MCP authoring

Every workflow primitive is reachable from an LLM via the wick MCP server. The catalog is self-documenting: every executor declares its schema next to its Execute method, so adding a new node type makes it appear in workflow_node_types and the canvas palette automatically — no separate registration.

The full op catalog with destructive flags and per-op semantics is the Workflow connector page. This page covers the authoring flow — how to drive that surface end-to-end.

Recommended sequence

When an LLM creates a workflow from scratch:

workflow_workspace            ← entry point, returns base_dir + node_types + trigger_types + templates
  → workflow_node_types        ← know what types are available
  → workflow_check_name        ← surface conflicts up front
  → workflow_create            ← scaffold from template
  → workflow_add_node × N      ← validates type + schema per node
  → workflow_connect × N       ← add edges, with case: for branch sources
  → workflow_validate          ← static + schema + cycle + guard dry-run
  → 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_set_triggers, workflow_update_node, …) the change lands in the workflow's draft slot. Call workflow_publish to promote — and always ask the user first.

Two ways to edit the same workflow

The canvas and MCP write through the same engine validator. The LLM scaffolds a graph over MCP, you tighten it in the canvas, both go through the validator before the body lands in the draft slot of the workflows table.

Canvas ─┐
         ├─→ engine validator ─→ body_draft (DB) ─→ publish ─→ body_published (DB)
MCP op ─┘

No separate "AI mode" / "canvas mode". The validator is the source of truth; both paths fail the same way on a broken edge or schema mismatch. Every save also appends a row to workflow_versions for history + compare + restore.

Discovery ops

These are the read-only ops the LLM should call before any edit, in this order:

Op	Why
workflow_workspace	One call returns base directory, every node type, every trigger type, every workflow template. Most efficient way to orient.
workflow_node_types	Full schema + example body + when_to_use per node type — what the LLM needs before workflow_add_node.
workflow_trigger_types	Same for triggers.
workflow_integration	Per-channel event + action catalog with match_schema / payload_schema / input_schema / output_schema. Use to know exact filter shapes for trigger.match and exact arg shapes for channel nodes.
workflow_connectors	Every connector module + its operations — needed for connector nodes.
workflow_skills, workflow_providers	Available skills + provider instances for agent nodes.

Diagnostics ops

Op	Use
workflow_describe	Human-readable summary: triggers, graph shape, deps, dangling-edge + template-reference warnings. Call before editing to orient on an existing workflow.
workflow_validate	Parse + validate. Errors decorated with did_you_mean / hint pointers when wick recognises the failure (lowercase key, misspelt match field, picker scalar vs object shape).
workflow_template_test	Render a Go template against a synthetic context. On missing-key errors, lists available keys at the offending path plus did-you-mean. Sample events: slack.message, slack.block_action, slack.view_submission, cron.
workflow_picker_resolve	Resolve a picker source (e.g. slack.channels) to [{id, name}]. Use when populating Match filter picker fields so the LLM passes valid IDs instead of guessing.
workflow_simulate	Dry-run with a synthetic event. No state persisted, no external calls. Returns per-node outputs + path_taken + final_result.
workflow_get_run_log (with diagnose=true)	One-shot summary of a failed run — classifies the error (template_missing_key, channel_action_missing, connector_op_missing, secret_leak, branch_no_edge, agent_session_invalid, provider_skill_missing) and surfaces available keys + suggested fix.

Test fixtures

Test cases live in the workflow_test_cases table — each case has a slug-safe name, a JSON body with the pinned event + per-node assertions. Author through these ops:

• workflow_save_test_case — create or update one case by name.
• workflow_list_test_cases — list case names + assertion counts.
• workflow_record_test — capture a real run's event + per-node outputs into a fresh case.
• workflow_capture_fixture — snapshot one node's output as a unit-test case.

The workflow_test runner re-executes the graph against the recorded payload and reports [{case, pass, error, diff}] plus coverage.

History, compare, restore

Every save appends to workflow_versions. AI surface:

• workflow_versions — list snapshots newest first; metadata only (kind, message, created_at, by).
• workflow_version_detail — fetch one snapshot's full body.
• workflow_diff_versions — return two snapshots' bodies side-by-side for diff rendering.
• workflow_restore_version — copy a snapshot back into the draft slot (no auto-publish).

Canvas lock + guard

• workflow_lock — freeze edits on a workflow. The engine ignores the flag (runs continue); every edit op rejects the write while locked. Use for production workflows.
• workflow_guard — separate from workflow_validate. Runs the safety rule pack (destructive shell, secret leak, unparameterized SQL, network allowlist, prompt injection patterns) and returns {ok, violations[], content_hash}. Call before workflow_publish on anything that touches data egress.

Execute step

• workflow_exec_node — run one node in isolation (n8n's "Execute step" pattern). Pass the node JSON, optional input, optional event, optional node_outputs map keyed by upstream node id. Returns {ok, latency_ms, output}. Nothing persists to runs/.

Watching live runs

workflow_watch is the streaming side: bounded read over recent runs returning only [run_id, workflow_id, status, started_at, ended_at, trigger_id]. With wait_seconds > 0 (server caps at 30s) it subscribes to the live event stream and returns the moment expect / stop_on_first is met — cheaper than workflow_get_run for the "what happened on the last cron tick?" pattern.

Multi-dim filter: workflow_id, trigger_id, node_id, status, since.

See also

• Built-in Workflow connector — full op catalog with destructive flags + per-op input details.
• MCP for LLMs — transport layer (wick_list / wick_execute).
• Canvas editor — visual counterpart to the MCP authoring flow.