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

lat.md/chat-commands.md

@@ -47,7 +47,7 @@ So `ensureClient` distinguishes two failures: a **genuinely absent** dashboard (
 
 On `message.complete` the desktop reconciles the text streamed via `message.delta` with the turn's `final_response`, because a last-turn-only final would otherwise clobber text streamed before a tool call (#746).
 
-[[src/renderer/src/screens/Chat/dashboardEventAdapter.ts#completeAssistantWithFinalText]] rewrites the last assistant bubble through [[src/renderer/src/screens/Chat/dashboardEventAdapter.ts#mergeStreamedWithFinal]], which compares whitespace-insensitively and: uses the final text when it already contains the streamed text; keeps the streamed text when it contains the final (preserving pre-tool-call content); stitches a re-streamed boundary by dropping the duplicated word-aligned seam (rejecting coincidental mid-word overlaps); replaces a garbled re-stream with the final text when the two converge on a substantial common suffix (a corrupted-prefix delta — e.g. a mangled CJK stream — that ends the same sentence as the clean final, rather than the disjoint pre-tool-call + answer pair); and otherwise concatenates the two with a blank-line separator so segments never run together. On the remote/SSH path deltas are not rendered (`renderAssistantDeltas: false`), so the bubble starts empty and the final text is used verbatim.
+[[src/renderer/src/screens/Chat/dashboardEventAdapter.ts#completeAssistantWithFinalText]] rewrites the last assistant bubble through [[src/renderer/src/screens/Chat/dashboardEventAdapter.ts#mergeStreamedWithFinal]], which compares whitespace-insensitively and: uses the final text when it already contains the streamed text; keeps the streamed text when it contains the final (preserving pre-tool-call content); replaces a **lossy chunk-dropped stream** with the final text when the streamed content is a subsequence of the final with substantial coverage (≥12 chars and ≥30% of the final — the upstream tagging alternate chunks as `reasoning` leaves the content stream a garbled subset like "! What are we working on?" for "Hey! What are we working on today?"; concatenating stacked the partial above the clean answer in one bubble); stitches a re-streamed boundary by dropping the duplicated word-aligned seam (rejecting coincidental mid-word overlaps); replaces a garbled re-stream with the final text when the two converge on a substantial common suffix (a corrupted-prefix delta — e.g. a mangled CJK stream — that ends the same sentence as the clean final, rather than the disjoint pre-tool-call + answer pair); and otherwise concatenates the two with a blank-line separator so segments never run together. On the remote/SSH path deltas are not rendered (`renderAssistantDeltas: false`), so the bubble starts empty and the final text is used verbatim.
 
 ## Streaming source-of-truth ref
 
@@ -61,6 +61,26 @@ Streamed reasoning and tool calls are folded into compact, collapsible transcrip
 
 [[src/renderer/src/screens/Chat/HistoryRow.tsx#ReasoningRow]] renders the `Thought` / `Thinking…` row and [[src/renderer/src/screens/Chat/HistoryRow.tsx#ToolActivityGroup]] folds a contiguous run of tool calls/results into one row titled by [[src/renderer/src/screens/Chat/HistoryRow.tsx#toolActivityGroupTitle]]. Each row is collapsed by default and borderless (Codex-style): dim at rest, it brightens and reveals an expand chevron beside the title on hover/focus, and clicking toggles the body open. While the turn is still streaming the leading icon is a `Grid` loader (purple for reasoning, blue for tools); once finished it shows the brain/tool glyph.
 
+### Reasoning reconciliation
+
+The live reasoning stream is best-effort — dropped delta chunks leave the streamed row garbled — while state.db holds the canonical text. The DB refresh must collapse the two, or the user sees both stacked in one Thought block.
+
+The observed symptom: a Thought block showing "moon-k3 … ous" (lossy live preview) above "moonshotai/kimi-k3 … nous" (canonical DB row) for the same thought.
+
+Because the garbled text can't match the DB row's text-based reconciliation key, [[src/renderer/src/screens/Chat/sessionHistory.ts#reconcileStreamedWithDb]] ends with [[src/renderer/src/screens/Chat/sessionHistory.ts#dropLossyStreamedReasoning]]: a streamed reasoning row is dropped when a DB reasoning row (`db-r-…`) in the **same turn** contains its normalized text as a **subsequence** (and is strictly longer). A dropped-chunks preview is by construction a subsequence of the canonical text, which cleanly separates "same thought, chunks missing" from a genuinely distinct live segment.
+
+#### Lossy live preview collapses into the DB row
+
+A streamed reasoning row whose garbled text is a subsequence of the same turn's DB reasoning row disappears from the merge; only the canonical DB text renders.
+
+#### Distinct live segments survive
+
+A second live reasoning segment that is not a lossy duplicate of any DB row in the turn (multi-segment thinking around tool calls) is kept alongside the reconciled first segment.
+
+#### Turn-scoped matching
+
+The subsequence check never crosses turns: a live preview in turn 2 is kept even when its text happens to be a subsequence of turn 1's canonical reasoning, so repeated questions can't cross-cancel live rows.
+
 ## Bubble hover timestamp
 
 Each user/assistant bubble reveals a relative "time ago" label on row hover, so the transcript stays uncluttered at rest but is still scrutable when a user wants to know _when_ something was said.

lat.md/hermes-account-login.md

@@ -92,6 +92,30 @@ account" card that opens it; once signed in it renders an identity card —
 avatar (or letter fallback), name/email, a "Connected" status line — with a
 Sign out action.
 
+## Auto-provisioned inference key and credits
+
+Signing in should yield model access without hand-copying keys: the desktop auto-issues a Hermes One Inference gateway key from the account and shows the account's AI-credit balance on the Providers card.
+
+[[src/main/hermesone-provision.ts]] is the convenience layer. `ensureHermesOneApiKey` checks the profile's `.env` for `HERMESONE_API_KEY`; when missing, it POSTs the backend's `/api/credits/keys` (bearer device-login token, key named `Hermes Desktop (<hostname>)` so the console list shows its origin) and persists the one-time raw `hs-live-…` key via `setEnvValue`. Setting that env var **is** what "adds the provider": the Hermes One card ([[provider-setup]]) and the active-model picker both key off it, so no store writes are needed. `fetchHermesOneCredits` GETs `/api/credits/balance` for the USD-denominated balance.
+
+It runs from two places, both idempotent: the `hermes-account-login` IPC handler right after a successful device login, and the Providers screen whenever it loads with a signed-in account (covering users who signed in before this feature). Both are **local-mode only** — the key lands in the local profile `.env`, which remote/SSH chat doesn't read, and provisioning there would strand an orphan backend key per visit. The account card renders a credits chip (`$X.XX credits`, Coins icon) next to Connected/Sync-on, backed by the `hermesone-credits` IPC; a `created` result makes the screen re-read env so the Hermes One card appears immediately.
+
+### Issues a key only when missing
+
+An existing `HERMESONE_API_KEY` is always kept — the backend shows a raw key exactly once, so re-issuing would orphan the old one. No backend call happens at all in that case.
+
+### Provisions and persists a fresh key
+
+With a signed-in account and no local key, one authenticated POST issues the key and it is written to the target profile's `.env` under `HERMESONE_API_KEY`.
+
+### Single-flight provisioning
+
+Concurrent ensure calls (post-login hook + Providers screen mount) coalesce into a single backend key issue, preventing orphan keys.
+
+### Credits for the account card
+
+The balance endpoint is called with the bearer token and returns the numeric USD credit balance; signed-out or malformed responses yield `null` so the chip simply hides.
+
 ## Tests
 
 Unit tests cover the two pieces that can break silently.