Agents in SoloMD

1. Inline Agent Panel

A right-side panel, peer to Outline / Backlinks. Streamed chat-with-vault, routed through the in-process MCP tools and the existing 14-provider AI stack. Citations resolve to [[wikilink]] chips that open the cited note in a new tab. Tool-call cards (Cursor-style) show every read and write inline — expand a card to see the full args and result.

Run history persists as plain markdown in <workspace>/.solomd/agent-runs/<ts>.md. Grep-able. Git-trackable. The opposite of a black box.

• Open with ⌘⇧A or the right-rail toggle
• Active note toggle injects the current document into the system prompt — opt-in, off by default
• Allow write toggle gates write_note and append_to_note per chat session
• 10 in-process agent tools — same surface as the bundled MCP server

2. Scheduled Recipes

A recipe is one YAML file under <workspace>/.solomd/agents/*.yml. No build step, no plugin manifest. Edit, save, run. Five trigger modes:

Example recipe:

name: Weekly review
trigger: schedule
schedule: "0 18 * * SUN"        # cron, UTC
match: "daily/**/*.md"
prompt: |
  Read this week's daily/ notes.
  Write weekly/{{date:YYYY-WW}}.md: themes / decisions / open threads.
allow-write: true
write-cap: 5
provider: claude

Triggers compose with the safety model below — every run begins on its own AutoGit branch agent/<recipe>/<run-id>, and writes are reviewable, not magical.

3. Trace view + replay

Every run (Panel chat or Recipe) emits <workspace>/.solomd/agent-runs/<run-id>/trace.jsonl — line-delimited so you can tail -f a long run while it's still going. Schema: one JSON object per step, with kinds run_started · prompt · model_call · tool_call · tool_result · git_commit · run_ended. Fields include timestamps, tool args, truncated results (≤ 2048 chars), token counts, cost estimate.

The UI renders these as collapsible step cards:

• Inline in the Panel after the chat
• Settings → Recipes → History (one entry per run)
• Step-detail drawer for the full result body
• Replay-from-step N mints a new run that re-uses every line up to step N and then re-runs the model. Useful when the model took a bad turn at step 7 and you want to retry the decision with a different prompt.
• New read_agent_trace(run_id) MCP tool — agents can now read their own traces, opening the door to self-correcting recipes.

4. Workspace Federation

The bundled solomd-mcp grew a multi-workspace mode in v4. Pass any number of --workspace flags to bridge multiple vaults in a single MCP session:

solomd-mcp \
  --workspace ~/Vaults/personal \
  --workspace ~/Vaults/work \
  --workspace ~/Vaults/research

Tool signatures gain an optional workspace param. Default = first-passed (back-compat). AutoGit branches stay isolated per workspace. Settings → Integrations adds named MCP profiles — workspace bundles you can one-click copy a Claude Desktop config block out of.

Today users open multiple Claude Code sessions to bridge two vaults. v4 makes one session enough.

5. Ollama first-class

We don't bundle a local LLM runtime. Ollama is already in our 14-provider list and does this well — re-implementing it would violate principle #7 (write less code we maintain forever). Instead, v4 polishes the integration:

• Auto-detect Ollama at localhost:11434 with a green status indicator in Settings → AI.
• Install Ollama button if it's not detected (links to ollama.com — we don't proxy the download).
• /api/tags lists installed models inline; empty list → Pull recommended (qwen2.5:1.5b, ~1 GB) button.
• 3 built-in presets: rewrite (qwen2.5:7b) · quick (qwen2.5:1.5b) · CJK-friendly (qwen2.5:14b).
• Recipes can specify provider: local as a default → cheap autonomous loops on-device.

11 starter recipes

Open Settings → Recipes → Browse cookbook to install any of these. The YAML copies to .solomd/agents/ and opens for editing. Re-installable with auto-suffixed names.

• 01-weekly-review.yml — Sunday 18:00 UTC, summarizes daily/*.md into weekly/YYYY-WW.md
• 02-todo-extract.yml — on-save scans for - [ ] items, accumulates into todo.md
• 03-translate-zh-to-en.yml — manual; round-trip translate the active note
• 04-cjk-proofread.yml — on-save Chinese / Japanese / Korean grammar pass via Ollama (provider: local)
• 05-citation-cleanup.yml — on-commit normalizes @citekey formatting
• 06-meeting-notes-summary.yml — on-tag-add #meeting extracts decisions + action items
• 07-link-suggester.yml — read-only, suggests [[wikilinks]] in the trace view (no writes)
• 08-daily-summary.yml — every night 22:00, write a TL;DR into today's daily note
• 09-orphan-notes.yml — weekly scan for notes with zero backlinks, list them in orphans.md
• 10-on-commit-changelog.yml — on-commit appends a one-line summary to CHANGELOG.md
• 11-tag-classifier.yml — on-save suggests #tags based on note content (read-only)

Safety model — the AutoGit sandbox

Every recipe run begins by:

1. Creating a fresh AutoGit branch agent/<recipe-slug>/<run-id> off the current HEAD.
2. Running the prompt + tool loop. All write_note / append_to_note calls land on this branch.
3. On success the run goes into Pending review with three buttons: View diff, Accept (fast-forward into main), Reject (delete the branch entirely).

A rejected run vanishes — no merge, no leftover commits in your history. An accepted run shows up in git log like any other commit.

The write-cap is enforced before dispatch, so a model that asks to write 100 files when the cap is 5 will get refusals starting at the 6th call, with no half-applied state. Default 5; hard upper bound 50.

Write your own

The full schema, every variable ({{path}}, {{date:YYYY-MM-DD}}, {{files}}, ...), every tool the agent can call, and the patterns that actually work in production live in docs/agents.md in the repo:

Read docs/agents.md →

Schema is the source of truth in app/src-tauri/src/recipes.rs — Rust types, not a paper spec.