fix(cli): make configure inputs a typed request so first-run setup can't crash

[elyasmnvidian]

The first time you run switchyard launch claude (or codex/openclaw) without a saved setup, it runs a one-time setup: it asks for your provider, model, and API key and writes them to a config file so later launches don't need those flags. That setup crashed before it could write anything:

Switchyard is missing claude launch defaults. Starting setup.
AttributeError: 'Namespace' object has no attribute 'disable_skill_distillation'

Why it happened

cmd_configure (the code behind switchyard configure) read its inputs straight off an argparse Namespace. Two callers built that namespace: the configure CLI command (argparse fills every field), and launch setup, which hand-built a namespace with ~30 fields. The two drifted — launch setup was missing three fields the command reads, so the command hit a missing attribute and stopped.

The fix

Give cmd_configure a real type. A new ConfigureRequest dataclass lists every field the command reads and its default, in one place. Both callers build that same type:

• the configure command via ConfigureRequest.from_namespace(args) (the one spot that touches the raw argparse namespace),
• launch setup via the constructor, passing only the handful of fields it knows and letting the rest take their defaults.

Now a caller can't leave out a field the command needs — a missing or misspelled field is a mypy error or fails right where it's built, instead of crashing mid-launch. The change also lets us drop ~11 defensive getattr(args, ...) reads and 15 dead fields that launch setup was setting for no reason.

What changed

• New switchyard/cli/configure_request.py — the ConfigureRequest dataclass and its from_namespace builder.
• cmd_configure and its helpers take ConfigureRequest instead of a bare namespace.
• Launch setup builds a ConfigureRequest instead of a hand-rolled namespace.
• switchyard configure behavior is unchanged.

Test

The old setup test replaced cmd_configure with a stub, so it never noticed the missing fields — that's how the crash shipped. Now test_configure_parser_builds_a_complete_request runs a real configure parse through from_namespace and checks the field that crashed is present, and the bootstrap test asserts the request setup builds works with the real configure helpers.

Live run — NVIDIA Inference Hub, real tokens

Real setup path with a fresh config dir (--reconfigure forces setup):

switchyard launch claude --reconfigure --model azure/anthropic/claude-haiku-4-5 --base-url https://inference-api.nvidia.com/v1 --api-key $NVIDIA_API_KEY --no-tui --no-model-discovery

Before (fix reverted):

Switchyard is missing claude launch defaults. Starting setup.
AttributeError: 'Namespace' object has no attribute 'disable_skill_distillation'

After:

Switchyard is missing claude launch defaults. Starting setup.
Saved Switchyard config to <cfg>/config.json
Saved Switchyard credentials to <cfg>/credentials.json
  switchyard  ready  →  azure/anthropic/claude-haiku-4-5
  ...
  READY
  requests : 1
  tokens   : 80,072 in  5 out

Setup finishes, writes the config and credentials, the proxy starts, and one claude turn reaches the model through the NVIDIA gateway for a real reply (80,072 in / 5 out).

Saved config.json (the API key goes to a separate credentials.json):

{
  "default_provider": "nvidia",
  "launch": {
    "claude": {
      "model": "azure/anthropic/claude-haiku-4-5",
      "route": {"model": "azure/anthropic/claude-haiku-4-5", "type": "single"}
    }
  },
  "providers": {"nvidia": {"base_url": "https://inference-api.nvidia.com/v1"}}
}

Fixes #60.

[coderabbitai]

Walkthrough

First-run launch bootstrap now initializes the routing-profile and skill-distillation fields expected by interactive configuration. A regression test verifies these fields and the related helper behavior.

Changes

Launch bootstrap configuration

Layer / File(s)	Summary
Bootstrap fields and regression coverage switchyard/cli/launch_command.py, tests/test_launch_claude.py	The bootstrap namespace now includes routing-profile and skill-distillation values, while the launch test validates the corresponding helper calls and defaults.

Estimated code review effort: 2 (Simple) | ~10 minutes

Poem

A rabbit hops through setup’s gate,
No missing fields to make it wait.
Profiles rest and skills stay still,
The launch now climbs the config hill.
“Bootstrap bright!” the bunny sings.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The change adds the missing bootstrap fields and regression coverage needed to prevent the first-run AttributeError in #60.
Out of Scope Changes check	✅ Passed	The added fields and test updates are directly tied to the launcher bootstrap crash and stay within the issue scope.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the main fix: typed first-run configure inputs to prevent launch bootstrap crashes.

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/test_launch_claude.py`:
- Around line 115-129: Extend the bootstrap namespace assertions in the launch
regression test to directly validate routing_profiles, skill_distillation, and
disable_skill_distillation on configure_args. Keep the existing helper
assertions, and assert each field has the expected default value consumed by
cmd_configure.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: bc4b681e-cc3c-44ad-b6c5-79f8e2348060

📥 Commits

Reviewing files that changed from the base of the PR and between 8b8dad8 and a61f93a.

📒 Files selected for processing (2)

• switchyard/cli/launch_command.py
• tests/test_launch_claude.py

[elyasmnvidian]

Tested this before and after the change — nothing regressed.

Full unit suite, before vs after (same machine and env, run separately so they don't contend for sockets):

• Before (origin/main, without this commit): 1996 passed, 2 failed
• After (this branch): 1997 passed, 2 failed

The two failures are identical on both branches, so they're pre-existing and unrelated to this change: a flaky routing e2e test (test_evict_reroute_e2e) and a verify test that picks up a real secrets file on this box (test_verify…secrets_file_wins). The one extra pass on this branch is the new regression test that runs a real configure parse through ConfigureRequest.from_namespace. Every test that was green before is still green.

Live runs with real tokens (model azure/anthropic/claude-haiku-4-5, one headless claude turn each):

1. First-run setup — the path this fix actually changes. Fresh config dir, --reconfigure forces setup:

Switchyard is missing claude launch defaults. Starting setup.
Saved Switchyard config to <cfg>/config.json
Saved Switchyard credentials to <cfg>/credentials.json
  switchyard  ready  →  azure/anthropic/claude-haiku-4-5
  ALPHA
  requests : 1
  tokens   : 52,039 in  5 out

Setup finishes, writes the config and credentials, the proxy starts, and the model replies for real.

2. Normal launch with credentials on the command line (setup is skipped) — confirms the common path still works:

  switchyard  ready  →  azure/anthropic/claude-haiku-4-5
  BRAVO
  requests : 1
  tokens   : 89,587 in  6 out

Both round-tripped real tokens through the gateway and returned the exact one-word reply asked for.