Skip to main content

Profile — .omao/profile.yaml

One per project, created by oma setup. All hooks, skills, and validation tools treat this file as single source of truth.

Format (v1)​

version: 1
created_at: "2026-04-30T02:00:00Z"

harness:
  primary: claude-code        # claude-code | kiro
  secondary: null             # or "kiro" / "claude-code"

aws:
  account_id: "123456789012"
  region: ap-northeast-2
  profile_name: default
  environment: sandbox        # sandbox | staging | prod

aidlc:
  entry_phase: inception      # inception | construction | operations
  strict_gates: false

approval:
  mode: interactive           # interactive | ci-auto-approve-safe | strict
  blast_radius_ceiling: single-account

budgets:
  default_monthly_usd: 200
  warn_at_pct: 80
  block_at_pct: 100

observability:
  mode: none                  # none (default, opt-in) | opentelemetry-only | langfuse-self-hosted | langfuse-managed
  endpoint: null
  trace_mcp: null             # { server_name: "langfuse", tools: ["mcp__langfuse__*"] } — trace-reading MCP server (for self-improving/continuous-eval/cost-governance feedback loops)

star_confirmed: false

Validation​

  • Schema: schemas/profile/profile.schema.json
  • oma setup always calls profile_validate immediately after write — installation does not complete with an invalid profile.
  • oma doctor's profile-valid probe re-validates per session.

Field effects​

  • harness.primary/secondary — Determines which install scripts to run.
  • aws.* — Seed budget's scope_ref, MCP server's region context.
  • aidlc.entry_phase — Phase where first /oma:autopilot enters.
  • approval.mode — ci-auto-approve-safe auto-approves within single-namespace blast radius. Otherwise human-in-the-loop.
  • approval.blast_radius_ceiling — If exceeded, enforce human approval + secondary review.
  • budgets.* — Seed .omao/ontology/budgets/default.json, and simultaneously set budget warning threshold for user-prompt-submit.sh.
  • observability.* — Opt-in; defaults to none. The ontology + harness core needs no backend. Set opentelemetry-only for a vendor-neutral OTLP target, or langfuse-self-hosted / langfuse-managed to use Langfuse (the ai-infra langfuse-observability skill then reuses the endpoint field).
  • observability.trace_mcp — (optional) Registers the trace-reading MCP server that agenticops feedback-loop skills (self-improving-loop, continuous-eval, cost-governance) will call. null disables trace-based feedback. Example: { server_name: "langfuse", tools: ["mcp__langfuse__get_traces", "mcp__langfuse__get_sessions"] }.

Manual editing​

  • Editable without re-running oma setup. However, re-validate with oma doctor after editing.
  • Empty values (null) may use safe defaults or error depending on field. See schema's required / default clauses.

Rationale for defaults​

  • Monthly budget $200 — Observed median cost for single developer using Claude Code + Claude Sonnet full-day. Scale up for teams.
  • blast_radius_ceiling: single-account — By default, cross-account or cross-region deployments route through human approval to control blast radius.
  • approval.mode: interactive — Safest for initial adoption. Recommend promotion to ci-auto-approve-safe only in CI.