fix(gateway): preserve multiplex profile in model picker

[cresslank]

What does this PR do?

Preserves one exact multiplex profile identity across the gateway /model picker: adapter selection, deferred callback credential resolution, and global config read/write.

On current main, the picker reads self.adapters directly, so a named-profile event can present through the default profile's adapter. The stored on_model_selected callback also runs after the routed profile scope has exited, and --global can seed a named-profile write from the default profile's config.

This resolves the adapter with _adapter_for_source(source), captures the routed profile home once, and re-enters that scope when the deferred callback runs. The same captured home selects the config file used for persistence.

Related Issue

Follow-up to #58587. Complements #61488 and #48187; it does not duplicate their session-default or context-length policy changes.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• gateway/slash_commands.py: route the picker through the profile-aware adapter registry.
• gateway/slash_commands.py: capture the routed profile home for deferred callback execution and profile-local config persistence.
• tests/gateway/test_model_picker_persist.py: exercise the production handler, profile secret scope, routed adapter registry, and profile-local config write.

How to Test

1. Use a temporary HERMES_HOME with plugins.enabled: [] so a user plugin cannot mask core behavior.

2. On unpatched current main, run the two new multiplex picker tests: both fail because the default adapter captures the picker.

3. On this branch, run:

   python -m pytest -q \
     tests/gateway/test_model_picker_persist.py \
     tests/gateway/test_multiplex_adapter_registry.py \
     tests/gateway/test_model_command_custom_providers.py \
     tests/gateway/test_model_command_flat_string_config.py

   Result: 23 passed.

4. Additional validation:

   ruff check gateway/slash_commands.py tests/gateway/test_model_picker_persist.py
   git diff --check origin/main...HEAD
   python scripts/check-windows-footguns.py --all

   All clean.

Not in Scope

• Messaging-platform /model session-default policy; fix(gateway): default /model to session scope on messaging platforms #61488 covers that.
• Route-scoped context-length resolution; fix(gateway): scope session-only /model context_length to configured route #48187 and the model.context_length override leaks onto sessions running a different model #62152 PR cluster cover that.
• Per-model/provider reasoning policy.
• First-completion backend readiness or rollback.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Fedora Linux 44

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — N/A; behavior is covered by code comments and tests
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — Windows footgun scan is clean
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

Unpatched current main + new regressions: 2 failed
Patched focused suite:                     23 passed
Ruff:                                      clean
Windows footguns:                          clean

[cresslank]

Rebased onto current main (9df5f879b) and force-pushed the PR branch as a single commit: 60dcc1d68. The rebase applied cleanly.

Validation after rebase:

• focused gateway model-picker suite with an isolated HERMES_HOME and plugins.enabled: []: 23 passed
• Ruff on both changed Python files: clean
• Windows footgun scan: clean (762 files)
• git diff --check: clean

All GitHub checks now pass (23 successful, 7 expected skips, 1 neutral OSV check), and the PR reports MERGEABLE / CLEAN. This should be current for review.

[teknium1]

Thanks for the focused multiplex picker fix. The current-main premise is real: gateway/slash_commands.py:1506 selects from the default adapter map, while secondary adapters live in the profile registry.

Problems

• Blocking: the initial picker setup remains outside the routed profile scope. Production dispatch calls _handle_model_command() directly at gateway/run.py:9930; _load_gateway_config() uses the default home unless an override scope is active (gateway/run.py:2335-2355). The new scope only wraps the deferred callback, so a named profile can still build its picker from default-profile config/providers.

Suggested changes

• Scope the initial config/provider-list work with the captured profile home as well as the callback.
• Update tests/gateway/test_model_picker_persist.py to invoke production dispatch without manually entering _profile_runtime_scope; the current test does so at PR line 310 and masks this gap.

Automated hermes-sweeper review.

[teknium1]

Capturing the routed home is not sufficient for the initial picker setup: production dispatch reaches this handler without _profile_runtime_scope (gateway/run.py:9930), so the subsequent _load_gateway_config() still reads the default home. Please scope the initial config/provider-list work with this captured home too; the deferred callback scope only fixes the later selection phase.

[teknium1]

Merged via PR #65700 — your commit was cherry-picked onto current main with your authorship preserved in git log (rebase merge). All three gaps you found were confirmed live: the picker's default-adapter-map straggler, the unscoped selection callback, and --global writing the wrong profile's config. The tests were excellent — real _profile_adapters resolution with byte-identical default-config assertions. Thanks!