allow
["ls", "-la", "src/"]
Read-only. It goes straight through.
decidespeak is a harness that reads, edits and runs commands in your project, and calls MCP tools. Every action clears a deterministic gate first, and nothing counts as done without evidence.
Every action meets the same consequence check. It is deterministic code, not a model, and it answers allow, confirm or deny before anything runs. Approve mode clears low-risk actions on its own; Strict asks for every one.
["ls", "-la", "src/"]
Read-only. It goes straight through.
["rm", "-rf", "build/"]
Destructive. It waits: allow once, allow from now on, or refuse.
["curl", "https://api.example.com"]
Your deny rule matches. Nothing can loosen it.
Deny wins. Allow, ask and deny rules are created only by you, and no setting, mode or model output can loosen a deny. Commands are argv arrays, never shell strings.
The mark
The logo is the product, drawn. The D is the gate and it does not move. The S is the action, and it only gets anywhere by passing through.
How it works
A confirm card shows the full content. Allow once, allow from now on, or refuse. Change the action and the old confirmation is void.
A file is snapshotted before it is overwritten, and the write clears on its own only if the snapshot succeeded. Edits can be undone.
On macOS, and on Linux with Landlock, writes stay inside the project and the network is denied. Reads are not confined. Where no sandbox exists, every command asks, and an unavailable sandbox is an error, never a silent downgrade.
Give it acceptance criteria and a turn needs deterministic checks plus an independent review bound to the artifact revision. Without criteria, the reply is still fact-checked against the record.
Task subagents get a required file scope, and the gate refuses edits outside it. Read-only ones run up to four at once. No subagent verifies its own work.
MCP servers are managed in the app. Every call runs behind schema validation, the gate, approval, a timeout and output limits.
Modes
Mode is separate from the permission mode. There is no chat-only mode: small talk is just the loop's first reply.
The main loop. Each round it decides whether the situation calls for the decision layer on top.
The same loop, read-only: the gate refuses writes and MCP calls. It ends on a plan card. Run it, keep planning, or throw it away.
The decision layer on every round: candidate directions, a routed choice, a coder subagent and a fresh-context reviewer.
A turn
Rounds, calls, time, failures and lack of progress all stop a task as paused or failed. Never as completed.
Defaults. Each is a setting.
Each step the model returns a batch of actions, a reply, or a question. A message you send mid-turn is picked up at the next step. Long conversations are compacted deterministically, and large results go to disk.
Read-only actions run in parallel; everything else runs one at a time, each through the gate. A command that runs past 30 seconds moves to the background, with live output and a cancel button.
Completion goes through deterministic checks, an independent review and an omitted-evidence check. A reply that claims a file changed when it did not is sent back, then replaced by an answer written from the record.
Only passing counts as done. Anything else goes back into the loop with the reasons attached. The event log is the only record, so a restart or a refresh restores every task.
Decision layer
Laya is an open-source, non-generative model: a state and typed questions in, every answer with a probability out, in one forward pass. Jev, hosted by TypeSafe, answers the same questions over the network.
{
"type": "choice",
"choice": "refactor first",
"probabilities": {
"refactor first": 0.71,
"patch in place": 0.22,
"ask the user": 0.07
},
"confidence": 0.71,
"action": { "act_probability": 0.68 }
}
Below the calibrated threshold (0.55 by default) you get a question, not a guess.
Registering another backend is a line in a table. The chat model is pluggable too: local Ollama, or any OpenAI-compatible server.
Laya reads 512 tokens (1024 multilingual) and truncates from the end. Choice questions degrade past roughly twenty options.
Most agents speak first and explain later. This one decides first, and speaks only once it is through.
Access is limited. Sign in to get the build and the documentation.