Implements #83. Supersedes #112 — reuses that PR's Slack drafts.* API layer, but takes the "compose" direction (option 2) from the issue instead of the plural message drafts subcommand.

What & why

The issue flagged that adding Slack-native draft ops under the existing draft name would confuse an LLM. #112's answer was a near-homograph pair — message draft (browser editor) vs message drafts (native) — one letter apart, opposite meanings. This PR instead gives each concept a distinct verb/namespace:

• message compose <target> [text] — the human-in-the-loop browser editor (renamed from message draft; behavior identical). "Compose" matches what it does.
• message draft <list|create|update|delete> — Slack-native drafts via the undocumented drafts.* session endpoints, so drafts show up natively in the user's Slack client (desktop + mobile).

agent-slack message compose "#general" "Here's my update"      # opens browser editor

agent-slack message draft list
agent-slack message draft create "#general" "Here's my update"
agent-slack message draft create "<msg-url>" "Looking into it"    # thread reply
agent-slack message draft create "W0123456789" "hi"              # DM draft
agent-slack message draft update "<draft-id>" "Revised text"
agent-slack message draft delete "<draft-id>"

• create accepts the usual targets: #channel/name, channel id, user id (opens a DM), or a message URL (drafts a thread reply — same semantics as message send).
• Draft text runs through the existing mrkdwn → rich_text conversion, so lists/bold/code/mentions render natively.
• update replaces the body but preserves the draft's destination and attached files unless overridden; the conflict-detection client_last_updated_ts is auto-fetched for update/delete when omitted (incl. Slack's 7-digit fractional-padding quirk).
• Native drafts require browser-style auth (xoxc/xoxd), like later/unreads; noted in the docs.

Broadcast semantics (thread replies)

--broadcast ("also send to channel") is handled carefully so it can't silently do the wrong thing:

• inherited on a text-only update (editing text keeps your broadcast choice), but reset on any --channel re-address so a reply is never ratcheted into a different thread's channel;
• clearable with a new --no-broadcast flag;
• rejected for DM targets (matching message send).

Refuses unsafe updates

update fails loudly instead of corrupting drafts it can't faithfully round-trip: a scheduled draft (would drop the send time) or a multi-destination draft edited text-only (would drop recipients). It points the user to edit in Slack or delete + recreate.

⚠️ Breaking change (intentional, no compat shim)

message draft <target> [text] no longer opens the editor — use message compose. Old usage errors cleanly (error: unknown command '#general'), and message draft --help points to message compose. Deliberate clean break (pre-1.0, v0.9.x); the bundled agent skill is updated in the same PR, so agents pick up the new surface immediately. No magic dispatch / dual-meaning draft.

Structure

• src/slack/drafts.ts — API layer (reused from feat(message): Slack-native drafts — message drafts list/create/update/delete #112).
• src/cli/message-draft-command.ts / message-draft-actions.ts — CLI layer (singular draft subcommand group).
• src/cli/compose-actions.ts — renamed from draft-actions.ts.
• Docs kept in sync: README, skills/agent-slack/SKILL.md, references/commands.md, references/targets.md, llms.txt.

Testing

• test/drafts.test.ts (API layer, from feat(message): Slack-native drafts — message drafts list/create/update/delete #112) + new test/message-draft-actions.test.ts (CLI action layer + a commander-level test locking --broadcast/--no-broadcast/omitted parsing): create, update read-modify-write, scheduled/multi-destination guards, broadcast re-address + DM guards, last_updated_ts auto-fetch/pad, findDraft malformed-duplicate handling, channel-name hydration.
• Full suite (284), typecheck, lint (0 errors), format, and build all pass. Verified CLI wiring for compose/draft/all subcommands, and that old message draft <target> usage errors cleanly.

Review

Reviewed by independent correctness + docs sub-agents, then an adversarial loop with GPT-5.6-Sol (xhigh) iterated to convergence. It surfaced and drove fixes for several real issues in the broadcast/thread/DM handling (broadcast one-way ratchets on re-address, DM broadcast via D… ids/URLs, and error-ordering that reported red-herring lookup failures). A couple of findings were pre-existing repo-wide concerns already owned by other PRs (rendered usergroup/broadcast text → #113) or resolved by rebasing onto current main (W-prefixed user IDs → #114).

Closes #83.

Credit: Slack drafts.* API layer originally by @mikaelq in #112.

🤖 Reviewed with GPT-5.6-Sol; generated with Claude Code.

Made with Orca 🐋