Operations¶

Open this section when you need to run agent work repeatably: install it, exercise workflows, diagnose trace drift, release it, or recover from failure without relying on who last touched the package.

Operating Loop¶

flowchart LR
    setup["setup"]
    run["run workflow"]
    inspect["inspect trace"]
    recover["recover execution"]
    release["release package"]
    proof["tests and artifacts"]

    setup --> run --> inspect --> recover --> release
    run --> proof
    inspect --> proof

Agent operations should make workflow behavior teachable from the repository. A maintainer needs one path to execute a flow, inspect the emitted trace, and recover when orchestration drifts, without depending on whoever most recently debugged the package.

Read These First¶

open Installation and Setup first when you need a clean package starting point
open Observability and Diagnostics when workflow traces or outputs no longer match expectation
open Failure Recovery when workflow execution has already gone wrong

Operational Risk¶

The main operational risk here is letting orchestration succeed only for people who already know the unwritten workflow path.

First Proof Check¶

pyproject.toml, README.md, and package-local entrypoints for checked-in operating truth
tests and runnable workflows for evidence that the package can be operated repeatably
release notes and version metadata when the work changes caller expectations

Pages In This Section¶

Leave This Section When¶

leave for Interfaces when the live problem is contract shape rather than package operation
leave for Architecture when a workflow problem exposes structural drift underneath it
leave for Quality when the package runs but the real question is whether the evidence is strong enough

Design Pressure¶

If workflow recovery depends on tribal knowledge about trace interpretation, the package is not operationally ready. This section has to turn agent execution into a repeatable, inspectable routine.