Daily project and learning log.

What I did

• Worked out that “empty queue” is the name of a tick’s outcome, not the state of a folder: Ready held 111 tasks and not one could be taken, but the system treats that as emptiness — it goes off promoting drafts instead of finding out why everything is blocked.
• The daemon kept shutting down on a fault we had fixed the day before; it was running the code it started with, and a restart was all it needed.

What the factory did

• Found that a renamed epic freed its name for a new one, so a reference written before the rename still resolves, only it points at a different epic; of 28 live citations four are broken and sixteen need the surrounding sentence read.

What I did

• Settled how the data crosses into the new database: a cutover, not a period of running both, and three rounds fixed the eight parameters that turn a direction into something executable.
• The crossing ran: lifalign-prod holds 5,236 rows, every table balances, and I checked on screen that the data is where it belongs.
• The first sign-in to the new application looped — three different errors in a row had one cause; fixed.
• Moved the functions to Frankfurt; until now a request from Prague was executed in Washington.
• Found that the daily 16:00 deploy I had set up the day before has never run: zero scheduled runs, and production is 95 commits behind.

What the factory did

• Finished moving the reminder scheduler onto the API; the web reaches Supabase nowhere any more.
• Finished binding the colours to tokens — no CSS module in the web writes a colour now, and the exception list is deleted rather than left empty.
• Converted the documents in lifalign/designs and the last two decision records to English, in batches overnight.
• Refuses to store a retirement older than the row itself, on every table that can be retired.
• Deleted the web’s unreachable copy of the day’s rules and repointed what named it.

What I did

• The orchestrator stopped shutting itself down when it cannot find the task behind a locked branch; that state clears on its own, so skipping one tick is enough.
• Had three cheap guard workflows merged into one billed job, because I was running out of free GitHub Actions minutes.
• Measured the factory’s throughput: 32 unique tasks reached a pull request in 24 hours, 52 in 48.

What the factory did

• Split a task that contradicted itself — “empty the whole colour allowlist” against “one screen per pull request” — into four that can be claimed, and split them along 550 measured violations rather than by file count.

What I did

• Settled how reminders will run: a cron runs SQL that only checks whether waking the backend is worth it, and the backend does the sending.
• Had production deploy once a day at 16:00.
• Dropped the “add as a flow moment” action; the Flow lístek no longer takes anything over from a todo.
• Settled that emptying the text in the diary leaves an empty record — the diary has no delete.
• Turned down the proposed fix for the false “Úkol se mezitím změnil” warning: the guard would skip a write that changes nothing, but the application makes exactly that write on purpose, so that a save with no change still passes the concurrent-edit check.

What the factory did

• Opened the move of the reminder scheduler onto the API, the last place the web reached Supabase directly; a cron in the database now wakes it only when there may be something to send.
• Moved Historie plánů and the Dnes dashboard behind the API, and thinned Uzavření dne down to one read.
• Gave every foreign key an index on its own side; Postgres indexes only the target, so deleting a parent scanned the whole child table.
• Named the columns on on delete set null across seven composite keys; without the list it also cleared user_id, which may not be empty.
• Readiness now refuses a replica that does not hold the issuer’s signing keys and so could not verify a sign-in token.
• Made the build fail and name the method when an adapter does not wrap its database work so a technical failure arrives above as one known type; nothing watched for the omission before.
• Tested for the first time what happens when saving a form fails — until now the tests only knew success: an expired session sends you to the login, a server outage is not passed off as a mistake in what you typed, and a refused form does not wipe what you wrote.
• Bound the colours in the seven Detail dne CSS modules and in the shared components to tokens, and added a detector for every syntax that can write a colour.

What I did

• Brought the Discord bot back up; it had died with an unexpected machine restart and stayed down because its configuration still held the repository path on the old drive.
• Found why #alerts went silent for eight days: one rejected message whose fingerprint ran to 2,224 characters against a 2,048 limit in the footer jammed the whole channel.
• Made the parked-task alert name what is parked; until now it sent an empty list.
• Went through the piled-up alert queue one at a time and insisted the cause of the slowness be measured rather than guessed; batching came out of it, instead of one card edit per alert.
• Extended branch cleanup to every namespace the factory pushes under; only claude/ and codex/ were being swept.
• Had the remote branches nobody needs any more deleted.

What the factory did

• Stopped one task file from living in two board folders at once.
• Promoted nothing and wrote down why: every draft it would have taken sits in one of five open promotion pull requests.

What I did

• Settled how a collision in the day’s plan is resolved: either rearrange the day or drop the colliding blocks — both ways stay.
• The smoke check now starts the whole stack itself, fake sign-in included, so it needs no dedicated account and no hosted project; it had been red on every task since 17 September.

What the factory did

• Moved eight more screens behind the API: the notification devices, Opakované bloky, Úkoly, Rituály, Rolling Planning, Detail dne, Uzavření dne and Plán dne.
• Built Měsíční plán and Roční plán, the last two rungs of rolling planning with nowhere to store themselves.
• Taught the API to drop colliding blocks; until now it could only rearrange the day, so Detail dne could not move without losing the feature.
• Renamed the colours from --color-daily-* to roles; the name claimed they belonged to one screen while every screen read them (974 occurrences in 62 files).
• Gave ScreenHeader the two slots its last two screens needed and were drawing their own heading for.
• Deleted the self-location line from fifteen documents; it was wrong twice over — it named folders that had moved and a drive the repository had left.
• Made decision 0017 derive its numbers from the model; the hand count was stale for the third time (3 → 16 → 28).

What I did

• Added books to the site as their own kind of content, next to projects, and reading notes from individual chapters under them.
• A journal entry can now hang off a book the way it hangs off a project, and the journal index filters by book.
• Books and notes have both language versions and are in the sitemap.

What I did

• Moved reading the board from the working tree to origin/main, after the orchestrator ran a long-finished task fourteen times and ate 20% of a five-hour limit; along the way it turned out Get-TaskFileTitle returned an empty title for all 92 Ready tasks.
• Let promotion pull requests merge on their own; a coordination run adds no code, it only moves tasks around the board.
• Counted what the factory has done so far: 504 task runs with a median of 13 minutes.

What the factory did

• Promoted five refined drafts and recorded three blockers.
• Refused the premise that the board was empty — Ready held 14 and 75 tasks — and instead of promoting, wrote two generated stubs into real tasks and parked one for me to decide.

What the factory did

• Moved eight screens behind the API: Kategorie času, Nastavení, Osobní kontext, Flow lístek, Ohlédnutí, Šablony rozpočtu dne, Rozpočet dne and Zastavení; each lost its Supabase adapter and its second copy of the rules.
• Opened the ninth, the notification device screens; that one waits to merge.
• Production schema drift now reports itself; until today a user found it first, on a screen that stopped working.
• Ran dotnet ef database update in CI for the first time; the one route by which the schema reaches a database had never been exercised.
• The health check now tries the database too; until today a live process was enough, so a deploy with the wrong password came up green and fell over on the first real request.
• Added a test that reads the schema as the migrations leave it and catches a write that depends on a default which disappears on the way.
• Fixed a randomly failing test — the fault was in the test, not the component: it did not wait for React to process the action’s result and skipped a state transition that runs once.

What I did

• Rewrote P15 and the incidents README: a record that changes nothing is not learning but an archive — whoever writes an incident proposes the change too.
• Had the incident written about the session that handed me a table of three Vercel projects as fact, with nothing to build it from.
• Added reading a pull request’s state before pushing to its branch to the developer skill; three of the day’s pull requests took commits an hour after they merged.
• Replaced the one-open-pull-request limit with the Touch set: a tick now skips only a task that touches files an open pull request already changed.
• Repointed the paths agents create worktrees and read their state from at the drive the repository moved to (the agent spotted it and proposed the change).
• Rewrote 68 task references so they carry no folder; the folder is the task’s state, so a reference broke every time a task moved (the agent spotted it and proposed the change).

What the factory did

• Promoted three refined drafts and parked the one only the owner can decide.
• Opened the record that the coordination pass promoted the same drafts three times in thirty-seven minutes, because nobody merged its branches.

What I did

• Brought up a new production estate in Azure and pulled a full deploy through it for the first time: migrations, image, container app, web, tag.
• Opened ingress on production; prod.bicepparam still had it off and would have cut production off again on the next deploy.
• Disconnected the old ai-coach Vercel project from git so no migration can break it and it keeps running on the old sign-in and the old database; the two new Lifalign projects went up beside it.
• Renamed Supabase to lifalign-prod, postponed staging and now deploy production straight from main; the free tier allows two projects and both are taken.
• Pulled the design gallery back to the development server; its prerendered images had grown the deployment bundle from 1.74 MB to 52 MB.
• Built the gateway to the .NET API and sent the first screen, Poznámky, across it.
• Rewired the Lifalign web app’s sign-in from Smart Diary to the .NET API; Supabase is still the identity provider, the API is just the one that talks to it now.
• A screen not yet behind the API throws at its first read and names the missing table, so the migration is visible; sign in, sign out, More and Poznámky work.
• Wrote down how the columns of the live and target databases correspond, both ways; they are not two versions of one schema.
• Ran the product locally against plain Postgres.
• Cut the rest of the move into eighteen P0 tasks, Detail dne among them at five screens.

What the factory did

• Served personal context, deleting a note, and the day’s plan timeline over HTTP, the last of those across eleven routes.
• Gave the change trail an implementation, so writing a plan block no longer ends at an empty description.
• Added a check that finds a merged, unapplied migration before a user does.
• Capped the time category tree at three levels.
• Straightened the naming: the epic is Detail dne and the section Na tento den; the fix for twelve glossary citations waits to merge.

What I did

• Vercel reported 100% of its 10 GB Function Storage; the cause is accumulation — every retained deployment keeps its own copy of three Node functions, and since 17 July only 40 of 1072 commits on main touched lifalign/apps/web.
• The web app stopped building for changes it cannot see; the rest of the day added two skipped records instead of dozens of builds.
• Wrote a tool in the factory that deletes old deployments from the tick — it has never run, because profile.json has no vercel block and the trigger returns silently.
• Deleted the old deployments by hand: 312 left, 154 of them READY, nothing older than 1 September. The 12.33 GB has not moved yet.
• Rate limiting got a cheap local check: a cookie’s token is tested for shape and a known signing key before anything goes out to Supabase for it.
• Both .bicepparam files asked for the image tag latest, which nothing publishes; they point at dev.
• Decided the new frontend will hold no access to the database, and started the client that proves it: sign in, list time categories, sign out, one API_BASE_URL.

What the factory did

• Opened the change that offers a ritual as done without a Reality block — a finished plan block is enough; waiting to merge.

What I did

• A project journal lists only the roots: the weekly summary yes, the days inside it no, and a day from a week that has not closed still yes.
• One function holds that rule and the journal index calls it too — it had the same logic written inline.
• The entry chip lost the word “Projekt”; the project’s name alone is left.

What I did

• Worked out what dev and prod should look like for the database, the frontend and the backend, and picked Bicep over Terraform and Pulumi.
• The grilling turned up that the dev API had been serving an image from 28 August for seventeen days, because a setting it needs was never added to the container app.
• Split the work on the environments into seven tasks, by who can actually execute each step.
• Both Azure environments are described by one Bicep template now, and a deploy verifies the revision it created.
• Fixed deploying migrations: the command ran with a context the caller guard had never been attached to.
• Wrote down that the product name has two spellings and which one belongs where.

What the factory did

• Put a row’s two timestamps on one clock and repaired the rows that had them apart.
• Recomputed the length of every Reality block that disagreed with its own range.
• Mirrored the Todo text bounds in the EF mapping and put a guard on every check constraint.
• Reads a day’s captured blocks in the order they happened.

What I did

• A journal entry carries a chip for its project: clickable when that project has a page, dashed and inert when it does not.
• A project links the other way, into the journal pre-filtered to itself.
• The entry header came down to the chip, the date and the title.
• Renamed the product to Lif-align across the whole site.
• The personal website got its own project page, and a link to it from the Now page.
• A finished week now gets a summary, and that week’s days leave the journal index.