Changelog

Source: CHANGELOG.md · repository commit df39fcf

Unreleased

Packaging and docs. npm metadata now points at the hosted documentation site, the README leads with the one thing you have to install (the CLI) and presents the bundled agent skill as the optional extra it is, and CHANGELOG.md ships inside the published package. Profile examples no longer suggest codex -m: a profile cmd must read the prompt on stdin, which that invocation was never verified to do (#71).
New: a bundled agent skill. chainq ships an Agent Skill at skills/chainq/ that teaches a coding agent how to author a flow. It exists to fix one specific failure: an agent asked for a flow would emit YAML whose real work sat in cmd steps — a task runner with chainq syntax, not a prompt chain. The skill’s rule is that the model’s work belongs in visible ai nodes while cmd stays a narrow deterministic boundary (a command may sit mid-chain; it may never carry reasoning or call a model for an ai node), and it guards the opposite failure too: a full chain is not a maximal one, and a stage earns a node only when its output is worth inspecting, reusing, caching, or tuning alone. Progressively disclosed — SKILL.md is the only file loaded up front, with references/ (authoring · patterns · flow syntax · CLI · troubleshooting) and three runnable templates/ opened on demand. Installed with the open skills CLI, which reads the same directory straight from the repository: npx skills add wahengchang/chainq --skill chainq (add -g for every project), covering Claude Code, Codex, Cursor, opencode and more from one command. .claude-plugin/ additionally publishes it as a Claude Code plugin. src/skills.test.ts keeps it honest in CI: every shipped template is parsed and validated by the real engine and must obey the skill’s own rule, every internal link must resolve, and every node type, command, and flag the skill documents must still exist in the product — so the skill fails the build instead of silently teaching agents something that no longer works. See docs/guides/agent-skill.md.

0.2.1

Docs: simpler Quickstart. The README now leads with a global install (npm i -g @wahengchang2023/chainq) so every command is just chainq …, and the happy path is three lines — install → chainq init → chainq ui — followed by a one-line chainq run to land the output. Fixed a stale “8 node types” claim (it’s 5: ai · cmd · assemble · input · write) and added the editor screenshot the README references. Docs-only — no engine change.

0.2.0

Breaking: removed the collection operators splitOut, aggregate, and merge. Array fan-out/fan-in inside a flow added more mental overhead than it earned, so the node model is now just ai · cmd · assemble · input · write. Combining two upstream streams (“fan-in”) is expressed with a multi-input assemble/ai node (from: [a, b]) that references both in its prompt — see examples/fan-in-merge.yaml. A flow that still uses a removed type no longer parses-and-throws: it loads, validate reports unknown node type "merge", run aborts before any model call, and the visual editor paints the node as a red ⚠ error node you can retype or delete. Covered by e2eCli/scenarios/validate.e2e.ts (unknown-type validate/run gate), src/engine/run.test.ts (runtime guard), and e2e/browser/node-types.spec.ts (error-node rendering).
Fix: a write node’s output path is now part of its cache key. Changing a write node’s path and re-running with --cache used to serve the old cached output and never write the new file; the path is now folded into the Merkle key. Covered by src/engine/cache.test.ts.

0.1.18

CLI: chainq run now shows you the result. It used to print only per-node status (✓ greet) to stderr — to see what a node actually produced you had to cat .chain/outputs/<id>.out. Now the chain result goes to stdout (every leaf node; multiple leaves each under a — <node> — header) while progress stays on stderr, so chainq run flow.yaml | jq pipes only the result. Two new flags follow the usual convention: -q/--quiet hides progress but keeps the result (and still shows failures), -s/--silent prints nothing at all (exit code only). Reads from the in-memory run result, not the cache file, so a cmd leaf prints too. Covered by e2eCli/scenarios/output.e2e.ts (default split / -q / -s / multiple leaves / long forms).
Editor: the middle column only appears for nodes that use it. It no longer shows a fake prompt field for the 5-of-8 node types that have no prompt (cmd, splitOut, aggregate, merge, input).
Editor: forcing a connection no longer lets a cycle through. /api/connect’s force path only relaxes pure edge removals, never an edge add that would create a cycle.
UI text is now English across the editor (e.g. lineage wires read “refs”).
New example flows: merge-join, split-aggregate, shell-command.

0.1.17

Editor: the node panel stays centered no matter how far you’ve scrolled the canvas. The floating node panel was positioned absolute inside the scrollable canvas, so it anchored to the canvas’s un-scrolled origin — pan or scroll over to a node on the right or bottom, open it, and the panel drifted off-screen (showing only a sliver, or nothing at all). It now uses position: fixed, anchoring to the viewport, so it opens dead-center every time. Covered by e2e/browser/modal-position.spec.ts (wide chain → scroll to the far corner → open the last node → assert the panel is centered and fully on-screen).

0.1.16

Editor: delete a connection straight from the canvas. Every data-flow wire now floats two buttons at its midpoint — + (insert a node, as before) and × (delete the connection; hover turns it red). Deleting a wire used to be impossible: it goes through /api/connect, and if removing it left a downstream prompt’s {{ $node["x"] }} reference no longer upstream, editFlow treated that as a newly-introduced error and refused to write the whole flow — the wire simply wouldn’t delete. Edge-delete now carries force (like node-delete): the removal lands, any step left with a dead reference comes back as a warning and is flagged red (⚠) on the canvas to fix, while adding a wire stays strict (cycles / bad wiring still rejected). Covered by e2e/browser/delete-edge.spec.ts, delete-edge-fanin.spec.ts, and delete-edge-fanout.spec.ts.
Editor: edit the default profile’s command from the toolbar pill. The ● claude -p · real pill in the top bar was a fixed label; click it to edit the default profile’s launch command (e.g. claude -p --model claude-sonnet-4-6) — the local CLI each ai step shells out to. Prompts always arrive on stdin, so the pill only holds the launch command. New /api/set-profile (comment-preserving; an empty command is rejected). Covered by e2e/browser/profile-field.spec.ts.

0.1.15

Editor: canvas interaction v2 — marquee select, Space-to-pan, click-selects / double-click-edits. Reworks the canvas gestures toward the Figma/n8n model so a big graph is easy to select and navigate (supersedes the 0.1.14 drag-to-pan + Shift+click model):
- Drag the empty canvas = rubber-band marquee select — every node the box touches joins the selection; Shift+drag adds to it; a plain press clears it. Multi-select is now grab-a-box, not one Shift+click at a time.
- Pan moved off plain-drag: hold Space then drag to pan (cursor grab/grabbing); mouse/trackpad scroll still pans natively. Space is ignored while a field is focused, so it still types in the prompt editor.
- A single click selects a node (ring, no panel); a double click opens the editor.
Covered by e2e/browser/canvas-interactions.spec.ts (marquee + Space-pan + single/double-click).
New examples/generate-json.yaml — a copy-paste recipe for producing a JSON file. input → three fields → an ai step with schema assembles them into a validated JSON object → a write node lands it as out/result.json. Shows the reliable way to emit JSON: let schema parse and validate the object instead of hand-building a { } string (which breaks on quotes/newlines). How-to in docs/scenario/creation/create-write.md (產生 JSON 檔); covered by e2e/browser/generate-json.spec.ts.

0.1.14

Editor: pan, multi-select move, and collapsible output — three canvas-readability wins so a big graph stays usable.
- Drag the empty canvas to pan the view (cursor grab/grabbing). Drags on nodes (move) and on ports (connect) are untouched.
- Shift+click nodes to build a selection group (accent ring, no panel opens); dragging any member moves the whole group by the same delta and saves once via /api/layout. A plain click drops the group and opens the panel.
- A finished node’s output is hidden by default so a tall result no longer buries its neighbours. The ×N items badge doubles as a show/hide toggle (▸/▾ when there’s no count) — click to reveal, click again to hide.
Covered by e2e/browser/canvas-interactions.spec.ts.

0.1.13

Editor: run buttons renamed to match n8n — “Execute step” / “Force execute”. The node panel’s ▷ Run to here / ↻ re-run (and the ▷ / ↻ icons on each canvas card) are now ▷ Execute step / ↻ Force execute, with tooltips that spell out the cache behavior: Execute step runs the node plus the upstream it needs and reuses cache (only what changed runs — n8n’s partial execution); Force execute ignores cache and really calls the model. Wording only — the partial-execution behavior is unchanged.

0.1.12

Timeout is now yours to set — a long ai step no longer dies at a hardcoded limit. An ai step writing a whole article used to be killed at a fixed ceiling (120s on the CLI, 300s in the editor — and the two disagreed). Now any ai/cmd step takes an optional timeout: (seconds) in the YAML, and a flow can set a defaults: { timeout: N } that every step falls back to. Resolution, most specific wins: a node’s own timeout → the flow default → a built-in 300s, identical on the CLI and in the editor. Set timeout: 1200 on the article step and it runs to the end. In the editor a node’s timeout hides behind a ◷ clock in its INPUT header (bare when unset, ◷ 1200s when set, click to edit), and the flow-wide default is the ◷ Timeout clock in the top bar. Covered by e2e/browser/timeout-field.spec.ts plus engine tests in dag.test.ts (parse/validate) and run.test.ts (per-node caps, flow default, precedence).

0.1.11

Editor: delete a node even when a downstream step still points at it. Deleting a node that another step’s from: still referenced used to be rejected outright (another step still depends on this — rewire it first) — the panel stayed, the node stayed, and it looked like the delete did nothing. Now the delete always lands: any downstream step left with a dangling reference is flagged red on the canvas (⚠ from: "X" does not exist) and named in a canvas message, so the breakage is loud and you can rewire it. Edits that would corrupt the YAML are still rejected (壞不落地 still holds for corruption). The deleted node’s reference is left in place on purpose — so the broken step shows an error to fix, rather than silently dropping the wiring. Covered by e2e/browser/delete-node.spec.ts (single dependent) and e2e/browser/delete-fanout.spec.ts (a fan-out middle node whose two downstream leaves both turn red). New examples/fan-out.yaml to try it.

0.1.10

Editor: reference wires — see which upstreams are data, which are just looked up. Every connection on the canvas used to look the same warm line, so a real $json data input and a {{ $('id') }} value lookup were impossible to tell apart. Now the canvas draws two kinds: data-flow wires (warm, solid — the $json main input) and reference wires (cool, dashed — a $('id') / $node["id"] value lookup). A 引用線 / reference wires toggle in the zoom toolbar hides the reference wires when you want a cleaner view. Covered by e2e/browser/reference-wires.spec.ts.
You can now reference a value from any step back, not just a direct from:. A prompt’s {{ $('id') }} / {{ $node["id"] }} can reach any ancestor — a node anywhere upstream — instead of only a node wired into from:. So you can pull a value from several steps back without adding it to the data flow. ($json still binds the primary from[0] input; referencing a non-ancestor is still flagged.) Spine-aligned ancestors pair by lineage; off-spine ones fall back to best-effort.
Editor: clicking an “earlier output” now inserts a reference without rewiring. This reverses the 0.1.7 behavior: a click on an earlier output used to append the node to from: as well. It now inserts {{ $node["id"] }} as a pure cross-step reference and leaves from: untouched — keeping “connected nodes” (data flow) and “earlier outputs” (references) distinct. Paired with the engine change above, the reference resolves at run time on its own.

0.1.9

Editor: stop a run in flight. A run had to finish on its own — a slow model call left you waiting, or killing the server. A Stop button now cancels the running chain: it aborts the in-flight node’s subprocess and skips everything downstream, so you can fix a prompt and re-run without waiting it out. The add-step control also moved into the canvas toolbar (next to zoom), keeping the canvas chrome in one place.

0.1.8

Editor: zoom the canvas to manage big graphs. A large flow ran off-screen with no way to see its whole shape. The canvas now zooms — − / + / fit buttons in the top bar, plus ⌘ + / ⌘ - and trackpad pinch. fit frames the entire graph; zoom is layout-only and never touches the flow.
Editor: the ai output-format box collapses. The output-schema selector (Text / JSON / List) now folds up, and collapsed it shows just the current format — so the output column stays tidy when you’re not editing the schema.

0.1.7

Editor: run your unsaved edits without saving first (draft model). A node’s edits used to live only in the textarea, while a run read the flow from disk — so a re-run executed the SAVED prompt and the stream’s re-render wiped what you typed. Now an edit is kept as a per-node DRAFT: a run sends it as an in-memory override (the file is never touched, same idea as the /api/render template override), the run stream redraws only the output region so the edit survives, and the draft PERSISTS across close / node-switch / raw toggle — no prompt on leave. Save writes it to the file; ↩ Reset discards it back to the saved value; a ● marker on the canvas node + a footer chip flag unsaved drafts. Browser-only, per flow, per session. Covered by e2e/browser/draft-run.spec.ts.
Editor: ai output-schema field editor in the output column. An ai node’s structured-output schema moved from a raw JSON textarea to a two-level editor that sits above the output (where the contract belongs): pick an output format (Text / JSON / List), and JSON exposes field → type rows with a live “model returns:” preview. List wraps the array in a reserved _list field (the engine forbids a bare top-level array) — pure UI sugar, zero engine change. Covered by e2e/browser/schema-editor.spec.ts.
Run state: “queued” vs “running” are now distinct. A run marked the WHOLE cone “running” at once, so a sequential chain looked like every node was executing simultaneously. The engine now emits an onStart the instant a node actually begins (past the cache/skip gates), streamed to the UI as a running record. The canvas shows the ONE executing node with the live spinner (solid accent + pulse) and the rest “queued” (dashed dim border, no spinner) — so you can see the queue advance one step at a time. Covered by run-state.spec.ts + server-stream tests.
Run state: a running node no longer shows its stale previous output. While a node is queued or running, the panel’s OUTPUT (and the canvas card) showed the LAST run’s cached result — looking like the current run already finished with wrong data. It now shows “running…” / “queued — waiting its turn…” instead, and the real output replaces it when the run settles (loadItems is guarded so the cached items can’t paint over the live indicator).
Editor: click an earlier output to wire it in + insert its reference. The node panel’s “earlier outputs” box (transitive upstreams not yet wired) was read-only — you had to drag a wire on the canvas, then come back and click the reference. Now one click on an earlier output appends it to from: (NON-primary, so $json stays the first input) AND inserts {{ $node["id"] }} at the cursor. Unsaved prompt edits are preserved across the wiring reload, matching the existing direct-input insert. Covered by e2e/browser/insert-earlier.spec.ts.

0.1.6

Docs release (no code changes).

README is now a scannable landing page — one-line pitch, a hero screenshot of the visual editor, a CLI + UI quickstart, and a “you want to… → go to” table linking the deeper docs. The engine internals moved behind those links.
Docs synced to current behavior — the cache walkthroughs use --cache (since run re-runs by default), dead fake/offline-profile references are gone, and the dated design/planning snapshots carry a “current behavior differs” pointer.

0.1.5

Editor: tighter node panel. The input column had large phantom gaps (the .mbody white-space:pre-wrap was rendering the source-code newlines between its blocks). Removed it there — so the input chips, schema, direct inputs and earlier outputs all fit without scrolling.
Editor: the type dropdown is context-aware. A start node (no upstream) only offers input / ai / cmd; a node with upstream offers the consumer types (ai / cmd / assemble / splitOut / aggregate / merge) but not the input trigger. No more nonsensical options like merge on a start node.
Editor: an input trigger no longer shows a prompt column. The hide had no effect because .mcol{display:flex} overrode .hidden{display:none} — made .hidden win (!important) and collapse the panel to INPUT + OUTPUT (two columns) for triggers, since a trigger has no prompt.

0.1.4

Editor: change a node’s type. The panel header has a type dropdown — switch a node between ai / cmd / input / merge / etc. It resets that node’s type-specific fields to the new type’s defaults and keeps its wiring (an input trigger drops its from). E.g. turn a step into an input node and add a default-valued parameter.
Editor: fix the doubled “no upstream” note. A start node now shows one clear line instead of two.

0.1.3

Editor: no more typing node ids. The from text box is gone. A node’s wiring now shows as chips (first = $json), each with a × to disconnect — so you wire on the canvas (drag a node’s ● / the + on a wire) and just see the result here.
Editor: the panel shows earlier steps’ outputs. Below the direct input(s), every earlier upstream step’s output is listed (read-only) so you can see all the data flowing toward this node.

0.1.2

Editor: the from field is clearer. It now reads “wired upstream(s)”, shows empty = start node, and a hint explains it auto-fills when you connect on the canvas — typing ids is optional, not required.

0.1.1

chainq run now re-runs every node by default. Pass --cache (alias --reuse) to reuse unchanged outputs; the partial-run modes (--from / --to / --steps / --pin) still reuse upstream cache as before.
Editor: invalid nodes are flagged. A node whose prompt references an unwired upstream (e.g. $('start') without start in from:) — or any other validation error — now shows a red outline and a ⚠ reason on the canvas and in its panel, so the problem is visible before you run.
Editor: insert a step between two nodes. Hover a wire and click the + at its midpoint to drop a new node onto that edge — the upstream and downstream are rewired automatically.

0.1.0

Initial release: prompt-chain runner for local CLI models (claude -p, codex -m), Merkle-cached engine, CLI + local 127.0.0.1 visual editor.