Plain-language model for AURA Harness.
A logical entity you run under AURA.
| Field | Role |
|---|---|
agent_ref |
Stable slug for humans and CI, e.g. acme/compliance-bot |
aura_id |
Internal ULID (default) or your supplied id |
name |
Optional alias for lookup |
ids |
Trailer for tenant, Skillware, external ids, and optional operator (manual or verified adapter) |
Legacy profiles with AURA-000n ids still load. See trust-paths.md and integrations/identity/README.md for optional verified operator adapters.
Profile fields also include skills, sequencer spec, observers, rules, optional spectrum (enforcement posture — see aura-levels.md), and optional capabilities[] (named intents + secret refs — see capabilities.md).
One run of an agent. Opens (with ingress), records events, closes, exports logs.
| Mode | When to use |
|---|---|
script |
Starts and ends in one go (default) |
task |
Ends when you call complete_goal() |
continuous |
Long-running until error or manual stop |
The runtime coat around your body — not the loop itself.
| Boundary | Meaning |
|---|---|
| Ingress | Context normalized at session open |
| Body | Your host loop (script, Skillware, framework) |
| Egress | Policy + audit before tools execute |
Anything that happens during a session: turn.start, tool.call, sequencer.step.start, membrane.ingress, etc.
Every event is appended to the audit trail with causal links (event_id, parent_id, trace_id, optional step_id).
The live, append-only record of a session. Official name for what the code calls the audit spine. Written to JSONL as events occur.
What you get when a session closes: JSONL audit file + conformance summary JSON (+ OTel JSONL by default). Ship to logs, observability, or storage.
With export=False, no files are written, but run.summary and run.audit_report are still built in memory when the context exits. Summary and OTel files commit atomically on disk export; a failed export leaves neither artifact updated.
After close, the session is sealed — further emit or approve calls raise SessionClosedError.
Rules, guardrails, and constraints the run must obey — on the agent profile, in YAML, or from adapters. Enforced during the run; checked again on close (conformance).
A constraint checked when relevant events are emitted.
Built-in types: max_tokens_per_step, confirm_before, allow_tools, deny_tools, sequencer_required, escalation_pause, capability_scope, schema_check.
Profile variables.schema_refs holds named JSON Schema specs; at spectrum high / full they auto-wire as schema_check rules unless spectrum.schema_enforcement: false. Mid profiles use explicit schema_check rules only.
On session close, AURA compares declared rules at open (and sequencer step order) vs observed events and writes a summary. The open-time open_snapshot_hash in the summary matches conformance when base rules are unchanged; runtime skill binds may update snapshot_hash for live constraint checks.
Prescriptive multi-step pipeline inside a session — not emergent model tool chaining. You declare steps upfront; AURA enforces order on close.
Parallel subscriber to the audit trail. Must not block the host. Used for metrics, alerts, or custom analytics.
How the body executes (Python script, Skillware host, future: LangGraph). AURA wraps; it does not own the loop.
Default: ~/.aura/ (override with AURA_HOME). Project-local: set storage: project in aura.project.yaml → .aura/ in project root.
→ onboarding.md — storage, identity, and first run · glossary.md