ArchieQueryStepDefinition

Archie-query step — runs a natural-language query through the V4 chat orchestrator and threads its answer into the workstream. Stuart 2026-05-19: 48 templates on dev use ``type='archie_query'`` today (most-common step type in the DB) but the write-side schema rejected it as 422. Adding the model unblocks PATCH for every workstream that mixes archie_query with other types — the read shape stored in JSONB is already what the V4 chat-orch executor consumes via ``step.get("instruction") / step.get("expected_outputs")``.

Field	Type	Description
`id`required	`string`	—
`name`required	`string`	—
`type`	`string`	—
`description`	`string \| null`	—
`depends_on`	`string \| object[]`	—
`stage`	`string \| null`	—
`timeout`	`string \| null`	—
`retry_policy`	`RetryPolicy \| null`	—
`error_handler`	`StepErrorHandler \| null`	—
`impact_level`	`"low" \| "medium" \| "high" \| "critical"`	Impact classification for steps. `lowmediumhighcritical`
`run_mode`	`"propose" \| "auto_run"`	Execution mode for steps. `proposeauto_run`
`requires_approval`	`boolean`	—
`metadata`	`object \| null`	—
`examples`	`StepExample[] \| null`	Attached example files for this step. Each entry pairs a platform.file_uploads.file_id with a role + tags + notes so the runtime can shape its instruction. Bounded at 20 to keep a reasonable upper limit on the prompt preamble.
`tags`	`string[] \| null`	Categorical labels on the step itself (e.g. 'pdf-read', 'coutts', 'fy25'). Free-form; recommended tag set is documented per workstream surface.
`instruction`required	`string`	Natural-language query the runtime sends to the chat orchestrator.
`inputs`	`object \| null`	Optional Jinja2 input bindings ({{ inputs.* }} / {{ outputs.* }}).
`expected_outputs`	`object[] \| null`	Output schema hints (name + type per entry).
`skill_slugs`	`string[] \| null`	Skills that should be available to the chat orchestrator for this query.
`confidence_threshold`	`number \| null`	Minimum confidence the runtime should treat as acceptable before auto-continuing (drives confidence-gated review).
`tool_hints`	`string[] \| null`	V4 tool slugs to bias the orchestrator toward (e.g. 'read_pdf_document').
`deterministic`	`boolean \| null`	When true, run this step's specialist call at temperature=0 with extended thinking disabled so the same input is extracted identically every run. For structured-extraction steps only.