fix: desktop model picker hid Anthropic despite valid Claude Code credentials

[adurham]

What does this PR do?

Fixes the desktop model picker hiding Anthropic even when the user has
valid Claude Code CLI or Hermes-managed OAuth credentials — while the CLI
/model picker shows it correctly on the exact same credentials.

Root Cause

_filter_explicit_provider_rows() in hermes_cli/inventory.py (the
desktop-only explicit_only=True picker filter) re-checks every row
against is_provider_explicitly_configured(), which deliberately
excludes CLAUDE_CODE_OAUTH_TOKEN / external Claude Code credential files
(#4210 — prevents auxiliary tasks from silently burning the user's Claude
Code subscription tokens without an explicit choice). That gate is
correct for aux-task consumption, but this desktop-only filter was also
using it to decide what to display, silencing a provider row that
list_authenticated_providers() (used by both the CLI and desktop
pickers, upstream of this filter) already surfaces.

Confirmed the premise directly by reading the code: list_authenticated_providers()
in hermes_cli/model_switch.py already has this exact fallback (~line
2126: if not has_creds and hermes_slug == "anthropic":, checking
read_hermes_oauth_credentials()/read_claude_code_credentials() from
agent/anthropic_adapter.py, with a comment citing the same #4210
rationale). This PR extends an already-established, already-documented
upstream pattern to the desktop-only filter that wasn't updated to match
it — not new credential-detection logic.

Changes Made

hermes_cli/inventory.py: when is_provider_explicitly_configured("anthropic")
is False, also check for valid external Claude Code / Hermes-PKCE
credentials before dropping the row from the desktop picker's display.
is_provider_explicitly_configured() itself is untouched, so aux-task
gating is unaffected — this only widens what the desktop picker shows.

Related Issue

No existing issue found (searched gh search issues for "desktop model
picker missing anthropic", "Claude Code credentials not shown picker",
"explicit_only provider filter" — no hits).

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

How to Test

1. Authenticate via Claude Code CLI or Hermes-managed OAuth (no
   ANTHROPIC_API_KEY set, no explicit Hermes-side Anthropic
   configuration).
2. Open the CLI /model picker — Anthropic/Claude models appear.
3. Open the Desktop model picker on the same credentials.
4. Before this fix: Anthropic is missing from the desktop picker.
5. After this fix: Anthropic appears, matching the CLI.
6. pytest tests/hermes_cli/test_inventory.py -q — 46 passed.
7. Also ran adjacent credential-detection test files to check for
   regressions: test_inventory_pricing.py,
   test_model_switch_custom_providers.py,
   test_model_switch_configured_provider_routing.py — 117 passed total,
   0 failed.

Checklist

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs/issues (none found)
• My PR contains only changes related to this fix
• I've run the relevant test suite and all tests pass
• I've added tests for my changes
• Tested on macOS

[teknium1]

Thanks for the focused fix. Current main still has the reported mismatch: hermes_cli/model_switch.py:2298-2308 includes Anthropic when external Claude Code/Hermes OAuth credentials provide an access token, while hermes_cli/inventory.py:581-613 subsequently removes that same non-explicit row for Desktop’s explicit_only response. Desktop defaults to that response mode at apps/desktop/src/hermes.ts:1448-1449.

The proposed display-only carve-out matches the existing CLI discovery behavior without changing the intentional auxiliary-task gate in hermes_cli/auth.py:1661-1731. The positive and negative inventory tests cover the new filter branch. GitHub currently reports the branch as conflicting with main, but the verified production change is narrow and mechanically salvageable.

Automated hermes-sweeper review.

[adurham]

Rebased onto current main (was showing as conflicting). The only real conflict was in tests/hermes_cli/test_inventory.py, where an unrelated test-pruning pass upstream had removed 3 pre-existing tests my diff's context lines touched (test_include_unconfigured_keeps_current_provider_visible_without_credentials, test_include_unconfigured_does_not_duplicate_configured_current_row, test_explicit_only_keeps_moa_when_raw_config_has_enabled_preset) -- resolved by not resurrecting them, just cleanly inserting this PR's 2 new tests. hermes_cli/inventory.py itself auto-merged with zero conflicts against the intervening refactor(cli): route every aux picker through one provider-inventory seam work.

Verified after rebase: tests/hermes_cli/test_inventory.py 16/16 passed, plus ran the adjacent inventory/aux-picker/model-switch suites for regressions (59/59 passed, 0 failed) given the seam refactor landed in between.

[teknium1]

Closing with a credit correction: you were the first to submit this fix — filed July 26, nearly a month before the equivalent #92391 (Aug 22). Our post-merge duplicate sweep surfaced yours only after #92702 had already merged the later PR's commit, which independently re-derived the same root cause and the same carve-out approach (display-only filter fix, is_provider_explicitly_configured() / #4210 aux-task gating untouched).

The fix that landed on main matches your diagnosis exactly, plus one addition neither PR had: pool-only OAuth entries in auth.json credential_pool.anthropic (the canonical wired device-flow token location) are also accepted now.

Apologies for the credit miss — your PR was the earliest correct diagnosis of this bug, and this comment is the record of that. Thanks for the clean work.