dsh-connect-agnes-token-plan 0.5.0

package/cordis.patch.yml

@@ -0,0 +1,147 @@
+# dsh-connect-agnes-token-plan — the bundle's own patch layer.
+#
+# Reads the Agnes Token Plan quota through the console API and shows it in
+# a config card on the Plugins page (`plugins.bundle.config` slot) — not in the
+# sidebar. Everything below is plain configuration: edit it here (or
+# override the same id in the profile's cordis.patch.yml) and reinstall/reload.
+#
+# The console token is NOT configured here. Open the panel and sign in once:
+# the Host logs in, keeps the account in the DSH credentials, and re-logs in
+# on its own when the token dies (Agnes issues no refresh token). See README.md.
+- insert:
+    - id: dsh-connect-agnes-token-plan
+      name: dsh-connect-agnes-token-plan
+      config:
+        # Console origin (the BACKEND host, not the console front-end: the
+        # front-end origin serves /api/* as a Next.js 404 shell). Override if
+        # you use a mirror or proxy.
+        consoleBase: https://platform-backend.agnes-ai.cn
+        # The OpenAI-compatible API base, used only for the optional model
+        # catalog (which models this key can call). Leave it alone unless the
+        # platform moves the endpoint.
+        apiBase: https://api.agnes-ai.cn/v1
+        # How many days of usage the chart covers (max 365). Days, not hours:
+        # the console's series endpoint takes start_date / end_date as DATES.
+        usageDays: 30
+        # Pseudo multipliers for the model roster: keys are case-insensitive
+        # SUBSTRINGS of a model id (first match in writing order wins), values
+        # are positive numbers. Empty by default — the factor used to scale a
+        # per-model credit figure, and Agnes publishes no per-model usage at
+        # all. What is left is your own annotation: a `×N` badge on a roster
+        # row. Rows without a match get no badge, never a guessed 1.
+        # Invalid entries (empty key / value <= 0 / non-number) are dropped.
+        trendMultipliers: {}
+        # How long the Host caches each console response, in seconds.
+        cacheSeconds: 60
+        # How often the panel asks the Host again, in seconds. The Host states
+        # this to the bundle on every snapshot, so the two cannot disagree about
+        # how fresh the screen is.
+        pollSeconds: 30
+        # Deadline for one console call, in milliseconds. The login flow has its
+        # own (loginTimeoutMs, below) because it carries a password and may be
+        # rate-limited, so the two are not the same number.
+        consoleTimeoutMs: 15000
+        # Renew the console token this many seconds before it actually expires,
+        # so a poll never races the expiry boundary. 120 = two minutes.
+        tokenSkewSeconds: 120
+        # Extra host names this Host answers as, beyond the loopback defaults
+        # (localhost / 127.0.0.1 / ::1). ADDED to them, never substituted — see
+        # docs/SETUP.md. Uncomment and list to serve the panel from another name.
+        # allowedHosts: []
+        # The login flow's overrides. Everything below is optional: a key left
+        # commented keeps the platform default. Agnes signs in with ONE request
+        # (POST {consoleBase}/api/user/login), so there is no OIDC walk and none
+        # of the SenseNova endpoint knobs exist here any more.
+        # loginPath: /api/user/login
+        #   -> the sign-in path, if the platform ever moves it.
+        # fallbackExpiresInSeconds: 604800
+        #   -> lifetime assumed when the access token is not a readable JWT
+        #      (7 days by default). Deliberately LONG: a token assumed dead
+        #      while still alive spends a real sign-in on every poll, and Agnes
+        #      locks an account after a few bad attempts.
+        # loginTimeoutMs: 20000
+        #   -> deadline for one sign-in attempt, in ms. requestTimeoutMs is its
+        #      older spelling and still wins when only it is set; the console
+        #      calls use consoleTimeoutMs above.
+        # requestTimeoutMs: 20000
+        # ------------------------------------------------------------------
+        # Vision step two: publish which of this key's models accept image
+        # input into this row's OWN settings namespace, for a later LLM
+        # connect plugin to read. It writes only its own row - never another
+        # provider's imageModelIds - so a miscalculated model list can only
+        # affect the panel, not DSH's model routing.
+        #
+        # `writeImageModelIds: true`  -> the Host syncs visionModels into
+        #   `imageModelIds` on this row's settings on every catalog poll
+        #   (opt-in, off by default; the panel's read-only info layer is the
+        #   safe shape).
+        # `imageModelIds: []`         -> the last written model-id list.
+        # `visionModels: []`          -> the last full vision identification
+        #   (ids + source markers), kept for a reader that wants the detail.
+        writeImageModelIds: false
+        imageModelIds: []
+        visionModels: []
+        # ------------------------------------------------------------------
+        # Step three ("one-stop service"): register the Agnes LLM
+        # provider DIRECTLY, with no hand-written llm-pi-ai row. When true
+        # the Host offers an OpenAI-compatible provider
+        # (`agnes-token-plan`, base apiBase above): the panel's API-key
+        # field stores the `sk-` key in the DSH credentials (the process
+        # environment still works as a fallback), the catalog poll builds
+        # the model list, and vision models accept image input
+        # automatically. The cached catalog and the model allow-list live
+        # in this plugin's own state file, never on this row. Off by
+        # default - registering a model source is a Host-wide change.
+        # This value is now the DEPLOYMENT DEFAULT only: the panel's
+        #「模型接入」switch (POST /api/<name>/provider) overrides it live,
+        # with no restart. See docs/PROVIDER-HOT-RELOAD.md.
+        registerProvider: false
+        # ------------------------------------------------------------------
+        # Draw absorption (docs/ARCHITECTURE.md §5.4, route B): when true AND
+        # the Host exposes a tools service, the agent gains an
+        # `agnes_draw_image` tool that POSTs to
+        # `{apiBase}/images/generations` with the panel-saved
+        # `AGNES_TOKEN_PLAN_API_KEY`. Draw models are identified by the
+        # catalog's own `output_modalities` when the platform declares it and
+        # by an `image` segment in the model id when it does not - the Agnes
+        # gateway declares NO modality field at all (docs/AGNES-API.md §7.1),
+        # so the name fallback is the path that actually fires here. The key is
+        # resolved per call, and a failed draw starts a 30s cooldown. Off by
+        # default - an agent tool is a Host-wide change; a Host without the
+        # tools service simply never sees it.
+        # drawEnabled: false
+        #   -> the opt-in switch (strict boolean).
+        # drawModelId: ""
+        #   -> preferred image model id; empty = first of the catalog
+        #      (e.g. Agnes-u1-fast).
+        # drawTimeoutMs: 120000
+        #   -> deadline for one image request (image models are slow).
+        # ------------------------------------------------------------------
+        # ------------------------------------------------------------------
+        # Video absorption (docs/AGNES-API.md §7.5): the same ladder as draw,
+        # with one protocol difference that decides the shape - video is an
+        # ASYNCHRONOUS TASK, so the tool creates a task and polls it. Off by
+        # default, like the draw tool.
+        # videoEnabled: false
+        #   -> the opt-in switch (strict boolean).
+        # videoModelId: ""
+        #   -> preferred video model id; empty = first V2.0-family model in the
+        #      catalog. The 2.5 family (`agnes-video-2.5`, `-2.5-flash`) is
+        #      NOT selectable: it takes `mode`/`seconds`/`size`/`aspect_ratio`
+        #      instead of `width`/`height`/`num_frames`/`frame_rate`, and
+        #      mixing the two schemas is a 400. The panel names the excluded
+        #      models so their absence is explained rather than mysterious.
+        # videoTimeoutMs: 600000
+        #   -> deadline for the WHOLE task (create + poll). Minutes, not
+        #      seconds; floor 30000.
+        # videoWidth: 1152
+        # videoHeight: 768
+        #   -> the V2.0 frame size. Passed straight through.
+        # videoNumFrames: 121
+        #   -> frame count; must be <= 441 AND satisfy 8n+1 (81 / 121 / 161 /
+        #      241 / 441). 121 @ 24fps is about 5 seconds. An illegal explicit
+        #      value is REFUSED, not clamped: silently shortening a clip the
+        #      caller asked for is worse than an error they can read.
+        # videoFrameRate: 24
+        #   -> frames per second, 1-60.
+        # ------------------------------------------------------------------

package/lib/agnescode-llm-adapter-DDvQ-xvK.js

@@ -0,0 +1,135 @@
+import { a as name } from "./host-config-BNGgm_j7.js";
+import { i as buildAgnescodeDescriptors, n as AGNESCODE_PROVIDER_ID, r as agnescodeRoster, t as AGNESCODE_DISPLAY_NAME } from "./agnescode-models-CUcL9UwK.js";
+import { n as buildRetryPolicyConfig, t as reclassifyStream } from "./llm-error-fix-DQUng8AL.js";
+import { createProvider } from "@earendil-works/pi-ai";
+import { openAICompletionsApi } from "@earendil-works/pi-ai/api/openai-completions.lazy";
+import { PiAiAdapter } from "@deepseek-ai/dsh-llm-pi-ai";
+import { resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
+
+//#region src/host/agnescode-llm-adapter.ts
+/**
+* The peer-dependent half of the directly-registered AgnesCode provider —
+* "third upstream provider" (ROADMAP §6.3).
+*
+* Mirrors `raccoon-llm-adapter.ts`'s shape: ONE `PiAiAdapter` carrying ONE
+* profile (`sensenova-agnescode` → the credential's per-account BFF base), an
+* INERT pi-ai auth plane (the AgnesCode JWT is resolved per request from the
+* plugin's own credential store, pi-ai never manufactures it), and no token
+* baked into the profile — the picker advertises models without one and a
+* request fails at resolve time, where the panel status is visible.
+*
+* IMAGE hooks are deliberately NOT wired in v1: every catalogue row declares
+* `model_type: "text"` with no image modality fact (ROADMAP §6.3), so there
+* is nothing honest to advertise — a vision answer here would be an invented
+* fact. If the platform later declares image input, this is the one place to
+* add the hooks (mirror `raccoon-llm-adapter.ts`).
+*
+* The descriptor mapping lives in the peer-free `agnescode-models.ts`, so this
+* module holds only assembly against the runtime and is exercised by the
+* wiring/e2e checks, exactly as `llm-adapter.ts` is.
+*
+* @module dsh-connect-agnes-token-plan/agnescode-llm-adapter
+*/
+/** Idle ceiling while one stream read is outstanding (dsh-llm-pi-ai default). */
+const STREAM_IDLE_TIMEOUT_MS = 3e5;
+/**
+* Inert pi-ai auth plane.
+*
+* Authentication goes through `resolveApiKey` (the stored AgnesCode JWT, read
+* per request from `agnescode-store.ts`) — pi-ai's own credential lifecycle
+* must never manufacture a credential for this route, so every ambient
+* question answers "nothing stored, nothing set".
+*/
+const INERT_AUTH = {
+	credentials: {
+		async read() {},
+		async list() {
+			return [];
+		},
+		async modify() {},
+		async delete() {}
+	},
+	authContext: {
+		async env() {},
+		async fileExists() {
+			return false;
+		}
+	}
+};
+/**
+* Assemble the AgnesCode adapter for one credential/catalogue snapshot.
+*
+* A fresh instance per rebuild is deliberate: `PiAiAdapter` memoizes the
+* profiles snapshot internally, so the caller REPLACES the registered
+* adapter when the credential or catalogue changes and emits
+* `llm/adapters-updated`.
+* @param {object} options - wiring.
+* @param {object[]} [options.rows] - the AgnesCode roster rows (`agnescodeRoster`).
+* @param {string} options.bffBase - the credential's pinned per-account BFF base.
+* @param {() => Promise<string>} options.resolveToken - resolves the live
+*   AgnesCode JWT per request (re-harvested by the caller when expired).
+* @param {(service: string) => unknown} [options.get] - service resolver
+*   (reserved for the image hooks when the platform declares image input).
+* @returns {{adapter: object, providerIds: string[]}} the adapter and the ids
+*   it owns.
+*/
+function createAgnescodeAdapter({ rows, bffBase, resolveToken, get } = {}) {
+	const models = buildAgnescodeDescriptors(rows ?? agnescodeRoster(null), { bffBase: String(bffBase ?? "") });
+	const provider = {
+		...createProvider({
+			id: AGNESCODE_PROVIDER_ID,
+			name: AGNESCODE_DISPLAY_NAME,
+			auth: { apiKey: {
+				name: "AgnesCode access token",
+				async resolve({ credential } = {}) {
+					const token = credential?.key;
+					return token === void 0 || token === "" ? void 0 : {
+						auth: { apiKey: token },
+						source: AGNESCODE_DISPLAY_NAME
+					};
+				}
+			} },
+			models,
+			api: openAICompletionsApi()
+		}),
+		getModels: () => models
+	};
+	const profiles = /* @__PURE__ */ new Map([[AGNESCODE_PROVIDER_ID, {
+		provider: AGNESCODE_PROVIDER_ID,
+		displayName: AGNESCODE_DISPLAY_NAME,
+		streamIdleTimeoutMs: STREAM_IDLE_TIMEOUT_MS,
+		retryPolicy: resolveRetryPolicy(buildRetryPolicyConfig(), `${name}.${AGNESCODE_PROVIDER_ID}.retryPolicy`),
+		configuredMaxTokens: /* @__PURE__ */ new Map(),
+		modelErrors: /* @__PURE__ */ new Map(),
+		piProvider: provider
+	}]]);
+	const inner = new PiAiAdapter({
+		profiles: () => profiles,
+		auth: INERT_AUTH,
+		resolveApiKey: async () => resolveToken(),
+		resolveAttachments: () => get?.("attachments"),
+		resolveImageAccess: void 0
+	});
+	return {
+		adapter: new Proxy(inner, { get(target, prop, receiver) {
+			const value = Reflect.get(target, prop, receiver);
+			if (prop === "stream") return (options) => reclassifyStream(target.stream(options));
+			if (typeof value === "function" && prop === "prepareCall") return (...args) => {
+				const prepared = value.apply(target, args);
+				if (prepared && typeof prepared.then === "function") return prepared.then((p) => p && typeof p.stream === "function" ? {
+					...p,
+					stream: (o) => reclassifyStream(p.stream(o))
+				} : p);
+				return prepared && typeof prepared.stream === "function" ? {
+					...prepared,
+					stream: (o) => reclassifyStream(prepared.stream(o))
+				} : prepared;
+			};
+			return value;
+		} }),
+		providerIds: [AGNESCODE_PROVIDER_ID]
+	};
+}
+
+//#endregion
+export { createAgnescodeAdapter };