mini-SWE-agent

A bash-first coding CLI with a human-started session and confirm mode by default. Confirmation isn't a sandbox: the normal local environment runs commands with your host-user access.

No product visual on file

Choose it when

You want a local coding session with confirm, yolo and human modes, plus a saved trajectory you can inspect.

Skip it when

You need default OS-level isolation or an always-on service with a named failure owner. This dossier covers the interactive mini CLI, not separately orchestrated batch work.

01 / Classification

Autonomy evidence

This is our classification, not a vendor badge. The record says what the product normally does. The gaps stay visible.

How we classify it

Tool use + multi-step plans, still human-gated per session.

Runs unattended

No — the current record treats it as session-gated.

Trigger

The operator starts mini and supplies a task. Confirm is the default; optional yolo or externally orchestrated headless runs don't establish a native scheduler or failure-owner service.

State

The message trajectory is linear. The last run is saved in the global config directory as last_mini_run.traj.json by default, with configurable output. This is an inspectable run record, not a guarantee of resumable memory.

Tools

Bash is the action interface. The default local environment executes separate host subprocesses; cwd selects a starting directory, not a filesystem, network or credential boundary.

Human control

Enter approves and rejection text rejects in confirm mode; /c, /y and /u switch modes, and Ctrl+C interrupts. Regex whitelists bypass prompts and one prompt can cover several commands. The human owner must review changes, grants and failures; no rollback guarantee is established.

02 / Operations

Control reality

Useful agents eventually touch something real. This is the operational picture we have, not a promise hidden in a sales page.

Where it runs

Normal mini runs locally. Container and other sandbox environments are optional alternatives, not default isolation for this CLI.

Data boundary

Local execution isn't local-only model processing. The message trajectory and tool observations feed the selected model. Provider retention and training treatment are unknown from these sources.

Identity & permissions

The operator configures model access and API credentials. Local commands inherit the process environment plus configured overrides and run with host-user reach; audit available credentials and network access before use. Model/API usage is separate from the MIT software license.

What it can write

Files reachable by the host user through shell commands · Local run trajectory JSON

Surfaces

Interactive mini terminal CLI

Export / lock-in

Source, filesystem changes and local trajectory JSON are inspectable artifacts. The default last-run file and configurable output aren't an account-portability or durable-memory guarantee.

03 / Receipts

Evidence ledger

Here is what this brief is built on. “Documented” means we found a source. “Tested” means this desk actually ran it. Those are different things.

Evidence status

documented

Source set

8 primary sources

Evidence review

2026-10-19

Sources on file

Corrections

Not recorded yet. If we change a material claim, the correction belongs here.

04 / Change log

Related coverage

The brief tells you what the product is. Payloads track the changes that may alter the decision.

No linked change coverage on file yet.