feat: interactive discli setup wizard

[DevRohit06]

Closes #35.

Replaces the five manual steps in README §Setup with one interactive command.

discli setup - press Ctrl-C at any point to stop.

1. Bot token
   Paste your bot token:
   OK - helperbot (ID: 1298374), in: Acme HQ
   Saved to ~/.discli/config.json

2. Invite the bot to a server
   Grant messaging? (Read and send messages, embeds, files, reactions, polls, threads) [Y/n]: y
   Grant moderation? (Delete and pin messages, kick, ban, timeout, rename members) [y/N]: y
   Grant voice? (Join voice channels, speak, move members between them) [y/N]: n
   Grant admin? (Create/edit channels and roles, webhooks, invites, emoji, events, audit log) [y/N]: n
   Open this URL to invite the bot:
   https://discord.com/oauth2/authorize?client_id=1298374&permissions=2815883772816454&scope=bot+applications.commands
   Privileged intents cannot be granted by a URL. If you need them,
   enable them under Developer Portal > Bot > Privileged Gateway Intents:
     Message Content - message text in `discli listen` and `discli serve`
     Server Members  - `discli member list` and name lookups

3. Permission profile
   * full: Full access to all commands
     chat: Messages, reactions, threads, typing, interactions only
     readonly: Read-only: list, info, get, search, listen
     moderation: Moderation, voice, and interactions. No structural or security changes.
   Profile [full]: moderation
   Active profile: moderation
   Note: `discli setup` requires the full profile, so re-run it with:
     discli --profile full setup

4. Voice (optional)
   Voice extras are not installed. Set voice up now? [y/N]: n
   Skipped.

5. Verify
   <doctor's own report>

Two gates before any prompt

enforce_profile() first. setup writes the token and sets the active profile, so it is config set + permission set fused and must be full-only. Without it a readonly agent runs the wizard and promotes itself to full — the same hole moderation had with permission set. Denying before prompting also matches confirm_destructive(): a forbidden command that answers "needs a terminal" first has told the caller to go find a terminal for something it was never allowed to do.

Worth knowing: --profile widens as well as narrows, so this stops an agent handed a restricted default, not one that can pass arbitrary global flags. Same guarantee config set already has, no stronger.

Then the terminal check. discli is built to be driven by agents over a pipe, so blocking on a prompt is a hang, not an error.

$ discli --json setup
setup is interactive and needs a terminal.
Non-interactive equivalents:
  discli config set token <TOKEN>
  discli permission set <profile>
  discli doctor --json
$ echo $?
1

The refusal goes to stderr, so stdout stays parseable.

Design notes

• The token is validated before it is saved. Saving first leaves a broken token in config.json that every later command picks up, so the failure surfaces far from the paste that caused it. The login also reports which bot the token belongs to — pasting a different application's token otherwise succeeds silently and quietly retargets every command.
• Nothing is persisted for voice. tts.py/stt.py read credentials from the environment only, so the step prints export/setx lines rather than writing a file discli would never consult. Platform-aware, and the placeholder is shell-quoted (<your-key> unquoted is a redirection).
• Reuse over reimplementation. run_rest_action/enforce_profile (client.py), save_config (config.py), DEFAULT_PROFILES/set_active_profile (security.py), _gather/_format_text/_voice_extras_installed/DISCLI_PERMISSIONS (doctor.py). Step 5 prints doctor's report verbatim rather than offering a second opinion.
• Privileged intents are called out explicitly, because no invite URL can grant them and a missing one fails the whole Gateway connection.

Tests

15 new tests in tests/test_setup.py; suite is 351 passed, 1 skipped.

Built test-first. Three times a test passed without proving anything, so the production code was inverted to watch it fail properly — the no-save-on-bad-token test only earned its keep once it printed {'token': 'bad3'}, and the bundle/doctor sync test once it was fed a bogus permission.

Two seams worth flagging for review:

• _is_interactive() exists as a function because CliRunner never supplies a tty stdin. Without that seam no prompt path is testable and the guard tests pass for the wrong reason.
• The wizard fixture pins _voice_extras_installed and security.PERMISSIONS_PATH; otherwise the tests branch on whether the developer happens to have voice installed and on their own ~/.discli.

setup is also added to test_only_full_may_reconfigure_discli, and BUNDLES is pinned against doctor's DISCLI_PERMISSIONS so the invite the wizard generates and the permissions doctor checks for cannot drift apart.

Reviewer judgment calls

• _active_profile_name() duplicates a few lines from cli.py's permission_show. Extracting it into security.py is the cleaner fix but widens this change; left local deliberately. Happy to change.
• The doctor helpers imported here are underscore-private. Imported as-is rather than renamed, to avoid churning doctor.py and its tests for no behavior change.
• docs/guides/cli-usage.mdx is untouched — it doesn't document doctor either, so this follows that precedent.

Out of scope

Creating the Discord application itself (no API for it), writing to shell profiles, and storing provider API keys.

Unrelated drift noticed, not fixed

CLAUDE.md describes tts.py as having "ElevenLabs and OpenAI implementations" — it also has Deepgram Aura (get_tts_provider accepts all three).

🤖 Generated with Claude Code

[DevRohit06]

Code review found 15 issues; 14 were real. All fixed in 44bc2b1, each with a test watched failing first. Suite is now 367 passed, 1 skipped (up from 351).

Bugs

	Fix
The invite failed the wizard's own step 5. The moderation bundle granted manage_messages without bypass_slowmode — precisely the regression PERMISSION_SPLITS reports. setup printed an invite URL and four steps later told you to re-invite the bot.	bypass_slowmode added to the bundle and to doctor's DISCLI_PERMISSIONS.
setup was schedulable. parse_action("setup") was accepted; a fire blocks forever on click.prompt in the scheduler's worker thread — last_run never written, entry wedged silently.	New _NEEDS_A_HUMAN set with its own message (_NON_TERMINATING's "runs indefinitely" is the wrong reason).
An active custom profile locked the prompt. The local name reader ignored the profiles map get_active_profile() honours, so the default was a click.Choice value Click rejects — and show_choices=False never revealed the valid set.	Custom profiles are selectable; reader moved to security.py.
9 of 16 tests failed under DISCLI_PROFILE or a restricted permissions.json, since enforce_profile denies setup before the terminal check.	Autouse fixture neutralises both.
Exited 0 while printing [FAIL] lines, so discli setup && echo ready said ready for a broken install.	Exits like doctor does.
discli setup > log hung silently. _is_interactive() checked stdin only, but click writes prompts to stdout — the exact failure the guard exists to prevent.	Both streams checked.

Contract violations against this branch's own README

• Voice prompts hardcoded default="none", so re-running offered to un-configure a working setup — against the documented "every step shows what is already configured". Now defaults to the exported DISCLI_TTS/DISCLI_STT.
• The skip existed only when extras were missing, so anyone who installed voice for voice play alone was marched through both provider prompts. The step is titled "(optional)"; now it behaves that way.
• Saving a token while DISCORD_BOT_TOKEN is set reported OK, but cli.py resolves the envvar first. Now warned at the point the choice is overridden, not only before the prompt.

Cleanups

_invite_url delegates to discord.utils.oauth_url. _identify caps the guild lookup at 10 instead of paginating every guild of a bot in hundreds to print one line. get_active_profile_name()/get_profiles() moved into security.py, replacing the third copy in setup and the inline copy in permission_show. Export dialect is chosen by shell, not os.name — which is "nt" in Git Bash too — and the Windows branch no longer says "add to your shell profile" for setx, which doesn't affect the current console.

The bundle/doctor sync test now asserts both directions and covers PERMISSION_SPLITS — the gap that let the bypass_slowmode bug through in the first place.

Not taken

The review flagged discli --profile full setup (printed as recovery advice at step 3) as a bypass of the full-only restriction. It is, but it is pre-existing — --profile widens for config set and permission set identically — and without it a user who picks a restricted profile is locked out of the wizard with no signposted way back. Keeping it. Worth a separate discussion about whether --profile should be allowed to widen at all.

Found, not fixed here

DISCLI_PROFILE=readonly uv run pytest tests/ fails 47 pre-existing tests across other files, same root cause as the isolation bug fixed here. Out of scope for this branch — the fix is an autouse fixture in conftest.py alongside no_network. Happy to open a separate issue.