Recipes

Workflow is the medium — a runnable portable graph. Recipe is a workflow packaged as a job: outcome first, contract explicit, guards on by default.

Examples teach one mechanic each (knot, ply, tripwire…). Recipes finish real work against your repository or session.

threadle recipes
threadle run repo-brief --dir ~/code/some-repo

In the app: Workflows → recipes → use (same catalog as examples/recipes/ in the repo). Teaching graphs sit in the examples table below.

	Examples	Recipes
Question	How does this mechanic work?	What do I get when this finishes?
Catalog	`threadle templates` · `examples/workflows/`	`threadle recipes` · `examples/recipes/`
UI	Workflows → examples	Workflows → recipes
Success	You understand the wire	You have a verdict, brief, triage list, …

Every recipe is a workflow (kind: "workflow"). Not every workflow is a recipe.

The contract

Every recipe states what you must supply before ▶ will succeed:

You supply	Meaning
model	Pick a model on each agent node (readiness blocks until you do)
`--dir`	Working directory for agents that read or run tools in a repo
session slot	Drag a session onto the empty ❝ context node and wire it in

Id	Outcome	Contract
`diff-review-panel`	Three-lens review → one verdict	model ×3 · `--dir` · `scope`
`second-opinion`	Cross-provider review of finished work	model · `--dir` · session slot · `concern`
`handover-brief`	Brief in the context library	session slot · `audience` (0 agent nodes — distill is the model call)
`test-triage`	Failures → root causes in one session	model · `--dir` · `failures`
`repo-brief`	One onboarding page	model ×3 · `--dir` · `focus`

Machine-readable catalog: examples/recipes/catalog.json in the repo. API: GET /api/graphs/recipes, POST /api/graphs/recipes/:id.

Diff review panel

Three reviewers, three lenses (correctness, security, performance), one synthesize knot. Agents read the diff with git — point --dir at the repo.

Guards: content tripwire parks under 200 characters, approval before accept, $2 spend ceiling, ply 3.

threadle run diff-review-panel \
  --dir ~/code/my-project \
  --param scope="the staged changes (git diff --staged)" \
  --approve-all

The interesting variant: point each labeled agent at a different provider.

Second opinion

Distill a finished session, frame it as another engineer’s work, hand it to an agent on a different provider — no sunk cost in the reasoning.

Default agent is Claude. If the session under review was already Claude, swap the node to Cursor or opencode.

# Clone in the UI, wire the session into ❝, pick a model, then ▶
# or:
threadle run second-opinion --dir ~/code/my-project

Session slots cannot be filled from the CLI alone — wire them on the canvas (or import, wire, then run).

Handover brief

Four hours of session → a page someone can act on. No agent node — distilling is the model call (cheapest recipe). Output lands in the library for re-inject or another workflow.

Wire the finished session into ❝. Thin-brief tripwire (400 chars) + approval/splice.

threadle run handover-brief \
  --param audience="a teammate picking this up tomorrow morning with no context"

Test triage

Paste failing output; get causes. An iterator splits on blank lines; a framing convert tells the agent to diagnose each failure in one accumulating session so later items can cite shared roots.

threadle run test-triage \
  --dir ~/code/my-project \
  --param failures="$(npm test 2>&1 | tail -60)" \
  --approve-all

Capped at 8 items and $3.

Repo brief

Three labeled explorers (architecture, build & test, risks) under ply 3 → synthesize → one onboarding page. Prefer read-only / ask agents.

threadle run repo-brief \
  --dir ~/code/unfamiliar-project \
  --param focus="the whole repository" \
  --approve-all

Habits every recipe keeps

Spend ceiling · approval before accept · content tripwire when empty output means “found nothing” · note that leads with outcome + contract · ply only for true parallel branches · role labels on parallel agents.

Adapt freely. Keep the habits.

Examples — teaching set
Portable graphs — how these files travel
Running — readiness, params, Issues
CLI — threadle recipes / threadle run <id>
Costs & billing — what a three-agent panel costs