Provider reconfig: agent-config sync, Hermes One auto-provisioning, provider UI redesign & chat stream fixes

[fathah]

This branch makes providers behave as one system across the desktop, the bundled hermes-agent, and the Hermes One backend — plus a light-based redesign of the provider surfaces and two chat-streaming correctness fixes the work surfaced along the way.

🔑 Provider routing fixes

Native provider keys were invisible in the model picker. providerRouteForEnvKey only knew keys carried by a PROVIDERS.setup card or a local preset; everything else fell to a dead custom + empty-base-URL route, which the active-model picker silently drops. A configured Nous Portal API key never appeared in the "Select default model" modal. A new NATIVE_ENV_KEY_ROUTES table maps each orphaned key to its hermes-agent slug (NOUS_API_KEY → nous, GLM_API_KEY → zai, KIMI_API_KEY → kimi-coding, MINIMAX[_CN]_API_KEY → minimax[-cn], NVIDIA, OPENCODE_ZEN/GO, HF_TOKEN → huggingface, PERPLEXITY → custom + base URL), verified against the agent's own plugins/model-providers/* registries. A test asserts no LLM-section key can ever land on the dead fallback again.

Session switches to custom-endpoint models routed to the wrong provider. /model <id> --provider custom lets the agent bind custom to the session's current base URL — so switching to hermesone-swift while a session sat on Nous sent the request to the Nous proxy (HTTP 404 … OpenRouter catalog). Fixed at both ends:

• the desktop mirrors a keyed Hermes One into the agent's config.yaml as a named providers: hermesone: entry (base_url + key_env), so the gateway always has a slug that carries the right endpoint;
• resolveDashboardProviderForModel now matches any gateway provider row by base URL (named user providers included, not just custom:* rows) before ever falling back to bare custom.
• the config.yaml scaffold's inline providers: {} is rewritten into block form on upsert — the line-based parser couldn't index it, which silently disabled the mirror on real configs (and appending would have produced a duplicate top-level key).

🔄 Desktop ↔ hermes-agent provider sync

Named OpenAI-compatible providers now sync both ways with the agent's config.yaml via a new text-based bridge (src/main/agent-config-providers.ts) that preserves comments and unrelated keys:

• Terminal → desktop: providers: dict entries and legacy custom_providers: list entries added via the CLI are imported on every provider-list / model-library read — cards, picker entries, and model rows appear without reseeding. A terminal entry's custom key_env value is aliased additively to the desktop's derived CUSTOM_PROVIDER_<NAME>_KEY so keys resolve everywhere. Previously custom_providers: was only honored once, at first models.json seed.
• Desktop → terminal: adding/editing a custom provider in the UI upserts a matching providers: entry (patching fields in place — a terminal user's extra fields like transport: survive); removing a provider deletes it from both config.yaml shapes so it can't re-import.

🪪 Hermes One account: auto-provisioned key + credits

Signing in should yield model access without hand-copying keys:

• On device login (and whenever the Providers screen loads signed-in with no key), the desktop calls the backend's POST /api/credits/keys with the bearer token and stores the one-time hs-live-… key as HERMESONE_API_KEY — which by itself makes the Hermes One card and picker entry appear. Idempotent (an existing key is never replaced), single-flight (concurrent calls can't double-issue), and local-mode only (a remote/SSH profile would strand orphan keys).
• The account card now shows the account's AI-credit balance ($X.XX credits chip via GET /api/credits/balance).

🎨 Provider UI redesign (light-based, stroke-free)

• Cards & Add-provider picker: titles show the plain provider name ("Hermes One", not "Hermes One API Key" — labels aren't suffix-strippable across locales); well-known providers are ordered first (Hermes One leads, AIML API demoted last) via a stable rank sort; search matches both the name and the full label.
• Provider config modal: brand logo tile + name + live "Key verified" pill in the header; masked API key (hs-••••••••3k8d) with Show / Replace and a "Used by N models" meta line; models as compact chips with an inline "+ Add model ID" pill (autocomplete, Enter/blur commits) collapsing behind "+N more" past 10; hairline header/footer dividers, fills instead of strokes, 11–13px control type.
• Select-default-model modal: rebuilt stroke-free (filled controls, elevation-only dropdown) and — functionally — the provider dropdown is no longer clipped by the modal's overflow: hidden, so long provider lists actually scroll.
• Model Registry modal: the search field now shares the header/body inset instead of running edge-to-edge.
• Profile switcher: the switch modal is now a fixed-size, command-palette-style picker — Cmd/Ctrl+P from anywhere, fuzzy search over name/id/model, running vs Stopped groups, keyboard navigation — in the same stroke-free material as the sidebar session menu.

💬 Chat streaming correctness

Both fixes key on the same insight: text assembled from dropped chunks is, by construction, a subsequence of the canonical text.

• Duplicate thoughts: a lossy live reasoning row ("moon-k3 … ous") could never key-match its canonical DB row ("moonshotai/kimi-k3 … nous"), so both survived the merge and stacked in one Thought block. Reconciliation now drops a streamed reasoning row whose text is a subsequence of a same-turn DB reasoning row — while keeping genuinely distinct live segments and never cross-cancelling between turns.
• Duplicate answers: when the upstream tags alternate chunks as reasoning, the streamed bubble ends up a garbled subset of the final text, and mergeStreamedWithFinal concatenated the two ("! What are we working on?\n\nHey! What are we working on today?"). A guarded subsequence branch (≥12 chars, ≥30% coverage) replaces the garbled stream with the final — the pre-tool-call + answer pair still stacks.

🧪 Tests & docs

~60 new tests: env-key routing invariants, provider ordering, the config.yaml bridge (parse/upsert/remove/{}-scaffold round-trips), import/mirror/delete sync, key provisioning (single-flight, failure paths), dashboard provider resolution (incl. the Nous-404 regression), and the two subsequence-based stream reconciliations. All 1795 pass; lat.md/ updated throughout (provider-setup, hermes-account-login, chat-commands, sidebar-navigation) and lat check is green.

[greptile-apps]

Greptile Summary

This PR unifies provider behavior across the desktop, Hermes Agent, and Hermes One. The main changes are:

• Two-way synchronization for named providers and agent configuration.
• Automatic Hermes One key provisioning and credit balance display.
• Native provider routing and custom-endpoint model selection fixes.
• Redesigned provider, model, profile, and sidebar surfaces.
• Chat-stream reconciliation for lossy reasoning and answer chunks.

Confidence Score: 5/5

The reviewed fixes appear safe to merge.

• Profile-specific provisioning no longer shares work across profiles.
• Provider configuration values preserve the reported quote and backslash cases.
• Empty profile searches no longer send an undefined selection.
• No additional blocking issue cleared the follow-up review bar.

Important Files Changed

Filename	Overview
src/main/hermesone-provision.ts	Provisioning now isolates concurrent work by profile and clears completed operations.
src/main/agent-config-providers.ts	Provider values now escape quotes and backslashes before being written to agent configuration.
src/renderer/src/screens/Chat/lossyText.ts	The shared matcher now requires ordered contiguous runs and minimum coverage before treating text as a lossy copy.
src/renderer/src/screens/Layout/ProfileSwitcher.tsx	Keyboard selection now ignores Enter when the filtered profile list is empty.

_{Reviews (3): Last reviewed commit: "Review Fixes" | Re-trigger Greptile}

[fathah]

@greptile-apps review