Background coding agents: how unattended work ships reviewed PRs
What happens between the task and the pull request, what the tools' own docs promise, and a ticket that passed review, bounced on merge conflicts, and merged overnight.
Why we built YATFA
YATFA is working even when you are away
You decide how far it goes before it waits for you.
23:00You log off. The night shift starts.
23:0001:0003:0005:0007:00
Acme/Activity
Tonight
What YATFA found and moved while you were away.
Small, safe work — tests, docs, small fixes — is approved by an agent at every setting. Operational changes always wait for a person.
It maintains knowledge autonomously
Your agents write memories during work — YATFA consolidates them into versioned articles.
Knowledge / This week
During the work
Written during a ticket
The provisioning job’s docblock promised a retry — the class had none.
→ folded into Background-Job Retry Integrity
Written during a QA pass
The “retry works” proof ran through a path that bypasses the retry machinery.
→ folded into Background-Job Retry Integrity
Around the work
Written during a review
A Cloudflare 403 from a write tool is transient — a byte-identical retry succeeded.
→ started a new article
Article · Background jobs
Background-Job Retry Integrity: the Promise, the Arm, and the Proof
When a background job fails, will it actually retry — and how do you prove it?
Grep the job class for retry_on whenever prose promises a retry — the machinery lives in class configuration, and the prose may live nowhere.+ added from a ticket · Sep 22, 2026
Prove a retry with perform_now and the test adapter’s enqueue count — subject.perform bypasses rescue handlers entirely.+ added from a QA pass · Sep 22, 2026
New article · started from one finding
Cloudflare 403 is transient — retry onceIt arrived with “do not retry” in the payload — and an identical immediate retry succeeded.
Knowledge / Background jobs / Background-Job Retry Integrity
Article · Background jobs
Background-Job Retry Integrity: the Promise, the Arm, and the Proof
Arm the aggregate error — the only arm that can fire past a rescue-record-re-raise conversion.
The final three are filed and unclaimed — retry_on count zero. Watch it.
Outdated — contradicted by the state on main
Also checked in this pass ✓ the aggregate-arm rule — still accurate ✓ the token-exchange escape hole — still accurate
Versions
v8Arm the aggregate error, not the base classSep 22, 2026
v7Grep for retry_on when prose promises a retrySep 22, 2026
v6The end-to-end retry QA methodSep 22, 2026
v1Founded from six same-day capturesSep 22, 2026
Documentation describes what the code does. These articles hold what the code can’t tell you.
Visual drafts show you the interface instead of describing it
Describe the screen you want. YATFA renders a few genuinely different directions from your running product. You pick one — and the pick is what gets built.
1
Your brief“People scroll past a chart to find three numbers. Make the numbers the first thing they see.”Already settled: keep the navigation · use the app’s own components
Your screen today
2Three directions, each made from your running app — compare them
3
Direction 1 of 3
Numbers first
Picked
Ticket
Code
Review & QA
Ready
The build runs on the render you pick. A direction you reject with a reason is never tried again.
A roadmap is a direction here — tickets are cut at execution time, not before
You describe what you want to see. YATFA turns your direction into proposals and tickets, cut slice by slice from the codebase as it stands.
main
Direction · written by you
Reports — one surface that explains what was delivered Reading main as it stands
- shift_reports_controller.rb
- time_window.rb — the resolver the sibling pages use
- reports/index — today: “this calendar month” only
Slice 1Proposal
Give the reports page any date windowPresets plus a custom from/to — through the resolver the sibling pages already use.
CodeReviewQAMerged
slice 1 merged
Slice 2Proposal · next
Move the delivery chart off the dashboardCut now — from code that already has the window control.
cut from here
Paused — nothing left to derive
Every slice this direction named has landed. What remains is a reading for you, not code.
CompleteResumeRewrite
From the blog
Testing AI-generated code: why green tests are not enough
Research on weak tests and false greens in AI coding benchmarks, what to verify beyond a passing suite, and a ticket where QA caught what specs missed.
Memory for AI coding agents: architecture and persistence
Why AI coding agents forget between runs, how context rot degrades long tasks, and how to build persistent, versioned codebase knowledge.