[Sadykhzadeh]

Closes #176. All changes are in core/src/connection.rs and its tests; session.rs is untouched, so this doesn't overlap the #140 stack.

Sign-in

• Which methods the server accepts: a Password host first sends a none request, as OpenSSH does, to learn the server's methods.

  • If password is offered, it is used.
  • Otherwise keyboard-interactive is used.

• Keyboard-interactive is answered the way OpenSSH answers a PAM password prompt:

  • The password goes only to a round with a single prompt that isn't echoed, and only once.
  • Rounds with no prompts get an empty reply.
  • A round asking for more (a one-time code, several fields, a visible prompt) is refused without sending the password. The error names the prompts: the server asks for more than the password ("Password:", "Verification code:"); Clavyn cannot answer additional sign-in prompts such as one-time codes yet.
  • At most 5 rounds, all within the existing auth timeout.

• A rejected password is not retried through keyboard-interactive. Both methods check the same credential, so a second attempt can't succeed. It would only count twice against MaxAuthTries and fail2ban.

• Why a sign-in failed: the message keeps the authentication rejected by server: prefix, then says one of:

  • the server rejected the password (it accepts: password, publickey)
  • the server does not accept password sign-in (it accepts: publickey)
  • the server accepted this step but requires another: publickey; Clavyn cannot complete multi-step sign-in yet

  The frontend only displays this text and never matches on it.

Connecting

• Dialing: the name is resolved with lookup_host, and each address gets its own 5 s in resolver order. The overall 20 s cap still covers resolve, dial and the handshake. The stream then goes to russh::client::connect_stream, so host key checking is unchanged.
• A failed dial lists every address and its reason, e.g. [2001:db8::1]:22: timed out after 5s; 192.0.2.7:22: Connection refused.
• TCP_NODELAY is set on the dialed socket. russh ignores config.nodelay for a stream it didn't open itself.

Tests

Each fails against main (or, for the resend test, against the earlier commit on this branch):

• Loopback sign-in tests (ssh_roundtrip.rs):
  • a_password_is_answered_through_keyboard_interactive
  • a_second_factor_prompt_is_refused_without_sending_the_password
  • a_rejected_password_is_not_sent_a_second_time
  • a_rejected_password_names_the_methods_the_server_accepts
  • a_server_without_password_sign_in_says_so
  • a_partial_success_is_reported_as_a_further_step
• Dialer tests (connection.rs):
  • a_silent_address_does_not_hold_up_the_next_one: a black-hole address, then a loopback listener; must finish in under 10 s.
  • dialed_streams_disable_nagle
  • a_failed_dial_names_every_address

Trade-offs and limits

• The none probe costs one round trip per password sign-in.
• One-time codes and other MFA prompts still can't be answered; the user gets a clear message instead of a generic failure. Answering them needs prompts in the UI, and the auth timeout would have to stop covering the time the user spends typing.
• Tested only against russh's in-process server, not a real OpenSSH or FreeBSD sshd.

Test plan

• cargo test -p clavyn-core: 121 + 5 + 32 + 15 passed
• cargo clippy -p clavyn-core --all-targets: nothing new
• node scripts/verify/check-comments.mjs: ok
• No UI change.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[greptile-apps]

TREX could not finish testing

• Plan what to test
• Set up the environment and test the changed flows
• Post test results

[qodo-code-review]

PR Summary by Qodo

Support keyboard-interactive SSH sign-in and bounded address dialing

🐞 Bug fix ✨ Enhancement 🧪 Tests 🕐 40+ Minutes

AI Description

• Support password sign-in through keyboard-interactive without resending rejected credentials or
 answering additional-factor prompts.
• Explain authentication rejections using the methods and next steps reported by the server.
• Bound each address dial, report individual failures, and disable Nagle for interactive
 connections.

Diagram

graph TD
  Resolve["Resolve host"] --> Dial["Bounded TCP dial"] --> SSH["SSH handshake"] --> Choice{"Host auth type"} --> Password["Probe password methods"] --> Outcome["Session or error"]
  Password -->|interactive only| Interactive["Answer safe prompts"] --> Outcome
  Choice -->|key host| PublicKey["Public key auth"] --> Outcome

Loading

High-Level Assessment

The following are alternative approaches to this PR:

1. Race addresses with Happy Eyeballs

• ➕ Can reach a working IPv4 address sooner when IPv6 connections stall.
• ➖ Adds concurrent connection management and departs from strict resolver-order dialing.

Recommendation: Keep the PR's sequential, per-address timeout approach: it fixes starvation within the overall connection cap while preserving predictable attempt order. Consider Happy Eyeballs separately if the five-second delay before a fallback address proves problematic.

Files changed (2) +474 / -16

Bug fix (1) +258 / -13

connection.rsBound address dialing and broaden password authentication +258/-13

Bound address dialing and broaden password authentication

• Resolves and dials addresses with individual five-second bounds, enables TCP_NODELAY, and passes the stream to russh under the existing overall timeout. Password hosts discover supported methods before authenticating, safely handle keyboard-interactive prompts, and report server-provided rejection details. Unit tests cover address fallback, socket settings, and dial failures.

core/src/connection.rs

Tests (1) +216 / -3

ssh_roundtrip.rsExercise SSH sign-in decisions against loopback servers +216/-3

Exercise SSH sign-in decisions against loopback servers

• Generalizes the in-process server helper and adds sign-in modes for keyboard-interactive passwords, additional prompts, rejected credentials, unavailable methods, and partial success. Roundtrip tests verify both successful sign-in and that unsafe or duplicate password responses are not sent.

core/tests/ssh_roundtrip.rs

[qodo-code-review]

Code Review by Qodo

View medium (1)

1. Password copies remain in memory after sign-in 🐞 ⭐

Description

keyboard_interactive_password creates an ordinary owned String with password.to_owned() and
passes it to russh without a way to wipe that copy. Every keyboard-interactive password response
therefore leaves an additional secret allocation to be freed without zeroization, unlike the
deliberately wiped private-key buffer used in the same connection path.

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Keyboard-interactive sign-in creates an owned password response that is dropped without zeroization.
## Fix Focus Areas
- core/src/connection.rs[394-415]
## Recommended Fix
Ensure the owned response is wiped after russh serializes it. Because the response API takes ownership of `Vec<String>`, use or add an API that zeroizes the consumed response rather than wiping only a separate local copy.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

Dismiss ↗ | View ↗