Why

The installable artifact is an agent skill, not a human command manual. Official Codex skill guidance uses progressive disclosure: metadata triggers the skill, SKILL.md supplies the core workflow, and references load only when needed. A duplicated command catalog and tables of contents consume agent context and drift from the installed CLI.

Reference: OpenAI — Build skills

What changed

• remove the reference TOCs and 227-line command/flag catalog
• route agents to installed CLI help for commands and flags, including an explicit version-skew stop instead of guessing or self-updating
• rewrite SKILL.md as concise execution, safety, scheduling, formatting, and progressive-disclosure guidance
• retain only non-obvious target, output, download, and recovery behavior in two short conditional references
• promote fragile contracts into authoritative CLI help and protect them with test/help-contracts.test.ts
• keep README, llms.txt, and AGENTS.md aligned with the skill
• rebase onto main and preserve the newly merged canvas-create trigger, mutation boundary, auth/source constraints, and output shape

The installed skill drops from 25.6 KB to 7.2 KB (about 72%) and from roughly 520 lines to 96, without removing CLI capabilities.

Zero-regression review

Nine independent review passes across three rounds covered realistic task routing, every Slack mutation family, base-to-head compatibility, and help/runtime alignment. Findings fixed before the final round include:

• compose-only requests never invoke Slack; message draft is treated as send-capable and blocked in CI without separate send authorization
• absolute scheduling and Later reminder timezone behavior
• URL/channel/user workspace selection, thread overrides, mark-read timestamp overrides, scheduled cancellation, reply broadcasts, and group-DM limits
• attachment formatting and send-result metadata
• failed-download paths and reply-after-search recovery
• canvas-create source, credential, channel-tab, and output contracts added by Add canvas creation from Markdown #116

The final three reviewers found no blockers and marked the PR merge-ready.

Validation

• official skill validator: pass
• bun test: 272 pass, 0 fail
• bun run typecheck: pass
• bun run build: pass
• bun run format:check: pass
• bun run lint: 0 errors (10 pre-existing warnings)
• git diff --check: pass
• GitHub CI Build, Lint, Test: pass