Files
windmill/system_prompts
Ruben Fiszel 74f579e6d9 feat(pipeline): local development for data pipelines (CLI --local + pipeline dev preview) (#9840)
* feat(pipeline): local development for data pipelines (CLI --local + pipeline dev preview)

Add the local edit→preview→run loop for data pipelines (folders of `// pipeline`
scripts), the analog of `wmill dev` / `wmill app dev`, usable from a code editor
or an agentic loop — without deploying.

No backend changes: full body inference comes from the same wasm the frontend
uses (windmill-parser-wasm-asset), which returns assets + pipeline annotations in
one call; local runs reuse runScriptPreview with _wmill_skip_asset_dispatch.

- localGraph.ts: wasm-backed working-tree → asset-graph builder (the enabler)
- pipeline show/run --local; new pipeline docs (PIPELINE.md/AGENTS.md) subcommand
- pipeline dev watcher + /pipeline_dev page (PipelineDevView) rendering the same
  PipelineGraphEditor from the pushed local graph, run via preview
- cascadeRun.ts: reusable run primitives extracted from the route page
- regenerated CLI agent docs

See docs/pipeline-local-dev.md for the full design, test steps, and handoff TODOs.
The live `pipeline dev` browser preview is implemented but not yet stack-verified.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(pipeline): improve local dev preview (run, activity, responsive)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(pipeline): dev-preview args, multi-root run, ws auto-reconnect

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): connect managed-materialize producer in local dev graph

The CLI pinned windmill-parser-wasm-asset ^1.728.1, which predates managed-materialize support (added in 1.733.1); the frontend already pins 1.740.0. The CLI's wasm therefore never emitted `// materialize`, so the producer had no output edge and showed disconnected from its `// on` consumers. Bump the CLI to 1.740.0 (matching the frontend) and translate the parsed materialize target into the producer's write edge + materialize_target, mirroring frontend resolveGraph.ts.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): harden local-dev CLI (bare-.sql crash, defaultTs, docs clobber)

Review fixes, complementary to the dev-preview/materialize/multi-root work already
on the branch (none overlap those commits):

- localGraph: a bare `.sql` (no dialect) made inferContentTypeFromFilePath throw and
  abort the whole graph build — and wedge `pipeline dev` at startup. Skip the
  unclassifiable file instead. Also map `bunnative` → parse_assets_ts and add
  ruby/rlang/nu/powershell to the `#`-comment fallback.
- show/run/docs/dev: thread the resolved `wmill.yaml` defaultTs into the graph
  builder so `.ts` infers under the workspace's runtime (bun vs deno) instead of
  always bun — `opts.defaultTs` was always undefined (no such CLI flag).
- dev: wrap the startup graph build so a half-written file can't abort the watcher.
- docs: don't clobber a user-authored AGENTS.md/CLAUDE.md — only (over)write the
  pointer when absent or already a generated `@PIPELINE.md` pointer.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): bind dev WS to loopback + local-graph regression tests

- pipeline dev WS broadcast the folder's full script source (scripts[].content + temp_script_refs) unauthenticated on 0.0.0.0:3201 — bind 127.0.0.1 so it's not LAN-reachable (webview localhost + SSH/devbox port-forward still work).

- Add regression tests for the just-landed local-graph fixes: bare .sql is skipped (was a build/dev-startup crash), defaultTs threads into .ts runtime inference (bun vs deno), and #-comment languages (ruby) use the # annotation fallback.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(pipeline): --frontend flag for pipeline dev page origin

wmill pipeline dev opens <remote>/pipeline_dev, but that route only exists in this build's frontend, so it 404s against a remote whose deployed frontend predates it. --frontend <origin> points the page at a locally-run frontend (REMOTE=<remote> npm run dev) while the API/token still target the remote — enabling the live preview against a real backend before the PR is deployed. No behavior change when omitted. Regenerated CLI agent docs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): WS session token + details-pane live-reload refresh

Addresses CI review (Codex/Pi/Claude):

- dev WS: a browser tab could open ws://localhost:<port>/ws and receive the folder's full source (browsers don't enforce same-origin on WS, loopback bind alone doesn't help). Gate the upgrade on an unguessable per-session token carried in the dev-page URL (verifyClient → 401 without it). Verified: no-token/bad-token connections get 401 with no bundle.

- details pane: scriptRes keyed on [workspace, selection, draftScript] didn't re-run on a pipeline dev live-reload (same selection), so the open pane showed stale source. Thread a localScriptsVersion (the pushed bundle) into the key. Verified: editing a selected node's file updates the pane source without reselect.

- docs/pipeline-local-dev.md: refresh the stale 'not yet exercised' status + done TODOs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): emit volume: annotation assets in local dev graph

Addresses CI review (Codex P1 / Pi P1): the wasm body parser doesn't surface `// volume: <name>` annotations — the frontend (infer.ts:parseVolumeAnnotations) and backend (asset_inference.rs) parse them separately and merge as rw volume assets. localGraph didn't, so a `# volume: cache` producer had no write edge and showed disconnected from its `// on volume://cache` consumer (and pipeline run --local wouldn't schedule downstream). Mirror the leading-comment-block scan (SQL excluded, matching both reference parsers) and merge into inferScriptAssets. Regression test added; verified producer -> volume://cache -> consumer connects.

Also (Codex P2): docs/pipeline-local-dev.md manual browser URL omitted the new ws_token param — without it the WS upgrade is rejected and the page sits disconnected. Doc now says to copy the URL the CLI prints (carries wm_token + ws_token) and recommends --frontend.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): runAll excludes event roots + review polish

Addresses CI review (Codex P1, Claude P2/P3):

- pipeline run runAll: derive the whole-pipeline selection from validStarts + descendants instead of all runnables, so an unqualified 'pipeline run <folder>' no longer fires event-trigger roots (kafka/mqtt/…) with empty args/side effects. Verified: a kafka root is excluded from the plan.

- cascadeRun.ts runBoundedCascade: use buildLineageDownstreamMap (read-aware) so a pure-reader runs after its producer, and return cyclic — parity with the route page's bounded run (the file is meant to be THE shared correct primitive).

- PipelineGraphEditor: storedRightPaneSize starts at 0 so the orientation-aware default (55% stacked / 40% side-by-side) actually applies on first open.

- localGraph fallbackParse (go/bash): scan only the leading comment header (no body-comment phantom triggers) and strip key=value options from the asset URI; regression test added.

- docs: reject '..' in the folder arg (it writes files under f/<folder>).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): route local previews to the // tag worker

Addresses CI review P1: the local graph/bundle dropped the parsed `// tag`, so a node annotated `// tag gpu` ran on the default worker in both `pipeline run --local` and `/pipeline_dev`, while the deployed pipeline routes it to that worker tag. Carry the tag through LocalScript / the pushed bundle / LocalScriptContent and pass it to runScriptPreview at all three launch sites. Verified: a duckdb node tagged `bash` produces a job tagged `bash`; regression test added.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): add asset partitions/schemas routes to OpenAPI, use generated client

The ducklake asset panels (PartitionStatusGrid, SchemaHistoryPanel) hit /assets/partitions and /assets/asset_schemas via raw fetch with cookie-only auth, because those backend routes were never added to openapi.yaml so the generated client had no methods for them. On /pipeline_dev (token-via-URL, no session cookie) the raw fetches 401'd. Add both GET routes + MaterializedPartition/AssetSchemaVersion schemas to openapi.yaml and call them through AssetService, which injects the bearer token, types, and cancellation automatically. Verified: Partitions + Schema tabs load in /pipeline_dev. (backfill stays a raw fetch — it's an EE-only route not in the OSS spec — with the token added inline.)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore(cli): regenerate bun.lock for windmill-parser-wasm-asset

package.json / package-lock.json carry windmill-parser-wasm-asset@1.740.0 but the tracked bun.lock (the CLI installs/builds/tests via bun) was stale, so fresh bun installs would resolve a different graph than the committed lock. Regenerated.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): show asset producer + its runs in the dev-preview panel

Selecting a ducklake/asset node in /pipeline_dev showed 'No producer for this asset' because selectionProducers wasn't passed (it's derived from the deployed graph on the route page, absent here). Compute it from the local graph's w/rw write-edges (incl. the // materialize target) and pass it through, mirroring the route page — so the panel shows the producing script and its (preview) runs, including data-test failures.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): carry annotation metadata onto local-graph runnables

The local graph emitted only path/usage_kind/in_pipeline/materialize_target per runnable, so /pipeline_dev and pipeline show --local weren't the same surface as the deployed graph for annotated scripts — missing the badges/lineage the shared canvas renders. Map the wasm-parsed partition_kind, freshness, tag, retry, data_tests, column_lineage, and materialize_strategy (derived append/merge/replace) onto each runnable, mirroring the deployed AssetGraphRunnableNode.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): exclude event handlers that are lineage descendants from runAll

The runAll guarantee ('never fires an event handler with empty args') only held for event ROOTS — validStarts excludes them, but runAll then unions in descendants(dag, start), so a kafka/mqtt/... handler that also reads an upstream pipeline asset (a lineage descendant of a valid start) still landed in the plan. Add eventTriggerScripts() and subtract it from the selection after the descendant union. +unit test.

Also: docs/pipeline-local-dev.md recipe used 'pipeline docs demo_pipeline' without --local (default queries the deployed graph → hits the empty hint); add --local.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): whole-pipeline run cuts at event handlers (drop their downstream too)

The prior runAll fix subtracted event handlers from the selection but left their downstream: for manual_root → asset_x → kafka_handler → asset_y → consumer, deleting only kafka_handler left consumer selected, and topoOrder then ran it as a root with missing/stale event-derived inputs. Replace the descendant-union+delete with reachableCutting(dag, validStarts, eventHandlers): traverse from valid starts but treat event handlers as cut points, so a node reachable ONLY through an event handler is dropped while one reachable via a non-event path stays. +unit test.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): recover // tag in the go/bash annotation fallback

The wasm path carries out.tag, but the go/bash fallback (and the wasm-error degradation path) only recovered pipeline + on, so a // tag gpu on a bash/go node — or a temporarily-unparseable ts/py/sql node — silently routed the local preview to the default worker while the deployed pipeline routes to the tag. Scan for // tag in fallbackParse too. +test.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(pipeline): extract shared assetProducers helper

The 'who writes this asset' write-edge derivation was copied verbatim in PipelineDevView and the pipeline route page — two copies that would drift. Extract assetProducers(graph, selection) into graphTraversal.ts and use it from both, keeping the dev view and route page in lockstep.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): only overwrite AGENTS.md/CLAUDE.md when it's the exact generated pointer

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): wire local-dev runs into the selected-node runs pane

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): exclude data_upload/webhook entrypoints from auto CLI runs

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(pipeline): --upload binds an object to a data_upload/webhook entry point

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(pipeline): add "Run + downstream" to the dev preview detail form

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): cut non-autorun triggers on all run paths; multi-binding --upload

Address CI review: apply the data_upload/webhook/event barrier cut to the
single-root and bounded (--from/--to) paths, not just whole-pipeline; accumulate
repeatable --upload bindings per script (were overwritten); scope dev upload keys
by script+param to avoid basename clobbering; drop <script> from help text.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): reseed dev run form when a local edit changes the script's args

The read-only pane is keyed on script.path only, so in /pipeline_dev the selected
node re-resolves on every WS bundle without remounting; PipelineScriptView cloned
script.schema once, so adding/removing args left the run form on a stale schema
(could run with missing inputs). Extract PipelineRunForm (owns the SchemaForm
clone) and key it on the serialized schema: a real arg change reseeds the form,
an unchanged re-resolve keeps in-progress input.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): don't cut a scheduled/manual root that also has a non-autorun trigger

Address Codex P1: the barrier set subtracted only --upload-bound scripts, so a
script with both `// on schedule` and `// on data_upload` resolved as the start
yet was also a barrier — reachableCutting skipped it, giving an empty run plan.
Subtract all valid starts (schedule/manual roots + bound handlers) from barriers:
a legitimately-scheduled root runs on its schedule path even if it also carries a
caller-input trigger; pure input-only roots stay cut. Adds a regression test.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): deployed non-autorun enrichment, s3:// storage, --to cut accounting, tag regex

Address CI review (Codex P1/P1/P2, Pi P2):
- Deployed `pipeline run` recovers marker-only data_upload/webhook/email triggers
  from script bodies (like the `show` path) so input-only entrypoints are cut
  instead of auto-run empty on the deployed graph.
- `--upload s3://<storage>/<key>` keeps the named storage (authority) instead of
  folding it into the key, matching the S3Object round-trip convention.
- Bounded `--to` targets cut by a barrier are reported in droppedEnds (+warning),
  not reachableEnds.
- fallbackParse `// tag` matches a single token (\S+), rejecting multi-word prose.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): header-only deployed marker scan, fail-closed enrichment, default-storage s3 keys

Address CI review (Codex P2, cubic P1/P1/P2):
- Deployed marker recovery scans the LEADING comment header only (shared
  recoverHeaderMarkers helper, reused by the show enrichment too) so a body
  comment `// on data_upload` can't inject a phantom trigger and over-cut.
- Deployed run enrichment fails CLOSED: a script-body fetch error aborts the run
  instead of silently letting an input-only entrypoint run with empty args.
- Revert `--upload s3://` to default-storage whole-path keys (matching pipeline
  `s3://` asset-URI semantics); named-storage authority-splitting broke nested
  default keys like `s3://raw/2026/events.csv`.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): reject trailing content on fallback native markers; trim s3:/// key

Address CI review (Codex P2, cubic P3):
- fallbackParse now requires a native marker (`// on data_upload`) to stand alone;
  a line with trailing content (`// on data_upload f/foo`, `# on kafka topic`) is
  rejected, matching the canonical parser and keeping local/deployed parity.
- s3UriKey trims a leading slash so the canonical empty-authority default form
  `s3:///key` doesn't leak a leading slash into the object key.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): persist dev WS token per-port so reconnect survives a CLI restart

Address Codex P2: the /pipeline_dev auto-reconnect reuses the ws_token from the
page URL, but `pipeline dev` minted a fresh random token each start, so a restart
on the same port left the open page rejected by verifyClient forever. Persist the
token per-port under the user-private config dir (0600) and reuse it on restart,
so an already-open page reconnects — matching the reconnect behavior's intent.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): scope persisted dev WS token by workspace+folder+port

Address cubic P2: keying the persisted token by port alone let a stale browser
tab from a previous folder's session on the same port reconnect and receive a
different folder's source. Scope the token file by workspace+folder+port so a
same-session restart still reconnects, but a different folder on the same port
gets a distinct token that rejects stale cross-folder tabs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): caller args can't override skip-dispatch guard; hash the dev token key

Address CI review (Codex P1, cubic P2):
- makeLaunch / CLI run build args with `_wmill_skip_asset_dispatch` LAST (and drop
  any caller-supplied copy) so a run-form/`--upload` arg can't re-enable backend
  asset dispatch while the client orchestrates the cascade (double-run / running
  deployed subscribers from a local preview). Adds a cascadeRun guard test.
- Dev WS token file key is a sha256 of NUL-delimited workspace+folder+port, so
  different folders (`a/b` vs `a_b`) can't collide onto the same token file.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(pipeline): canonical s3://storage/key --upload parsing; scope dev token by remote+root

Address Codex P1/P1:
- Restore canonical S3Object URI parsing for `--upload` s3 sources, matching the
  frontend's `parseS3Object` (`s3://<storage>/<key>`, empty authority ⇒ default,
  `s3:///key`/`s3:///nested/key` for the default store). `s3://secondary/k.csv` →
  `{ s3: "k.csv", storage: "secondary" }` so a named-storage object is read from
  the right store. (This is the canonical convention; the default-storage nested
  key is served by the `s3:///` form.)
- Scope the persisted dev WS token by remote+workspace+root+folder+port (was
  workspace+folder+port), so two profiles on different remotes (or local checkouts)
  with the same workspace/folder/port don't share a token — a stale tab can't
  reconnect across a workspace/remote boundary.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-01 11:41:36 +02:00
..

System Prompts

This directory contains the single source of truth for AI system prompts used by both the frontend copilot and CLI guidance.

Structure

system_prompts/
├── base/              # Core instruction templates (manually written)
│   ├── flow-base.md   # Shared OpenFlow structure guidance
│   └── flow-cli.md    # CLI/local-agent workflow guidance for write-flow skill
├── languages/         # Language-specific instructions (manually written)
└── auto-generated/    # Auto-generated files (DO NOT EDIT)
    ├── sdks/          # SDK documentation
    ├── cli/           # CLI command documentation
    ├── prompts.ts     # TypeScript exports
    └── index.ts       # Helper functions

Usage

Regenerating Prompts

When SDK methods or the OpenFlow schema change, run:

python system_prompts/generate.py

To also refresh the standalone skills in a Claude plugin checkout:

python system_prompts/generate.py --plugin-dir ~/windmill-claude-plugin

--plugin-dir accepts:

  • the windmill-claude-plugin repo root
  • a plugin root such as plugins/windmill
  • a direct skills/ directory

To regenerate the public docs repo (consumed by context7):

python system_prompts/generate.py --context7-dir ~/windmill-cli-docs

--context7-dir writes a fully-rendered snapshot (AGENTS.md, cli-commands.md, skills/<name>/SKILL.md, README.md, manifest.json with the Windmill version) with all template placeholders resolved — suitable for ingestion by docs aggregators. In CI this runs from .github/workflows/publish-cli-docs.yml on every release tag. The generator refuses to wipe the target directory unless it's empty or has a context7 marker (context7.json, manifest.json, or a windmill-cli-docs git remote), so a typo can't delete unrelated files.

This will:

  1. Parse TypeScript and Python SDK files to extract function signatures
  2. Parse the OpenFlow YAML schema
  3. Parse the CLI commands
  4. Assemble complete prompts from markdown files
  5. Generate TypeScript exports in auto-generated/
  6. Optionally refresh plugin-ready standalone SKILL.md files in the target directory

Scope

These system prompts contain ONLY:

  • How to write Windmill scripts (language syntax, conventions, SDK usage)
  • How to structure Windmill flows (OpenFlow schema, module types, data flow)
  • Resource type handling, S3 operations

They DO NOT contain:

  • Tool usage instructions (edit_code, set_flow_json, etc.)
  • IDE/editor specific commands
  • Testing tool invocations

Tool instructions are added separately by the frontend and CLI.

CLI-only workflow instructions live in base/flow-cli.md and are included in the generated write-flow skill for wmill init. They are intentionally excluded from the frontend flow chat prompt.

Integration

Frontend

Uses Vite path alias $system_prompts pointing to auto-generated/:

import { FLOW_GUIDANCE } from "$system_prompts/flow";
import { getLangContext } from "$system_prompts/languages";

CLI

Generates /cli/src/guidance/skills.gen.ts with embedded skill content for wmill init.

Editing Guidelines

  • Edit markdown files in base/, languages/
  • Never edit files in auto-generated/ directly
  • After editing, run generate.py to update exports