Handled errors were reported with Honeybadger.notify at ~80 call sites, so they never reach Sentry; this PR replaces them with one API, Observability::Errors.report, which forwards to Honeybadger verbatim and mirrors to Sentry, leaving each vendor with exactly two ingestion paths.

flowchart LR
    U[Unhandled exception] --> HM[Honeybadger middleware] & SM[sentry-rails middleware]
    C["Handled rescue: Observability::Errors.report(...)"] --> HB["Honeybadger.notify (args verbatim)"] & S["Sentry capture_exception / capture_message"]

Call sites change shape only in the method name; all Honeybadger-specific options pass through unchanged:

# before
rescue => exception
  Honeybadger.notify(exception, context: {message: "No code sample found in S3 with with args: #{s3_args}"})

# after
rescue => exception
  Observability::Errors.report(exception, context: {message: "No code sample found in S3 with with args: #{s3_args}"})

observability/errors.rb is now requirable without Rails, so the 13 call sites in root lib/ (DCDO adapters, buffer, image moderation, etc.) use the same API as dashboard code. 79 of 80 Honeybadger.notify sites are converted; the one skip is notify_cronjob_error's internal call, which reports to a different Honeybadger project. The 7 previously Sentry-only Observability::Errors.capture_* sites migrate to report and therefore now also report to Honeybadger.

Links

• Jira: N/A

Testing story

• Observability engine suite: 27 runs, 0 failures, including new unit tests for the exception, string, and options-only shapes, verbatim kwarg pass-through, and vendor-absent no-ops.
• 11 non-Rails lib/test suites pass with their pre-existing Honeybadger.expects(:notify).once.with(...) assertions unchanged, demonstrating the Honeybadger call is byte-equivalent at converted sites.
• Standalone proof without Rails loaded: bundle exec ruby -e "require './deployment'; require 'observability/errors'; ..." exercises all three call shapes and a real DCDO adapter rescue path (Rails remains undefined).
• Targeted dashboard suites over all converted files pass; every residual failure reproduces byte-identically with the change stashed (unbuilt asset pipeline in the worktree, known omniauth LTI failures).
• Rubocop clean on all 68 touched files.

Deployment notes

• Sentry issue volume will be noisy at first: ~25 message-style sites interpolate runtime values and capture_message groups on text. Fingerprinting is a follow-up.
• Watch Honeybadger volume on the 7 newly dual-reporting paths (notably the Delayed Job deadlock rescue and the EC2 pricing-client rescue).
• Post-deploy check: confirm one handled report appears in Sentry (the wrapper's Sentry branch is unit-tested only; no test environment runs with Sentry enabled).

Privacy and security

Context hashes previously sent only to Honeybadger are now mirrored to Sentry as extra data. Sentry remains configured with send_default_pii = false; no new data is collected, but existing report contexts gain a second destination.

🤖 Generated with Claude Code