Workflow

workflow exposes the workflow engine itself as a connector. Every operation in the workflow design (Tier 1 / 2 / 3) is reachable via wick_execute so any MCP-aware LLM — Claude Desktop, ChatGPT, Gemini — can introspect node types, scaffold a graph, edit files, validate, simulate, run tests, replay, and request review without native file access.

This is the LLM-facing surface for the Workflows feature. If you're building a workflow in the canvas, you don't touch these directly — the canvas calls them under the hood. If you're driving wick from outside (Claude Desktop, a custom MCP client), these are the ops you'll see in wick_list.

Source	internal/connectors/workflow/
Key	workflow
Icon	⚙
Tier	runtime (wired with the workflow engine at boot)
Fixed	✅ — single row, no per-instance credentials

Configs

Empty (type Configs struct{}). The connector talks to the in-process workflow engine, not an external API. The row exists so it appears in /manager/connectors/workflow/{id} for tag-based access control + audit log.

Operation catalog

Operations are grouped by tier — Tier 1 is read-only introspection (every LLM can call freely), Tier 2 is graph editing, Tier 3 is running + tests. Destructive ops are flagged.

Tier 1 — introspection (read-only)

Op	Purpose
workflow_workspace	Entry point. Returns {base_dir, node_types[], trigger_types[], templates[]}. Always call this first to orient the LLM.
workflow_node_types	All node types with schema + example YAML + when_to_use.
workflow_trigger_types	All trigger types with schema + example.
workflow_channels	Configured channel integrations + trigger/action schemas (legacy specs).
workflow_integration	Full per-channel event + action catalog from the integration registry — match_schema, payload_schema, input_schema / output_schema per action, destructive flag. More complete than workflow_channels.
workflow_node_detail	Per-type detail — input schema, output fields, examples, pair-with hints, common pitfalls.
workflow_connectors	Every connector module + its operations (for type:connector nodes).
workflow_skills	AI provider skills available for type:agent nodes (optional provider filter).
workflow_providers	Configured AI providers (claude / codex / gemini) with capabilities + status.
workflow_list	Every workflow with id, name, enabled, version.
workflow_check_name	Returns {available, conflict_id}. Call before workflow_create so the LLM surfaces a friendly error.
workflow_get	Full workflow definition: triggers, graph, env schema.
workflow_list_files	Files in the workflow folder (workflow.yaml, nodes/*.md, __tests__/, …).
workflow_read_file	Content of one file. Replaces native file tool for remote AI.

Tier 2 — editing (mostly idempotent, some destructive)

Op	Destructive	Purpose
workflow_create	no	Scaffold a new workflow from a template (empty, support-triage, incident-response, daily-digest). New workflows are disabled — admin enables explicitly.
workflow_write_file	no	Write any file in the workflow folder.
workflow_delete_file	yes	Delete one file.
workflow_delete	yes	Delete the whole workflow folder + unregister scheduled triggers.
workflow_add_node	no	Add a node via declarative patch — validates type + schema.
workflow_update_node	no	Merge-patch one node's fields (prompt, config, on_failure, …).
workflow_delete_node	yes	Remove a node and every edge that references it.
workflow_connect	no	Add an edge. Pass case label for classify / branch / switch sources.
workflow_disconnect	no	Remove an edge.
workflow_move_node	no	Update canvas position (x, y pixels). No effect on execution.
workflow_set_triggers	no	Replace the trigger list.
workflow_toggle	no	Enable / disable. Disabled workflows skip cron/channel/webhook but still respond to workflow_run_now.
workflow_publish	no	Promote workflow.draft.yaml → workflow.yaml and re-register the workflow with the router. Required after every edit. The connector description tells the LLM to ALWAYS ask the user before publishing edits.
workflow_discard_draft	no	Throw away the draft and revert to the published version.
workflow_has_draft	no	{has_draft: bool}.

Tier 3 — running + tests + diagnostics

Op	Destructive	Purpose
workflow_validate	no	Parse + validate: cycle detect, schema check, guard dry-run. Errors decorated with did_you_mean / hint pointers (lowercase key, misspelt match field, picker scalar vs object shape, …).
workflow_template_test	no	Render a Go template against a synthetic context. On missing-key errors, lists available keys + did-you-mean. Sample events: slack.message, slack.block_action, slack.view_submission, cron.
workflow_picker_resolve	no	Resolve a picker source (e.g. slack.channels) to [{id, name}]. Use when populating Match filter picker fields so AI passes valid IDs instead of guessing.
workflow_describe	no	Human-readable summary: triggers, graph shape, dependencies (channels/connectors/providers), dangling-edge + template-reference warnings. Call before editing to orient the LLM.
workflow_simulate	no	Dry-run with a synthetic event. No state persisted, no external calls. Returns per-node outputs + path_taken + final_result.
workflow_test	no	Run __tests__/ fixtures (optional name-prefix filter). Returns [{case, pass, error, diff}] + coverage summary. Requires runner-enabled module.
workflow_test_coverage	no	Which nodes were hit and which are untested across all fixtures.
workflow_record_test	no	Generate a __tests__/ fixture by capturing a real run's event + per-node outputs.
workflow_capture_fixture	no	Snapshot one node's output as a unit-test fixture in __tests__/nodes/.
workflow_run_now	no	Enqueue a manual run (bypasses Enabled check). Returns {run_id}.
workflow_get_runs	no	Recent run IDs + status + started_at (default limit 20).
workflow_get_run	no	Full run state — node outputs, events, path_taken, status, cost.
workflow_get_run_events	no	Raw events.jsonl stream for a run — every node_started / node_completed / node_failed / edge_traversed entry.
workflow_watch	no	Bounded read over recent runs. wait_seconds > 0 subscribes to the live event stream and returns when expect / stop_on_first is met (server caps at 30s). Multi-dim filter — workflow_id, trigger_id, node_id, status, since.
workflow_get_run_log	no	One-shot summary of a run — status, error, completed/failed/skipped nodes, per-node duration, total duration. Pass diagnose=true to classify the error (template_missing_key, channel_action_missing, connector_op_missing, secret_leak, branch_no_edge, agent_session_invalid, provider_skill_missing) and surface available keys + suggested fix.
workflow_copy_run_to_editor	no	UI parity with the Copy to editor button — loads run state, saves the current published workflow as draft, writes runs/<run_id>/mocks.json so the Execute step can prefill from real data.
workflow_replay_run	no	Re-enqueue a run with the same trigger event as a past run.
workflow_list_test_cases	no	__tests__/*.json fixtures with name + assertion count + last result.
workflow_save_test_case	no	Create or update one __tests__/<name>.json fixture (slug-safe names only).
workflow_delete_test_case	yes	Delete one fixture.
workflow_request_review	no	Notify admin that the workflow is ready for approval. Workflow stays disabled until admin enables it. Returns {url}.

Authoring flow from MCP

Recommended sequence when an LLM creates a workflow from scratch:

workflow_workspace
  → workflow_node_types
  → workflow_check_name        ← surface conflicts up front
  → workflow_create
  → workflow_add_node × N
  → workflow_connect × N
  → workflow_validate          ← static + schema checks
  → workflow_simulate          ← dry-run with synthetic event
  → workflow_test              ← if __tests__/ fixtures exist
  → workflow_request_review    ← admin enables manually

After every edit (workflow_add_node, workflow_connect, workflow_write_file workflow.yaml, …) the change lands in workflow.draft.yaml. Call workflow_publish to promote — and always ask the user first.

See also

• Workflows — the user-facing feature this connector drives.
• Connector Module — the contract this connector reuses.
• Wick Manager — sibling connector for inspecting jobs / tools / lifecycle.