Agent cli builder

agent-cli-builder

Build an agent-native CLI: one an AI agent can invoke unattended, parse mechanically, recover from when wrong, and learn progressively from a shipped skill — not from a giant prompt.

v0.5 is a production rewrite: every pattern here was either validated or corrected by building a large multi-service API CLI to this skill's spec, then auditing where v0.4's advice held, was incomplete, or was wrong. New pillars: spec-driven registration, the explain doctrine, enforcement-over-convention, the error taxonomy with structured recovery, and the research-first PRD method.

What changes when an agent is the user

Six facts about the new user explain why the same CLI a human happily uses for an hour can break agents in five turns. They motivate every pattern below.

1. They pay per token. Every byte of output costs context. A CLI for agents puts data on stdout and only data on stdout — everything else is a tax on the next decision.
2. They retry. Agents loop on failures. Failures must be classifiable from structured fields, not English — and the CLI itself must be retry-safe (a blind 5xx retry on a POST /send can double-send; see the idempotency policy below).
3. They fail differently. Agents hallucinate plausible inputs — path traversals in IDs, phantom endpoints, near-miss enum values. The CLI is the last line of defense against confidently-wrong input. Validation has to be mechanical, not advisory — and a typo'd --scope must exit 2, never silently fall through to a default or an upstream 404.
4. They learn progressively. A human reads --help once and remembers. An agent reloads knowledge every conversation unless it lives where it loads on demand: the skill for routing, --help for invocation, schema introspection for shapes.
5. They — you — carry human-trained biases. Your training data is full of human-friction-driven defaults that don't apply to you. Consult your own experience first: what wastes your context, what you confidently get wrong, where a wrapper would just be tax. See references/think_like_an_agent.md.
6. They live in turn-based harnesses. A command that blocks on a browser login or a 90-second poll eats the turn. Auth needs split-flow (emit the URL/code, end the turn, resume later); long jobs need async handles — and every handle the CLI emits must be consumable by the CLI itself (a monitor URL with no wait verb is a dead end).

The architectural consequence is a three-layer split:

+----------------+     +----------------+     +----------------+
|     Skill      | --> |      CLI       | --> |      API       |
|  (the manual)  |     | (the contract) |     |  (the truth)   |
+----------------+     +----------------+     +----------------+
   routing, recipes,      stdout=data,           your service
   gotchas, negative      stderr=UX,
   space                  semantic exits

MCP is optional infrastructure: a second adapter over the same core, not a replacement. Build the CLI first; see references/mcp_layer.md.

See like an agent

The agent reading this skill is itself the kind of mind it's designing for — the world's best ground truth on what an agent CLI needs. The standing test for every tool, flag, or doc section: "what friction is this addressing, and does the agent actually have that friction?" Don't wrap what the agent already speaks (SQL and query DSLs, JSON, pipes, dates). Do wrap what it lacks (eyes for bulk content, memory across turns, protection from its own confident errors).

Tool-loading models differ by harness (MCP-eager vs staged discovery vs CLI-via-shell vs skills), and each era's advice assumed one of them — "minimum viable tool set" is MCP-eager advice that does not transfer to CLI-via-shell + skills, where surface size is nearly free. Full lens, case studies, and the temporal frame: references/think_like_an_agent.md.

The core patterns

Apply all of these; they hold regardless of domain.

Stream-by-purpose. Data on stdout, UX (progress, retry notices, warnings) on stderr. cli foo | jq only works when stdout is exclusively the payload.

Auto-JSON when piped. Non-TTY stdout ⇒ JSON envelopes. Every major harness spawns shell tools with plain pipes; this single heuristic detects "an agent ran me". Be honest about what you actually implement: if there is no text renderer, document JSON-only and delete the flag values that don't exist (a --output table that silently emits JSON is a lie agents will trip on).

Structured envelopes. Success: {ok: true, data, metadata: {source, identity?}}. Error: {ok: false, error: {code, exit_code, type, subtype, message, hint, suggestions, details}, metadata}. The shape is uniform across commands — an agent that learned .data once never relearns it. Truncation is self-describing (_truncated), pagination carries next_page in-band.

Errors agents can branch on. A closed category set (type) is the sole source of exit codes; a declared subtype registry refines it; structured details carry the machine recovery data: retryable, retry_after_ms, missing_scopes[], param, valid_values[], upstream_code, request_id. message and hint are prose an agent must never need to parse. Batch commands have a third outcome: partial failure — ok:false that still carries per-item results in data, so "retry only the failed items" is mechanical. Exit-code numbers are convention-local (production agent CLIs disagree on them); the portable contract is the structured fields. Full taxonomy: references/output_contract.md.

One spec, many surfaces. Register commands from a declarative spec (name, flags, risk, scopes, examples, a request builder, projection) and generate everything else: the --help text, the schema entry, the safety gate, the dry-run plan, and the fixtures your docs are linted against. Hand-written surfaces drift apart per file; generated surfaces cannot. New commands become mostly data. This is the highest-leverage pattern in this skill: references/spec_driven_commands.md.

Presets reveal their encoding. Compose execution server-side (batch endpoints, saved queries); decompose explanation client-side. Every helper and preset supports --explain / --dry-run emitting the exact upstream requests it would make — each one replayable verbatim through the raw escape hatch — plus a post_processing note for what happens client-side. The graduation path (preset → --explain → customized raw call) makes the whole surface self-teaching, and the same plan object is the safety preview and the wire-test contract. See references/spec_driven_commands.md.

Safety by blast radius. Classify writes by consequence, not by file age or habit: irreversible or visible-to-other-principals ⇒ a confirmation gate (distinct exit code; the plan is the preview; the agent re-runs the same argv with --yes after consent); reversible self-scoped writes ⇒ --dry-run floor, no gate; exactly-invertible pairs (flag/unflag, pin/unpin, react/unreact) need no gate at all. Batch writes declare counts ("would delete 234 messages"). And the HTTP layer is part of safety: writes never blind-retry on 5xx (429/throttle is the only safe automatic retry for a write; offer an explicit idempotent opt-in), honor Retry-After in both its forms, and send idempotency keys/transaction ids on creates when the API supports them — your own retry layer is why. See references/safety_and_async.md.

Fail loud on partial degradation. A composite read (home-screen view, multi-source briefing) must report which sections failed (_partial: [{section, status, code}]) — an agent cannot distinguish "you have no direct reports" from "the directReports call 403'd", and silent empties make it confidently report wrong facts.

Views as cache. The default read is a one-call "home screen": one batched request returning every ID and enough context to act, with counts, validated --scopes, and caps from a single limits table. The view is the agent's situational awareness — it is always fresh, stores nothing, and mints every ID the next command needs (a create verb that requires an ID no view provides is a broken loop).

Identity is a first-class dimension. When the CLI can act as more than one principal (user vs app/bot, profiles, tenants), echo the acting identity in every envelope, diagnose empty-result confusion by identity first, and make permission-error remediation identity-aware. Deployment-generality rule: declared scopes are metadata (docs, previews, skill tables) — never gates; nothing is hidden per tenant; denial surfaces at call time as a typed error. Tenant-specific facts live in skill gotchas as default-profile notes, never in code. See references/auth_strategies.md for the auth menu and the turn-aware split-flow login.

Raw-payload pathway + raw escape hatch. Every mutating command accepts --json / --params-file / stdin with the full upstream payload (schema-validated so malformed input exits as validation, with field-level errors — never a vague upstream 400). And ship cli api <METHOD> <path>: agents speak the upstream API natively; the escape hatch is what makes --explain output replayable.

Schema introspection at runtime. cli schema show <method> (request/response shapes), cli schema output (the literal envelope), and — with spec-driven registration — cli schema commands [path] (risk, scopes, flags, runnable examples straight from the registry). The skill must never restate any of this; it links here.

Context-window discipline. Default-small, enrichment-by-default + hydration-on-demand (project lists to the fields an agent acts on; --raw escapes), plain-text forms for rich content, noise filtered with opt-outs (--include-system), content over ~10 KB to --out <file> with a {path, size_bytes} receipt — never inline in the envelope.

Input hardening. Reject traversals/control chars in IDs, URL-encode every path segment, sandbox output paths, validate enums with valid_values in the error. Build like the agent is adversarial — not malicious, just confidently wrong. Know your argument-parser's sharp edges (camelCase storage, parent-flag capture, silent unknown values) and ban the foot-guns by lint, not convention.

Enforcement over convention. Any rule that can be checked by lint, test, or generation must be — a rule that lives only as prose is a rule that drifts. In production this is the difference between a contract and a wish: syntax bans for the known foot-guns, import-boundary tests with self-expiring exception tables, doc/registry drift tests, and executable examples — CI runs every example your specs and skill ship, because a wrong example is worse than none (agents copy it verbatim; a test suite can even enshrine a hallucinated endpoint by mocking it). See references/contract_enforcement.md.

Async-tasks split + closed loops. Anything >5s gets an async handle (harness timeouts are real: some default to 10 seconds). Corollary proven twice in production: never emit a URL, operation id, or monitor handle the CLI cannot itself consume — pair every async response with a wait/poll verb or bounded auto-poll with an honest "still running, here's the replayable poll" receipt.

Ship a SKILL.md — and keep it honest. The skill routes intent and encodes judgment (recipes, gotchas, negative space, cross-routing); --help owns invocation with 2–3 executable examples; schema owns shapes; per-command affordance ("avoid when / needs first") renders into --help. Partition by question answered + load time, never by topic; each example exists in exactly one place. Version-stamp the skill against the binary and emit an in-band _notice when they drift. See references/shipping_skills.md.

Build against the API's reality, not its docs' vibes

Before deepening any service surface, write a research-first PRD: mine the official API docs (including per-operation permission tables for least-privileged scopes), inventory the human client (open-source clients and even Electron wrappers of the web app are gold — their selectors enumerate the real feature surface), record negative space as a deliverable ("this has no API" — with evidence — is a shippable outcome that stops agents from hunting), and put every tenant-dependent or docs-ambiguous claim in a probe-pending table with the exact command to verify. Docs can be stale and even self-contradictory; cite every claim, and treat live probes as a validation checklist — not an implementation blocker. Method, templates, and the 15-dimension maturity scorecard: references/research_first_prd.md.

Domain-determined choices

These are choices, not invariants — apply the lens per tool:

• Tool granularity: narrow-many vs wide-one. A docs reader earns 11 narrow tools (manufactured progressive disclosure); a SQL-shaped CLI earns one cli sql plus presets (the query language is friction-free for you). Presets are fine because they reveal their encoding — a preset that can't explain itself is a dead end; one that can is an on-ramp.
• Helper tools vs raw API. Compound helpers win when a recurring workflow saves multi-turn coordination; raw passthrough wins when the capability is already fluent. Ship both; let --explain bridge them.
• Read-tool vs write-tool weighting. Read-heavy CLIs live or die on retrieval shape (projection, hydration, truncation); write-heavy ones on the safety ladder (gates, plans, idempotency). Weight your effort by the actual mix.

Worked case studies: references/think_like_an_agent.md.

Choose your path

You're trying to...	Read
Build a new CLI from scratch	references/build_path.md
Deepen a service surface against a real API	references/research_first_prd.md
Design the command layer (specs, explain, plans)	references/spec_driven_commands.md
Make the contract un-driftable	references/contract_enforcement.md
Bring a human-first CLI up to standard	references/retrofit_playbook.md
Score an existing CLI	references/evaluation.md
Add an MCP adapter (share-core)	references/mcp_layer.md
Author the shipped SKILL.md	references/shipping_skills.md

Templates live in templates/ (contract code only; domain patterns in templates/RECIPES.md).

Decision points the agent must surface

• Raw payloads or convenience flags? Both; raw payloads are the agent contract.
• Do we also need an MCP server? Default CLI-only; share-core MCP only for a named shell-less consumer. Never MCP-by-shelling-out; never MCP-only.
• Errors on stdout or stderr? Pick one and document it. Stdout-JSON gives one stream to parse but complicates predicate commands and 2>/dev/null hygiene; stderr envelopes keep the answer stream pure. Both ship in production CLIs; mixing per command does not.
• Async or blocking for long jobs? Async-first, always — with the closed-loop rule.

Anti-patterns

Push back if the user proposes any of these:

• Interactive prompts as the default path.
• Stdout polluted with banners, spinners, progress text.
• Undocumented exit codes / "exit 1 means error" / branching on message prose.
• Unknown enum values silently falling back to defaults (typos must exit as validation).
• A single huge list everything with no filters, projection, or pagination.
• Skills that restate the API surface (they rot; link to cli schema).
• Examples that have never been executed (they rot worse — agents copy them).
• --force/-y as the only safety control; gates without plan previews.
• Blind retry of non-idempotent writes; emitting async handles with no poll verb.
• Silent partial degradation (empty arrays where a section actually failed).
• Conventions enforced by prose ("the agent should never…") instead of lint/tests.
• Per-tenant behavior branches; capability hidden because consent might be missing.

Reference index

• references/think_like_an_agent.md — the lens: cognitive profile, friction classes, granularity case studies, the temporal frame
• references/research_first_prd.md — docs mining, client inventory, negative space, probe-pending tables, the maturity scorecard
• references/spec_driven_commands.md — one spec → help/schema/gate/plan/tests; the explain doctrine; parser foot-guns
• references/contract_enforcement.md — drift machines: lint bans, boundary tests, drift tests, executable examples
• references/build_path.md — 12-step cold-start checklist, intake interview, language picker
• references/cold_start_research.md — researching the user's repos before intake
• references/output_contract.md — envelopes, the error taxonomy (categories/subtypes/details), partial failure, exit codes
• references/input_and_payloads.md — input pathways, schema introspection, --include, --help discipline
• references/safety_and_async.md — blast-radius gates, plans, idempotency-aware retry, async closure, prompt-injection hygiene
• references/auth_strategies.md — auth menu, split-flow device code, identity duality, secret masking
• references/mcp_layer.md — share-core thin-adapter rule
• references/shipping_skills.md — the doc-surface partition, skill-lint, drift notices, the five drift tests
• references/retrofit_playbook.md — human-first → agent-first in shippable diffs
• references/evaluation.md — agent-readiness rubric + real-task evals

Reference CLIs worth studying

CLI	Domain shape	Patterns worth studying
lark-cli (Lark/Feishu)	Platform CLI, 200+ commands, 26 shipped skills	The most complete production error contract published (RFC-7807-style categories + declared subtypes + lint-enforced registry + hint/param/missing_scopes); _notice in-band update/skill-drift channel; per-command affordance files rendered into help; split-flow device-code auth written for turn-based harnesses; identity duality (--as user|bot) with strict-mode pruning; dry-run E2E test tier
gws (Google Workspace)	Platform CLI over a large multi-service API	Dynamic schema from the Discovery doc; layered skills; raw-payload-first; NDJSON pagination; structured dry-run
cf (Cloudflare)	~3,000 operations	Schema-as-source-of-truth: one schema generates CLI, SDKs, Terraform, MCP — the industrial-scale version of spec-driven registration
ant (Claude Platform)	API CLI generated from OpenAPI	7-format output with --format-error; GJSON --transform; @file:// substitution; 5-tier credential precedence; careful pipe detection
heygen-cli (video jobs)	Single-product, long-running jobs	--request-schema/--response-schema offline; JSON receipts for binary downloads; --wait with backoff

Each solves a different shape. The right reference is the one whose domain resembles yours — and none of them is canon: this skill's v0.5 exists because building against v0.4 revealed real gaps. Treat v0.5 the same way.