A Brain, Some Hands, and a Brake Pedal
· 3 min read
yoke llm ai elixir erlang architecture programming opinion hacking
The assistant had just announced that everything was fixed. The tests disagreed. Three files lay open on the desk like contradictory witness statements, and the terminal had the exhausted glow of a bar that should have closed an hour ago.
“Perhaps,” I suggested, “we could try engineering.”
I have already argued that the model is an engine, not a car. Yoke is my attempt to supply the steering, the instruments, and the brake pedal. Written in Elixir on Erlang/OTP, it treats an agent session as a running system, not a disposable chat window with permission to vandalise a repository.
Keep the Brain Away From the Power Tools
Yoke calls the split Brain/Hands . The Brain owns conversation and agent state in an isolated GenServer. The Hands execute tools separately. OTP supervision contains a failed tool process instead of letting it take the conversation down with it.
flowchart LR
U["You"] <--> B["Brain<br/>conversation and decisions"]
B <--> H["Hands<br/>supervised tool execution"]
H --> L["Local workspace"]
H --> R["Remote Erlang node"]
H --> D["Docker container"]
Where the model runs and where the tools run are separate decisions. Local inference does not mean remote execution; remote execution does not require outsourcing the Brain.
Tool batches can run concurrently, with per-file write locks. Larger workflows give parallel subtasks their own Git worktrees and branches. Two assistants should not share an unguarded working tree any more than two surgeons should share a blindfold.
Give It a Map, Not Another Shovel
Dumping a repository into a context window is not understanding. It is fly-tipping with an API bill.
Metastaticsupplies a shared MetaAST representation through language adapters. Ragex builds on that foundation alongside native analysers: symbols, call relationships, semantic and hybrid search, impact analysis, and structural refactoring. Yoke brings those tools into the conversation through MCP.
flowchart LR
S["Source code"] --> M["Metastatic<br/>shared MetaAST"]
S --> N["Native analysers"]
M --> R["Ragex<br/>graph, search, refactoring"]
N --> R
R --> Y["Yoke<br/>task-relevant context and tools"]
Now “who calls this?” has something better behind it than a confident guess. A rename can follow references rather than impersonate a search-and-replace accident. Support varies by language and operation; a diagram is not a universal-correctness certificate.
The point is not to make the model read everything faster. It is to stop making it read everything.
Regret Should Have an Address
There are two different kinds of recovery, and confusing them is how a demo becomes an incident report.
Conversation snapshots
preserve messages and provider settings. Save one with
/checkpoint before-refactor
;
/undo
restores conversation state. It does
not
rewind your working tree, reverse a shell command, or unsend an email.
Code transactions belong to Ragex’s editing tools: pre-validate a multi-file change, keep backups, apply coordinated edits, and roll back on failure. Use that path when a change must travel as a group. It is not a magic transaction around arbitrary shell side effects, and syntax validation still does not replace tests.
Meanwhile, sessions persist as readable LMML narratives. Resume with
yoke -c
, compact an overgrown conversation with
/compact
, and inspect
/tokens
or
/cost
instead of treating the invoice as a surprise ending.
I want an assistant that can recover from a bad idea. Mine included.
A Conversation Is an Interface
The terminal need not be a punishment for refusing an IDE subscription.
-
Type
@for an in-place, gitignore-aware file finder ; attach a file or URL without conducting a copy-and-paste séance. -
Open
/explorerto inspect settings, rules, conversations, practices, jobs, skills, diagnostics, and analytics. Actual controls, not a treasure hunt through dotfiles. - Compose multiline prompts, search input history, use an external editor, and answer structured questions when the agent needs a decision rather than a licence to improvise.
-
Drop into
!!for plain shell work, then return to the conversation. Not every command needs a language model supervising it like an expensive crossing guard.
Make the Discipline Portable
Scoped rules and learned language practices keep conventions out of the daily nagging ritual. Skills, MCP servers, and hot-reloadable plugins extend the toolbox. Workflows can encode branching, tests, documentation, linting, and commits; review and audit tools put evidence back into the loop.
Provider profiles switch model and endpoint together. Built-in local profiles stay on loopback and do not silently fall back to a remote service. “Local” should describe where the request goes, not the vendor’s feelings about privacy.
None of this makes autocomplete infallible. It makes its mistakes less lonely, less expensive, and easier to inspect.
That is the harness I want: a Brain that keeps the thread, Hands that work within boundaries, and a brake pedal that is not merely painted on.
Happy hacking.