Skip to content

Workflow recipes and resources

Besides tools, the Orcho MCP server publishes two discovery surfaces:

  • Workflow recipes — published prompts that guide a client through a multi-step job. The server does not execute them: the client reads the steps and calls the tools itself, in order, with its own judgment at each gate.
  • orcho:// resources — addressable, read-only artifacts. A client can fetch one slice of run state by URI instead of calling a tool and parsing a full snapshot.

Both exist so a client does not have to reverse-engineer multi-call workflows from individual tool descriptions.

If you have not configured MCP yet, start with Start with MCP. For the tool-by-tool lifecycle surface these recipes compose, read MCP control surface.

Each recipe is an MCP prompt. Clients that support prompts (Claude Code, Cursor, and others) surface them as slash commands; invoking one injects the step-by-step guidance into the conversation.

RecipeWhat it walks the client through
orcho_getting_startedFirst-contact walkthrough: workspace check, profile choice, first run, watching progress, the QA gate flow, and reading final evidence.
orcho_plan_then_implementTwo-step pipeline: spawn a planning-profile run, review the parsed plan at the pause, then spawn a from_run_plan child run on the feature profile.
orcho_followup_from_planSpawn an implementation run that inherits the parsed plan from an existing parent run, skipping the planning block.
orcho_review_paused_runInspect a run paused on a phase handoff and propose a decide action: continue, retry_feedback, continue_with_waiver, or halt.
orcho_halt_with_reasonTerminate a paused run via a phase-handoff halt with a structured audit-trail note explaining the reason.
orcho_resume_failed_runClassify a stopped run’s state first, then resume from checkpoint when resumable — or route to from_run_plan when it is not.
orcho_observe_active_runFollow an in-flight run with a resilient loop: short bounded watch, event-summary fallback on disconnect, reconnect from the cursor.
orcho_inspect_delivery_gateClassify a stopped run’s post-release delivery or correction gate and resolve it through orcho_delivery_decide, never by applying the retained diff by hand.

Tools-only clients discover the same catalogue without prompt support: the orcho_workflows_list tool and the orcho://workflows resource both return the recipe names, descriptions, and arguments.

Every resource is a read-only projection of durable workspace or run artifacts. Most return JSON; the exceptions are noted.

URIContent
orcho://workspaceResolved workspace dir, runs dir, and recent project paths.
orcho://runsRecent pipeline runs, newest first.
orcho://runs/{run_id}/metameta.json for a single run.
orcho://runs/{run_id}/metricsmetrics.json: token counts, durations, per-phase split.
orcho://runs/{run_id}/eventsFull events.jsonl stream (newline-delimited JSON).
orcho://runs/{run_id}/summaryLatest bounded event summary for a run.
orcho://runs/{run_id}/parsed_plan.jsonDurable parsed-plan artifact.
orcho://runs/{run_id}/evidenceComposed evidence bundle for a run.
orcho://runs/{run_id}/diff.patchCaptured unified diff (text/x-patch).
orcho://runs/{run_id}/phases/{phase}/diff.patchCaptured per-phase diff (text/x-patch).
orcho://profilesPipeline profile catalogue.
orcho://profiles/{name}A single pipeline profile by name.
orcho://projects/{project_b64}/skillsAgent Skills discovered for the project (same registry as orcho_skills_list).
orcho://workflowsWorkflow recipe catalogue (the table above, as data).
orcho://docs/getting-startedThe getting-started walkthrough as markdown — same content as the orcho_getting_started prompt, for URI-based discovery.

Resources are addressable, cacheable reads. A URI names one artifact, so a client can refresh exactly the slice it cares about — re-reading orcho://runs/{run_id}/meta is cheaper than calling orcho_run_status and receiving the whole snapshot again.

Tools are parameterised queries and actions. Use a tool when you need arguments beyond the URI (an event cursor, an evidence slice name, a bounded watch), when you want typed next_actions attached to the answer, or when you are changing state rather than reading it.

A practical split:

  • polling one known artifact repeatedly — resource;
  • asking “what happened since sequence N” — tool (orcho_run_events_summary);
  • deciding, resuming, halting, delivering — always a tool.