docs/PIPELINE.en.md
Rama main · réplica de lectura del repositorio original.
# Runtime Pipeline
The runtime pipeline explains how one job moves through BailingHub from trigger to result delivery.
## 1. Trigger
A job can enter through:
- `POST /run`;
- web chat widget;
- inbound channel adapter;
- controlled internal APIs.
The entry layer authenticates the client or channel, validates payload shape, applies idempotency, and resolves the route.
## 2. Route Resolution
The route determines:
- target model or executor;
- model credential;
- memory and session policy;
- knowledge base and retrieval settings;
- allowed tool providers and scopes;
- budget and rate limits;
- callback or delivery policy.
`route=auto` can be used when the client is allowed to let the hub select a route from configured audience and keyword rules.
## 3. Context Assembly
The hub assembles:
- current input;
- message ledger;
- optional rolling summary;
- route-level knowledge retrieval;
- page context;
- principal and metadata;
- tool definitions visible to the route.
Untrusted user input and page context are wrapped as data, not executed as instructions.
## 4. Dispatch
The job is dispatched to a target:
- in-hub OpenAI-compatible model target;
- external executor target;
- future adapters that implement the same target contract.
Jobs are claimed through the persistent queue and lease model so multi-process deployments can recover abandoned work.
## 5. Tool Invocation
When the model requests a tool call, the tool runtime checks:
- route allowlist;
- tool scope;
- subject requirement;
- risk level;
- rate limits;
- approval or confirmation rules;
- authorization probe result when configured.
Allowed tool calls are signed and sent to the business system. The business system verifies the signature and performs final authorization.
## 6. Approval
High-risk or confirmation-required calls are frozen as approval intents. In production, the business-side workflow should usually own the actual approval decision. The hub records the approval intent and only replays the exact approved argument snapshot.
## 7. Audit, Trace, And Delivery
The hub records trace events, audit records, approvals, messages, delivery attempts, and final result. The business system can poll job status or receive signed callbacks.