Synod

Synod

Repository: https://github.com/Slaymish/Synod

A council of LLM "value agents" reads your journal, surfaces tensions you haven't resolved, and never suggests compromises. Runs entirely inside Obsidian; LLMs are reached via local Ollama, OpenRouter, or any OpenAI-compatible endpoint (e.g. llama-swap).

What it does

╭─────────────────────╮     ╭──────────────────╮     ╭────────────────╮
│ Import journal      │────▶│ Per-value agent  │────▶│ Compiler       │
│ (Rosebud / vault    │     │ (parallel, fully │     │ (3 passes)     │
│ folder / journals)  │     │  isolated)       │     │                │
╰─────────────────────╯     ╰──────────────────╯     ╰────────┬───────╯
                                                              │
                                                              ▼
                                              ╭──────────────────────╮
                                              │ Bulletin (.md in     │
                                              │ vault) + side panel  │
                                              │ status               │
                                              ╰──────────────────────╯

Three importers

Pick from the Import journal entries command (or the Import entries button in the side panel). Each importer remembers its own folder / date-format settings between runs, and the modal opens on whichever importer you used last:

The Rosebud picker also accepts drag-and-drop. Imports are deduped on SHA-256 of the normalised user text, so re-running an import never creates phantom duplicates.

Editable system prompts

On first load the plugin writes default prompt files into <vault>/Synod/_prompts/ (path configurable):

• value-agent.md — per-value agent system prompt
• compiler-finder.md — Pass 1 tension finder
• compiler-validator.md — Pass 2 tension validator
• compiler-report-validator.md — Pass 0 evidentiary gate
• schwartz-extractor.md — value-discovery prompt

Edit them like any markdown note. The next pipeline run reads the changes. Use {value_name} / {value_definition} placeholder syntax.

Settings

Every knob is in Settings → Community plugins → Synod:

• Provider: Ollama / OpenRouter / llama-swap, with separate models for the agent role and the compiler role, plus a Test connection button.
• Output: vault root folder, mirror entries / values to vault toggles.
• Prompts: folder location, "re-create defaults" button.
• Schedule: bulletin interval (hours), run-on-startup, minimum entries.
• Prompt budgets: char-per-call ceilings for extractor and value agents.
• Tensions: severity threshold, escalation cooldown.
• Active values: list with Active toggle and Delete.

Status & progress

Open the side panel via the ribbon icon (or Synod: Open status view). The view shows the live phase (ingesting → extracting → agents → compiling → writing → done), a progress bar, last-run timestamp, counts, and a tail of the log.

Privacy & network

• Default backend is Ollama on localhost. Nothing leaves your machine.
• OpenRouter / llama-swap modes only send data to the endpoint you configure in settings. The plugin makes no other network calls.
• All journal text, value definitions, and reports are stored locally:
  • structured records in the plugin's data.json,
  • human-readable bulletins / value notes inside your vault.
• No telemetry. No background uploads. No bundled remote code.

Build from source

npm install
npm run build      # bundles to main.js
npm test           # pure-function smoke tests (parsers, hashing, packing)

To install into a vault during development:

ln -s "$(pwd)" "/path/to/Vault/.obsidian/plugins/synod"

Then in Obsidian: Settings → Community plugins → enable Synod.

Architecture notes

• Storage: JSON via Plugin.loadData()/saveData(). No SQLite, no vector DB. Entries fit in context for typical journal sizes; if/when they don't, plug a vector store into agents/value-agent.ts.
• Isolation contract: runValueAgent() takes only one value's (id, name, definition) plus the shared corpus. The Compiler is the only module that sees all reports; its output is never fed back into the agents.
• LLM single-flight: llm/index.ts funnels generation through one Promise so single-GPU backends don't trample each other when value agents fan out in parallel. Hosted backends can lift this gate.
• Prompt budgeting: char-based proxy for tokens. Long entries are split into labelled [date — part k/N] blocks across LLM calls — never silently truncated.

License

MIT — see LICENSE.