Pipeline

[MODE:SIDEEFFECTS] Tier: meta. The most powerful Fibe tool.

Executes an ordered list of tool steps in one MCP request. Each step's output is normalized to plain JSON and stored in bindings keyed by step.id; later steps reference prior outputs with JSONPath strings ("$.create_pg.id").

When to use

• Create+wait+link greenfield flow.
• Roll out a template change and immediately wait for running.
• Bulk operations (for_each across a list returned by a prior step).
• Eliminating round-trip latency — saves 100s of ms per chained step.
• Atomic-ish flows where mid-failure must surface partial results for cleanup.

Step shapes

Single tool call:

{ "id": "pg", "tool": "fibe_resource_get", "args": { "resource": "playground", "id": 123 } }

Parallel block — independent steps run concurrently:

{ "parallel": [
  { "id": "a", "tool": "fibe_resource_list", "args": {"resource":"prop"} },
  { "id": "b", "tool": "fibe_resource_list", "args": {"resource":"marquee"} }
]}

Fanout (for_each) — iterate a list with as alias:

{
  "id": "rolls",
  "for_each": "$.list.data",
  "as": "pg",
  "steps": [
    { "id": "rollout", "tool": "fibe_playgrounds_action",
      "args": { "playground_id": "$.pg.id", "action_type": "rollout", "confirm": true } }
  ],
  "collect": "$.rollout"
}

JSONPath bindings

• "$.step_id" — entire step output.
• "$.step_id.field" — specific field.
• "$$." — escapes to literal "$.".
• Bindings live in a shared map; nested objects/arrays are walked recursively.

Top-level inputs

Field	Type	Default	Notes
steps	array	required	At least 1; max 25 by default (FIBE_MCP_PIPELINE_MAX_STEPS)
return	string | object	bindings map	JSONPath or object literal projecting the final return
dry_run	bool	false	Validates refs+schemas without running
cache	bool	true	Whether to cache for fibe_pipeline_result
idempotency_key	string	—	Threaded as a per-step idempotency-key (sha256 of key:step_id); Rails caches responses 24h

Output

{
  "steps": { "create": {...}, "wait": {...} },
  "result": <projected via "return">,
  "status": "completed",
  "pipeline_id": "<id>",        // present unless cache=false
  "truncated": false             // when cached payload exceeded buffer
}

On failure mid-run:

{
  "steps": { "create": {...} },                // partial bindings
  "result": null,
  "status": "partial",
  "error": { "step_index": 1, "step_id": "wait", "tool": "...", "message": "...", "code": "...", "status": 422 },
  "completed_step_ids": ["create"],
  "pipeline_id": "..."
}

completed_step_ids is your cleanup hint — those resources exist; the failed step never started its mutation.

Step-level options

Field	Notes
on_error	"abort" (default) or "continue" (record {error:...} in bindings, keep going)
input_path	JSONPath that filters resolved args before calling
output_path	JSONPath that projects the result before storing in bindings

Per-step idempotency

With pipeline-level idempotency_key, every step ID gets sha256("<key>:<step_id>") as its Idempotency-Key header. Retrying the whole pipeline does NOT recreate completed resources — Rails returns the cached prior response.

Gotchas

• No nested pipelines. A step calling fibe_pipeline rejects with "nested fibe_pipeline is not allowed".
• for_each requires both as and steps; the value must resolve to a JSON array (not an object).
• Maximum total for_each iterations across a pipeline: FIBE_MCP_PIPELINE_MAX_ITERATIONS (default 50).
• Destructive tools still need confirm:true even inside a pipeline (unless server is --yolo).
• Step output is JSON-marshaled before bindings — typed Go structs become plain maps. Field names follow JSON tags, not Go field names.
• Steps without an id produce no binding (useful for fire-and-forget side effects within a parallel block).

Recipes

Create greenfield, wait, get URL:

{
  "steps": [
    { "id": "gf",   "tool": "fibe_greenfield_create",
      "args": { "name": "demo", "wait_timeout": "10m" } },
    { "id": "urls", "tool": "fibe_local_playgrounds_info",
      "args": { "view": "urls", "playground_id": "$.gf.playground.id" } }
  ],
  "return": "$.urls"
}

Roll out every running playground in a marquee:

{
  "steps": [
    { "id": "list", "tool": "fibe_resource_list",
      "args": { "resource":"playground", "params":{"marquee_id":42,"status":"running"} } },
    { "id": "rolls", "for_each":"$.list.data", "as":"pg",
      "steps":[{"id":"r","tool":"fibe_playgrounds_action",
                 "args":{"playground_id":"$.pg.id","action_type":"rollout","confirm":true}}],
      "collect": "$.r"
    }
  ]
}

Related

• fibe_pipeline_result — re-query cached results within 5 min.
• fibe_call — single hidden-tool invocation; pipelines wrap many.