Files
obsidian-vault/work/projects/cpm-m3-telemetry-privacy-triage.md
T

8.8 KiB

TD: Pixel Definition: CPM Failure Diagnostics (macOS/iOS)

Author: Alex M
Reviewer: TBD
Stakeholders: Michal, Konrad, Russell (from the first CPM Tech Design; reviewers for this proposal to be confirmed)
Project: CPM: root-cause the stuck embedded extension messaging (macOS/iOS)
Previous Tech Design: CPM breakage pixels

Background & Requirements

The first CPM pixels detect initialization failure, a stuck messaging episode and recovery. They do not explain whether a failure coincides with background-process termination, a context-load error or incorrect tab wiring.

Terminating the extension's background process can leave CPM unable to respond, even after page reloads or opening new tabs. To investigate these failures in production, the iOS and macOS apps will attach process, extension and tab state to the existing failure pixels. This reporting does not depend on the extension responding.

Problem Statement

Distinguish background-process/context failures from tab-local problems and identify failed extension reload operations. Collect the diagnostic state alongside existing failures, without page content, URLs or persistent identifiers.

This proposal covers diagnostic collection only; it does not change recovery behavior or introduce an experiment.

Keep the existing CPM failure/stuck detection. A shared in-memory recorder collects lifecycle events, critical-memory-pressure notifications and current WebKit/tab state. When a failure or stuck pixel fires, capture native state synchronously on the main actor and attach a best-effort snapshot. Callbacks from an earlier context load cannot enter the current timeline. No additional JavaScript probes or periodic telemetry uploads are introduced.

A forwarding delegate proxy observes background-process termination and responsiveness callbacks. The original delegate owns the proxy; the proxy references it weakly. A remote flag enables proxy installation on the current view; disabling restores the original delegates on all surviving observed views, including views retained after unload. It does not disable the entire diagnostics recorder.

Pixel Trigger Frequency
debug_web_extension_cpm_initialization_failed_after_<reason> Existing initialization failure; add diagnostic parameters. Daily
debug_web_extension_cpm_messaging_stuck_<reason> Existing failure confirmed by a later eligible navigation; add diagnostic parameters. Daily + count, once per episode
debug_web_extension_reload_failed New: extension unload/load fails during a reload operation. Applies to all supported extension types. Daily + count

<reason>: session_restoration, tab_crash, extension_reload, other. PixelKit applies standard frequency/platform suffixes. Standard metadata includes appVersion; macOS also declares pixelSource and channel.

Recovery pixels remain unchanged. The new reload-failure pixel reports a failed reload operation, unlike the existing cpm_messaging_extension_reload_failed, which reports CPM failure after a successful reload.

Failure / stuck parameters

Unavailable optional facts are omitted. Boolean values are true / false.

Parameter What is collected
context_loaded Context is loaded — boolean.
bg_view_alive Current background WKWebView exists — boolean.
bg_process_alive Background process identifier is nonzero — boolean; PID is not sent.
bg_process_responsive WebKit's process-responsiveness verdict — boolean.
network_restarted Network PID differs from the value at context load — boolean; PIDs are not sent.
tab_in_context Failing tab is in WebKit's open-tabs set — boolean.
tab_controller_match Tab and context use the same extension controller — boolean.
tab_has_ext_scripts Registered scripts include the expected extension world — boolean, not proof of execution.
handler_registered Autoconsent native-message handler is registered — boolean, not proof of message delivery.
critical_memory_age Age of last observed critical pressure: none, 1m, 5m, 30m, over_30.
bg_view_creations Creations since context load: 0, 1, 2, 3_to_5, over_5.
bg_view_leaked_count Retained views beyond the expected current view; same buckets. Not proof of a leak.
context_errors none or accumulated error descriptors: raw domain/code pairs, including the immediate underlying error. No localized text; no explicit payload cap.
bg_events Recent lifecycle sequence with relative ages in whole seconds.

The recorder keeps 40 events in memory; the pixel includes at most the last 12 and 255 characters, dropping oldest entries first. Format: event_name@-seconds, relative to snapshot time.

Events: load, view, dealloc, unresponsive, responsive, proxy_on, proxy_off, died_<reason>, error_<descriptor>, context_unload_failed_<descriptor>, extension_files_remove_failed_<descriptor>. Termination reasons: memory_limit, cpu_limit, requested_by_client, crash, shared_crash_limit, unknown. The error_ events represent background-load failures. Unload failures during reload, data clearing or uninstall, and file-removal failures during uninstall, are also recorded for the tracked CPM extension. A failed unload preserves the tracked context; file-removal failures can still be recorded after successful unload. Existing standalone error pixels are unchanged; these timeline entries only travel with a later failure/stuck snapshot. Event labels are lowercased, character-filtered and capped at 64 characters.

Example: view@-60,died_crash@-42,error_background_failed_to_load@-41.

Reload-failure parameters

Parameter Values
extension_type embedded, darkMode, adBlocking, searchToken, unknown
reload_trigger data_clearing, scriptlet_update, explicit
reload_phase unload, load, lightweight_load, full_load, fallback_load

The three fields above are allowlisted enums. The pixel also sends standard error parameters:

Parameter Values
d (errorDomain) Original NSError domain, without category mapping.
e (errorCode) Original numeric NSError code.
ud (underlyingErrorDomain) Immediate underlying NSError domain, when present.
ue (underlyingErrorCode) Immediate underlying numeric code, when present.

Original domains/codes are preserved without category mapping. Localized descriptions, file paths and other userInfo contents are not serialized. The underlying error chain is not traversed beyond the immediate error.

Testing

  • Unit tests: parameter serialization, buckets, timeline limits, attachment to failure/stuck pixels, reload-error mapping, delegate forwarding, ownership and runtime flag changes.
  • Integration validation on iOS/macOS: force background-process termination, check the emitted diagnostics, test proxy on/off and extension reload failures. Confirm unknown SPI values are omitted.
  • Privacy validation: inspect actual payloads and run both platform schema validators. Confirm URLs, identifiers and error messages are absent. Review exact-second timelines and error-domain/code values before release.

Additional Considerations (if applicable)

Privacy

New pixel and additional parameters require Privacy Triage. No intentional collection of browsing URLs, search queries, page content, message payloads, tab/context IDs or PIDs in these payloads. Exact-second timelines and unrestricted error-domain/code combinations can be high-cardinality; bucketing/allowlisting needs a decision. data_clearing discloses the operation preceding a failed reload.

Local context-error diagnostics log only original and immediate underlying error domains/codes, without localized descriptions or other userInfo contents.

Security

Uses guarded private WebKit selectors for process/script facts. Validate supported OS versions; missing facts must not block pixel delivery.

Site Breakage

No intended page-behavior changes. However, the investigation recorded a changed reproduction with the delegate proxy installed; validate the current proxy on/off before treating observations as an unaffected baseline.

Experimentation

None. Recovery changes and experiments are outside this proposal.

Operational

Update the existing CPM Grafana board for the new fields and reload failures. Agree a diagnostic removal/review date; none is enforced currently. Failure-only snapshots do not provide a denominator of all process deaths or prove an all-tabs outage. Rolling timelines can lose the trigger; snapshots are captured synchronously at failure time.

Localization / Internationalization

None.