TL;DR: AdaL 1.8.4 adds write-like-human, a capability that rewrites anything meant for a person to read so it stops sounding like a model wrote it. BYOAK now takes a base URL alongside the key, so AdaL can run against your own gateway. And /usage shows your weekly usage and when the window resets.
Everybody can tell
Your agent writes a lot of text that other people read. PR descriptions, commit messages, release notes, docs, outreach, posts. Most of it arrives with the same fingerprints: an em dash in every other sentence, three parallel examples where one would do, a closing line that repeats the paragraph above it, and the construction that gives the whole thing away, "it's not just X, it's Y".
Readers notice. A PR description that reads like a model wrote it gets skimmed instead of read, and the reviewer trusts the summary less than they would trust two plain sentences from you.
The capability
write-like-human is a capability, so it follows the 1.8.0 rule: it is not a mode you switch into. Ask AdaL to write like a human, or hand it a draft to fix, and the capability loads itself mid-session.
What loads is not a tool. It is an editing procedure with a catalogue of patterns, ordered by how strongly each one gives the text away:
- Staging instead of stating. A sentence that signals importance rather than adding a fact. The not-X-but-Y contrast, and the one-line closer that restates the paragraph.
- Rhythm by rule. Triads and dashes applied everywhere, whether or not the meaning asks for them.
- Inflation. Ordinary facts dressed up as pivotal or expert-backed.
- Formatting by rule. Bold and title case on every item in a list.
- Leftovers. Chat wrappers and drafting scaffolding that were never meant for the reader.
The strongest patterns justify an edit on a single sighting. Weaker ones need company in the same passage before the agent touches them, because a careful writer does occasionally make that choice on purpose.
Two rules hold the whole thing together. Every sentence that survives has to give the reader something they did not already have. And the rewrite may not add a fact, name, number, date or citation that was not in the source. Shape edits are where claims usually go missing, so the procedure checks the rewrite against the original for dropped claims before it returns anything.
It also knows what it is editing. Give it a writing sample and it matches your sentence length, punctuation and openings, sample over rules, so if you use dashes it keeps them. Without a sample it takes the register from the kind of text: essays keep opinion, uncertainty and asides, reference and legal text stays plain, and social posts drop to casual lowercase.
Two things worth knowing. It applies to text meant for other people, not to AdaL's replies to you in the CLI. And it survives compaction, so a long session that started with a rewrite still has it at the end.
Point AdaL at your own endpoint
BYOAK used to take a provider and a key. Now it takes a URL as well:
/byoak add anthropic sk-ant-your-key https://gateway.example.com/anthropic
/byoak add qwen your-key http://localhost:8000/v1
Requests for that provider go straight to your URL with your key. They do not pass through AdaL, and they take effect on the current model with no restart. The /byoak dialog shows the URL next to the key, so you can always see where your requests are going.
This is the piece that was missing for teams with a company LLM gateway, and for anyone running a model on their own hardware. Your endpoint has to serve the model id AdaL sends, which pass-through gateways and proxies do. A server that hosts its own models under different names is not covered yet.
Full setup is in the BYOAK docs.
Know where you stand
/usage now answers the question you actually have: how much of this week is left. It shows your plan, your weekly usage, when the window resets, and your credit balance, in the CLI and the desktop app.
Warnings changed too. They used to fire once and never again, which meant the one that mattered arrived while you were mid-task and got dismissed. Now they arrive at 70%, 80% and 90%, so the first one lands early enough to be useful.
Also in 1.8.4
- A clearer student discount. Fewer steps for eligible students to claim it.
- Clickable pull requests in the footer. The PR number in the footer is a link now.
- Steadier tool calls. Improved tool-call stability across several models, and ChatGPT Plus and Pro models can run tools again.
- Older x64 machines. AdaL launches on processors that previously refused to start it.
Update AdaL and try the first one on something you were about to send anyway. Ask it to write like a human, then read the before and after next to each other.
Full notes are in the changelog, and how capabilities load is in the docs.