Blog on Dippin

What's New in Dippin v0.28

Tue, 19 May 2026 00:00:00 +0000

Issue #39 closes a parser gap. Tracker’s runtime already supported marker_grep, route_required, and output_limit for stdout-driven routing, but .dip source couldn’t reach them — authors following tracker’s TRK101 lint recommendation hit unrecognized tool field "marker_grep" from dippin instead of a cleaner pipeline.

Three new tool-node fields

tool RunTests
 marker_grep: "^(tests_pass|tests_fail)$"
 route_required: true
 output_limit: 65536
 timeout: 2m
 command:
 go test ./... 2>&1

marker_grep — regex matched against captured stdout. The last match populates ctx.tool_marker, so routing edges can reference it instead of parsing raw ctx.tool_stdout.
route_required — when true, the node fails (EventToolRouteMissing) if the command emits no _TRACKER_ROUTE=<value> sentinel line. The matched value populates ctx.tool_route.
output_limit — per-node override of the engine’s captured-stdout byte cap, for tools that need a larger or smaller window than the global default.

Why it matters

Without these fields, authors working around tracker’s TRK101 truncation foot-gun had to enumerate every possible stdout marker as a conditional edge. That made dippin coverage flag false-positive non-exhaustive routing on nodes that were actually fully covered. Typed marker_grep is the canonical fix; now dippin accepts it.

What's New in Dippin v0.27

Mon, 18 May 2026 00:00:00 +0000

The LLM market moves. Since the last verification pass roughly a month ago, two providers shipped new flagships, one cut prices fleet-wide, another retired its previous-generation IDs and started silently redirecting calls, and a fifth doubled the price of three older models. v0.27.0 is the catalog refresh that catches up.

New IDs are now recognized by dippin lint. New prices flow through dippin cost. Stale IDs come out so authors get a DIP108 nudge instead of routing requests at a model that doesn’t quite exist anymore.

What's New in Dippin v0.26

Fri, 15 May 2026 00:00:00 +0000

v0.26.0 is a small release. One new keyword — requires: — in the workflow header. The shape is intentionally simple: a comma-separated list of identifiers the workflow needs in order to run. The motivation is less simple, and it shows up most visibly the first time a workflow crashes 20 minutes in because git wasn’t on the runtime’s PATH.

The problem

Workflows quietly assume things about the environment they execute in. A pipeline that ends with git commit && gh pr create assumes git and gh exist. A workflow that calls a jq filter inside a tool node assumes jq is installed. An agent that uses an MCP filesystem server assumes the runtime has that server configured. None of those assumptions live in the .dip file — they live in the author’s head, until they don’t.

What's New in Dippin v0.25

Mon, 11 May 2026 00:00:00 +0000

v0.24.0 introduced .dipx — a deterministic, hash-verified bundle format for shipping workflows. The shape was right. The implementation was mostly right. But shipping a format means it leaves your machine and starts getting opened, packed, and inspected by other code — Tracker, CI scripts, forensic tools — and that’s where the gaps showed up.

v0.25.0 is the follow-up release that closes them. The format itself (v1.1 in the spec) is unchanged in any way that affects bundles you’ve already packed. What changes is the contract between the library and its callers: cancellation works end-to-end, inspect actually does what its flags claim, error sentinels match the spec, and the exit codes you script against are the ones the spec promises.

What's New in Dippin v0.24

Fri, 08 May 2026 00:00:00 +0000

You’ve spent a week stitching together a workflow. It’s six .dip files: an entry pipeline, three reusable subgraphs, two interview loops your teammate built and you finally got around to wiring in. It runs locally. Now you need to ship it to Tracker.

Today, that means tarring up the directory, scp’ing it, and hoping nobody renames a file in transit. There’s no way to ask “is this the same workflow tree I packaged on Friday?” without diffing source files by hand. And if your CI runner unpacks the tarball into a directory that contains a leftover phases/old_review.dip from yesterday’s build, your pipeline runs against the wrong subgraph and you don’t notice until the bill arrives.

What's New in Dippin v0.21–v0.22

Wed, 22 Apr 2026 00:00:00 +0000

Two releases, five days apart. v0.21.0 shipped budget caps for workflows that get chatty with the API, plus proper timeout support on human nodes. v0.22.0 added a new node kind — manager_loop — for supervising a child sub-pipeline without writing bespoke orchestrator glue.

Here’s what changed.

Human nodes that don’t block forever

v0.21.0 — Human-gate workflows have always had a hole: if nobody ever answers, the pipeline just sits there. You could hack around it by adding a separate timeout branch in your orchestrator, but the human node itself had no concept of a deadline. Now it does:

What's New in Dippin v0.23

Wed, 22 Apr 2026 00:00:00 +0000

v0.23.0 is a small, security-flavored release. Two new defaults: fields give you first-class control over what tool nodes are allowed to shell out — something you previously could only reach by dropping to raw DOT or patching the Workflow.Vars map in Tracker directly.

The problem

Dippin tool nodes execute real shell commands. Tracker’s runtime reads an allowlist and denylist at start time to decide which commands are permitted. Until this release, those lists lived in graph.Attrs — DOT-level attributes. If you wanted to set them from a .dip file, your options were:

What's New in Dippin v0.17–v0.20

Fri, 17 Apr 2026 00:00:00 +0000

Four releases shipped in two weeks. Here’s what changed and why.

Conditional nodes: pure routing without LLM calls

v0.17.0 — The biggest addition. Before this, routing decisions required an agent node with auto_status: true — which meant burning an LLM API call just to check a condition. Now there’s a dedicated node kind:

conditional CheckOutcome
 label: "Route by Result"

Conditional nodes have no prompt, no model, no provider. They evaluate outgoing edge conditions and route accordingly — zero token cost. In DOT export they render as diamond shapes, matching the visual convention for decision points.

Conditional Edges: Routing Pipelines with when

Tue, 31 Mar 2026 00:00:00 +0000

In the last post we showed how Dippin handles prompts. Now let’s make pipelines that think. Linear workflows are fine for simple tasks, but real AI pipelines need to branch: retry on failure, escalate when something looks wrong, skip steps that don’t apply. Dippin’s when keyword is how you express all of that.

1 The Linear Baseline

Let’s start with a document review pipeline. It drafts, reviews, and publishes – in that order, every time, with no branching. Here’s the whole thing:

Cost Estimation: Know Before You Run

Tue, 31 Mar 2026 00:00:00 +0000

You’ve got a branching pipeline with conditional edges. Before you run it against real APIs, let’s find out what it’ll cost. Dippin’s cost command gives you a per-run estimate from the .dip file alone – no API calls required.

1 dippin cost

Let’s use complexity_cleanup.dip from the examples directory – a real pipeline with tool nodes, retry loops, and a mix of models. Run dippin cost on it:

$ dippin cost complexity_cleanup.dip
═══ Cost Estimate ═════════════════════════════════════════════
 Min Expected Max
 ──────────────────────── ──────── ──────── ────────
 TOTAL $0.65 $0.65 $2.66

─── By Provider ─────────────────────────────────────────────
 anthropic $0.65 $0.65 $2.66

─── Top Cost Drivers ──────────────────────────────────────────
 Refactor $1.46 (max) anthropic/claude-opus-4-6
 Triage $0.50 (max) anthropic/claude-opus-4-6
 QualityGate $0.49 (max) anthropic/claude-opus-4-6
 Done $0.21 (max) anthropic/claude-sonnet-4-6
 RemeasureCognitive $0.00 (max) /

The output has three sections worth understanding. The top section shows three columns: min, expected, and max. Min is the cost if every conditional branch takes the cheapest path. Max is the worst case if every node runs every retry. Expected sits in the middle – the most likely path through the pipeline.

Multi-line Prompts Without Escaping

Tue, 31 Mar 2026 00:00:00 +0000

If you’ve been authoring AI pipelines in Graphviz DOT, you know the pain. Every multi-step workflow lives in a .dot file, and every string – prompts, commands, instructions – has to fit on a single quoted line. What that looks like in practice is this, a real tool_command from a sprint management pipeline:

tool_command="set -eu\nmkdir -p .ai .ai/drafts .ai/sprints\nif [ ! -f .ai/ledger.tsv ]; then\n now=$(date -u +%Y-%m-%dT%H:%M:%SZ)\n printf 'sprint_id\\ttitle\\tstatus\\tcreated_at\\tupdated_at\\n001\\tBootstrap sprint\\tplanned\\t%s\\t%s\\n' \"$now\" \"$now\" > .ai/ledger.tsv\nfi\nprintf 'ledger-ready'"

If you’ve authored AI pipelines in Graphviz DOT, this looks familiar. And painful. Every newline is \n, every quote is \", every tab is \\t. Reading this is archaeology. Editing it is anxiety.

CI Integration: Lint, Test, Format

Fri, 27 Mar 2026 00:00:00 +0000

Your Dippin workflows deserve the same CI treatment as your application code. Here’s how to set up GitHub Actions to validate, lint, test, and format-check your .dip files on every push. We’ll build a workflow YAML from scratch, then look at how dippin-lang dogfoods its own CI.

What to Check in CI

A Dippin CI pipeline has four stages. Each catches a different class of problem:

validate

→

lint

→

test

→

fmt --check

Stage	Command	What it catches	Exit code on failure
Validate	`dippin validate`	Structural errors (DIP001-DIP009)	1
Lint	`dippin lint`	Semantic warnings (DIP101-DIP125)	0 (warnings don’t fail)
Test	`dippin test`	Scenario test failures	1
Format	`dippin fmt --check`	Non-canonical formatting	1

Exit code contract

dippin validate and dippin test exit 1 on failure, making them suitable as CI gates. dippin lint exits 0 even with warnings -- warnings are advisory, not blocking. To fail CI on warnings, use dippin check, which combines validate + lint and can be configured to fail on warnings.

Editor Setup: LSP, VS Code, and Tree-sitter

Fri, 27 Mar 2026 00:00:00 +0000

Dippin ships with a built-in Language Server Protocol (LSP) server and a VS Code extension. Together they give you real-time diagnostics, hover documentation, go-to-definition, autocomplete, and syntax highlighting – the same checks you’d get from dippin lint, but live as you type.

The LSP Server

The LSP server is built into the dippin binary. No separate install needed:

$ dippin lsp

Starts the server on stdio (stdin/stdout), speaking JSON-RPC 2.0 per the LSP spec. Any editor with LSP support can connect.¹

Getting Started with Dippin

Fri, 27 Mar 2026 00:00:00 +0000

Dippin is a domain-specific language for authoring AI pipeline workflows. It replaces Graphviz DOT as the format for defining multi-step LLM pipelines, with first-class support for multi-line prompts, conditional routing, parallel execution, and a full validation toolchain. You’ll go from zero to a validated, linted, and visualized workflow in under 10 minutes.

1 Install the Toolchain

Dippin is a single Go binary. Install it with go install:

$ go install github.com/2389-research/dippin-lang/cmd/dippin@latest

Or via Homebrew (macOS and Linux):

Migrating from DOT to Dippin

Fri, 27 Mar 2026 00:00:00 +0000

If your team has existing pipelines authored in Graphviz DOT, Dippin provides an automated migration path. dippin migrate converts DOT files to Dippin syntax, and dippin validate-migration verifies structural parity between the original and the converted file.¹

Why Migrate?

DOT was never designed for AI pipelines. The pain points stack up:

Problem in DOT	How Dippin fixes it
Prompts need `\n` escaping and `\"` quoting	Indentation-based multi-line blocks, no escaping
No validation beyond syntax	34 diagnostic codes: structural errors + semantic warnings
No testing framework	Scenario testing with `.test.json` files
No formatting standard	Idempotent `dippin fmt` with canonical field ordering
No cost estimation	`dippin cost` and `dippin optimize`
Node types inferred from shape attribute	Explicit `agent`, `tool`, `human` keywords

The Migration Process

We’ll migrate consensus_task.dot, a 17-node multi-model consensus workflow with conditional edges and restart loops.

Scenario Testing with .test.json

Fri, 27 Mar 2026 00:00:00 +0000

AI pipelines are non-deterministic – LLM outputs vary between runs. But the structure of your pipeline is deterministic: given a particular outcome at each node, the same path should always be followed. Dippin’s scenario testing lets you inject context values and assert on execution paths, giving you deterministic tests for non-deterministic systems.

The Core Idea

A .test.json file sits next to your .dip file. It holds an array of test scenarios, each declaring what context values to inject and what execution behavior to expect (visited nodes, path ordering, status).