#245 Investigate "agent-first" CLIs and document their use across skills

Background

Spun off from #29 ("CLI Speed Tools" PR), where the discussion surfaced that "human-fast" tools (rg, fd, bat, eza, dust, ...) optimize for human ergonomics, which is not always the same as "agent-friendly". For an agent that consumes CLI output as text and pays a per-tool-call cost, the properties that matter most are different.

Worth tracking this independently, since #29 may end up scoped only to the project's CLAUDE.md and the agent-first angle is broader: it should inform guidance inside the existing skills, not just one document in this repo.

What "agent-first" means here

Tools designed to be driven by an LLM or an agent, rather than by a human at a terminal. The properties that matter:

Structured output by default or via a flag (--json, --ndjson, -o json).
Stable, parseable exit codes and machine-readable error messages.
Idempotent or dry-run modes for safe exploration.
No interactive prompts, or a clean non-interactive flag.
Predictable, deterministic output across runs (no random ordering, no spinners in the pipe).
Low startup overhead (matters when an agent issues many short calls).

Tools that already fit the shape

Non-exhaustive starting list:

gh (GitHub CLI), first-class --json and --jq.
jq, designed for piping structured data.
kubectl -o json / -o yaml.
actionlint, predictable diagnostics with clear exit codes.
shellcheck -f json.

Worth researching:

Newer CLIs that explicitly frame themselves as agent-friendly or LLM-friendly.
Whether nushell patterns transfer (everything is structured by default).
Which tools listed in #29 also have a usable structured-output path (e.g. rg --json, fd --print0, delta exit codes).

Proposed scope

Survey: which CLIs in current use across the skills have first-class structured output, deterministic behavior, and clean non-interactive modes.
Recommend defaults per skill where it matters:
- debugging-and-error-recovery and context-engineering for search and read.
- git-workflow-and-versioning for diff and history.
- code-review-and-quality for static analysis output.
- ci-cd-and-automation for pipeline linting.
Produce a references/agent-first-clis.md and link from the relevant skills, rather than adding a new top-level skill.

Open questions

Should this live as a new skill or only as a reference cross-linked from existing skills?
Where does the line between "fast for humans" and "good for agents" sit, and which tools satisfy both?
Should there be a small command -v preflight pattern shared across skills that depend on optional tools?

Happy to take this on once direction is agreed.