[Improvement] Crash capture: hand macOS diagnostic reports to the default agent

[splaice]

Quattro feature

Omarchy 4 watches systemd-coredump; on a crash you get a notification, and clicking it hands the core dump to your default agent with a shipped diagnose-crash skill guiding the analysis. Toggleable, experimental-labeled.

Omacase fit

The macOS analog is clean: crash/diagnostic reports land as .ips files in ~/Library/Logs/DiagnosticReports. A watcher + omacase notify + the default agent (issue #22) reproduces the flow without root or private APIs.

Proposed shape

• Opt-in extra: omacase extras crash-capture [on|off|status] — a LaunchAgent with a WatchPaths on ~/Library/Logs/DiagnosticReports (no polling), invoking omacase crash-report <newest .ips>
• omacase crash-report: dedupe against state, omacase notify with the crashed process name; on invocation, open the default agent in an overlay with the report path and a shipped diagnose-crash skill (agent-skills dir per harness) that teaches: parse the .ips JSON header, symbolication caveats, look up the binary, propose next steps
• Depends on: [Improvement] Default-agent abstraction: omacase agent + launch keybind + shell alias #22 (default agent), [Improvement] Shared LaunchAgent helper for generated plists #15 (shared LaunchAgent helper) — sequence after both

Scope

Future release. Label it experimental like upstream does.

🤖 Generated with Claude Code

[splaice]

Implementation plan

Self-contained; needs only this comment and the repo. Depends on #22 (default agent); use #15's launchagent_install if landed. Label the feature experimental in all user-facing text, as upstream Omarchy does.

Background

macOS writes crash/diagnostic reports as .ips files (JSON header + payload) into ~/Library/Logs/DiagnosticReports — no root needed to read your own. launchd WatchPaths can start a job when that directory changes (event-driven, no polling). Goal: crash happens → notification names the process → one command hands the report to the default agent with a guiding skill.

Change 1 — the extra (lib/extras.sh + registry if #16 landed)

omacase extras crash-capture [on|off|status]:

• on: install LaunchAgent org.omacase.crash — WatchPaths = ~/Library/Logs/DiagnosticReports, ProgramArguments = $OMACASE_ROOT/bin/omacase crash-report --scan, plus ThrottleInterval 10 (WatchPaths can fire in bursts). Stage+install via the [Improvement] Shared LaunchAgent helper for generated plists #15 helper or the launcher.sh dance.
• off: bootout + rm. status: agent present + last-seen marker age.

Change 2 — lib/crash.sh (new): omacase crash-report

• --scan (what the agent runs): list *.ips in DiagnosticReports newer than the state marker $OMACASE_STATE/crash-last-seen; for each new one, parse the first line (it is a standalone JSON object: app_name, timestamp) with python3; skip our own noise (optionally ignore-list in ~/.config/omacase/crash-ignore); notify --title "Crash" --event warn "<app> crashed — omacase crash-report" (use --event if [Improvement] OmacaseNotify — themed, fully-owned notification system #28 has landed, else --sound); update the marker. Never abort — launchd context, best-effort everywhere.
• No args (interactive): show the most recent report(s) (--list for a table), and omacase crash-report open [n] → launch the default agent in the overlay (omacase agent launch machinery from [Improvement] Default-agent abstraction: omacase agent + launch keybind + shell alias #22) with an initial instruction referencing the skill + the .ips path. If the default agent supports one-shot (agent prompt), offer omacase crash-report diagnose [n] to fire it directly.

Change 3 — the diagnose-crash skill

Ship assets/skills/omacase-diagnose-crash/SKILL.md: how to read an .ips (line 1 = summary JSON; body = legacy-format crash report), what to extract (crashed thread, exception type, binary images, termination reason), symbolication caveats on ad-hoc-signed binaries, and next-step suggestions (check brew for the app, look for known issues, offer to file upstream). Install: symlink into each harness skills dir that exists (~/.claude/skills/, ~/.codex/skills/, ~/.grok/skills/), exact-target-guarded both directions — create only if absent or already ours; uninstall removes only links pointing at our store. This does NOT conflict with the issue-#3 rule (Omacase must not manage the herdr skill — herdr owns that one; this skill is Omacase's own, and ownership is provable by link target). Add the install step to the crash-capture on path, not global install (the skill is useless without the feature).

Change 4 — tests

1. Scan logic: temp HOME with fixture .ips files (hand-write minimal first-line JSON {"app_name":"TestApp","timestamp":"…"}); stub notify; first --scan notifies once per new file and advances the marker; second --scan notifies zero.
2. Ignore list respected.
3. Skill links: temp HOME with only ~/.claude/skills present → link created there only; re-run idempotent; uninstall removes only our link; a pre-existing foreign omacase-diagnose-crash dir is left untouched (warn).
4. Wiring greps: usage/completion/extras list mention crash-capture.

Verification / done criteria

Suite green; shellcheck clean. Live: enable; kill -SEGV a scratch process (e.g. run sleep 100 & then SEGV it — macOS writes a report within seconds); notification appears; omacase crash-report lists it; open drops into the agent with the report path. Disable; verify agent gone.

Branch + PR; body says Closes #24.