zcode-acp-server 0.65.0 → 0.66.0

package/docs/CLI.md

@@ -136,10 +136,10 @@ Without credentials, the default multi-provider mode silently shows GLM only
 ## Ollama Cloud setup
 
 Ollama Cloud exposes an (undocumented) usage endpoint at
-`ollama.com/api/usage` that the CLI queries with your API key. Create a key
-at [cloud.ollama.ai](https://cloud.ollama.ai) → API keys. The key comes from
-two sources, with the **config file taking precedence** over the environment
-variable:
+`ollama.com/api/usage` that the CLI queries with your API key (the rolling
+`30d` window). Create a key at [cloud.ollama.ai](https://cloud.ollama.ai) →
+API keys. The key comes from two sources, with the **config file taking
+precedence** over the environment variable:
 
 - **Config file** (`~/.config/zcode-acp/config.json`, `quota` section):
   ```json
@@ -150,14 +150,17 @@ variable:
   export OLLAMA_API_KEY="your_ollama_api_key"
   ```
 
-The card shows whatever usage windows the account's plan exposes — legacy
-plans get `5h` + `Week` bars, current credit plans get a `Month` bar. The
-server returns fractions only, so reset moments are derived client-side
-(epoch-aligned 5h buckets, Monday 00:00 UTC weeks, and for monthly the
-`/api/me` billing period end or — on new credit plans, which return only a
-`CreatedAt` — the next subscription-day anniversary). Without a key, the
-default mode silently skips the Ollama section; `zcode-acp quota oc` without
-a key prints a setup hint.
+The API was rewritten upstream (2026-10): it now reports **spend analytics**
+(`usage_usd`, request/token counts) and no longer exposes usage fractions or
+plan limits. The card therefore derives the `Month` percent from the 30d
+spend ÷ the plan's monthly credit (`/api/me` → `Plan`; credits from
+ollama.com/pricing: Pro $60, Max $300, Team $1000 — unknown plans show the
+spend line only, no bar) and shows a spend detail line
+(`$1.57 of $60 · 246 req (30d)`) under the bars. The monthly reset is derived
+from the `/api/me` billing period end or — on new credit plans, which return
+only a `CreatedAt` — the next subscription-day anniversary. Without a key,
+the default mode silently skips the Ollama section; `zcode-acp quota oc`
+without a key prints a setup hint.
 
 ## Hub and server subcommands
 

package/docs/REMOTE-CLIENTS.md

@@ -419,9 +419,15 @@ section — so clients can reproduce the CLI layout exactly:
       "ollama": {
         "kind": "success",
         "windows": [
-          { "key": "session", "label": "5h", "usagePercent": 31 },
-          { "key": "weekly", "label": "Week", "usagePercent": 67.5 }
-        ]
+          { "key": "monthly", "label": "Month", "usagePercent": 2.6,
+            "resetsAt": 1729071480308 }
+        ],
+        "usage": {
+          "plan": "pro", "usageUsd": 1.5747, "creditUsd": 60,
+          "requestCount": 246, "inputTokens": 14294081,
+          "cachedInputTokens": 12922504, "outputTokens": 940919,
+          "window": "30d"
+        }
       }
     } }
 ```
@@ -438,12 +444,14 @@ section — so clients can reproduce the CLI layout exactly:
   means the user never set OpenCode Go credentials — omit the section, like
   the CLI does.
 - `ollama` (`kind`: `success` | `not_configured` | `auth_error` |
-  `unavailable`): on success, `windows` carries whichever entries the
-  account's plan exposes — legacy plans: `session` (`5h`) + `weekly`
-  (`Week`); current credit plans: `monthly` (`Month`) — each with
-  `usagePercent` (0–100) and, when available, `resetsAt` (epoch ms,
-  derived client-side: epoch-aligned 5h buckets / Monday 00:00 UTC weeks /
-  the `/api/me` billing period or subscription-day anniversary for monthly).
+  `unavailable`): since the 2026-10 upstream API rewrite the endpoint reports
+  spend only, so on success `windows` carries the derived `monthly` (`Month`)
+  entry — 30d `usageUsd` ÷ the plan's monthly credit, absent when the plan's
+  credit is unknown — with `resetsAt` (epoch ms, from the `/api/me` billing
+  period or subscription-day anniversary) when available. The additive
+  `usage` object carries the raw analytics (`plan`, `usageUsd`,
+  `creditUsd?`, `requestCount`/token totals, `window: "30d"`) for clients
+  that want to render spend directly; older clients ignore it.
   `not_configured` means no
   Ollama API key is set — omit the section, like the CLI does.
 - Provider failures are per-section `kind` strings, not JSON-RPC errors —

package/docs/REPLAY-GUIDE.md

@@ -128,7 +128,7 @@ update's `_meta`:
   too. These updates never collide with live tool_call ids (they carry the
   `histfold_` prefix).
 
-## Turn running state (`replayMeta.turnActive` + `$/zcode/turnState`)
+## Turn running state (`replayMeta.turnActive` + turnState notifications)
 
 A turn may already be in flight when you attach — started by the editor or
 another remote client (the bridge runs it to completion regardless of who
@@ -137,16 +137,26 @@ prompted). Two signals cover it:
 - **Attach snapshot**: `session/load`'s `replayMeta.turnActive` (boolean) —
   `true` when any turn for this session is running at attach time. Seed your
   spinner/running state from it.
-- **Out-of-band updates**: the notification `$/zcode/turnState` with params
+- **Out-of-band updates**: the turn-state notification with params
   `{ sessionId: string, running: boolean }` — emitted when a turn starts and
   when it ends (including failures, e.g. a failed subscribe). On preemption (a
   new prompt interrupts an in-flight one) the old turn's exit reports
   `running: true`: the preempting turn took over, so the session is still busy.
 
+The notification is **dual-emitted under two names** (#311): the deprecated
+`$/zcode/turnState` (LSP-style prefix) and the spec-compliant
+`_zcode/turnState` (ACP reserves `_`-prefixed names for extension
+notifications). Subscribe to exactly ONE of them — a client handling both
+sees every transition twice. New clients should use `_zcode/turnState`; the
+`$/` spelling remains for existing consumers (martty, zcode-acp-remote) and
+will be removed once they migrate. Both names are advertised in
+`initialize`'s `agentCapabilities._meta.zcode.turnState`.
+
 The client that sent `session/prompt` already knows its own turn via the
 request/response; these signals exist for the OTHER attached clients
 (re-attached mobile, second editor). Unknown notifications are ignorable —
-clients that don't handle `$/zcode/turnState` lose nothing (Zed ignores it).
+clients that don't handle the turn-state notifications lose nothing (Zed
+ignores them).
 
 ## Scroll-up pagination
 
@@ -202,8 +212,9 @@ Never parse the cursor — it is opaque. (For the curious it round-trips
 - [ ] `session/load` params include `cwd` and `mcpServers` (even `[]`).
 - [ ] Message list keyed/deduped by `messageId`; tool cards by `toolCallId`.
 - [ ] Pagination pages prepended, live updates appended.
-- [ ] Running state seeded from `replayMeta.turnActive` and updated from
-      `$/zcode/turnState` notifications (covers other clients' turns).
+- [ ] Running state seeded from `replayMeta.turnActive` and updated from the
+      turn-state notifications — `_zcode/turnState` (or the deprecated `$/`
+      spelling; pick ONE) — covering other clients' turns.
 - [ ] `"cursor expired"` handled by full re-attach.
 - [ ] Cursor stored per session, never parsed, never persisted across app
       runs (it is only meaningful to the bridge that minted it).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zcode-acp-server",
-  "version": "0.65.0",
+  "version": "0.66.0",
   "description": "Agent Client Protocol (ACP) server bridging headless ZCode to editors like Zed and JetBrains.",
   "type": "module",
   "license": "Apache-2.0",