Message Chain

💡 apparently the new theme is Locations in San Francisco

additional features needed:

• re-enable the angle-bracket quoting. ⚙️( this was accidentally ⚔️( inadvertently) removed during one of the re-writings.)
• generate annotations for 汉子 inline. the initial ones will just be the 拼音, which can be generated by dictionary lookup.
• generate annotations for <<

⚙️ Cabrillo Street is one block north of Golden Gate Park.

There are quite a few features to consider:

• Message editing / corrections. I say 汉子 when I mean 汉字. How does it get fixed?
• as teased earlier, switch to a YACC style parsing of the text between each "----" line.
• change the "mode" options to Light, Dark, and Print
• bold/italic formatting. probably italic and bold.
• revisit the "email" aspect. Probably rename the database table to be AtacamaMessage rather than Email. But, also, fix and start running the "read an IMAP server every few minutes" service.
• adjust the "hover-text" so it works better on mobile.
• another round of CSS tuning.
• the "two-window" editor; plaintext on the left, (partially) rendered text on the right. ⚙️( Some projects will do an AJAX request to render it on the server; others will use a separate JavaScript parser to render it client-side. I will do whichever one is easier for the machine to write.)

Separately, there are the various LLM-related features I have been considering.

• enforcement of Guild Law ⚙️( no discussion of politics or religion without an asserted exception)
• auto-teal 💡( i haven't determined the syntax; my first instinct is to use the @ character.)

💡 also I need a README section for the website. because there has to be somewhere to explain tone sandhi.

⚔️ well, actually, the wikilinks work for now. And but there should be a sigil.

and: authentication. 🔥( that isn't "do you know the secret password"). which is the gateway to multi-player features ⚙️( such as comments).

One of the skills involved in pair-programming with the machine is to know when to do a major compaction.

It is, roughly, the milestones of a waterfall project-management approach. 🤖( the Waterfall methodology follows a linear, sequential approach to project management. It is structured around predefined phases, such as requirements gathering, design, development, testing, deployment, and maintenance. Each phase must be completed and approved before moving to the next, which makes Waterfall highly predictable and easier to plan. However, this rigidity can pose challenges in adapting to changes or incorporating feedback once the project has moved past certain phases. Waterfall is best suited for projects with clearly defined objectives, stable requirements, and minimal expected changes. ) 💡( It is not exactly that; Waterfall in the literature is something of a straw-man. What it is, is a series of changes based off an old version of the code (version 3). Once the cumulative changes become too much to describe, one must make a new version of the code (version 4) and start over.)

Some of the features in the last iteration:

• Moving web handlers to "blueprint" files rather than server.py. 💡( I find myself wanting to re-invent the header file. I only want to have the machine read 3-5KLOC, and using header files can decrease the LOC by a large factor.)
• CSS/JS fixes. Color-text is now default invisible, the sigil appears in a box. 🔥( there is code trying to rotate the sigil when it is clicked on. It doesn't work. I don't know if I want to enable it.)
• Login fixes. Users can login now, as long as they are the one user on the Google Auth whitelist.

The new "lexer" was written by Claude. But it fails many tests, so it is not enabled yet. 💡( this is a feature that will benefit from the extra screen-space of AppleVision.)

We also need:

• Print mode (and dark mode).
• various "two-column" views
• other features, as needed.

Yesterday was a rest day.

Today:

• use all the tools on-hand to get the machine to write the lexer/parser correctly.
• add some extra syntax (triple-angle-brackets are different from double-angle-brackets).
• more auth tools

and then, time-boxing into "LLM auto-generation only".

neither Claude nor ChatGPT can write the lexer correctly.

i think it is close enough that it can be fixed without having to explain what to do.

although, "write the parsing rules in plain English" might be a useful thing to have it do now while I can correct it.

I have time-boxed out of the "new lexer/parser". Neither ChatGPT nor Claude can write it correctly.

💡 they seem to "understand" what is going on. but the output code doesn't solve the main problems.

initial features for the afternoon:

• User tracking
• React? 💡( React is a front-end programming library. It makes some thing easier, but comes with a base complexity cost. However, as all the LLMs already know react, it doesn't hit the cognitive-load limits that quickly)
• A few more landing pages. "Chain view", for one.

Ideally, by the end of the day, Atacama will be at a point that I don't need to touch it for a week.

The machine keeps getting "session" (for database connections) and "session" (for HTTP request information) confused.

If there is another round of "rewrite this file", I will have to tell it to use "db_session" for all database connections.

The machine continues to be less-and-less accurate.

I ask it to write a function, and my acceptance criteria is Is this good enough that a second pass might fix it correctly.

Both the "lexer" and "migrate to React" are too complicated for Claude today.

🔥 at least neither is necessary today.

💡 But, "nested parentheses" will still not work. Triple-angle-bracket quotes will not work. And, other than "request LLM content", all new syntax is prohibited. I can adapt around those problems for two weeks.

⚔️ well, actually, "make Atacama better" might not be an active concern of mine in 2 weeks. or, maybe the machine will work then.

https://earlyversion.com/messages/10/chain exists. It has ⚔️( had) an ugly header that hovers, and it doesn't have inbound links. But it works enough for today.

💡 Also, the main landing page is now somewhat less filled with useless information.

The only feature remaining before I can pivot to other tasks is LLM response generation. With a sigil (that I cannot say here).

⚔️ well, actually, there will need to be a solution to "clicking on every sigil is a lot". but, that is deliberately delayed until Thursday.