audit-agentic-cli

Make a CLI an agent can drive without human babysitting. The contract: pure machine-readable stdout, semantic exit codes, non-interactive defaults, structured errors with retry guidance, and — for workflows that produce repairable artifacts — an iterative feedback loop the agent can finish on its own.

When To Use

Trigger this skill when the user is doing CLI-only agent-readiness work:

• "make my CLI agent-friendly" / "audit CLI for LLM use"
• the agent keeps failing to parse a CLI result
• stdout mixes JSON with progress text or spinners
• every failure exits 1 with no structured error body
• a command hangs waiting for an interactive prompt
• the workflow validates, rejects, repairs, and resubmits a generated artifact
• a fresh CLI needs agent-first constraints from day one
• help text is unreliable as a contract for an agent caller

Do not use this skill for:

• Surface choice (CLI vs MCP vs hybrid). Make the surface decision first using your normal architecture process. Only after the surface decision is fixed as CLI does this skill apply.
• MCP server audits, schemas, transports, auth, or context budgets. Route to audit-agentic-mcp for audit/architecture, build-mcp-server-sdk-v1 / build-mcp-server-sdk-v2 / build-mcp-use-server for implementation, test-by-mcpc-cli for live MCP verification.
• Vendor CLI operations when a vendor skill exists. Use run-railway instead.
• Generic shell scripting unrelated to the CLI's agent contract.

Five Core Audit Checks

Before recommending features, verify these. If any fail, fix them first.

1. --json (or equivalent) exists and stdout is pure machine output — no progress, no banners, no prose.
2. Logs, spinners, banners, and progress go to stderr.
3. Exit codes distinguish success, usage error, auth, not-found, conflict, validation, and transient/retryable failures — not all 0 or 1.
4. Headless runs never block without an explicit --no-input/--yes path or a clear non-interactive error.
5. Error responses include a stable error code and a retryable boolean (or equivalent guidance).

When a command must emit a non-JSON artifact (patch, diff, query, prompt, translation), keep stdout one canonical artifact format and put status on stderr or in a sidecar — never mix prose into the data channel.

Severity Rubric

Classify findings in agent terms, not human-UX terms.

Minimum Output Envelope

Aim for one stable JSON envelope across all commands.

Success:

{ "ok": true, "result": {}, "error": null, "schema_version": "v1" }

Failure:

{
  "ok": false,
  "result": null,
  "error": {
    "class": "validation",
    "code": "MISSING_FLAG",
    "message": "Flag --target is required.",
    "retryable": false,
    "suggestion": "Run `mycli deploy --target <name>`."
  },
  "schema_version": "v1"
}

Stdout = data channel. Stderr = operator channel. Exit code = semantic class. All three must agree.

Build Order (Repair Existing CLI)

When fixing an existing CLI, do the work in this order. Skipping ahead wastes effort.

1. Pure JSON output and stdout/stderr separation.
2. Structured errors and semantic exit codes.
3. Non-interactive flags and safe defaults (--no-input, --yes, --dry-run).
4. Iterative feedback loops for any command whose output the agent might need to repair.
5. Discovery: help text, examples, documented exit codes and output fields.
6. Async / JSONL / job-style flows for long-running operations.

Audit Workflow

Before recommending changes:

1. Inspect top-level help and each relevant subcommand. Use scripts/audit-cli-help.sh (read scripts/audit-cli-help.md first) to capture standard flags, examples, and missing affordances.
2. Look for machine-readable output flags: --json, --output json, --quiet, --fields, --jq.
3. Test stdout/stderr separation on safe read-only commands.
4. Classify exit codes across success, usage, auth, not-found, validation, conflict, and transient paths.
5. Probe non-interactive behavior on prompt-prone commands.
6. Inspect destructive flow safety: --dry-run, --yes, --force, explicit confirmation semantics.
7. Verify help documents realistic examples, output fields, and exit codes.
8. If auditing a CLI you've snapshotted before, diff with scripts/diff-cli-help.sh (read scripts/diff-cli-help.md first) to spot command/flag drift.

Iterative Repair Loop

Use this pattern when an agent generates an artifact, the CLI validates it, and the agent must repair on rejection.

input artifact -> CLI validation -> structured diff/errors -> agent repair -> retry -> accepted -> finalize

Typical fits: translation/localization batches, code generation with validation gates, manifest/config/migration generation, bulk import or sync.

Each iteration must tell the agent:

• current stage
• accepted / rejected / partial
• what failed, with identifiers or locations
• whether the failure is retryable
• the next command to run (next_action.command)
• progress complete vs work remaining

One-shot read/list/get/status commands do not need this — use the simpler envelope.

For the full design pattern and a worked case study, read references/iterative-cli.md.

Design Rules

• One stable envelope shape across all commands.
• Stdout = data. Stderr = operator. Exit code = semantic class.
• Command names and field types stay consistent across releases.
• Destructive flows are explicit: --yes, --force, --dry-run.
• --help documents examples, output fields, and exit codes.
• For repairable workflows, return enough structure for an agent to fix and retry without human interpretation.

Output Contract (Deliverable Shape)

Pick one based on the task.

Audit report: scorecard by audit dimension, severity-ranked findings, command evidence (exact safe command + observed stdout/stderr/exit), why it matters for agents, recommended fix, verification command.

Refactored CLI contract: proposed command grammar, flags and non-interactive behavior, stdout/stderr rules, JSON envelope and stream format, exit-code map, migration notes for existing users and scripts.

Iterative repair-loop design: phases and command family, artifact channel, validation response shape, retry budget, next_action.command, finalization criteria.

Local Exemplars

• linear-cli — strong agent-ready CLI: JSON everywhere, --dry-run, --id-only, non-pager flags, exit-code contract, bulk-mutation gates.
• run-railway — drift-aware CLI: installed-help snapshot, upstream-vs-local distinction, refresh scripts, version-drift routing.

Reference Routing

Read the smallest reference set that matches the task.

These are canonical homes — do not duplicate their content here.

Bundled Scripts

Read-only discovery support. Do not treat them as agent-readiness proofs.

These cover discoverability and drift only. They do not replace stdout/stderr, exit-code, non-interactive, and destructive-flow checks — run those by hand.

Finish Criteria

Do not call a CLI agent-ready until:

• The five core audit checks pass.
• The JSON envelope is stable enough to script against across commands.
• --help documents the command contract (examples, output fields, exit codes).
• A non-interactive run can succeed or fail deterministically with no human keystroke.
• Any iterative workflow returns enough structured feedback for an agent to repair without human interpretation.