Omniscio

Interaction trace (slow-click, jank, and navigation latency log)

A single log file capturing every slow click, keypress and rendering stall while Omniscio is open, kept as a continuous tape across launches with a worst-offenders summary at the bottom. It exists so a report of "clicks feel laggy" travels as one attachable file anyone can read.

On this page

What it is

A dedicated log file — interaction.log — that captures every slow interaction (click, keypress, tap) and rendering-thread stall that happens while Omniscio is open, in one continuous self-contained timeline. It exists so a user reporting "clicks feel laggy" or "the UI freezes for a second when I switch sessions" can attach exactly one file to a bug report, and any agent (or outside AI like ChatGPT) reading that pasted file can immediately tell which interactions were slow and where the time was spent — without seeing the rest of the codebase.

The file lives next to the regular logs:

  • Windows: %APPDATA%\omniscio\profiles\<your account>\logs\interaction.log
  • macOS: ~/Library/Application Support/omniscio/profiles/<your account>/logs/interaction.log
  • Linux: ~/.config/omniscio/profiles/<your account>/logs/interaction.log

Unlike startup.log (which rotates per-launch), interaction.log is a continuous tape across launches — the same file keeps growing every time you open the app, capped at 1 MB. When it reaches the cap, it rotates to interaction.log.1 → … → interaction.log.5 (oldest dropped). This is intentional: a "lag that only shows up sometimes" needs a tape that spans days, not a fresh file each launch.

It is separate from main.log on purpose: main.log is a rolling firehose of every log line during runtime; interaction.log is a focused timeline of just the perf-sensitive events, with a TOP 10 worst-offenders summary at the bottom so even a non-technical reader can see the pattern.

Where to find it

The file sits in your account's logs folder, next to the other logs — on Windows under %APPDATA%\omniscio\profiles\<your account>\logs\interaction.log, on macOS under ~/Library/Application Support/omniscio/profiles/<your account>/logs/, and on Linux under ~/.config/omniscio/profiles/<your account>/logs/. The account part of the path differs per sign-in, so take the exact folder from Settings → Diagnostics → Log Files rather than typing it. There is no in-app viewer; you open it, or attach it to a bug report.

How it behaves

How to use it

To attach it to a bug report

  1. Open Settings → Diagnostics. You'll see a card titled Interaction Trace with a single button: Reveal Interaction Log.
  2. Click it. Omniscio flushes a fresh TOP 10 summary block, then opens your file manager with interaction.log highlighted.
  3. Drag the file into the bug report (or paste its contents into a chat).

The file is plain text, opens in any editor. The very top has a self-describing header so an LLM reading a pasted copy with no other context understands what each section means.

What you'll see in the file

═══════════════════════════════════════════════════════════════
 Omniscio INTERACTION TRACE
 …self-describing header (purpose, Version, Electron, Mode, PID)…
═══════════════════════════════════════════════════════════════

── Launch resumed at 2026-05-18T15:42:11.123Z (PID 12345) ──
2026-05-18T15:42:14.901Z [event] name=click duration=423ms inputDelay=32ms processing=280ms target=button.send iid=7
2026-05-18T15:42:15.044Z [loaf] duration=240ms blocking=180ms scripts=3 slowest=120ms at DOMWebSocket.onmessage
2026-05-18T15:42:18.221Z [phased-nav] activation=7 total=850ms session=sess-abc
   nav-to-mount                    12ms
   mount-to-fetch                  4ms
   fetch                           380ms
   build-turns                     18ms
   commit-to-paint                 436ms
…

────────────────────────────────────────────────
 SUMMARY (flushed at 2026-05-18T16:01:33.444Z)
────────────────────────────────────────────────

TOP 10 SLOWEST INTERACTIONS
   1. click        423ms (input=32ms processing=280ms) target=button.send at=2026-05-18T15:42:14.901Z
   2. keypress     308ms (input=18ms processing=210ms) target=textarea#composer at=2026-05-18T15:43:02.118Z
   …

TOP 10 LONG ANIMATION FRAMES
   1. 240ms blocking=180ms scripts=3 slowest=120ms at DOMWebSocket.onmessage at=…
   …

TOP 10 PHASED-NAV
   1. activation=7 850ms session=sess-abc at=2026-05-18T15:42:18.221Z
      nav-to-mount                    12ms
      mount-to-fetch                  4ms
      fetch                           380ms
      build-turns                     18ms
      commit-to-paint                 436ms
   …

The TOP 10 sections are the most useful for triage — they sort by duration so you can immediately see the worst offenders without scanning the whole tape. Phased-nav entries break down a single session cold-mount into five named phases — nav-to-mount → mount-to-fetch → fetch → build-turns → commit-to-paint — so you can see where the time went between the user clicking a session and the first paint.

How to read each line

  • [event] — a click / tap / keypress that took longer than 200 ms end-to-end. inputDelay is wall-clock time from event hit to JS handler start; processing is the synchronous JS work the handler did. target is a short CSS-selector-ish description of the element. iid is the browser's interaction id (groups multiple events from one logical interaction).

  • [loaf] — a "long animation frame" (>100 ms of render-thread time). blocking is time the main thread couldn't yield to input. scripts is how many JS callbacks ran in that frame. slowest names the worst single callback, formatted <ms>ms at <invoker> — the separator is at, never @ (a @ joins into an email-shaped string that the log scrubber rewrites to [EMAIL_REDACTED], destroying the one field that says what blocked the frame; guarded by loaf-attribution-avoids-email-shape.test.ts). Burst protection coalesces multiple LoAFs in the same 1-second window into a single entry with a coalesced=N suffix.

  • [phased-nav] — a session/route switch broken down into named phases (currently click-to-state-set → state-set-to-mount → mount-to-first-message). These are emitted explicitly by the navigation code rather than auto-detected, so they are the ground truth for "where did the second go."

  • [composer-keys] — once a minute, the message box's keystroke-to-paint histogram: n= keys, counted= how many the page saw, max= the slowest, and b= the per-bucket counts. The session message box and the Team Chat message box are tallied separately; a Team Chat line ends with surface=team-chat (an untagged line, or surface=session, is the session box). It exists because every line above is only recorded at 200 ms or more, so no honest p95 of ordinary typing could be computed from them. A zero-count line at least every five minutes says the tracer was alive and nothing was typed. Read by the perf:lag scorecard — see perf-lag-scorecard.md.

When the window stops painting — the [present-stall] lines (in renderer-diag.log)

Sometimes the window freezes although nothing above looks slow: the keystrokes show inputDelay and processing near 0, and the [loaf] line reads blocking=0ms scripts=0. That is a presentation stall — the app handled the input, but the finished frame did not reach the screen. A small probe watches for exactly this and writes to renderer-diag.log, in the same folder:

  • [present-stall] — one line per stall you could feel: after a text edit or a click, the next frame was still missing half a second later. It carries late (how late the frame was), input (text or pointer), verdict, then the readings taken at both ends of the stall: gpuCpu (the graphics process's CPU time, and its share of one core), gpuFaults (its page faults), gpuWs (its memory in use), gpuIo / gpuPage (its current disk and memory priority), kernel (the share of all CPU time the machine spent inside Windows itself), dwm (alive when the Windows desktop compositor drew during the stall; quiet is ambiguous, because it only draws when something on screen changes), window and beginLag (how much of the stall was measured), live (running sessions), episode, and dropped (stalls skipped by the six-a-minute cap).
  • The verdict: gpu-busy — the graphics process worked hard through the stall; gpu-blocked — it barely ran while the machine was in a kernel storm, so the machine starved it; gpu-idle — it barely ran on a calm machine, so nothing asked it to draw; inconclusive — a reason= says why (a reading failed, the window went hidden, the stall never ended). A value that could not be read is written n/a, never 0.
  • [present-stall-count] — at most once a minute, and only for a minute in which you typed or clicked: composerKeys (characters typed into a message box), otherInputs, and late (late frames). A log with count lines and no stall lines means "watched, and nothing stalled"; no count line at all means nobody was typing.

The probe runs only in the main desktop window, and only while Settings → Diagnostics → "Performance & interaction tracing" is on. It reads nothing until a frame is already late, and every reading happens off the app's main thread. It sees frames whose drawing started late; a frame drawn on time but shown late is not caught.

How it works

Two parts: the renderer observer that detects slow events, and the main-process recorder that writes them to disk.

Renderer (observer)

Lives in interaction-trace.ts. The useRendererDiagnostics hook (useRendererDiagnostics.ts) — called once from App.tsx at boot — runs an effect that calls installInteractionObserver({ reporter }):

  • Wraps two PerformanceObserver streams:
    • event (Event Timing API, Chrome 85+) — fires for every event whose duration ≥ 200 ms.
    • long-animation-frame (LoAF API, Chrome 123+) — fires for every animation frame whose duration ≥ 100 ms.
  • Coalesces LoAF bursts in a 1-second window (setTimeout flush) so a misbehaving render loop can't flood the log.
  • Caps every user-derived string at 200 chars before forwarding (DoS protection — a giant aria-label on a slow button can't blow the IPC budget).
  • Returns a { stop, reportPhasedNav } handle. The reportPhasedNav shim is also exposed at module-level via phased-nav-reporter.ts so navigation code can fire phased-nav events without prop-drilling.

Each detected entry goes through ipc.invoke(IPC.RENDERER_DIAG_REPORT, payload) — the same channel the existing longtask/snapshot diagnostics use. The Zod schema is a discriminated union so every payload shape (longtask / snapshot / interaction-event / long-anim-frame / phased-nav, plus freeze attribution and the two presentation-stall reports) stays independently validatable.

What turns the observer on. Every stream above — [event], [loaf], [composer-keys] and [phased-nav] — comes from Settings → Diagnostics → "Measure how the window feels" (rendererHarmHeartbeatEnabled), which is ON by default: they are what the interaction-latency, long-frame and blocked-seconds readings are measured from, so they cannot be opt-in. The separate "Performance & interaction tracing" switch (diagnosticTracingEnabled), OFF by default, adds only the per-frame attribution (cb= / kd= / pdTop= on a freeze line) and the [present-stall] lines below — the parts that run code inside every frame and every click. Turning that one off leaves everything measured here running.

Which pushes a freeze applied. On the web/phone surface a freeze line can also carry pushes=session:status-changed*12,session:output*3 — the push channels the frame's push-apply flush applied. It rides the ordinary freeze line (no tracing switch needed), is matched to the frame's own window rather than to "the newest flush", and is bounded to the three busiest channels with the newest 16 flushes retained. Desktop never runs the push-apply batch, so a desktop line never carries it.

Kill switch: set AMC_DISABLE_INTERACTION_TRACE=1 in the environment before launch and the observer never installs (returns a no-op handle). The reveal button still works — it'll just open whatever file is on disk.

Main process (recorder)

Lives in interaction-trace.ts. Initialized once from index.ts immediately after initStartupTrace():

  • initInteractionTrace() — opens interaction.log. If the existing file is already over 1 MB, rotates it first; otherwise appends a ── Launch resumed at … ── marker and continues writing.
  • recordInteractionEvent / recordLongAnimFrame / recordPhasedNav — three append helpers, each also maintaining a rolling in-memory TOP 10 sorted by duration. Per-record format is one line per entry; phased-nav writes its line plus an indented breakdown per phase.
  • flushInteractionSummary() — appends a fresh SUMMARY (flushed at …) block with all three TOP 10 sections. Called from two sites: (a) gracefulShutdown() in index.ts so the final file always ends with an up-to-date summary, and (b) the reveal handler itself, so clicking the button always reveals a file with current data even if the user has been running for hours.

All disk writes are wrapped in try/catch — a failed appendFileSync must never disrupt the renderer or the IPC handler. If the log directory can't be created, the recorder silently no-ops.

The IPC wiring is in diagnostic-handlers.ts:

  • RENDERER_DIAG_REPORT — routes by payload.type: longtask, freeze-attribution and snapshot become lines in the non-blocking renderer-diag.log (buffered and flushed off the main thread; snapshot only in dev builds), the three perf-trace variants → the recorder (does NOT write to main.log, which would balloon), and the two presentation-stall variants → the presentation-stall probe.
  • INTERACTION_LOG_REVEAL — flushes summary, then shell.showItemInFolder(getInteractionLogPath()).

Why a separate file (not just main.log)

main.log would be wrong for three reasons:

  • LoAF can fire hundreds of times per second during pathological renders — interleaving that into main.log would crowd out actionable lines.
  • A user reporting "clicks have felt laggy for a few days" needs a tape that spans days, but main.log rotates by size — your bad-Tuesday afternoon may be gone by Thursday.
  • The TOP 10 summary at the bottom only makes sense if every line in the file is the same shape — interleaving with unrelated log lines defeats the summary.

interaction.log solves all three: only perf-sensitive events, 1 MB cap with 5 rotated backups, and a self-describing header + TOP 10 sections that make it interpretable in isolation.

Limitations

  • Sub-200ms events are invisible, with ONE exception: the session and Team Chat message boxes' keystrokes, which are recorded as a once-a-minute histogram from one frame (16 ms) up (see [composer-keys] above). Every other click and key press still needs 200 ms to appear.
  • LoAF requires Chrome 123+. Electron 44 bundles a Chromium newer than 123, so it's covered, but if a future Electron downgrade removes the API the observer silently no-ops on that stream (Event Timing still works).
  • Phased-nav events are explicit, not auto-detected. Navigation code has to call reportPhasedNav(...) from phased-nav-reporter.ts. If a nav site doesn't, it just won't appear in the trace.
  • Continuous tape means cross-launch context lives in one file. If you only care about the most recent launch, look for the latest ── Launch resumed at … ── marker and read forward from there.

Last verified 2026-10-06