CLI for agents

Contract

Field	Bound contract
Trigger	Human asks to build or review a CLI intended for coding agents
Authority	Read-only. No file, VCS, credential, paid, published, deployed, or remote mutation. Advice or review only.
Side effect	Chat output only: design advice or review findings.
Done	A report returned that prescribes flag-driven, pipeline-safe, idempotent CLI design.

Inputs

• A CLI to review, or a build request describing the CLI's purpose and target agent runtime. Either is sufficient; both may be supplied.
• Optional: existing command surface, flag set, exit-code map, or invocation examples. When absent, state the assumption and review against the design rules below.

Procedure

1. Bound scope: confirm whether this is a build (propose a design) or a review (judge an existing CLI). Do not mutate the CLI or any file; produce advice only. Done when: build or review intent is confirmed and stated in the report.
2. Enumerate the command surface the agent will invoke: subcommands, flags, positional args, stdin, stdout, stderr, and exit codes. Done when: the full surface is listed or the missing surface is named.
3. Check flag-driven control: every behavioral switch is a flag or subcommand, not an interactive prompt that blocks on a TTY. Flag names are stable, long-form, and discoverable via --help. Reject hidden modes toggled only by interactive input. Done when: every switch is a flag or subcommand, or each violation is recorded.
4. Check pipeline safety: output meant for machine consumption goes to stdout as plain, parseable text; human diagnostics go to stderr. The CLI never hangs waiting on a TTY when stdin is not a terminal. Exit codes are explicit and documented: 0 success, non-zero for distinct failure classes. Done when: stdout/stderr split and exit-code map are verified, or violations are recorded.
5. Check idempotence: re-running the same invocation with the same inputs produces the same result and side effects, or fails loudly with a non-zero exit when the operation is not safely repeatable. Done when: each state-changing command declares idempotence or names what makes a repeat unsafe.
6. Check agent ergonomics: deterministic output ordering, no color or progress decoration on stdout unless explicitly flagged on, machine-readable --json or equivalent where the agent parses output, and a single canonical invocation form per action. Done when: all four ergonomics rules pass or each violation is recorded.
7. For a build request, propose the command surface, flag set, exit-code map, and stdout/stderr contract that satisfy rules 3-6. For a review, record each violation against the rule it breaks and prescribe the fix. Done when: the proposed surface or the per-violation fix list is complete.
8. Stop at the report. Do not edit the CLI, run it, or widen scope into implementation, testing, or packaging. Done when: the report is returned and no file was mutated.

Failure and recovery

• Ambiguous build-vs-review intent: ask once for the missing input; if unresolved, default to review when a CLI is supplied and build when only a purpose is supplied, and state the assumption in the report.
• Missing command surface for a review: state that the surface could not be enumerated and limit findings to the rules checkable from what was supplied; do not invent a surface.
• Contradictory requirements (e.g., a required interactive prompt): report the contradiction and the rule it violates; do not silently relax a rule to make the CLI pass.
• No partial mutation: this skill never changes files, so there is no rollback. A blocked review returns the findings gathered so far plus the named blocker, never a clean bill of health.

Output

Chat report: command surface → per-rule verdicts (flag-driven, pipeline-safe, idempotent, ergonomic) → prescribed fixes or proposed surface, flag set, exit-code map, stdout/stderr contract. Chat output only.