audit-clerk-skill

Clerk CLI Skill Audit

Cross-check skills/core/clerk-cli/ in this repository against the actual Clerk CLI source and propose precise edits where the skill has drifted. The CLI is the source of truth; the skill is maintainer documentation that must track it.

This task needs a complete command inventory, not a quick grep pass. Build the source model first, then compare it to the skill's current claims.

Inputs

• CLI source of truth: a clerk/cli checkout containing packages/cli-core/src/commands/**, plus packages/cli-core/src/cli.ts, cli-program.ts, mode.ts, and any referenced files in packages/cli-core/src/lib/.
• Target skill: skills/core/clerk-cli/SKILL.md and skills/core/clerk-cli/references/*.md.

Source Checkout Resolution

Resolve the CLI source checkout in this order:

1. Use --source <path> when supplied. The path must contain packages/cli-core/src/commands/.
2. If the current repository itself contains packages/cli-core/src/commands/, use the current repository root.
3. If CLERK_CLI_REPO is set, use that path. Do not print the variable value if it contains sensitive path fragments.
4. In Conductor, look for exactly one sibling workspace matching ../cli/*/packages/cli-core/src/commands/. If multiple match, ask which one to use.
5. If no checkout is available and network access is acceptable for the task, clone the public CLI repository into .context/clerk-cli-source:

mkdir -p .context
cd .context
git clone --depth 1 https://github.com/clerk/cli.git clerk-cli-source

If none of these work, stop and ask the user for a CLI source path. Do not audit against the installed clerk binary alone, because the binary may be stale and does not expose every source-level branch.

Workflow

1. Inventory the CLI

Walk packages/cli-core/src/commands/ and build a structured inventory. For each top-level command and subcommand capture:

• Full command path, such as clerk config patch or clerk api ls.
• Purpose, from the command description or help string in source.
• Flags, including short form, type, default, and destructive behavior.
• Exit codes beyond the default if the command overrides them.
• Agent-mode branches gated on isAgentMode(), CLERK_MODE, TTY detection, or related helpers.
• Whether the command mutates remote state and needs --dry-run, --yes, or production-targeting guidance.

Read packages/cli-core/src/commands/<name>/README.md as secondary context only. Use those files to flag mocked or stubbed commands and to cross-check API endpoint claims in references/recipes.md. Do not propose copying command READMEs into this repository; they include internal implementation detail and would bloat the skill.

Also capture cross-cutting behavior:

• Runner preference logic, including lockfile to package-runner mapping.
• Auth, key resolution, --app, and --instance targeting.
• doctor checks and --json output shape.
• Agent-mode behavior for prompts, JSON defaults, browser opening, OAuth callbacks, deploy handoff, and sandbox warnings.
• clerk init --prompt or equivalent handoff behavior if the skill mentions it.

Prefer reading source over running the binary. When runtime behavior is unclear, use tests under packages/cli-core/src/**/*.test.ts or packages/cli-core/src/test/ as supporting evidence.

2. Extract the Skill's Claims

Read skills/core/clerk-cli/SKILL.md and each file under skills/core/clerk-cli/references/. Extract every concrete claim:

• Commands in the core command table and invocation guidance.
• Flags named in prose, tables, and examples.
• Agent-mode behavior bullets.
• Exit code, error format, and JSON output claims.
• Cross-references between SKILL.md and references/*.md.

3. Diff Source Against Skill

Produce a structured diff with four buckets:

1. Missing from skill: commands, subcommands, flags, or behaviors that exist in source but are not mentioned anywhere in the skill.
2. Stale in skill: claims in the skill that no longer match source, including renamed flags, changed defaults, removed commands, shifted exit codes, or reworked agent-mode branches.
3. Thin in skill: commands mentioned but under-specified relative to their real complexity or footgun surface.
4. Over-specified: details the skill encodes that clerk <cmd> --help, generated examples, or referenced files cover better.

For buckets 1 through 3, cite both the source file and line and the target skill location.

4. Decide Placement

Route changes by maintenance value:

• Core loop, mental model, and safety stay in SKILL.md.
• Per-command flag details, recipes, and edge cases belong in references/recipes.md or a new reference file.
• Agent-mode branches belong in references/agent-mode.md; SKILL.md gets only a concise summary.
• Auth, key resolution, and targeting belong in references/auth.md.

Prefer clerk <command> --help over duplicating generated help. Flags with destructive semantics, hidden interactions, or agent-mode divergence deserve skill coverage. Self-explanatory flags should usually be removed from prose and delegated to --help.

Treat skill shrinkage as a valid proposal when it reduces drift risk. The goal is an accurate, durable skill, not a larger one.

5. Propose Edits

Emit a review-ready proposal. For each change include:

• Path, such as skills/core/clerk-cli/SKILL.md or skills/core/clerk-cli/references/agent-mode.md.
• Severity: drift, gap, or polish.
• Why, with a source citation such as packages/cli-core/src/commands/<cmd>/<file>.ts:<line>.
• Target location in this repository.
• A unified diff when practical, otherwise a concise before and after block.

Group the proposal by target file. Do not rewrite accurate neighboring sections just because they are nearby.

6. Apply or Hand Back

Default behavior: present the proposal and stop for review.

If invoked with --apply, apply drift and gap edits directly, then list polish suggestions for review. After applying edits, run available formatting or validation commands for this repository. If no formatter exists, validate Markdown and JSON structure with lightweight checks.

Guardrails

• Never invent flags. If a flag appears in tests but not in the command parser, mark it for human review.
• Preserve the existing clerk-cli skill's terse, third-person voice.
• Do not use em dashes in proposals or edits.
• Keep skills/core/clerk-cli/SKILL.md near the 500-line guidance. Move detailed material to references/ instead of bloating the main skill.
• Do not treat the installed clerk binary as authoritative over source. Use --help only to confirm generated presentation when source and tests leave ambiguity.
• Do not commit or print secrets. If the audit touches env guidance, preserve the repository's 1Password and no-plaintext-secret rules.

Output Shape

Return the proposal as:

# clerk-cli skill audit - <YYYY-MM-DD>

## Summary
<counts per bucket, plus the largest drift>

## skills/core/clerk-cli/SKILL.md
### <section name>
- [drift|gap|polish] <one-line description>
  - source: packages/cli-core/src/commands/<...>:<line>
  - target: skills/core/clerk-cli/SKILL.md:<line>
  - change: <diff or concise before/after>

## skills/core/clerk-cli/references/<file>.md
...

## New files
...

## Open questions
...

Keep the result skimmable so a maintainer can approve, reject, or apply each entry independently.