#10415 feature: Add an "agent mode" to the CLI Issue created by edgarrmondragon Feature scope CLI (options, error messages, logging, etc.) Description AI coding agents (Claude Code, Codex, Cursor, etc.) increasingly run

meltano

through a pipe. The default CLI output is built for people at a terminal, so agents pay for it in tokens and can misread partial output as complete. DuckDB recently shipped an "agent mode" for the same reason: https://duckdb.org/2026/10/09/agent-mode Problems today •

meltano run

output is noisy (colors, progress, long tap/target stderr), and there's no end-of-run summary of records per stream, state written, duration, or which block failed. • Machine-readable output is inconsistent:

--format json

exists on

config

,

job list

,

plugin list

,

logs

,

schedule list

,

state list

and

hub

, but

select

uses

--json

. • Errors are plain text, so agents have to parse them. • There's no compact way to get an overview of a project. Proposal A presentation-only mode that changes output defaults and never changes semantics (exit codes, state, behavior). 1. Detection. Enable the mode when an agent is detected and stdout is not a TTY and no explicit format is given. • Add

--agent

/

--no-agent

and

MELTANO_AGENT_MODE=0|1

overrides. • Print one stderr line saying why the mode is on and how to disable it. • Explicit settings (

--log-format

,

--format

,

cli.log_format

, etc.) always win. 2. Logging. Default to JSON logs with no color and no progress animation (mostly a default change, since

cli.log_format

exists). 3. Run summary. Print a compact summary at the end of

meltano run

. Truncate only terminal output, state clearly that it was truncated and give the total, and keep full logs available via

meltano logs

. 4. Structured output. Default list-style commands to JSON and make

--format

/

--json

consistent across commands. 5. Structured errors. Emit

MeltanoError

as JSON on stderr, including the existing next-step hint. 6. Docs. Add a short page (and possibly

meltano --help agent

) describing the mode and the most useful commands for agents. Agent detection `agent-detector` (MIT, dependency-free, Python 3.9+, 17 agents,

AI_AGENT

override) covers detection. It returns the agent, a confidence level and the matched signal, never env values. Considerations: • v2.0.0 is marked "Alpha" on PyPI, and the agent list will keep changing, so we should decide between depending on it and vendoring a small helper. • Wrap it in a single function (e.g.

meltano.core.agent.detect()

) so that decision is reversible. • Require at least

medium

confidence, plus non-TTY stdout, before enabling the mode. • Detection is best-effort and must never gate behavior beyond presentation. Suggested first slice Detection helper +

--agent

flag, JSON-log/no-color defaults, and the

meltano run

summary. Structured errors and

--json

consistency can follow (and are useful on their own). meltano/meltano