generated: '2026-09-19'
method: searched
source: >-
  openapi/luthersystems-com-agentsearch-openapi.yml, https://agentsearch.luthersystems.com/api-docs and
  /llms.txt, the AgentSearch agent card (pricing.rateLimit), the live InsideOut tools/list descriptions
  (mcp/luthersystems-com-insideout-mcp-tools.json — tfplan, tfdeploy, tfdestroy, stackrollback,
  stackversions, tflogs) and the InsideOut SKILL.md / SECURITY.md.
description: >-
  Two surfaces with different semantics. AgentSearch is a read-only, anonymous HTTP/MCP/A2A API — no
  writes to protect, page/page_size pagination, undocumented error envelope beyond one 404, best-effort
  per-IP rate limiting with no documented headers. InsideOut is a session-scoped MCP surface whose writes
  are Terraform jobs: it has a first-class dry run (tfplan -> plan_id -> tfdeploy), a single-flight rule
  per session, versioned stack designs with rollback, and a destroy operation — but it documents no
  idempotency key and no time window for any reversal.
cross_links:
  authentication: authentication/luthersystems-com-authentication.yml
  errors: errors/luthersystems-com-problem-types.yml
  lifecycle: lifecycle/luthersystems-com-lifecycle.yml
  rate_limits: rate-limits/luthersystems-com-rate-limits.yml
  mcp: mcp/luthersystems-com-mcp.yml
auth:
  agentsearch: none (public)
  insideout: 'session token embedded in session_id (from convoopen); browser-based cloud credential connection for deployment'
pagination:
  style: offset (page / page_size)
  params: [page, page_size]
  defaults: 'page_size 100 (llms.txt); browse tool default page 1'
  response_fields: 'page of AgentSummary records plus pagination metadata (legacy card text); not schematised in the OpenAPI'
  streaming: 'tflogs cursors with last_event_id / tail on InsideOut'
field_expansion: none
metadata: none
request_id: not documented (no X-Request-Id or trace header described)
versioning:
  agentsearch: unversioned paths, product version 0.1.0
  insideout: '/v1/insideout-mcp path; A2A v0 (0.3) and root (1.0) endpoints coexist'
error_envelope:
  http: 'undocumented beyond 404 on getAgent (no schema)'
  jsonrpc: 'JSON-RPC 2.0 error {code, message, data[google.rpc.ErrorInfo]} on the InsideOut A2A endpoint; MCP errors as JSON-RPC errors or tool-level codes (tf_job_conflict, auth_required)'
rate_limit_signaling:
  documented: '~60 requests per minute per IP, best effort (agent card pricing.rateLimit); api-docs: "no rate limit beyond best-effort fairness"'
  headers: not documented
  status_on_exhaustion: not documented
idempotency:
  coverage: none
  header: null
  retention: null
  scope: []
  note: >-
    AgentSearch exposes no mutating operation in its OpenAPI or tool list (POST /api/search and
    /api/found-agent compute and return; the only write, POST /api/submit-agent, is documented outside the
    spec with no idempotency mechanism). InsideOut's mutating tools (convoopen, tfgenerate, tfdeploy,
    tfdestroy, stackrollback, submit_feedback) document no idempotency key; the closest mechanism is the
    single-flight rule — a second tfdeploy/tfplan/tfdestroy while a job is in flight returns
    tf_job_conflict with the live job_id instead of starting a duplicate, and force_new=true overrides it.
    That is duplicate suppression for one job at a time, not replay protection, so coverage is none.
dry_run:
  status: documented
  mechanism: 'tfplan runs terraform plan without applying; returns job_id, plan_id, project_id; tfdeploy accepts plan_id to apply exactly that plan'
  scope: [tfdeploy]
  evidence: live tfplan and tfdeploy tool descriptions
reversibility:
  status: documented
  note: >-
    Reversal paths exist for the two InsideOut write surfaces but NO time window is stated anywhere, so
    the grade is documented (reversal path) rather than verified (path + window). AgentSearch is read-only
    (na). Enterprise-platform reversals (Terraform state, phylum migrations) are customer-operated and
    not on a public surface.
  surfaces:
  - write: tfdeploy (deploy infrastructure to the user's AWS/GCP account)
    reversal: tfdestroy
    reversal_operation: tfdestroy
    window: not stated
    conditions: 'session must have a prior successful deployment; single-flight rule applies; long-running job monitored via tfstatus/tflogs'
    docs: https://github.com/luthersystems/insideout-agent-skills/blob/main/SKILL.md
  - write: convoreply / tfgenerate (advance or change the stack design)
    reversal: stackrollback
    reversal_operation: stackrollback
    window: not stated
    conditions: 'creates a draft version copying a previous version''s components/config/pricing (single-draft rule); use stackversions to list targets'
    docs: live tool description (mcp/luthersystems-com-insideout-mcp-tools.json)
  - write: tfplan
    reversal: na
    note: A plan applies nothing; nothing to reverse.
  - write: submit_feedback / POST /api/submit-agent
    reversal: none documented
  - write: AgentSearch reads
    reversal: na
    note: Read-only API.