approvals-and-policy/approvals.md.
API key (praxa_sk_) — The fabric’s tenant-scoped credential. In the live
partner preview it authorizes the admitted minimal task, run, cancellation,
SSE, and usage routes according to its endpoint scopes. Organization keys and
request-defined policy are not available. See
keys-and-access/api-keys.md.
Approval — A fail-closed pause before an action executes. The first public
form is an exact digest-bound browser action; it is source-ready and pending
production activation. Other product approval gates are not automatically
public /v1 capabilities. See
approvals-and-policy/approvals.md.
Custom-agent deployment key — The existing deployment credential scoped to a single agent. Available today, used by agent builders on the custom-agent platform, and unaffected by the fabric’s rollout: it keeps working and migrates onto the tenant model later rather than being replaced outright.
Bundled inference — A funding mode: your run executes on Praxa’s own provider keys, billed as credits against fixed per-run pricing. The default. Compare BYO.
BYO (bring your own keys) — A planned funding mode in which a tenant would
connect its own provider credentials. It is not available on the public
execution boundary. See
how-to/bring-your-own-keys.md.
Canonical payload (canonical action envelope) — The exact provider, action, and arguments recorded for a pending approval. Execution accepts only a byte-equivalent approved envelope; anything altered, expired, replayed, or attached to a different run fails closed.
Deny — A policy engine outcome: the action does not run, under any circumstance, without a policy change.
Event (run event) — One append-only, sequenced public projection of run
progress. The current durable projector emits lifecycle events; exact approval
events are source-ready pending activation. Rich tool/output events are not
currently projected. Streamed from GET /v1/runs/:id/events and resumable by
sequence. See api/events.md.
Execute — The core primitive. Today POST /v1/execute accepts a strict
minimal mode:"task" envelope. Agent, context, tool, policy, and delivery
overrides are unavailable, and mode:"turn" is unserved and rejected. See
api/execute.md.
Funding mode — How inference is funded. Bundled inference is the current
public mode; BYO is planned.
Human approval — A policy engine outcome: the action pauses until a person explicitly approves it.
Idempotency key — A value sent with /v1/execute so a retried request never executes the same task twice.
Last-Event-ID — The header used to resume a run’s event stream exactly where it left off, by sequence number, instead of replaying from the beginning or missing events in between.
Mode: task — The active execution shape: the call returns a UUID run_id
and work continues durably server-side. Track it through run reads, resumable
SSE, or automatic signed webhooks.
Mode: turn — An unserved candidate shape. It is not part of the public
execution schema: /v1/execute returns 400 invalid_request when sent this
value, and /v1/chat is not routed.
Peer review — A planned unified-policy outcome between automatic execution
and human approval. It is not configurable through the public API.
Policy — A reserved request field for future execution constraints. The
current minimal handler rejects it; authority and admission are server-derived.
Policy engine — The planned unified system for resolving action policy.
Praxa currently has multiple mature fail-closed product gates, but the unified
configurable public engine is not available. See
approvals-and-policy/policy-engine.md.
Policy pack — A planned versioned bundle of unified policy rules. Public
configuration and mature receipt linkage are not available.
Principal — The identity behind a call: an API key, a signed-in user, or a desktop session. One resolver handles all three; downstream authority is derived the same way no matter which kind made the call.
Receipt — A planned mature evidence record covering execution, policy,
approvals, and verification. Current run/event projections and internal task
events are narrower and are not that public receipt product. See
security/receipts-and-audit.md.
Reservation — A hold placed on funds or budget before a paid step executes, so a budget ceiling stops new work before it happens instead of producing a surprise charge afterward.
Risk tier — The planned unified-policy outcome vocabulary: auto, peer
review, human approval, or deny. It is not a current request override.
Run — One durable task admitted through /v1/execute and identified by a
UUID run_id, followed through a terminal projection. It does not currently
guarantee a mature receipt.
Scope — Which endpoint classes a key can call. The exact current values are
execute:write, runs:read, runs:write, and usage:read; there are no
tool, budget, policy, or webhook-specific scope literals. See
keys-and-access/scopes-and-tenancy.md.
Tenant — The ownership and billing boundary for keys, runs, events, and
usage. The live public execution preview is limited to personal tenants;
organization execution and organization keys are unavailable. See
keys-and-access/scopes-and-tenancy.md.
Webhook delivery — An HMAC-signed, retried partner-preview notification of
a run event. Automatic task-event projection and bearer endpoint/delivery
management are active for admitted personal tenants. See
how-to/receive-webhooks.md.