Four releases shipped in two weeks. Here’s what changed and why.
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.
The scaffold template (dippin new conditional) now uses the real conditional kind instead of the agent workaround.
v0.20.0 — Parameterized workflows were always possible, but the variables lived outside the .dip file — you’d patch them into the DOT export manually. Now there’s a vars block:
workflow SemportRust
goal: "Port $source_ref into $target_name"
start: Start
exit: Exit
vars
source_ref: "references/claude-agent-sdk-python/src"
target_name: claude-agents-rs
target_module: "claude-agents-rs/src/"
Vars export as DOT graph-level attributes (what Tracker reads via graph.Attrs), round-trip through parse/format/export/migrate, and the parser catches duplicate keys. DOT migration now captures unknown graph attributes into Workflow.Vars instead of silently dropping them.
v0.20.0 — DIP125 (“tool command binary not found on PATH”) used to parse shell commands with regex and first-token extraction. This produced false positives on common patterns:
hint[DIP125]: tool command binary "COUNTER='.ai/count.txt'" not found on PATH
hint[DIP125]: tool command binary "count=$(cat" not found on PATH
Now the linter uses mvdan.cc/sh/v3 (the parser behind shfmt) to walk the actual shell AST. Variable assignments, command substitutions, command -v checks, and preamble commands are all handled correctly. The exact script from the bug report now produces zero false positives.
v0.19.0 / v0.19.1 — Two new fields on agent nodes:
agent Worker
backend: claude-code
working_dir: .ai/worktrees/feature-branch
prompt: Implement the feature.
backend selects the execution backend (native, claude-code, acp). working_dir sets an isolated working directory — essential for git-worktree-based parallel execution. Both were previously silently dropped by the parser, causing hours of debugging for users.
As part of this fix, unrecognized fields on any node type now emit a parse diagnostic instead of being silently discarded. The error message suggests params: for agent/subgraph nodes that support it.
v0.17.0 — The --extra-models flag on lint and doctor lets you extend the DIP108 model catalog at runtime:
dippin lint --extra-models "custom-corp:my-model-v1,my-model-v2" pipeline.dip
No more false DIP108 warnings for private or newly-released models.
v0.20.0 — A new lint rule catches a common confusion. When you set max_retries in defaults and your workflow has restart: true edges but no max_restarts, the linter now warns you. max_retries controls per-node LLM retries; max_restarts controls the global loop restart budget. Setting the wrong one silently does nothing useful.
The tree-sitter grammar now works with tree-sitter 0.24+, and there’s a new Zed editor extension at editors/zed-dippin/ with syntax highlighting and LSP integration. Install via Zed’s “Install Dev Extension” command.
go install github.com/2389-research/dippin-lang/cmd/[email protected]
Or via Homebrew if you’re on the tap.