Weekend Creative Agent Challenge: Commit Muse
Every night an agent reads what I actually did on GitHub and writes the day up in English and Japanese — no button, no prompt. The hard part wasn't the model; it was that GitHub had quietly stopped telling me what my own commits said.
Series: Weekend Challenge (1 article)
- 1Weekend Creative Agent Challenge: Commit Muse This article
I have a bad habit. At the end of a long day of merging pull requests and
chasing dependency alerts, I could not have told you what I actually did. The
work was real, the commits were there, but the day itself left no impression.
GitHub remembers every SHA and none of the story.
So I built something that remembers the story for me — and, more importantly,
something I never have to open.
Commit Muse is an always-on agent. At 22:00 JST it wakes itself up, reads
what I did on GitHub that day, and writes a short journal entry about it in
English and Japanese. There is no generate button on the site. There never was.
By the time I look, the writing is already done.
The vision: the tool you never open
Last weekend's challenge was about building a creative app. The thing about a
creative app is that it still waits for you. You open it, you type a prompt, you
press the button. The creativity is on demand, which means it is on your
schedule, which means most days it does not happen at all.
Turning it into an agent inverts that. The agent has its own schedule and its
own opinion about what today was about. My job is only to live the day; its job
is to notice it. The website is not an app — it is just a quiet room where the
writing accumulates.
Two rules the agent holds itself to, and they turned out to matter more than
anything else I built:
- It never invents work. Every concrete claim in an entry has to trace back to a real commit, pull request, or branch in that day's digest.
- It would rather write nothing than write something broken. An empty day is better than a fabricated one.
How I built it
The part I got wrong first
My first version read
/users/{user}/events/public , counted the commits inside
each PushEvent , and handed them to the model. It confidently reported "0
commits" on a day where I had pushed fourteen times.GitHub has trimmed those payloads. A
PushEvent now carries only before ,
head , ref , and a couple of ids — no commit messages at all. A
PullRequestEvent gives you an action and a number, but no title. All the
content has been removed from the content feed.The fix reframed the whole data layer: treat the events feed as an index of
where to look, not as the material itself. So the agent now groups events into
JST calendar days, and for each repository I pushed to that day it goes and asks
the repository directly:
GET /repos/{repo}/commits?author={me}&since={day}T00:00:00+09:00&until={day}T23:59:59+09:00That is where the real subjects live. Pull request titles come from one more
lookup each, capped at six a day so a backfill cannot burn the rate limit. Every
one of those calls returns
None instead of raising on a 4xx — a rate limit or
a repo that went private should never take the whole night down.The part I got wrong second
With commit messages flowing again, I hit a subtler version of the same problem.
I was fetching them with
?author=midnight480 , which seems obviously correct.It silently skipped the commit I had just pushed.
GitHub only links a commit to an account when the commit's email is a verified
address on that account. My local
user.email was a placeholder, so GitHub
listed the commit but attributed it to nobody — and ?author=<login> filtered
out my own work in my own repository.So the agent stopped delegating the question of identity. It fetches the day's
commits unfiltered and decides for itself:
def _is_theirs(commit, user): login = ((commit.get("author") or {}).get("login") or "").lower() if login: return login == user.lower() # GitHub knows: trust it author = commit.get("commit", {}).get("author", {}) name = (author.get("name") or "").lower() # GitHub doesn't: look at the commit email = (author.get("email") or "").lower() return user.lower() in (name, email.split("@")[0])The effect was not marginal. On 2026-08-11 the agent had been seeing 5 commits.
With identity matched on the commit itself, it sees 12 — including the real
ones like
fix(ci): remove continue-on-error from SARIF upload step , which had
been invisible behind a wall of merge commits.An agent that reads your work is only as good as its definition of "your".
Making a small model behave
I used Amazon Nova Lite through the Bedrock
converse API. It is cheap, it
is fast, and it is perfectly good at this — as long as you do not trust it.I compared it against Nova 2 Lite on the same day of real data. Nova 2 Lite was
more terse and drifted into first person ("I merged a pull request") even though
the system prompt is written in second person, and it left an English word
untranslated in the Japanese output. Nova Lite matched the brief better, so the
newer model lost on merit. Worth actually measuring rather than assuming.
Everything the model returns goes through a validator before it is allowed to
exist:
- it must parse as JSON and carry all six required fields
- the English must clear a word floor, and the Japanese roughly twice that in characters
- the title must not match a recent title
That last check earned its place. I originally asked the model in the prompt not
to repeat recent titles, and handed it the list. It cheerfully produced "A Quiet
Night's Prep" twice in one backfill. Asking does not work; rejecting does. In my
final run the guard fired twice and both retries came back with something new.
Three strikes and the day stays empty.
The floor that moves
My first validator held every entry to the same length, and it started throwing
away quiet days — days where the honest answer really is three sentences. So the
minimum scales with the material:
material = digest["commit_count"] + len(digest["highlights"])min_words = 60 if material >= 3 else 35A day with five commits owes you five sentences. A day with nothing to say owes
you three. A fixed word count is precisely how you end up with an agent that
pads.
Learning what I like
The only thing a human can do on the site is vote 👍/👎 and, optionally, pick a
preference. That preference goes into a persistent style memory in DynamoDB and
is folded into the next prompt.
It works, and I can show it. Here is the agent on 2026-08-11, before any
feedback:
...You also added a branch to address specific Snyk issues in the astro-notion-blog. It's impressive to see you handling these tasks with such precision and care.
I sent one thumbs down with "no flattery". Same day, same digest, regenerated:
...The astro-notion-blog repository also saw attention with a merge fixing 24 vulnerabilities, ensuring a safer blogging experience. While the backlog-tools-portal had fewer interactions, the overall day was filled with steady progress.
No fine-tuning. Just memory, fed back into the prompt.
The vulnerability I shipped, and what fixing it taught me
The first version of that feedback box was a text field. You typed whatever you
wanted, it was stored verbatim, and the next night it went into the prompt under
the heading "What this reader has told you, in their own words."
I ran a security review over my own code before writing this section, and that
design does not survive it. The endpoint is a public, unauthenticated HTTP API —
its URL is in the
config.json that the site itself fetches, so it is trivially
discoverable. The handler validated vote against ["up", "down"] and then
accepted any 200-character string as the note. The prompt read the last six.So six requests own the entire preference section of the prompt. Not "influence"
— own. And the output of that prompt is not private: it is written to
data/entries.json and published on a page that carries my name and keeps 30
entries of history. An unauthenticated stranger could steer what my journal says
about my work, and I would not find out until I read it.The comment I had written above that handler said: "if someone found this
endpoint, the worst they could do is tell it to write shorter sentences." That
sentence was the bug. I had reasoned about what a reasonable user would send
and written the conclusion down as a security property.
The obvious fix is to put an authorizer on the route. I did something else,
because I think it is the better lesson: I deleted the text field.
Feedback is now a closed vocabulary — eight slugs,
shorter , less_flattery ,
more_technical , and so on. The handler rejects anything else with a 400. The
slugs are not what goes into the prompt either; the agent looks each one up in a
dictionary that lives in its own source file and inserts its own sentence:PREFERENCES = { "less_flattery": "Do not compliment the developer. Report the day, do not praise it.", "more_technical": "Be more technical: name files, tools, and specific errors.", ...}load_style drops any slug that is not a key, so even a stored value that
somehow bypassed the handler cannot put a single character into the prompt.Authentication would have hidden this problem — the injection channel would
still exist, just behind a login. Removing the free-text field means there is no
channel at all, and the feature the reader actually wanted still works. When
untrusted input has to reach an LLM prompt, narrowing the vocabulary beats
guarding the door.
Starting life with something to show
A fresh deployment has an empty page, which is a sad first impression for an
agent whose whole pitch is "it's already done." So the agent has a backfill
mode: hand it
{"backfill": 6} and it walks back over the last six days that
had activity, oldest first, and writes any entry that is missing. Oldest first
matters — each entry can then see the ones before it and avoid repeating their
angle.AWS services and architecture
EventBridge Scheduler ──▶ Lambda (muse) ──▶ GitHub REST API (22:00 Asia/Tokyo) │ ├──▶ Amazon Bedrock (Nova Lite) ├──▶ DynamoDB (entries + style memory) └──▶ S3 (private bucket) │ CloudFront (OAC) ─┘──▶ reader
reader ──▶ API Gateway (HTTP) ──▶ Lambda (feedback) ──▶ DynamoDB style memory- Amazon EventBridge Scheduler — the only thing that starts a run, with a timezone-aware cron so 22:00 stays 22:00 across daylight saving.
- AWS Lambda (arm64, Python 3.13) — the agent, plus a separate tiny handler for feedback. No dependencies beyond the SDK, so there is no build step.
- Amazon Bedrock / Nova Lite — called through a cross-region inference profile, which is how the Nova models are invoked on demand.
- Amazon DynamoDB — one table,
USER#<login>as the partition key,ENTRY#<date>andSTYLEas sort keys.
- Amazon S3 + CloudFront — the bucket blocks all public access and has no website hosting enabled. CloudFront reaches it through Origin Access Control, and the bucket policy only trusts requests carrying this distribution's ARN.
- Amazon API Gateway (HTTP API) — exactly one route,
POST /feedback, and it accepts only a vote and one slug from a fixed list. No route anywhere in this system can make the agent generate, and no caller-supplied text is stored anywhere the prompt can reach.
Everything is one SAM template with per-function least-privilege roles. The
Bedrock policy names the specific model and inference profile ARNs, not
* .One more thing, because the input is my own commit messages: emails and anything
shaped like a credential are scrubbed before they reach the model or the public
page. My first redaction rule was "any token of 40+ characters," which promptly
redacted a legitimate branch name (
snyk-fix-799929ff... ). Broad redaction
destroys meaning; the rule is now a short list of specific credential shapes.What I learned
Trimmed APIs change your architecture, not just your parser. The moment
GitHub stopped shipping commit messages in the events feed, "read the feed"
became "use the feed to decide what to read." Half my build time went here.
Validation is where agent quality actually lives. Every genuine improvement
in output came from a rule that rejects a reply, not from a nicer prompt. Prompt
instructions are suggestions; validators are guarantees. The title-dedup check is
the clearest example — the same instruction in the prompt was simply ignored.
"Filter it server-side" is a decision, not a shortcut. Every filter you hand
to someone else's API is a definition you have accepted without reading it.
?author=<login> sounds like "commits by me" and actually means "commits GitHub
has successfully linked to me," and the gap between those two swallowed more than
half my history.Newer is not better by default. Nova 2 Lite lost to Nova Lite on my actual
data, on the specific things I cared about. Ten minutes of comparison beat an
assumption.
A comment asserting a security property is not a security property. I wrote
"the worst they could do is tell it to write shorter sentences" above a public
unauthenticated endpoint that fed an LLM prompt whose output gets published under
my name. I had modelled the polite user, not the attacker, and then written my
conclusion down as if it were a fact. Reviewing my own code against that comment
is what found it.
Least privilege is easy when you design for it up front. Writing the read
path and the write path as separate functions with separate roles cost nothing at
the start and would have been an unpleasant refactor later.
An agent's honesty is a feature you have to build. "Never invent work" is not
a sentence in a prompt — it is a digest the model cannot see past, a validator
that rejects thin output, and a quiet-day path that says "nothing happened"
without embarrassment.
The first thing it wrote about me
I deployed it, made the repository public, and let the agent run over the day it
was built. It had no idea it was writing about itself:
Today, midnight480 worked diligently on the 'aws-weekend-challenge-20260821' repository, starting with an initial commit to lay the foundation. They also added 'Commit Muse,' an intriguing agent designed to journal their GitHub activities. This repository has been made public, marking a significant milestone. [...] It's a quiet yet purposeful day, with each step meticulously documented.
Tomorrow it will do that again, and I will not be there for it. That was the
whole point.
Links
- Live app: https://d2zeae0indfv6t.cloudfront.net
Built with: Amazon Bedrock (Nova Lite) · AWS Lambda · Amazon DynamoDB · Amazon
EventBridge Scheduler · Amazon S3 · Amazon CloudFront · Amazon API Gateway · AWS
SAM · AWS IAM
Series: Weekend Challenge (1 article)
- 1Weekend Creative Agent Challenge: Commit Muse This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article