Skip to content

Commit Timing ​

This lens answers: when does the code change, and does any code change off-hours more than this team's own habits explain? It also shows whether someone's weekend or late-night commits repeat, or came from one busy stretch.

It only knows when a commit was stamped, on the author's own machine. It doesn't know when anyone worked, how long, or how they felt, and a machine set to the wrong time zone looks like night work. Off-hours commits can't show stress or burnout: read every number here as something worth asking about, never as a verdict on a person.

Reading the numbers ​

What you seeWhat it means
"8% of 938 commits off-hours (late night or weekend)"The share of counted commits made on a Saturday or Sunday, or from 11pm to 5am on any day, in the author's local time. Counted means commits by people touching 20 files or fewer. A commit that's both late night and weekend counts once.
"92% of commits land on weekdays between 9am and 7pm"The share of the punch card's commits in those hours. This and the off-hours share usually won't add up to 100%: weekdays from 7pm to 11pm and from 5am to 9am are neither. A team that always commits at 9pm on weeknights has no off-hours at all.
"the busiest hour is Fri 11am"The fullest single cell of the punch card, over the whole period.
"120 bot commits and 30 AI-agent commits are not counted."Commits left out of the punch card, because their hours are nobody's. Merge commits are left out too, and aren't in this count.
"services/payments 41 of 120 (34%) changes off-hours more than chance explains at the repo's 8%."41 off-hours commits out of 120 is too many to be luck if this package ran at the repo's own rate. It says the package is unusual for this team, not why.
"No package changes off-hours more than chance explains; 4% of commits are off-hours repo-wide."Every tested package is in line with the team's own rate. It doesn't say the hours are good: a team that works every weekend has a high rate, and nothing stands out against it.
"services/web keeps the most regular hours: 3 of 300 commits off-hours (1%)."The package with the lowest share, among those with 30 or more commits and under half the repo's rate. Here "regular" means steady working hours, which is the opposite of a person being "regular" below.
Chance allowsThe most off-hours commits that chance explains for a package of this size at the repo's rate. One more makes it above chance (see How it's measured).
"weekends in 11 of 30 active weeks"Of the 30 weeks this person committed at all, 11 had a weekend commit. Shown in the accent, this person counts as regular: at least a third of their active weeks, over at least 6 active weeks.
"late nights in 41 of 50 active weeks (71 nights)"The same measure for late-night commits. The 71 is distinct nights, and a 1am commit counts toward the evening before.
"weekends in 1 of 29 weeks · 2 late nights"The Off-hours row on a person's card, in every lens. It's flagged when they count as regular.
"Nobody works weekends or late nights regularly"Nobody who can be judged: people with 8 or more commits and 6 or more active weeks. Someone with 5 active weeks, all of them weekends, isn't called regular, because that's too short to call.
A different total in each placeThe inspector's headline, the punch card's heading and the time strip can show three totals, such as 1,200, 1,206 and 1,356. The headline leaves out commits touching more than 20 files, the punch card keeps them, and the time strip also includes bots and AI agents.

The table's tiers, which you can filter by, place each tested package: Below the repo's rate, Above the rate, within chance, Above chance and Above chance, twice the rate. They use the exact shares, so a package and the repo can both print 4% and land on different sides of the rate. A package exactly at the rate is in the first tier.

Four example repositories ​

These repositories are invented, but each result is what GitRelic reports for commits like these. Each one reads differently, and the last two show why a high number needs a second look.

A Go service spread over four time zones ​

4% of 1,200 commits off-hours (late night or weekend)

No package changes off-hours more than chance explains; 4% of commits are off-hours repo-wide.

86% of commits land on weekdays between 9am and 7pm; the busiest hour is Wed 4pm. 120 bot commits and 30 AI-agent commits are not counted.

Nobody works weekends or late nights regularly (in a third or more of their active weeks). The most off-hours:

Tariq Aziz: weekends in 11 of 49 active weeks · 0 late nights

What it means: every commit is placed by its author's own clock, so people eight or more hours apart land on the same grid, and the punch card looks like one office. A 10am commit in Tokyo and a 10am commit in Berlin fill the same cell. Nothing here says where anyone is, or whether their working days overlap.

The percentages also leave a gap: 86% in working hours and 4% off-hours leave 10% unaccounted for. Those are weeknight commits around 8pm, which fall in neither group. Check them on the punch card, because no number here will point them out.

What to look at: nothing is asking for attention. To see how the team's time zones spread, the offset is at the end of each date in git log --format=%aI.

A Python repo with one package on its own schedule ​

8% of 938 commits off-hours (late night or weekend)

services/payments 41 of 120 (34%) changes off-hours more than chance explains at the repo's 8%.

92% of commits land on weekdays between 9am and 7pm; the busiest hour is Fri 11am.

services/web keeps the most regular hours: 3 of 300 commits off-hours (1%).

Nobody works weekends or late nights regularly (in a third or more of their active weeks). The most off-hours:

Omar Haddad: weekends in 14 of 47 active weeks · 0 late nights

PackageCommitsOff-hoursOff-hours %Chance allowsAbove chanceTier
services/payments120413416yesAbove chance, twice the rate
services/notifications123253Above the rate, within chance
services/api50025554Below the repo's rate
services/web3003135Below the repo's rate

What it means: this is a signal about code, not about anyone. services/notifications has a 25% share and isn't flagged: with 12 commits, 3 off-hours is well within luck. services/payments has 34% over ten times as many commits, which luck doesn't explain. Nobody is regular, so the work is spread across people.

What to look at: select services/payments, and the punch card redraws as "services/payments: 120 commits, 34% off-hours". Here the off-hours sit in two cells, Saturday 7am and 8am, and hovering the 8am cell says "by Omar Haddad 7, Priya Nair 7, Jonas Lind 6". The same hours, shared around, looks like a schedule rather than anyone working late. Ask what happens on Saturday mornings: a planned window is a fine answer, and Fix rate shows whether changes there need fixing more often.

A Java repo with one crunch and one repeating pattern ​

8% of 830 commits off-hours (late night or weekend)

services/billing 24 of 120 (20%); services/search 22 of 110 (20%) change off-hours more than chance explains at the repo's 8%.

One person works off-hours regularly (in a third or more of their active weeks).

Tomas Berg: weekends in 11 of 30 active weeks

What it means: the two packages look identical here, and they aren't. All 24 of services/billing's off-hours commits came from a single weekend, before a release. Its author's card reads "weekends in 1 of 29 weeks · 2 late nights" (three late commits, two of them the same night), and they aren't on the list at all. services/search has 2 commits on each of 11 Saturdays, a pattern that keeps coming back.

The package test counts commits, not weeks, so it can't tell these apart. That's what the person measure is for.

What to look at: treat these as questions about how the work is arranged. For services/search: is there a weekend on-call rota, a job only one person can run, or a choice someone enjoys? For services/billing: was that weekend planned, and did the release need it? Neither answer comes from this lens, and neither is a finding about the person.

A Rust repo whose clocks are wrong ​

20% of 800 commits off-hours (late night or weekend)

crates/engine 161 of 400 (40%) changes off-hours more than chance explains at the repo's 20%.

61% of commits land on weekdays between 9am and 7pm; the busiest hour is Tue 4pm.

crates/cli keeps the most regular hours: 0 of 400 commits off-hours (0%).

2 people work off-hours regularly (in a third or more of their active weeks).

Felix Brandt: late nights in 41 of 50 active weeks (71 nights)

Noor Aziz: late nights in 40 of 50 active weeks (65 nights)

What it means: this looks alarming, and it comes from the clocks. crates/engine's authors work ordinary days eight hours behind UTC, in a development container set to UTC, so every commit records an offset of +00:00. Their 3pm reads as 11pm, and their Friday 5pm reads as Saturday 1am. crates/cli's authors really are at +00:00, so the two look the same to GitRelic, and it has no way to warn you.

What to look at: the shape on the punch card gives it away. Select crates/engine, and the day runs from 5pm to about 3am without a dip. Midnight is as busy as any hour of it, Saturday has commits only from 12am to 3am, Sunday has none, and Monday's early hours are empty, because that's Sunday afternoon for the authors. Real late nights usually land on top of a normal day rather than replacing it.

To confirm, read the offsets with git log --format=%aI -- crates/engine: a +00:00 or Z on commits from people who aren't at UTC settles it. Setting the machine's time zone fixes new commits. Old commits keep what they recorded, and GitRelic has no setting to correct them, so read those packages and people with that in mind.

How it's measured ​

GitRelic reads the non-merge commits in the analyzed period, which is the last 12 months by default (--since all looks further back). Each commit gets one time: its author date, with the UTC offset git recorded for it. That gives the author's local day and hour, so rebasing a commit later doesn't move it.

If the offset is wrong, so is everything after it. GitRelic takes the recorded offset at face value and can't detect a mistake: a machine set to UTC puts its commits on UTC hours, and a laptop still on its home time zone while traveling keeps home hours. The weekend is always Saturday and Sunday.

Who is counted where ​

A commit is sorted by its author's email address (see what counts as AI and what counts as a bot). A person's commit with an AI tool in a Co-authored-by: trailer is still that person's commit.

PeopleAI agentsBots
The punch card and the 9am–7pm shareEvery commitLeft out, and counted in the "not counted" lineLeft out, and counted in the "not counted" line
The off-hours rate and the package testCommits touching 20 files or fewerLeft outLeft out
A selected package's or file's hoursCommits touching 20 files or fewerLeft outLeft out
A selected person's hoursEvery commit they made, however largeEvery commit it madeCan't be selected
Off-hours, by person8 or more commits, and 6 or more active weeksNever listedNever listed
The Off-hours row on a person's card8 or more commits8 or more commits, and never flagged as regularNo card
The time stripEvery commitEvery commitEvery commit

Someone with several email addresses counts as one person, when GitRelic can tell they're the same (Contributors has the rules).

The package test ​

  • The rate is the repo's own. It's every counted commit, once each, including commits to files since deleted.
  • A package's count has stricter rules. A commit counts once in each package it touched, and only files that still exist in the analyzed commit count.
  • A package is one of the repository's own declared workspaces, else packages/*, apps/*, libs/* or services/*, else the top-level folder. Files at the top of the repository are (root).
  • The test asks: if every commit in this package had the repo's rate of being off-hours, how likely is at least this many? Under 2.5% is above chance. It's the same test Fix rate uses.

So the bar depends on size. At the repo's 8.3% (78 of 938), 12 commits with 3 off-hours has a 7.2% chance, so it's within chance; 4 would be 1.4%, so chance allows 3. For 120 commits, 16 off-hours is 4.1% and 17 is 2.1%, so chance allows 16, and 41 is far beyond it.

Packages with fewer than 10 commits aren't tested and aren't in the table, though their commits still count toward the rate.

Regular people ​

Weeks start on Monday, in the person's local time, and an active week is one with any commit. Weekend weeks and late-night weeks are counted separately, and either can make someone regular on its own: at least a third of their active weeks, over at least 6 active weeks. So 2 of 6 counts, and 1 weekend week plus 1 late-night week in 6 doesn't. Everyone is held to the same fixed bar, with no chance test and no comparison to the team.

Late nights are distinct nights, so an 11pm commit and a 12am commit straight after it are one night.

The list under Off-hours, by person names every regular first, then fills up to five with the highest share of off-hours weeks, among people with 8 or more commits and 6 or more active weeks. Being on the list doesn't make someone regular: only the lines in the accent do.

The stress score in --json ​

The --json report also holds a stress score for each author with 8 or more commits, AI agents included and bots not: late-night percent × 0.6 + weekend percent × 0.4, rounded. The dashboard doesn't use it. It's a share of commits, so it can't tell one crunch from a habit, and despite the name, it measures timing and nothing else.

In the dashboard ​

Find this lens under Team & Activity. On the Overview it reads "8% off-hours" and lists its packages, above chance first. It earns a badge when at least 2 packages are above chance and those are at least 10% of the tested packages.

  • The punch card is the main picture. Days run down, Monday first, and hours run across. Each cell is that day and hour's commits, shaded on a log scale so a few late commits still show beside a busy afternoon. Off-hours cells take the risk color, over a faint band marking the weekend's rows and the 11pm–5am columns. The rest stay grey.
  • Select a package or file, and the grid redraws from its own commits, with off-hours in the accent. Its heading, such as "services/payments: 120 commits, 34% off-hours", can be higher than the package's table row: it counts commits to files since deleted, and everything under the folder, including any workspace nested inside it.
  • Select a person, and the grid redraws from every commit they made, with off-hours in the accent.
  • Hover a cell for its day and hour, and its commits and share of the repo's. With a selection, you also get how many of them belong to it ("20 commits touching services/payments", worded the same way for a person's own commits). Who committed there is named too ("by Omar Haddad 7, Priya Nair 7, Jonas Lind 6"), except on a person's own hours.
  • Against chance (the corner toggle) draws every tested package as a dot, by its commits (across, log scale) and its off-hours share (up). The line is the repo's rate, and the shaded band is what chance allows, wide on the left and narrowing as packages grow. Dots above the band are above chance. Hover a dot for "3 of 12 commits off-hours (25%)" and "Chance allows up to 3 at this size". Click one to select its package.
  • The time strip under the picture shows all commits per week and lights the weeks a selection touched. It doesn't separate off-hours commits.
  • The inspector, with nothing selected, shows the headline and findings quoted above, the five packages with the most off-hours (above chance first), and Off-hours, by person. Select anyone listed to see their hours.
  • The table (T) lists every tested package: commits, off-hours, the share, what chance allows and whether it's above chance, highest share first, with the tiers above.

If no package has 10 commits yet, Against chance says so, and the rest of the lens still works.

What it can't tell you ​

  • Whether anyone is stressed, burned out or overworked. A commit's time isn't time worked. Late commits can be a choice, a flexible schedule or a different time zone, and someone worn out by long days of meetings shows nothing here. Treat it as a reason to talk, and never as a measure of a person's effort or performance.
  • Whether the clock is right. Machines set to UTC, travel and wrong clocks all look like real hours, and nothing warns you.
  • When the work happened. A commit has one time. A branch squashed into one commit shows as one commit at one moment, and work done at 2am and committed at 10am shows as 10am.
  • About evenings and early mornings. Weekdays from 7pm to 11pm and 5am to 9am aren't off-hours. They show on the punch card, but no number flags them.
  • About your calendar. The weekend is always Saturday and Sunday, so a team with a Friday and Saturday weekend reads wrongly. Public holidays aren't known, and neither are part-time schedules.
  • Whether a habit is everyone's. The test compares each package with the repo's own rate, and that rate includes the package. If one package holds most of the commits, it mostly sets the rate it's tested against. If the whole team works weekends, nothing is above chance.
  • Whether one package above chance matters. Each package is tested on its own, so in a repo with 40 tested packages, as many as one could land above chance by luck alone. A repeating pattern in the cells is stronger evidence than the flag.
  • Fix rate — the same chance test, on changes that needed fixing.
  • Contributors — who works on what and when, and how email addresses become one person.
  • AI & Co-authors — which addresses count as AI agents and bots, and how much AI is in the commits.

Released under the MIT License.