FlowAId 0.7.0

• More list endpoints are paged. GET /v1/agents, /v1/api-keys, /v1/credentials,
  /v1/knowledge/sources, /v1/notifications, /v1/saved-views, /v1/workflows/:id/versions and
  /v1/workspaces/:id/members now answer { items, next_cursor } (keyset pagination with limit, at
  most 200, and cursor) instead of a bare array, like the other lists. Callers that read these
  responses as arrays must read items and follow next_cursor; the web app reads every page.
• Fixes from a page-by-page audit of the app.
  • A run's graph draws every connection: branch cases get their own outlets, and data connections
    come from the workflow's bindings when the version has no stored plan.
  • A new decision step uses the TypeSafe key the server already has (or a secret the workflow
    already declares) instead of starting with an unbound-key error.
  • ⌘K in the builder opens Add node whenever focus is on the canvas or nowhere in particular, and
    closing the palette hands focus back to the canvas.
  • Replay starts with "Reuse recorded results" (free) selected; running every step again is a
    choice the confirm button names.
  • Template cards say when a key the server has still needs a saved credential, and tell two
    secrets of the same kind apart.
  • Add schedule checks each cron field and whether it ever fires, and shows the next run times;
    missing required input fields are flagged at the input step.
  • New knowledge source rejects addresses that are not http(s) URLs and does not count an
    embedding model without a key as a finished choice.
  • Expose workflow no longer suggests exposing before a deploy that would switch it off again.
  • Evaluation cases need the workflow's required inputs in the form view; the set page names the
    workflow it tests.
  • Edit agent asks before discarding changes. New workflow only needs a name for Blank, and keeps
    an unsent "Describe it" text across a reload.
  • Missing runs, tasks and workflows say so and link back to their list; connection failures say
    the API cannot be reached; other errors keep their technical details one click away.
  • The Publish dialog lists problems in plain words with "Show node", and explains a protected
    environment when it is ticked. Settings shows a template copy's description.
  • Wording: steps rather than nodes in runs and human tasks, OpenAPI document versions, "an
    OpenAI credential", "TypeSafe", step requirements in words where nothing on the form is marked.
• Clearer step settings in the builder.
  • Inputs and output results that are built from several fields (object and array bindings) now
    open as editable Fields / List instead of an empty value box, so opening a step no longer hides
    its wiring and typing into it no longer replaces it by accident.
  • The value source switch reads Value · Reference · Template · Expression, with a one-line
    explanation of the current choice; step inputs and output results have visible labels.
  • "Execution policy" is now "Errors & limits": what happens when the step fails, attempts, time
    limit, cost and token limits as plain controls (with their defaults shown), and the full policy
    still editable as JSON.
  • Output steps explain the outcome label and "End the run early".
  • Workspace settings say what the retention, queue and budget values actually do today, show
    "Unsaved changes", and ask before a link or reload discards them.
  • Environment tags read dev · stg · prod; optional boolean criteria are no longer announced as
    required; "Add a key for LLM" instead of "llm".
  • Template fields speak the compiler's grammar everywhere: the footer counts the references a
    template really reads (start.order_id, $vars.limit), holes that name a missing step,
    output or setting are underlined with the fix, parse errors show under the field, single-line
    template fields complete references too, and the value switch's Template mode no longer offers
    the old picker whose input.* / nodes.* insertions did not compile.
  • Unsaved edits in Settings and workflow settings survive switching tabs; tabs with unsaved
    edits carry a dot, sections have Discard, and leaving the page asks first.
  • The mouse wheel zooms the canvas toward the pointer instead of panning it; pan with space-drag,
    the middle or right button, or the minimap.
• Step-by-step guidance on every page in the navigation.
  • Each page opens with "Start here": what you can do there, when to use it, what it needs (checked
    against the workspace: keys, tools, published versions, your role), how to start and what you
    get. It collapses to one line, remembers that, and still flags anything missing.
  • Creating an agent, knowledge source, evaluation set, workflow, credential, API key,
    notification channel, webhook, schedule, MCP server, OpenAPI import, exposed tool or a copy of a
    template is a guided flow: steps tick off from what the form actually holds, "All fields" shows
    the same form at once, and a review lists what blocks saving and what is worth knowing. Unsent
    drafts of new items are kept in the browser tab (never their secrets).
  • Monitoring pages explain how to read them: the Overview's insights, a run and its trace, a human
    task ("Before you answer"), an evaluation report, a knowledge source's test search.
  • Publishing reviews the draft's problems first; nothing is deployed unless ticked. Replay from
    the runs list asks before starting. Starting an evaluation or creating a web source happens
    only from its labelled button, not from Enter.
  • Fixes found on the way: API key example requests include the environment, correct external
    secret reference formats, credentials can be limited to chosen workflows, MCP tool policies can
    be edited, tool lists refresh after discovery or import, and evaluation gate labels are right.
• Pickers for steps that point at something in the workspace.
  • The Agent step's preset is chosen by name from the workspace's agent presets instead of typed
    as an id.
  • MCP tool and MCP prompt steps list the connected servers, then the chosen server's tools or
    prompts; OpenAPI steps list the imported toolsets, then the chosen toolset's operations. A
    list that depends on another field says to choose that one first, and a list that fails to
    load says why and can be refreshed.

Images

• ghcr.io/morohn/flowaid-api:0.7.0
• ghcr.io/morohn/flowaid-worker:0.7.0
• ghcr.io/morohn/flowaid-web:0.7.0

Verify with gh attestation verify oci://ghcr.io/morohn/flowaid-api:0.7.0 --owner MoRohn.