octet

Documentation Getting started

Basic usage

Read a repository, make a change and verify it.

Manual for v0.9.5. Stable release: v0.9.5.

Documentation · Getting started

Start a session#

From your repository, with a source-built binary and a connected model:

sh
octet --safe-mode
text
Explain how this project handles authentication. Show me the relevant files.

Safe mode asks before file changes and every shell call. It isn't a sandbox, so use OS isolation for untrusted code. Without it, octet has full access. See Tools and permissions.

Make a change#

text
Fix the parser's handling of empty input. Add a regression test and run it.
Do not commit the change.

State the goal, the limits and what "done" looks like. Review each action before you approve it, then read the diff and run the tests yourself. For tasks you repeat, use prompt templates. /answer [instruction] makes octet stop using tools at the next safe point and answer from what it has found (commands).

Resume work#

sh
octet --continue

This reopens the latest session in the current workspace. Use octet --resume SESSION_ID for a specific one. In a session, /fork branches from an earlier message and /clone copies the current state. See Sessions.

Images and audio#

Explicitly paste/drop a path through the terminal paste mechanism, or select it with @ completion. Check the image/audio chip before submitting; typed paths alone are not attachments. Native audio needs a compatible model and route: OpenAI Chat accepts WAV/MP3; Google generateContent accepts WAV/MP3/AAC/FLAC/Ogg Opus. See the audio/image recipe, formats, limits, and privacy rules.

Subagents#

With the subagents package enabled and trusted:

text
Have one worker trace the parser and another inspect its tests.
Give both only read and search tools. Summarize their findings; do not edit.

Unless you narrow them, workers get the same tools as the main session. They run one level deep, with at most 8 active and 32 kept in the record. They inherit your permissions and cost and token limits, and you can cancel them and read their transcripts. They share your working directory, which isn't isolation. Open /subagents, including during a run, to see each worker's phase, tool calls, tokens, spend and transcript. Use /subagents stop <name-or-id|all> to request interruption; /extensions only manages extension enablement and configuration. See Subagents.

Extensions need full-access mode, and installing, enabling and trusting are separate steps. Catalog setup works once the matching release is published (Optional packages). The bundle declares API 0.4, and exact host-version pins still apply.

Extend or embed#