> ## Documentation Index
> Fetch the complete documentation index at: https://headgate.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Retry only a workflow's failed subgraph



## OpenAPI

````yaml /api/headgate.openapi.yaml post /workflows/{id}/retry
openapi: 3.1.0
info:
  title: headgate control API
  version: 0.1.0
  description: >
    The web UI is one client of this control API and gets no

    privileged access — asynqmon reads Redis directly, which is why its
    compatibility

    note is three minor versions stale.


    Both the Go and Rust implementations serve this spec, and the conformance
    suite

    asserts identical responses. Every list endpoint is bounded to prevent
    inspection

    from becoming an unbounded store operation: asynq's

    GetQueueInfo is O(number of groups) and has pinned Redis CPU for seconds in

    production. Monitoring caused the outage.


    DERIVED FROM THE ARCHITECTURE, NOT FROM A PREDECESSOR'S UI. The first draft
    of this

    spec was the asynq Inspector surface -- list by state, act on one job, pause
    a queue

    -- with new nouns bolted on. That inherited three problems: it omitted bulk

    operations, history, and enqueue (which asynq and apalis-board respectively
    DO have),

    and more importantly it had no endpoint for the question this system's own
    design

    creates. When dequeue is an admission decision, the operator's first
    question is not

    "what is in the queue" but "why is THIS job not running" -- see
    /jobs/{id}/admission.

    No predecessor needs that endpoint because no predecessor has a gate.
servers:
  - url: /api/v1
security: []
paths:
  /workflows/{id}/retry:
    post:
      summary: Retry only a workflow's failed subgraph
      parameters:
        - $ref: '#/components/parameters/Id'
        - $ref: '#/components/parameters/IdempotencyKey'
      requestBody:
        required: true
        content:
          application/json:
            schema:
              type: object
              required:
                - expected_revision
              properties:
                expected_revision:
                  type: integer
                  minimum: 1
                recoveries:
                  type: array
                  maxItems: 999
                  description: >-
                    Explicit repair instructions for failed nodes that cannot be
                    retried as-is.
                  items:
                    type: object
                    required:
                      - node
                    properties:
                      node:
                        type: string
                        minLength: 1
                        maxLength: 128
                      payload:
                        type: string
                        contentEncoding: base64
                        description: Required when repairing an undecodable node.
                      schema_version:
                        type: integer
                        minimum: 1
                        description: Required when repairing an undecodable node.
                      release_quarantine:
                        type: boolean
                        default: false
                        description: Must be true to release a quarantined fingerprint.
      responses:
        '200':
          description: Accepted revision and generation
        '409':
          description: Revision conflict
components:
  parameters:
    Id:
      name: id
      in: path
      required: true
      schema:
        type: string
    IdempotencyKey:
      name: Idempotency-Key
      in: header
      required: true
      schema:
        type: string
      description: >
        Required on every mutating request. A double-clicked Retry, or a proxy
        retrying a

        POST, must not enqueue the job twice. Keys are retained long enough to
        cover a

        client retry window and the response is replayed verbatim.

````