Advanced recipes

Use these recipes after the guide library. Each recipe assumes that the basic workspace checks pass.

Registry latest installs the published 0.4.2 line. Install it with the public launchers from Getting started. The README release table owns publication truth.

Run commands inside a workspace unless you use --root.

Scaffold a new workspace

Use Create a Vivary workspace for the complete procedure.

create-vivary init my-workspace --preset writing      # coding | second-brain | knowledge-work | writing
create-vivary doctor my-workspace                      # validate it

Check a legacy or current workspace without changing it

create-vivary doctor my-workspace --json

Doctor is read-only in both human and JSON modes: it exits 0 when the report has no errors and 1 when it has any; warnings do not alter that result. Doctor 0.4.2 uses compatibility.schema_version = 2. Thin workspaces report workspace_contract = "thin-v0.3"; older full workspaces report workspace_contract = "legacy-full" plus legacy_layout = "legacy-v0.1" or "indexed-v0.2+". Legacy layouts remain readable without silent migration. Published releases through v0.3.1 can also lack newer privacy ignore rules. Without declared semantic memory, Doctor reports those gaps as warnings and names each line to add. A published semantic-memory profile keeps heartbeat-reports/* strict and leaves its newer *.vivary-tmp gap as an upgrade warning; a current semantic-memory profile keeps every privacy rule strict.

Use adoption only to preview the new thin contract:

create-vivary adopt my-workspace --preset writing

adopt is a dry run unless --yes --plan <plan_hash> is supplied. It does not migrate or remove legacy modules, templates, or skills. If .vivary/memory.toml says enabled = true, it must name a real provider: provider = "none" is misconfigured. Use --trend only when you intend to write local Doctor runtime state.

Bring an existing repo’s docs under tropo

cd my-repo
tropo init --packs repo-graph          # scaffold a tropo.toml (use a starter pack)
tropo check --lenient                  # see what's there without failing
tropo fix --dry-run                    # preview redundant-frontmatter removal

Iterate tropo.toml until tropo check is clean, then drop --lenient to make it a gate.

Add a typed document

The folder is the type. Drop a file in the right folder; tropo check tells you exactly what metadata is required.

cat > decisions/0002-pick-postgres.md <<'EOF'
---
status: accepted
date: 2026-06-14
related_modules: [billing]
---
# Use Postgres for billing

Rationale...
EOF
tropo check decisions/0002-pick-postgres.md

In a typed document, or an untyped document disallowed by policy, a field that just repeats what Tropo derives (id, title, dates) is noise — tropo fix removes it. Permitted untyped documents retain matching fields because their host format may require that metadata.

Add or change a type

Edit tropo.toml:

[types.runbook]
folder   = "runbooks"
required = { owner = "string" }
optional = { related_modules = "ref-list" }

Nested tropo.toml files tighten rules for a subtree (they can add requirements, never loosen inherited ones).

See what a change would touch (blast radius)

tropo blast billing                 # everything that (transitively) refs "billing"
tropo blast billing --depth 1       # direct dependents only
ozone impact billing                # same, with the review layer's framing

Run this before editing a load-bearing node — it’s the impact a text diff can’t show.

Review the graph before a gate

ozone review                # advisory: unverified changes, broken edges, orphans
ozone review --pack context-budget   # advisory: context bloat and routing surfaces
ozone review --pack all     # run every deterministic review pack
ozone review --strict       # gate: exit 1 if any warning (use in CI / pre-merge)

tropo check validates each document; ozone review checks the relationships between them. Use both before you merge.

Use --pack context-budget before a release or after adding repo-level docs/contracts. It flags missing modules/*/index.md routers, legacy modules/*.md files that coexist with directory indexes, oversized public routing surfaces, exact duplicated routing blocks, and wording that tells agents to bulk-load whole repos or docs trees. It does not read private USER.md, MEMORY.md, memory/**, or heartbeat reports.

Simulate a change

cat > plan.toml <<'EOF'
remove = ["old-module"]
retype = { draft-note = "decision" }
EOF
tropo plan plan.toml        # shows nodes/edges added, removed, newly-broken

Visualize the graph

tropo view --out graph.html             # the whole graph, self-contained HTML
tropo view blast billing --out impact.html   # one blast radius

Open the HTML in any browser — no editor, no server, no plugin. (Obsidian fans: see OBSIDIAN.md.)

Coordinate multiple agents

Opt into the coordination field in tropo.toml:

packs = ["repo-graph", "coordination"]

Claim a work item before editing, then inspect the board and conflict surface:

exo claim local-ci-baseline --agent connie
exo board
exo conflicts
tropo check

exo claim writes only to work items under changes/. It refuses to run unless assignee is declared by the effective tropo config, rejects symlinked or out-of-workspace work item files, and rewrites the workspace file without mutating hard-linked targets outside it. Single-agent workspaces stay free of coordination fields they do not use. exo roles still lists the bounded contracts to hand workers.

Set up LanceDB storage (embedded backend)

Install the embedded extra and migrate your existing workspace:

pip install vivary-tropo[embedded]
tropo migrate --from file --to embedded --root my-workspace --dry-run   # preview
tropo migrate --from file --to embedded --root my-workspace --yes        # run

Or configure storage at init time:

create-vivary init my-workspace --preset coding --storage embedded --yes

Query the knowledge graph

tropo query "CI baseline" --root .                   # text search, top-10 results
tropo query "auth" --root . --k 3 --json             # top-3, machine-readable
tropo find "what should I read for auth?" --root . --budget 1200 --json

tropo query and tropo find search analyzed typed graph nodes directly: id/title, frontmatter, path, body, and outbound edge context. They do not require LanceDB. Embedded storage is a separate opt-in backend for migrated node rows and future local retrieval work.

Use modes in this order:

Need	Command
Read-this-first context packet	tropo find "question" --root . --budget 1200 --json
Deterministic filtered graph search	tropo query "question" --root . --type decision --json
Local fuzzy ranking without providers	tropo query "question" --root . --mode vector --json
Optional provider-backed recall	tropo query "question" --root . --mode semantic --json

vector is still local and dependency-free; it only uses the explicit [storage.embedding] provider = "local-hash" policy and falls back to text when that policy is absent. semantic is different: it requires a configured optional memory provider such as vivary-memory-cognee.

Let an agent prepare a workspace

Let the agent preview each write. Keep provider installation and authority changes at a human gate.

# Preview the five-file seed.
create-vivary init my-workspace --preset coding --no-wizard --dry-run --json

# Discover optional capabilities.
create-vivary capabilities --preset knowledge-work --json

# Create the approved file-backed seed.
create-vivary init my-workspace --preset coding --no-wizard --json

# Preview an optional storage change.
create-vivary wizard my-workspace --storage embedded --dry-run --json

# Preview optional semantic-memory policy.
create-vivary init my-workspace --preset knowledge-work --memory local --dry-run --json
create-vivary init my-notes --preset second-brain --memory cognee --no-wizard --dry-run --json

--json gives machine-readable output. Plain --auto keeps file storage and installs no provider. Size and local privacy hints never select embedded storage. Use --storage embedded only after approval for the provider installation. Use --yes only after that explicit selection.

Semantic memory stays separate from storage. --memory local writes local policy. --memory cognee writes Cognee policy without installing Cognee or indexing content.

If you explicitly install the optional Cognee adapter, dry-run before provider writes:

vivary-cognee doctor --root . --json
vivary-cognee index --root . --dry-run --json
vivary-cognee index --root . --yes --json
vivary-cognee recall "which notes explain the auth decision?" --root . --json

The adapter sends privacy-filtered typed Tropo node packets and ignores recall hits that do not map back to known Vivary node ids. Real provider writes/recalls are blocked until memory.cognee.allow_network = true; if memory.cognee.api_key_env is set, that environment variable must also exist. For a local provider that intentionally needs no API key, set memory.cognee.allow_without_api_key = true.

Use Vivary in CI

CI is just the gate, run on the exit code:

- run: pip install vivary-tropo vivary-ozone
- run: tropo check                 # strict by default — warnings fail
- run: ozone review --strict       # relationship gate
# Optional, once adopted:
- run: ozone review --pack all --strict

Run Vivary as a CI gate

A full copy-paste GitHub Actions job: checkout, install the CLIs, then run doctor and the two graph gates against the exit code.

name: vivary
on: [push, pull_request]

jobs:
  vivary-gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
      - run: pip install vivary-tropo vivary-ozone create-vivary
      - run: create-vivary doctor . --json
      - run: tropo check --root .
      - run: ozone review --strict --root .

create-vivary doctor . --json validates the workspace contract, privacy, startup reachability, graph health, backend/memory status, and transaction recovery; it exits non-zero on any error. tropo check validates each typed document; ozone review --strict checks relationships between them (broken edges, orphans, unverified changes) and fails the build on any warning. Add --trend to the doctor step once you want drift tracking. Thin workspaces write .vivary/runtime/doctor-state.json, which is ignored; cache it explicitly if you want deltas across CI runs rather than a “first recorded run” every time.

Pull logs without a GUI

Vivary visibility is just local receipts plus a tiny reader; no service, account, or dashboard is required.

export VIVARY_RECEIPT_LOG=.vivary/receipts.jsonl
create-vivary doctor . --json
tropo check --root . --json
ozone review --strict --root . --json

vivary logs .vivary/receipts.jsonl
vivary logs .vivary/receipts.jsonl --failed --tail 20 --json
vivary logs email .vivary/receipts.jsonl --to support@example.com --out .vivary/support.eml

The email command creates a local draft or mailto: URL only. It never sends the email for you, and it only includes whitelisted receipt fields.

First run in an agent (bootstrap)

Open the workspace in Claude Code or Codex and say “bootstrap the workspace” — the strato skill interviews you and fills SOUL.md / USER.md / STATE.md. See SKILLS.md.

Publish your own Vivary-based tool (gated)

Publishing (PyPI/npm), creating orgs/repos, pushing, and opening PRs are human gates — one explicit approval per item. The workspace contract enforces this; don’t batch them.