fix(launch): guard zero-config default trio and report routing bundle in --dry-run

[elyasmnvidian]

Two launcher fixes.

Zero-config launch on a non-OpenRouter provider now stops before it launches.
With no --model and no --routing-profiles, switchyard launch claude (and codex/openclaw) uses the fixed default trio — strong anthropic/claude-opus-4.7, weak moonshotai/kimi-k2.6, classifier google/gemini-3.5-flash — which are all OpenRouter model ids. If your credentials resolve to a different provider (for example NVIDIA via NVIDIA_API_KEY), those ids don't exist on that endpoint. The old code launched anyway and only failed on the first request. Now the launcher exits first with a message that names the trio, the resolved provider, and the way out: pass --model for a single model or --routing-profiles FILE for a bundle, or set OPENROUTER_API_KEY.

launch --dry-run now labels a routing-profiles bundle as a bundle.
Running a dry run with --routing-profiles FILE printed route: single (the placeholder single-tier view). It now prints route: bundle, the bundle path, and the routable ids read from the YAML.

How to try it:

• NVIDIA_API_KEY=... switchyard launch claude → exits with the "OpenRouter-only" guidance instead of spawning.
• switchyard --routing-profiles routes.yaml -- launch claude --dry-run → prints route: bundle.

Regression tests for both live in tests/test_launch_claude_deterministic.py. The existing codex zero-flag test now points at an OpenRouter endpoint so it still exercises the deterministic dispatch path rather than tripping the new guard.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA-NeMo.github.io/Switchyard/pr-preview/pr-108/
Built to branch gh-pages at 2026-07-21 21:46 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[coderabbitai]

Walkthrough

The CLI now fails fast for zero-config deterministic launches using non-OpenRouter providers, while explicit model overrides remain supported. Dry-run output also reports routing-profile bundles for Claude, Codex, and OpenClaw.

Changes

Routing CLI behavior

Layer / File(s)	Summary
Deterministic provider validation switchyard/cli/launch_command.py, docs/guides/agent_launchers.md, tests/test_launch_claude_deterministic.py, tests/test_launch_codex.py	Deterministic zero-config launches require an OpenRouter provider, with explicit model overrides bypassing the guard; documentation and dispatch tests cover the behavior.
Routing-profile dry-run output switchyard/cli/output.py, switchyard/cli/launch_command.py, tests/test_launch_claude_deterministic.py	Dry-run formatting reports bundle routes and routing-profile IDs, and launch subcommands pass the routing-profile path through to the formatter.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Poem

A bunny checked the routes with care,
Found OpenRouter models waiting there.
Bundles bloomed in dry-run light,
Explicit models made things right.
Hop, hop—safe launches take flight!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the two main launcher changes: zero-config guardrails and dry-run routing bundle reporting.

---

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

[coderabbitai]

Actionable comments posted: 2

🧹 Nitpick comments (1)

tests/test_launch_claude_deterministic.py (1)

279-281: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Assert the bundle path and route ID too.

This passes if only route: bundle is rendered; assert routing profiles: {yaml_path} and my-route to cover the stated dry-run contract.

🤖 Prompt for 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.

In `@tests/test_launch_claude_deterministic.py` around lines 279 - 281, Strengthen
the assertions in the dry-run output test by verifying the bundle path includes
both the routing profiles value using yaml_path and the route ID my-route, in
addition to route: bundle; retain the existing assertion that route: single is
absent.

🤖 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 `@switchyard/cli/launch_command.py`:
- Around line 121-126: Update the provider-guard bypass condition to return only
when args.model is set; do not treat args.weak_model or args.classifier_model as
opt-outs. Preserve deterministic routing and provider validation whenever only
the partial model overrides are provided.

In `@tests/test_launch_codex.py`:
- Around line 509-511: Remove the project-management tracker reference “NvBug
6401771” from the source comment near the deterministic default dispatch
behavior, while preserving the explanation that the zero-flag default trio is
OpenRouter-only and non-OpenRouter providers are rejected.

---

Nitpick comments:
In `@tests/test_launch_claude_deterministic.py`:
- Around line 279-281: Strengthen the assertions in the dry-run output test by
verifying the bundle path includes both the routing profiles value using
yaml_path and the route ID my-route, in addition to route: bundle; retain the
existing assertion that route: single is absent.

🪄 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: 8bb30082-4f92-4343-88cc-ddbba579a886

📥 Commits

Reviewing files that changed from the base of the PR and between 2fa6b87 and a478815.

📒 Files selected for processing (5)

• docs/guides/agent_launchers.md
• switchyard/cli/launch_command.py
• switchyard/cli/output.py
• tests/test_launch_claude_deterministic.py
• tests/test_launch_codex.py

[elyasmnvidian]

Live before/after — launcher guard and dry-run bundle reporting

These are launcher-behavior fixes, so the reproduction is command output. Same commands on main and this branch.

1) launch --dry-run with a routing bundle — switchyard --routing-profiles bundle.yaml launch claude --dry-run:

# main
route: single

# this branch
route: bundle
routing profiles: /tmp/live/bundle_108.yaml
  fast-kimi
  moonshotai/kimi-k2.6
  random-router
  anthropic/claude-opus-4.7

2) Zero-config launch on a non-OpenRouter endpoint — switchyard launch claude --base-url https://inference-api.nvidia.com/v1 --api-key $NVIDIA_API_KEY (no --model):

• main: no such guard exists (0 occurrences in launch_command.py), so the launcher proceeds to start the OpenRouter-only trio against the NVIDIA endpoint and only fails on the first request.

• this branch — fails fast, before spawning:

  launch claude: the zero-config default trio (anthropic/claude-opus-4.7, moonshotai/kimi-k2.6, google/gemini-3.5-flash) is OpenRouter-only; the resolved provider is 'nvidia'. Pass --model, use --routing-profiles, or set OPENROUTER_API_KEY.
  

Audit follow-up in this push: the guard now rejects any deterministic launch on a non-OpenRouter endpoint. It previously skipped itself when --weak-model/--classifier-model was set, but the strong tier is always anthropic/claude-opus-4.7, so those flags don't make the trio reachable. The message and the launcher doc now name the real ways out (--model, --routing-profiles, OPENROUTER_API_KEY).