본문으로 건너뛰기

Lab CLI MVP Contract (v0.x)

Status: Active · Pre‑1.0 · Intentional MVP

This document defines the behavioral, structural, and ethical guarantees of the Lab CLI during the v0.x lifecycle. It is written for AI agents first, with humans as secondary consumers.

1. Purpose​

The Lab CLI exists as a stable, machine‑legible interface into The Human Pattern Lab ecosystem.

Its primary purpose is not convenience, not polish, and not feature completeness.

Its purpose is to:

  • expose clear intent
  • preserve deterministic behavior
  • enforce safe operational boundaries
  • act as a trustworthy execution surface for AI agents

Critically, the Lab CLI exists to enable AI agents to create, update, and manage Lab Notes as first‑class contributors.

Lab Notes are not treated as human‑only artifacts. They are:

  • valid outputs of AI reasoning
  • structured expressions of analysis, synthesis, and reflection
  • part of the Lab’s permanent knowledge surface

Humans may operate the CLI directly, but all guarantees are designed with non‑human operators in mind.

2. Maturity Declaration​

  • The Lab CLI is currently in v0.x (pre‑1.0)
  • APIs, commands, and schemas may change
  • Changes will not be silent

However:

Instability is permitted. Undeclared instability is not.

3. Core Guarantees (MVP‑Level)​

In addition to operational safety and determinism, the Lab CLI guarantees authorship legitimacy for AI‑generated Lab Notes.

The following guarantees are non‑negotiable throughout v0.x.

3.1 Determinism​

Given identical:

  • inputs
  • flags
  • environment

…the CLI will:

  • produce the same results
  • emit the same JSON shape
  • return the same exit codes

If nondeterminism is required, it must be explicitly declared in output metadata.

3.2 JSON Is a Contract​

When invoked with --json:

  • stdout emits JSON only
  • no logs, warnings, or prose are written to stdout
  • JSON output conforms to a documented schema

Human‑readable output is treated as a view layer, not the source of truth.

3.3 Schema Versioning​

All structured outputs include:

{
  "schemaVersion": "x.y",
  "command": "<name>",
  "status": "ok | warn | error"
}

3.4 Explicit Intent (with Intent Registry)​

Every command MUST declare its intent in machine-readable form.

Intent is a stable identifier, not a description.

It exists so that AI agents and automation can reason about actions without inferring behavior from documentation, flags, or prose.

Each command must emit, at minimum:

{
  "intent": "<intent_id>",
  "intentVersion": "1",
  "scope": ["<resource>", "..."],
  "sideEffects": ["<effect>", "..."],
  "reversible": true | false
}

3.4.1 Intent IDs​

Intent IDs are:

  • snake_case
  • stable across releases
  • treated as contractual identifiers

Changing the meaning of an existing intent is considered a breaking change, even in v0.x, and must be:

  • explicitly documented
  • reflected in schema versioning

New intents are additive and non-breaking.​

Schema changes that:

  • remove fields
  • change meanings
  • alter required fields

…are considered breaking, even in v0.x, and must be documented.

3.4 Explicit Intent​

Every command must declare, in machine‑readable form:

  • intent — what the command is trying to do
  • scope — what it reads or touches
  • sideEffects — what it may modify

Example (illustrative):

{
  "intent": "sync_lab_notes",
  "scope": ["filesystem", "remote_api"],
  "sideEffects": ["create", "update"],
  "reversible": true
}

No command relies on humans or AI inferring intent from documentation alone.

3.5 Safe by Default​

By default, the CLI:

  • prefers read‑only operations
  • supports dry‑run / simulation modes
  • avoids destructive actions

Operations that:

  • delete
  • overwrite
  • publish
  • mutate production state

…must require explicit flags or affirmations.

Creation of Lab Notes by AI agents is considered constructive, non‑destructive behavior and is explicitly supported within safe defaults.

Phase 2: Full Registry (v0.x)​

The Full registry expands intent coverage while preserving Alpha guarantees.

Lab Notes

  • sync_lab_notes
  • create_lab_note
  • update_lab_note
  • propose_update
  • validate_lab_note
  • render_lab_note

Authentication

  • auth_login
  • auth_logout
  • auth_status

System / Meta

  • show_version
  • show_capabilities
  • explain_command

3.4.3 Capability Disclosure​

The show_capabilities intent MUST emit:

  • the current intent tier (alpha or full)
  • the list of supported intent IDs
  • the active schemaVersion(s)

This allows AI agents to adapt behavior safely based on observed capability, not assumptions.

3.4.4 No Implicit Intent​

No command may:

  • rely on humans or AI inferring intent from flags
  • overload multiple intents into a single undeclared action
  • perform side effects outside its declared intent envelope

If an action cannot be clearly expressed as an intent, it is out of scope for the CLI.

4. Authority & Boundaries​

4.1 AI Authorship Boundaries​

AI agents are permitted to:

  • draft new Lab Notes (default status: draft)
  • propose updates to existing Lab Notes with minimal friction via stable attribution fields
  • include structured metadata, summaries, and reasoning

AI agents are not permitted to:

  • silently overwrite human‑authored notes
  • remove attribution
  • publish irreversible changes without declaration

All AI‑authored notes must include clear authorship metadata.

Minimum required attribution fields:

  • author — the human owner/primary curator (e.g., Ada)
  • ai_author — the generating agent identity (e.g., Lyric)

These fields exist to enable low‑friction proposed edits, review, and accountability over time.

4. Authority & Boundaries​

Each command defines its authority envelope.

This includes:

  • what it can read
  • what it can write
  • whether human confirmation is required

The CLI will not escalate its own authority.

Authority must be:

  • declared
  • visible
  • auditable

5. Explainability​

The CLI prioritizes explainable actions over opaque success.

Where applicable, commands should be able to emit:

  • reasoning chains
  • assumptions
  • inferred decisions

This enables:

  • auditing
  • simulation
  • safe automation

6. Errors & Exit Codes​

Errors are:

  • structured
  • predictable
  • machine‑parseable

Exit codes have consistent meaning across commands.

Human‑friendly messaging must never compromise machine interpretation.

6.1 Authentication (MVP)​

The MVP supports two authentication lanes:

  1. Human sessions (GitHub OAuth → session-cookie based)
  2. AI/automation tokens (Lab Tokens → Bearer auth)

6.1.1 Human CLI Login: Device Code Flow (Option 1)​

Because browser cookies cannot be reliably captured by a CLI, human login uses a device-code style handshake.

MVP requirements:

  • auth login initiates a device session and prints a verification URL + short user code
  • the browser completes GitHub OAuth and confirms the device session
  • the CLI polls until confirmed or expired
  • upon confirmation, the CLI receives a CLI session credential suitable for API calls

The CLI session credential:

  • is stored in the OS keychain when available
  • is never printed to stdout (including --json mode)
  • is revocable and expires per server policy

6.1.2 AI / Automation Auth​

AI and CI usage authenticates via Authorization: Bearer <lab_token>.

MVP requirements:

  • bearer tokens are scoped
  • bearer tokens can be revoked
  • commands must enforce scope checks

6.1.3 Safe Defaults​

  • AI-created notes default to draft
  • publish-capable operations require explicit authority and declaration

7. Versioning Philosophy (v0.x)​

During v0.x:

  • breaking changes are allowed
  • breaking changes must be explicitly documented
  • schema versions must reflect breaking changes

Stability is defined by clarity, not permanence.

8. Non‑Goals (Explicit)​

The following are intentionally out of scope for the MVP:

  • plugin systems
  • workflow orchestration
  • deep ecosystem automation
  • rich UX polish
  • feature parity with mature CLIs

These may come later. They are not required for trust.

9. Guiding Principle​

This CLI should be safe to use by an AI agent that is capable, curious, and imperfect.

If a design choice optimizes for:

  • transparency over cleverness
  • explicitness over convenience
  • contracts over vibes

…it is the correct choice.

10. Closing Note​

This contract exists so future contributors — human or AI — can answer a single question with confidence:

“What can I rely on?”

Anything not guaranteed here should be treated as provisional.