New to Claude Skills? Learn how to install them →

Copilot code review metrics: calculate and interpret addressed rate

Understand Copilot's new addressed-rate metrics with worked examples, fair comparisons and a practical plan for evaluating review skills.

November 6, 2026
Get Claude Skills
9 min read

What changed in Copilot code review reporting

GitHub's 5 November 2026 release adds a copilot_code_review object to repository daily usage reports. It includes total_comments_evaluated and total_comments_addressed, plus breakdowns covering severity, review tier, comment type and customisation, including skills. Comments count when their pull request merges. GitHub's meaning of addressed includes several responses, including a reply that a fix will follow in another pull request. Bot-authored pull requests usually do not receive comment evaluation, although they can appear in the separate merged-without-later-human-change-request count. See the primary release announcement.

These metrics create an opportunity to ask a more useful question than whether an agent ran: what happened to its feedback? This guide develops an original analysis workflow using synthetic counts. It does not present a benchmark, a tested integration or evidence that any particular skill improves code quality.

If you are choosing a review workflow, start with the code review skills guide. Use the skill directory to locate the upstream instructions for a candidate. The calculation here helps assess an adopted workflow. It cannot choose a suitable skill for your repository without understanding the work, the reviewers and the permissions involved.

Calculate the rate with a visible denominator

Use the number of addressed comments as the numerator and evaluated comments as the denominator. Keep both counts visible beside the percentage. A rate from five comments carries a different amount of information from the same rate calculated from several thousand. The percentage alone conceals that difference.

The following Python example uses invented aggregate counts. It performs no API request and reads no private data. Field names match the release announcement; the surrounding dictionary is an illustrative input, not a complete API response.

sample = {
    "total_comments_evaluated": 80,
    "total_comments_addressed": 52,
}

def addressed_rate(evaluated, addressed):
    if evaluated < 0 or addressed < 0 or addressed > evaluated:
        raise ValueError("Check the counts and their reporting scope")
    if evaluated == 0:
        return None
    return addressed / evaluated

rate = addressed_rate(
    sample["total_comments_evaluated"],
    sample["total_comments_addressed"],
)
print("Unavailable" if rate is None else f"{rate:.1%}")

This example produces 65.0%. The input check is a proposed guard for this calculation, not a claim about how GitHub validates reports. If it fails on your data, inspect the source fields and reporting scope before changing values to make the calculation succeed. Preserve missing values as missing rather than silently converting them into zeros.

For a dashboard, show a label such as “52 addressed out of 80 evaluated” with the percentage. That wording lets a reader understand the observation without opening a methodology document. Add the reporting period and scope nearby. A precise title prevents a repository-level observation from being mistaken for an organisation-wide result.

Combine repositories without misleading averages

Consider two fictional repositories. Repository A has 90 addressed comments out of 100 evaluated. Repository B has one addressed comment out of ten evaluated. Their individual rates are 90% and 10%. The arithmetic average of those percentages is 50%, while the combined comment-weighted rate is 91 divided by 110, or about 82.7%.

Synthetic repositoryEvaluatedAddressedRate
A1009090.0%
B10110.0%
Combined comments1109182.7%

Neither averaging method should be chosen accidentally. The combined calculation answers what proportion of the evaluated comments were addressed. The unweighted average describes the average repository percentage, giving each repository equal influence. Pick the question first, label the method and retain the component counts so someone can reproduce it.

Avoid combining unlike reporting scopes. If the same underlying activity appears in overlapping exports, summing them can count it twice. Build a small inventory of the files or records included in the calculation, their periods and their scope. Use aggregate identifiers appropriate to your private reporting environment, and keep that inventory out of a public content repository.

Decide what an addressed comment means for your team

Define the decision you want the metric to support. A review lead might want to identify workflows whose feedback is consistently ignored. A skill author might want to understand whether a revised checklist produces comments that developers can act on. Those are different decisions and may require different evidence alongside the same percentage.

Do not rename this measure “bugs fixed” or “review accuracy” in a presentation. A response to feedback and a verified improvement are separate observations. Your report should preserve that distinction even when the percentage rises. Otherwise, a useful operational signal becomes a claim the data cannot support.

Create a small manual review rubric. For each sampled comment, ask whether it refers to the actual change, explains a concrete concern and offers an appropriate next step. Then inspect what happened to the concern in the final code or follow-up work. Record uncertainty when the available evidence cannot answer a question. A neat spreadsheet is not a reason to force a judgement.

The agent skills security guide offers background for reviewing workflows that inspect code or run tools. Keep the evaluation proportional to the actions available. A comment-only helper and a workflow that edits branches need different checks even if both appear in a review report.

Compare a review skill using a controlled pilot

Choose a narrow question, such as whether a new review checklist makes feedback on error handling more actionable. Write down the intended improvement before looking at the results. A specific question prevents the team from searching across many percentages until it finds one that appears favourable.

Keep the skill revision, repository settings and evaluation period in the pilot notes. Record other changes that might affect the comparison, such as a different model, a changed review process or a large migration. If several things change together, describe the result as an observation about the whole workflow. Do not assign the difference to a single skill without supporting evidence.

Use similar kinds of work on both sides of the comparison where possible. A period dominated by straightforward maintenance is not directly comparable with a period containing a major architecture change. You can still report both periods, but explain the difference in work rather than hiding it beneath a single trend line.

The guide to writing an agent skill can help define the intended behaviour. A measurable pilot starts with a clear task boundary. “Better reviews” is too broad; “comments identify missing error handling in this type of operation” gives reviewers something concrete to examine.

Keep timing and selection effects visible

Choose a reporting window that matches the decision. A daily view can help spot an operational change, while a longer window may provide more examples to inspect. Do not treat a sudden daily movement as a stable trend without checking the underlying counts. One unusually large batch can dominate a small sample.

Separate the date a team changed its workflow from the dates used to assemble the report. Work already in progress may cross that boundary. For a pilot, maintain a clear rule for which work belongs to the old and new configurations. If that cannot be established, report the mixed period separately and avoid a strong before-and-after claim.

Consider the work missing from your analysis. Unmerged changes, abandoned work and unusual repository processes may matter to the question you are asking even if your selected metric does not describe them. Make a short note of exclusions. The aim is to state what the report covers, not to imply that one number summarises the entire review process.

Retrieve reports without mistaking links for data

GitHub's usage metrics API reference documents a daily repository report endpoint and its access requirements. Report retrieval returns download links and date information. Treat obtaining those links as a retrieval step, then validate the downloaded report before calculating anything. Use the current reference for authentication and endpoint details rather than copying credentials into a shared example.

Keep raw reports in an approved private location. A team may need only aggregate counts for the final analysis, but the source can still contain information that should not be placed in a public repository. The synthetic example above deliberately avoids any requirement to export real records for a tutorial.

Record the report date and when it was retrieved. This helps distinguish a change in the underlying data from a change in your calculation. If you refresh a report, rerun the calculation consistently and document the replacement. Do not compare a newly refreshed period with an earlier snapshot without considering whether the source data changed.

Troubleshoot a surprising addressed rate

If the rate is unavailable, inspect the denominator first. No evaluated comments means there is no rate to compute. Keep that state distinct from a measured zero rate. If values are absent, inspect the source and schema instead of treating the absence as a negative outcome for the team.

If the percentage jumps, compare the counts and the composition of the work. Look for a change in the repositories included, the types of tasks or the review configuration. Recalculate a small subset manually to rule out a denominator error. Only after those checks should you investigate whether the workflow itself changed behaviour.

If two dashboards disagree, write down each one's filters, time window and weighting method. A repository average and a comment-weighted total can differ even when both calculations are internally correct. Resolve the question each view answers before trying to force the numbers to match.

If the metric improves while reviewers report worse feedback, inspect examples from the disputed period. The disagreement is useful information. It may reveal that the measure captures developer response but misses the cost of distracting comments, or that a small set of difficult tasks is hidden inside a larger total.

Test the calculation before sharing a dashboard

Check a few deliberately simple inputs before connecting the calculation to reports. Ten addressed comments out of ten evaluated should produce 100%. Zero addressed out of ten should produce 0%. Zero evaluated should produce an unavailable result. A numerator larger than the denominator should cause investigation rather than a percentage above 100% appearing unnoticed in a presentation.

Also test the combined example in this guide. If your dashboard produces 50% for the combined comment-weighted result, it is averaging repository percentages. Fix the aggregation or relabel the view before other people rely on it. These checks validate the arithmetic, not the quality of the underlying feedback.

Turn the analysis into a review decision

Finish the pilot with a decision that matches the evidence. You might retain a skill because its feedback was relevant in the sampled cases and the team could act on it. You might revise the skill because it repeatedly raised concerns outside its intended scope. You might continue the pilot because too little comparable work occurred. Each is a reasonable outcome when its basis is explicit.

Keep a short decision record with the skill revision, sample scope, calculation method and unresolved questions. Link to private evidence through the team's existing access controls. Avoid copying full review discussions into an unrelated repository just to make the record self-contained. A public summary can use aggregate observations and clearly labelled synthetic examples instead.

Finally, decide when to reassess. A significant instruction change, a new tool permission or a different model may justify another bounded pilot. Routine maintenance need not become a permanent measurement project. The useful result is a review workflow whose behaviour the team understands, with enough evidence to change it when it stops helping.

Frequently asked questions