June 1, 2026 · 6 min

A weekly git digest for solo projects

One of the strange things about working alone is that nobody asks you what you shipped this week. There’s no standup, no Friday retro, no manager who skims your PRs at end of sprint. You either keep your own status, or you don’t.

I gave up keeping my own about a year in. What I do instead is have my Hermes box read the git history of my projects every Sunday and send me a digest. It’s mostly for me. Occasionally it’s for an investor email I would have otherwise put off until Tuesday.

What it looks like

The message lands at noon every Sunday:

week of 31 May - 6 Jun

meragpt/webapp  (14 commits, +1,247 -312)
  - Hermes v0.16 rollout + golden image v6
  - per-box flag for desktop OAuth (shipped behind
    feature flag, no customers enabled yet)
  - docs/messaging with real v0.16 screenshots
  Real change of the week: every customer is on v6
  with their data preserved. Image v5 deleted.

infra/edge-proxy  (1 commit, +61 -14)
  - carve-out for Hermes auth paths when flag is on

Quiet weeks: openclaw-workspace, agent-test.

Things you said you'd do and didn't:
  - "write the SQL standup bot" (mentioned Mon, no
    commit)
  - the marketing site refactor (touched zero files)

Next week, in your own words from Monday: "finish
the desktop coexistence and ship the connect-desktop
docs for real."

The bottom two sections are why this exists. The first one (“real change of the week”) is the agent summarising what actually moved. The other two are accountability. The agent reads back what I said I’d do, and notes what I didn’t. It’s gentle but it’s real.

The setup

Your box can’t read your local repo directly, but it can read a GitHub repo through the GitHub MCP plugin (one click in MCP → install). Sign in once, give it read access to the repos you care about, and the agent can run a git log on demand.

For the accountability bit, the agent needs to know what you said you’d do. That comes from one of two places:

Your Hermes chat sessions from the past week. If you talk to your agent about what you’re working on, that’s the source.
A weekly journal file you keep in memory. Less automatic, but works for people who don’t live in the chat tab.

The cron

CRON → CREATE. Schedule: 0 12 * * 0. Prompt:

Read my GitHub activity over the last 7 days across
all repos I have write access to. For each repo with
non-trivial activity:

- Total commits and rough +/- line counts
- 2-4 bullets of what actually changed (not a commit
  list; a human summary)
- One sentence: the real change of the week, if there
  was one

List quiet repos in a single line.

Then, read my last 7 days of chat sessions. If I said
I'd do something specific and didn't ship it, list it.
Be conservative; only flag the ones that are clearly
work I committed to.

End with one paragraph: what I said on the most recent
Monday about what this week would be.

Send to Telegram. Under 2000 chars. Voice: blunt,
no congratulations, no "great job".

That last line on voice matters. Without it, every digest ends with “Great work this week! Looking forward to seeing what you ship next!” which is the exact opposite of useful.

Tuning notes

Be specific about “non-trivial”. Without a definition, the agent counts merge commits and version bumps. I added a clause that filters those out: skip commits whose message starts with “Merge branch” or “Release”.

Define the accountability scope. The first week, the agent listed everything I’d ever mentioned wanting to do, which was demoralising. I narrowed it to “things I said I’d do this week” and the list got short and useful.

Cap the repos. If you have access to a hundred repos through your org, the digest is unreadable. Either list them explicitly in the prompt or filter to repos with commits in the last 7 days.

What it costs

Around 5 cents per digest depending on how busy the week was. About $2.50 a year. The bigger model bill is from running the digest twice when I want to A/B the wording.

The variant I’ve actually built

The exact prompt above is the public version. The one I use also reads PR review comments and counts how many times I said “TODO” in a commit message vs. actually finished the TODO. Both are noise but they’re funny noise, and the agent can’t see them anywhere else.

If you’re thinking about adding your own variants, a useful constraint: don’t add anything to the digest that isn’t already in a place you could look up yourself. The digest is for compression, not for new information.