Auto mode
Let a System 1 model (TypeSafe Jev) pick the agent, model, and reasoning effort for every launch, guided by live benchmarks and your own automode.md.
Every agent picker in Pragma — new session, new worktree, fanout attempts, board drafts, review fix-its, and design mode — has an Auto row at the top. Choose it and Pragma asks a System 1 model which agent, model, and reasoning effort fit the prompt you wrote. The picker just says Auto: nothing is asked while you type. The pick is made when you submit, from the final prompt, and the launch goes ahead with it in well under a second. To choose yourself, pick an agent and model from the same menu instead.
What a System 1 model is
A System 1 model is a fast classifier rather than a chat model. It does not write text: it
reads a state and answers typed questions with calibrated probabilities. Pragma uses
TypeSafe's Jev by default, and any API that speaks the same
systemone contract works.
The same connection also powers AI merge-conflict resolution in the pull request view.
Setup
Open Settings → AI (global scope), or the card at the top of the AI step in onboarding:
- Base URL or endpoint — leave it empty for TypeSafe (
https://api.typesafe.ai), or paste another Jev-compatible URL. A full endpoint URL such as OpenRouter'shttps://openrouter.ai/api/alpha/decisionsis used exactly as written; a bare origin gets/v1/systemoneappended. - Model — leave it empty to use the endpoint's default:
~typesafe/jev-lateston OpenRouter andjev-latesteverywhere else. Set it to pin a release such astypesafe/jev-1.13. - API key — a TypeSafe key from the TypeSafe console, or your OpenRouter key when using OpenRouter.
- Test connection, then Save.
The key is stored in an owner-only file in Pragma's app data directory, the same way as the GitHub token. It never goes back to the UI, and it is only sent to the base URL you set.
Agent progress in the sidebar
The same System 1 connection estimates how far each running agent is. After every new message an agent writes, Pragma sends the model the agent's original prompt (plus your latest follow-up, when there is one), the last message the agent wrote, and the tools it called most recently, and asks two questions at once: how much of the task is done, and which activity — planning, exploring, coding, testing, debugging, verifying, reviewing, documenting, waiting, or wrapping up — it is in. Each agent's line under its worktree row shows the answer as a verb and a progress bar.
Requests are coalesced: a burst of messages costs one request per agent, and a failed request (a bad key, a rate limit) pauses estimates for a minute rather than retrying on every message. Without a System 1 key the sidebar shows the plain status, with the bar held at 10% until the agent finishes.
How a pick is made
Pragma makes one request per pick. In it, the model sees:
| Input | Source |
|---|---|
| Your prompt | The launch prompt (or an empty-prompt note: Auto still picks your best default). |
| Where it runs | Project, worktree, and branch. |
| Harness benchmarks | Terminal-Bench results per agent × model: accuracy, minutes per task, tokens and cost per solved task. |
| Model benchmarks | The same catalog Pragma's own AI helpers use (modelgrep / Artificial Analysis): intelligence, coding, speed, latency, price. |
| Your preferences | automode.md, project over global (below). |
Alongside "which agent?", the same request asks "which model?" once for every candidate agent, plus how hard the task is. All of these questions are evaluated in parallel, so the model answer for the winning agent costs no second round-trip. Task difficulty then maps onto that model's reasoning levels: a trivial change gets the lowest effort, and very hard work gets the highest.
Benchmark data is cached for a day under ~/.pragma/cache/. If the leaderboards cannot be
reached, Auto still works from your preferences and whatever data is cached.
Terminal-Bench does not cover every harness. Agents without harness results are judged on their
models' benchmarks and on your notes, so automode.md matters most for them.
automode.md
Your own rules, in two places:
~/.pragma/automode.md— global.<project>/.pragma/automode.md— per project. It wins over the global file.
Edit either one in Settings → AI, switching the scope at the top.
---
agents:
include: [claude-code, codex]
exclude: [cursor]
models:
include: ["codex/gpt-6*"]
exclude: ["*haiku*"]
priority: accuracy
---
Use Codex for quick scripted fixes and CI failures.
Use Claude Code with Opus for multi-file refactors and anything touching Rust.Frontmatter holds hard filters. They are enforced in Pragma before the model is asked, so an excluded agent can never be picked:
| Key | Meaning |
|---|---|
agents.include / agents.exclude | Agent ids — the short id (claude-code, opencode, codex) or the full plugin-qualified one. * matches anything. A bare list (agents: [codex]) means include. |
models.include / models.exclude | model for any agent, or agent/model for one agent. An agent-scoped include only narrows that agent. |
priority | accuracy, speed, efficiency, or balanced: the tie-breaker when nothing else decides. |
When the files are merged, excludes from both scopes apply, a project include replaces the
global one, and a project priority overrides the global one.
The body is free text passed to the model as your priorities. The project notes come first and are marked as overriding the global notes. Write it the way you would brief a colleague: say when to use which harness, which models to avoid for which work, and what you care about.
Fanouts
Fanout attempt rows default to manual picks. Attempts exist to differ, and Auto on every row would pick the same thing each time. You can still set any single row to Auto.