How it works
mini-trellis connects project documents, agent skills, and local conversation history. Markdown files hold the knowledge; the CLI finds relevant passages; your coding agent uses and updates them during development.
Knowledge lives in the project
| Content | Location | Example |
|---|---|---|
| Conventions | .trellis/spec/ |
A callback requires an idempotency key. |
| Research | .trellis/research/ |
Why a retry policy was chosen, with sources and alternatives. |
| Journals | .trellis/workspace/<name>/ |
What changed, what was checked, and links to the topic. |
| Existing documentation | Root README files, docs/, doc/ |
Current architecture and user guides. |
| Conversation evidence | The agent’s local session store | The original discussion behind a decision. |
Existing project documents remain in place. A journal links to the current topic instead of copying the whole explanation. See writing knowledge.
The startup entry points to knowledge
Supported integrations give the agent a short orientation: journal location, spec indexes, and research entry points. It can then retrieve the documents relevant to the work.
Injection timing depends on the agent. Claude Code uses SessionStart; OpenCode V2 rebuilds the short entry for each outgoing model request; DeepSeek Harness uses project instructions and skills. The platform matrix describes the differences. Startup size is therefore a measured property of a particular integration and version, rather than a universal constant.
Recall first, expand when useful
mini-trellis mem recall "callback retries"
mini-trellis mem read "<ref-from-recall>"
recall searches project Markdown and supported local transcripts. It returns short excerpts with source references. read expands a reference. The search is local and lexical; it uses words and identifiers in the material. No embedding service is involved.
Update the knowledge as the project changes
When a rule changes, update the existing document. Research can retain the previous reasoning and rejected approaches. For a whole document replacement, a superseded_by link records the relationship and helps recall identify the current note.
The agent’s remember, update-spec, and session-insight skills guide these edits. Their roles are explained in Commands and skills. Planning, coding, and review use your existing development process.
Local storage and sharing
The CLI reads local files. Knowledge can be reviewed and shared through Git, while conversation history remains in each agent’s local store. A coding agent may send the excerpts it reads to its configured model provider. See configuration and sharing for files to include or ignore.