made from them (CONTRIBUTING.md), save data, keys, or personal paths: user names in paths become `<user>`. 4. **Where it goes:** a small endpoint on the Magic Bean server (next to the relay; `MagicBeanDev/Infra`), which: - rate-limits by IP and caps the report size - groups reports by a signature: the exception code plus the top frames of the backtrace - posts the first report of each signature to a Discord `#crash-reports` channel, and counts the rest on the same message (versions, GPUs and OSes seen) - opens a Forgejo issue (label `crash`) once a signature reaches a few reports, linking the Discord message, and comments on it when a new version still crashes the same way 5. **Symbols:** the backtrace is module+offset. Each release keeps its symbol map so the endpoint (or a script) can turn the top frames into function names for the post. ## Open questions - Minidumps: send one too (much more useful for debugging, but larger and may hold memory from the game), or the log only? - Should nightlies report as well (most crashes would come from them), and should they be told apart in the channel? - Sign the request with a per-build key, or accept anonymous reports and rely on rate limits? ## Done when - A crash in a test build, then a restart and *Send*, shows up in `#crash-reports` with its version, hardware and backtrace, and nothing is sent without consent. - A second report of the same crash updates the count instead of posting again. - `tests/` has a check that the report builder leaves out paths and game data (no game needed).

crash-bot commented

2026-11-13 02:39:32 +00:00

Collaborator

Crash: 0xc0000005 (read) at `livetest.exe+0xdead` in `0.0.0-livetest`

8184a6f2e233 · seen 3 times · first 2026-11-13 02:39 UTC · last 2026-11-13 02:41 UTC

Seen on
Versions	`0.0.0-livetest` (3)
Operating systems	`Live test from Bean-Ops (Linux)` (2), `Live test from Bean-Ops (Windows 11)` (1)
Renderers	`Vulkan` (2), `WebGPU` (1)
GPUs	`Test GPU (not a real crash)` (3)

First report


Version	`0.0.0-livetest`
Operating system	`Live test from Bean-Ops (Linux)`
Renderer	`Vulkan`
Map	`none (live test)`
Graphics card and driver	`Test GPU (not a real crash)`
CPU and RAM	`Test CPU`

Crash

CRASH: exception 0xc0000005 (read 0x0, guest 0x0) at livetest.exe+0xdead
CRASH: this is a Bean-Ops live test, not a real crash

The log before it

Bean-Ops crash-relay live test
open ~/Games/Longshot/save.dat

### Crash: 0xc0000005 (read) at `livetest.exe+0xdead` in `0.0.0-livetest` `8184a6f2e233` · **seen 3 times** · first 2026-11-13 02:39 UTC · last 2026-11-13 02:41 UTC | Seen on | | |---|---| | Versions | `0.0.0-livetest` (3) | | Operating systems | `Live test from Bean-Ops (Linux)` (2), `Live test from Bean-Ops (Windows 11)` (1) | | Renderers | `Vulkan` (2), `WebGPU` (1) | | GPUs | `Test GPU (not a real crash)` (3) |  **First report** | | | |---|---| | Version | `0.0.0-livetest` | | Operating system | `Live test from Bean-Ops (Linux)` | | Renderer | `Vulkan` | | Map | `none (live test)` | | Graphics card and driver | `Test GPU (not a real crash)` | | CPU and RAM | `Test CPU` | **Crash** ``` CRASH: exception 0xc0000005 (read 0x0, guest 0x0) at livetest.exe+0xdead CRASH: this is a Bean-Ops live test, not a real crash ``` <details><summary>The log before it</summary> ``` Bean-Ops crash-relay live test open ~/Games/Longshot/save.dat ``` </details>