Measure code quality with the quality report

agentx quality report looks at the pull requests (PRs: proposed code changes) merged into your main branch and tells you, week by week and part by part of the code:

• how many PRs merged, and how many of them were fixes,
• how many fixes repaired something merged a day or three days earlier (a regression: something that used to work and broke),
• how long work took, from the issue being opened to the PR being merged,
• how many minutes of automated checks (CI: continuous integration) the PRs used,
• what the AI models cost per merged PR, from AgentX's own records.

Use it to see whether a change in how your agents work makes things better or worse. Save one report as a baseline, change something, and compare the next report with it.

Before you start

• The GitHub CLI (gh) is installed and signed in. The report reads GitHub through it, with your own access. It only reads; it never changes anything on GitHub.
• For model cost, run the command in the folder where AgentX runs, so it can read AgentX's task records in .agentx/db.sqlite. Without them, cost shows as unattributed.

Run a report

1. In a terminal, go to a checkout of the repository you want to measure.
2. Run agentx quality report --since 2026-01-05 --until 2026-01-18.
3. Read the three tables: Total, Week of and Area.

Leave out --since and --until to get the last three weeks up to today. If you run it outside a checkout, add --repo your-org/your-repo.

A three-week report takes under a minute. Most of the time goes to reading CI timings, one GitHub call per CI run. Those timings are kept in .agentx/quality/ci-runs/, so the next report is faster. Add --no-ci to skip CI minutes.

What each column means

These definitions are exact: the numbers are counted this way, and two reports are only comparable because they are.

• Release PR: a PR made by release-please, the tool that publishes versions. Its title starts with chore(main): release (or chore: release), its branch starts with release-please--, or it has an autorelease: label. Release PRs are left out of every row. The report says how many it left out, and what the fix share would be with them counted.
• Agent PR: a merged PR that is not a release PR, and has an agent: label (such as agent:coder) or a branch name starting with claude/. The week and area rows count agent PRs. Add --scope all to count every non-release PR there instead. The Total table always shows both.
• PRs: merged PRs, by the day they merged, in UTC (world time, no time zone).
• Fixes: PRs whose title starts with the type fix, as in fix: …, fix(scope): … or fix!: ….
• Fix share: fixes divided by PRs.
• Reg ≤24h (/100) and Reg ≤72h (/100): regression fixes, with the number per 100 merged PRs in brackets, so weeks of different sizes compare. A regression fix is a fix whose title or description mentions #N (or a link to pull request N), where N is a PR that merged at most 72 hours before the fix merged. Reg ≤72h counts all of them; Reg ≤24h counts those within 24 hours, so it is part of the 72-hour number. A PR merged up to 72 hours before --since still counts as the PR being fixed.
• Fix review/new test/UI: three counts among the fixes, split by /. The first counts fixes whose title or description mentions a review: the word review, reviews, reviewer or reviewers, as in "(#12 review)", "the review of #12" or "reviewer asked". It does not count reviewed or reviewing, "after review" (as in "Merge after review"), or review followed by only, before or needed (as in "for review only"), because those describe the process, not feedback being answered. The second counts fixes that add a new test file: a file under test/, a file named like name.test.ts, or a file in a __fixtures__ folder. Editing an existing test does not count, because almost every fix does that. The third counts fixes that change a screen: a file under src/daemon/ui/, src/web/, src/boards/ or apps/, or src/daemon/board-dashboard.ts.
• Area: the folder directly under src/ with the most changed files in the PR, such as daemon or channels. Files directly in src/ count as src. A PR that changes nothing under src/ is other. Ties go to the name first in alphabetical order. Only the first 100 changed files are read.
• Lead issue p50/p90: for PRs that close an issue (GitHub's "closing issues" link, such as Closes #12), the time from that issue being opened to the PR being merged. With several issues, the oldest one counts. p50 is the median (half the PRs were faster); p90 means 9 in 10 PRs were faster. Both use the nearest-rank method.
• Lead PR p50/p90: the same for PRs that close no issue, from the PR being opened to it being merged.
• CI min: the sum of job durations of ci.yml runs on the PR's commits, including re-runs. It is — with --no-ci.
• Cost/PR: the model cost of the AgentX tasks that built the PR, divided by the PRs that have a cost. A task built a PR when it pushed the PR's branch (with git push naming the branch, or a GitHub tool naming it) or opened the PR. A task that only read, mentioned or answered on a PR does not count. A task that built two PRs is split between them. Costs use the same prices as the Costs page.
• matched: the attribution rate: the share of PRs with a cost. When no PR in a row has one, the cell says unattributed, never $0.00.
• external: PRs written in a Claude cloud session (their description links to claude.ai/code/session_…). They are billed outside AgentX, so they are counted apart and left out of the attribution rate. This holds even when an AgentX task later pushed the branch, for example to rebase it or fix a merge conflict: that task did not build the PR.
• Local follow-up: the cost of those later AgentX tasks on external PRs. It is never part of Cost/PR. When there is any, a line under the Total table shows it, such as Local follow-up on external PRs (rebases, conflict fixes), not in Cost/PR: agent PRs: $1.20 on 3 PRs. With --public, the line shows only the number of PRs.

Model cost only covers tasks recorded on the machine where you run the report. Point --db at another copy of .agentx/db.sqlite to read a different machine's records.

Compare with a baseline

A baseline is a saved report you compare later reports with. It is kept in .agentx/quality/baselines/ on your machine, which git ignores.

1. Run agentx quality report --since 2026-01-05 --until 2026-01-18 --save-baseline before.
2. Make the change you want to measure, and let it run for a while.
3. Run agentx quality report --since 2026-01-19 --until 2026-02-01 --compare before.
4. Read the Compared with baseline lines under the tables: the change in PRs, fix share (in percentage points), regression fixes, median lead time, CI minutes and cost per PR (in %).

Share a report in public

Cost amounts stay on your machine. To post a report somewhere others can read it, add --public. It shows cost only as a percentage change against the baseline you compare with, and leaves every amount out, in the text and in the JSON.

1. Run agentx quality report --compare before --public.
2. Copy the output where you want to share it.

Without --compare, a public report says no baseline in the cost column.

To post on GitHub (an issue or a PR comment), add --format markdown as well. The tables then show as real tables there, instead of lines of text that wrap.

1. Run agentx quality report --compare before --public --format markdown.
2. Paste the output into the GitHub comment box.

Hints such as "No AgentX trace store at …" and "Saved baseline …" are printed on the error stream, not with the report. They can name folders on your machine, so they never end up in what you paste or post.

Use the JSON output

Add --format json (or --json) to get the same report as JSON, for scripts and for posting week by week. The shape is stable: fields are only added. A rename or removal changes the schema value.

{
  "schema": "agentx.quality-report/v1",
  "repo": "your-org/your-repo",
  "base": "main",
  "since": "2026-01-05",
  "until": "2026-01-18",
  "generatedAt": "2026-01-19T08:00:00.000Z",
  "scope": "agent",
  "public": false,
  "totals": { "agent": "Row", "all": "Row" },
  "weeks": ["Row plus \"week\": the Monday the week starts, YYYY-MM-DD"],
  "areas": ["Row plus \"area\": the area name"],
  "excluded": { "releasePrs": 12, "fixShareWithReleases": 0.31 },
  "comparison": "only with --compare, see below"
}

Each Row has these fields:

Field	Meaning
mergedPrs	PRs in the row.
fixes	Fixes among them.
fixShare	fixes / mergedPrs, from 0 to 1. null when there are no PRs.
regressionFixes24h, regressionFixes72h	Regression fixes within 24 and 72 hours.
regressionFixes24hPer100, regressionFixes72hPer100	The same per 100 merged PRs, to one decimal. null when there are no PRs.
fixesMentioningReview, fixesAddingTests, fixesTouchingUi	Fixes that mention a review (rule above), add a test file, or change a screen.
leadTime.issue, leadTime.pr	Each { "n", "medianHours", "p90Hours" }. Hours are null when n is 0.
ciMinutes	CI minutes, or null with --no-ci.
cost.perMergedPr, cost.total	Dollars. null when nothing matched. Left out with --public.
cost.attributedPrs	PRs with a cost. Never counts an external PR.
cost.externalPrs	Cloud-session PRs, whether or not an AgentX task later pushed them.
cost.followUpPrs	External PRs that an AgentX task later pushed (local follow-up).
cost.followUpTotal	Dollars of that follow-up, not part of perMergedPr. null when there is none. Left out with --public.
cost.attributionRate	attributedPrs / (mergedPrs - externalPrs), from 0 to 1, or null.
cost.status	attributed, unattributed, or external when every PR in the row is external.
cost.changePct	Change of perMergedPr against the baseline, in %. Only with --compare (and null in public output without it).

With --compare, comparison holds baseline, baselineSince, baselineUntil, totals.agent, totals.all and areas (one entry per area found in both reports). Each entry has mergedPrs, fixSharePts, regressionFixes24h, regressionFixes72h, regressionFixes24hPer100, regressionFixes72hPer100 (null against a baseline saved before these fields existed), issueLeadMedianHours, prLeadMedianHours, ciMinutes, costPerMergedPrPct and attributionRatePts: the current value minus the baseline value.

All options

Option	Default	What it does
--since <date>	20 days before --until	First merge day, YYYY-MM-DD.
--until <date>	today (UTC)	Last merge day, YYYY-MM-DD.
--repo <owner/name>	the checkout's origin	The GitHub repository to read.
--base <branch>	main	The branch PRs merged into.
--scope <scope>	agent	Which PRs the week and area rows count: agent or all.
--no-ci	—	Skip CI minutes.
--db <path>	.agentx/db.sqlite	AgentX task records to read cost from.
--public	—	Show cost only as % change against --compare.
--save-baseline <name>	—	Save this report as a baseline. Letters, digits, ., - and _.
--compare <name>	—	Show changes against a saved baseline.
--format <format>	text	text (terminal tables), markdown (tables for a GitHub comment) or json.
--json	—	Same as --format json.

Check it worked

1. In a terminal, run agentx quality report --no-ci.
2. Check the first line names your repository and the dates you expect.
3. Check the Total table shows a number of PRs above 0, unless nothing merged in those dates.
4. Run agentx quality report --save-baseline test, then agentx quality report --compare test, and check the Compared with baseline lines show +0 changes.

If something is wrong

• gh auth login or "not logged in": the GitHub CLI is missing or signed out. Install it, then run gh auth login in a terminal.
• "Could not tell which GitHub repository this is": you are outside a checkout, or its origin is not on GitHub. Add --repo your-org/your-repo.
• Cost says unattributed everywhere: the report found no AgentX task records, or no task pushed or opened these PRs. Run it from the folder AgentX runs in, or add --db with the path to .agentx/db.sqlite. PRs built by a Claude cloud session show as external instead.
• "GitHub lists only 1,000" warning: a week had more CI runs than GitHub returns in one list, so that week's CI minutes are lower than the real number. Report shorter windows, or treat CI minutes for that week as a minimum.
• It takes longer than a minute: most of the time is CI timings. The next run is faster, because finished runs are remembered. Add --no-ci if you don't need CI minutes.
• The numbers differ from a count by hand: check the definitions above. Merge days are in UTC, release PRs are left out, and the rows count agent PRs unless you add --scope all.
• "No baseline": save one first with --save-baseline <name>, in the same folder. Baselines live in .agentx/quality/baselines/.