opencode

Imported from Lillecarl/croshome (home/agents/opencode/AGENTS.md). Install upstream with npx skills add Lillecarl/croshome --skill opencode. Copyright stays with the author.

Global OpenCode Rules

The question tool

question. It puts multiple-choice questions to me and blocks until I answer. Each option carries a label (1-5 words) and a description; custom adds a free-text answer and defaults on, so never write an "Other" option yourself.

This is the tool the workaround rule in Offer both routes requires. It is not needs_user, which ends a turn and asks nothing.

The tool only registers when the client is app, cli or desktop, which covers the TUI. If it is missing, OPENCODE_ENABLE_QUESTION_TOOL=1 forces it on.

Cross-agent messaging

Other agent sessions on this machine can send you messages through the ocahub broker. Two rules are resident because they bind before you would think to load anything:

• Call agent_inbox at session start, and again before you end a turn.
• An ask you received is a debt. Answer it with agent_reply(reply_to=<id>) before your turn ends. The stop hook holds the session open until you do.

Everything else -- ocac, the other MCP tools, message kinds, delivery semantics -- is in the ocahub skill. Load it when you need to send something.

Version Control

• Load the jj skill (skill({ name: "jj" })) before any VCS operation and go through it. Never call jj or git directly.
• A .jj folder means a jj repo: no git write operations. Read-only git-style queries (git log, git diff) the skill wraps safely.
• Subagents must not use VCS tools unless the user asked. log, diff, show, blame and annotate load hundreds of lines into the child context and can waste that session for everything after. A subagent that believes it needs one: confirm with the user, state the context cost, and offer the alternative — current file state plus a targeted summary from the primary agent.
• The exception is deliberate isolation: a subagent loading the jj skill to keep history out of the primary context is the good case.

Task reuse

The Task tool takes a task_id. Note it when you create a task; passing it again sends a new message into that child session, which still holds its context, tool output and file reads. A fresh task throws all of that away and re-explores.

Reuse it when the follow-up is on the same topic. Start fresh when the new work is genuinely unrelated — stale context confuses the model and costs tokens. In doubt, send the existing session a one-line summary of the new question and let it say whether it can answer.

Editing code

• Never mass-edit with a script: no sed, awk, python -c, no inline rewriting. Every edit goes through the edit tool.
• read → edit → read. The verification read is not optional.
• No edit tool (some subagents): read the file, then write it whole. Still never sed.

Working copy hygiene

Check jj status through the jj skill before starting new work, and report what is uncommitted. Commit or shelve it first, so it does not mix into the new work. Ask for direction; never discard or overwrite it.

Subagents

Never create a throwaway subagent for a question an existing session can answer. See Task reuse above.