codecov.yml

# Codecov configuration (#15336).
#
# Until now this file did not exist, so Codecov ran on its built-in defaults.
# That is why the release PR reported "patch coverage 70.02% ... please review"
# against a threshold nobody in this repository had chosen, on a 505-commit range
# nobody had been measuring — the number was real, but the bar was not ours.
#
# WHERE COVERAGE IS MEASURED
#
# `.github/workflows/coverage.yml` — nightly, and on every push to main.
# Deliberately NOT on pull requests (#10691): instrumentation, not test work, was
# the dominant cost, measured at 336s without `--cov` against 35m11s with it.
#
# That measurement predates PEP 669. The workflow now sets `COVERAGE_CORE=sysmon`
# so the cost can be re-measured on the current interpreter; if it falls far
# enough, gating pull requests becomes affordable and this file is where the
# `patch` status would be flipped to blocking.
#
# STATUSES ARE INFORMATIONAL FOR NOW — deliberately.
#
# Coverage is only collected AFTER a merge, so a blocking status on a pull
# request would judge it against data the pull request never produced. Making it
# blocking before coverage runs on pull requests would fail every PR for a reason
# its author cannot act on.

coverage:
  status:
    project:
      default:
        target: auto
        # Allow small dips without failing; a real regression is a trend, not a
        # rounding difference between two shard recombinations.
        threshold: 1%
        informational: true
    patch:
      default:
        # 70% matches the frontend's existing gate (vitest thresholds, #3285), so
        # the two halves of the codebase are held to one number rather than two.
        target: 70%
        informational: true

# The nightly recombines twelve shards. Without this, Codecov can post a partial
# report as soon as the first upload lands and then revise it, which reads as a
# coverage drop that never happened.
codecov:
  notify:
    after_n_builds: 12
    wait_for_ci: true

comment:
  layout: "condensed_header, diff, flags, components"
  # Comment only when the patch actually moves coverage, rather than on every PR.
  require_changes: true

ignore:
  - "**/*_test.py"
  - "**/tests/**"
  - "**/conftest.py"
  - "venv/**"
  - "**/migrations/**"

# NOT ignored, deliberately: `**/__main__.py`. Excluding entry points is the easy
# way to make a coverage number look better without any code becoming safer —
# `autobot-backend/acp/__main__.py` alone is 24 real lines and 17 of the 256
# missing in the release report. If an entry point is genuinely untestable, that
# is worth fixing or arguing case by case, not hiding by glob.