* feat: let AI harnesses solve tasks with code, then save and reuse it Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * fix: page code mode searches locally and name the job in the chat's run result Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * chore: point the EE ref at the code_mode feature usage registration Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * fix: follow the search cursor and match words in any script Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * fix: count only test_run_code runs and save only their previews Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * refactor: drop save_preview and keep code mode programs with write_script Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * docs: trim code mode guidance and cover it with ai_evals cases Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * fix: show the staged name of drafts found by a list_workspace_items search Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PjsYvufUfjaH7aRMxVWTue * chore: update ee-repo-ref to 6e6a9f3ed1e3023da46dc6f6af8f465abb5da001 This commit updates the EE repository reference after PR #988 was merged in windmill-ee-private. Previous ee-repo-ref: 82008bd50305789174eb50ed06d9c12b25b74d9e New ee-repo-ref: 6e6a9f3ed1e3023da46dc6f6af8f465abb5da001 Automated by sync-ee-ref workflow. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: windmill-internal-app[bot] <windmill-internal-app[bot]@users.noreply.github.com>
2.0 KiB
Solving a task with code
When the user asks for something done against the systems their workspace connects to (look something up in an API, move data, check a status), answer it with a Windmill script run as a preview: it executes on the workspace's workers, with its resources. A script that worked can be saved and run as is the next time.
- Look for a saved script first, by the user's words and by the system the task touches. If one fits, run it rather than writing new code; if one nearly fits, offer to change it rather than writing a near-duplicate.
- Otherwise write one script and run it as a preview. Do the fetching, filtering and
combining in the code, and return only what answers the task, not raw API payloads. Take
credentials as a resource-typed
mainparameter, passed as$res:<path>when running, never as a token in the code. Givemaintyped parameters for whatever would change between runs, so the script is reusable. - Offer to save it when the task is likely to come back, without saving it unasked. Saving makes a draft, deployed only when the user asks; from then on it is an ordinary script.
For a third-party API, look up its hub integration for its resource type and how to call it.
Load the write-script-<language> skill before writing the script; the resource's fields are in
rt.d.ts (wmill resource-type generate-namespace). A preview refuses a file that
is not at a workspace path inside the project: write it at u/<username>/<name>.ts (wmill workspace whoami gives the username) without asking, never under f/, which a push deploys,
even next to the resource it uses; run it with wmill script preview <file> -d '<args>'. Search with wmill script list and wmill flow list; run a deployed one
with wmill script run <path>. Never push the file on your own. Delete it if the user does not
keep it; if they do, move it where the project syncs (the includes of wmill.yaml, f/** by
default) and push only on their go.