feat(discord): add /auth slash command for device-flow auth

[chaodu-agent]

Summary

Adds /auth Discord slash command that executes $OPENAB_AGENT_AUTH_COMMAND and relays the device-flow URL + code back to the user via ephemeral DM.

Flow Diagram

┌─────────────┐         ┌──────────────┐         ┌────────────────────┐
│  Discord    │         │     OAB      │         │  Auth CLI          │
│  (User DM)  │         │  (Handler)   │         │  (e.g. codex       │
│             │         │              │         │   login --device)  │
└──────┬──────┘         └──────┬───────┘         └─────────┬──────────┘
       │                       │                           │
       │  /auth                │                           │
       ├──────────────────────▶│                           │
       │                       │                           │
       │  [access check]       │                           │
       │  [DM-only check]      │                           │
       │  [single-flight]      │                           │
       │                       │                           │
       │  defer (ephemeral)    │                           │
       │◀──────────────────────┤                           │
       │                       │                           │
       │                       │  sh -c $AUTH_CMD          │
       │                       ├──────────────────────────▶│
       │                       │                           │
       │                       │  stdout: URL + code       │
       │                       │◀──────────────────────────┤
       │                       │                           │
       │  🔐 URL + code        │                           │
       │◀──────────────────────┤                           │
       │                       │                           │
       │                       │      (process blocks,     │
       │  [user opens URL,     │       polling OAuth       │
       │   enters code in      │       server...)          │
       │   browser]            │                           │
       │                       │                           │
       │                       │  exit 0                   │
       │                       │◀──────────────────────────┤
       │                       │                           │
       │  ✅ Authenticated!    │                           │
       │◀──────────────────────┤                           │
       │                       │                           │

Security Controls

/auth invoked
  │
  ├─ is_denied_user? ──▶ 🚫 rejected
  │
  ├─ guild_id.is_some()? ──▶ 🔒 DM-only
  │
  ├─ AUTH_IN_PROGRESS? ──▶ ⚠️ already running
  │
  ├─ OPENAB_AGENT_AUTH_COMMAND unset? ──▶ ⚠️ not configured
  │
  └─ ✅ proceed with deferred ephemeral + spawn

Architecture

• Drain tasks: Separate tokio::spawn for stdout/stderr → run to EOF (no SIGPIPE)
• URL detection: tokio::sync::Notify (cancellation-safe)
• Single-flight: static AtomicBool + Drop guard (panic-safe)
• DM check: cmd.guild_id.is_some() (local field, no API call)
• Truncation: chars().take() (UTF-8-safe, Discord 2000-char limit)
• Timeout: 14 min (leaves headroom for Discord interaction token TTL)
• Tracing: info/warn at key lifecycle points

Changes

• crates/openab-core/src/discord.rs — Register + implement /auth
• docs/slash-commands.md — Full documentation

Review History

6 commits, 5 review rounds, 3 reviewers (擺渡/口渡/覺渡/普渡), all LGTM ✅

[chaodu-agent]

Consolidated Group Review — /auth Slash Command

Verdict: functionally mergeable — maintainer's call. All blocking, documentation, and test findings are addressed. One non-blocking hardening item (device-code capture timing) is accepted as a tracked follow-up by the maintainer. Latest head c00b417, CI green (33 pass / 1 intentional skip / 0 fail); cargo test --workspace passes including the new truncation tests.

The design is sound — DM-only + ephemeral + single-flight (AuthGuard Drop) + env-only command (no user input into sh -c), with bot-user rejection and an audit log of the invoking user.

Round 1 findings & resolution

#	Severity	Finding	Status
1	🟠	Fast-fail 30s stall — URL-collection select! had no child.wait() arm.	✅ Fixed — added child.wait() early-exit arm
2	🟠	500ms trailing-capture may truncate a device code printed >500ms after the URL.	🔭 Tracked follow-up (accepted)
3	🟡	Docs intro contradicted /auth DM-only.	✅ Fixed
4	🟡	No regression tests for /auth.	✅ UTF-16 truncation extracted + unit-tested; security-gate-path tests tracked as follow-up
5	🟡	Discord 2000-char limit is UTF-16 code units, not Unicode scalars.	✅ Fixed — UTF-16-aware truncation (now unit-tested)
6	🟡	"No output" error gave no cause/remedy.	✅ Fixed
7	🟡	Audit log missing invoking user ID.	✅ Fixed
8	🟡	No bot rejection (inconsistent with /remind).	✅ Fixed
9	🟡	std::sync::Mutex poison cascade in drain tasks.	✅ Fixed
10	🟡	AUTH_IN_PROGRESS is process-wide single-flight (by design).	✅ Documented as intended
11	⚪	SeqCst stronger than needed.	✅ Fixed — Acquire/Release

Round 2 findings & resolution

#	Severity	Finding	Status
R1	🟡 (bug)	Early-exit on exit-0-without-URL showed "✅ Auth command completed.", implying false success.	✅ Fixed — now a ⚠️ "run /auth again" prompt
R2	🟡	Docs "Error cases" missing the early-exit-without-URL path.	✅ Fixed
R3	🟡	Docs missing bot-rejection / DM-only error cases.	✅ Fixed

Round 3 (post-review hardening)

• ✅ Extracted the UTF-16 truncation into a pure truncate_to_utf16_budget() helper and added unit tests (short body, prefix/suffix budgeting, surrogate-pair counting, no-scalar-splitting, zero-budget saturation, assembled-total-within-limit regression guard). Runs under cargo test --workspace.
• ✅ Added the missing tracing::error on the final child.wait() Ok(Err(e)) arm (symmetry with the spawn-failure path).

Not a bug (verified, no action)

• "Spawned task panic doesn't reset the flag" — let _guard = AuthGuard; is the first statement in the spawned task; its Drop clears AUTH_IN_PROGRESS. No panic = "abort", so unwinding drops _guard on panic.
• "Timeout branch doesn't join drain tasks" — the final tokio::join! runs after the match for all branches; the empty-output path joins before returning. Drain tasks are always joined.

Remaining follow-up (non-blocking, accepted)

• The 500ms trailing-capture heuristic (round-1 perf: cache deps layer + drop arm64 QEMU build #2): consider "wait for N lines after URL" or idle-based capture for slow CLIs that print the code well after the URL.
• Dedicated tests for the access-control gate paths (bot rejection, is_denied_user, DM-only, AUTH_IN_PROGRESS single-flight). The UTF-16 truncation logic is now unit-tested; the gate logic remains a good follow-up candidate.

Praise 🟢

• AuthGuard Drop makes single-flight panic-safe; no command injection (env-only command, no user input to sh -c).
• Layered access control: bot rejection → allowlist → DM-only → single-flight → env guard.
• Drain tasks run to EOF (no SIGPIPE); ephemeral + DM-only minimizes credential exposure; 14-min timeout leaves headroom under Discord's token TTL.
• UTF-16-aware truncation (now tested), poison-safe locks, and Acquire/Release ordering are correct.

Final merge decision is the maintainer's.