@caupulican/pi-adaptative 0.81.38 → 0.81.39

package/dist/core/models/prism-llamacpp-lifecycle.d.ts

@@ -0,0 +1,89 @@
+/**
+ * Core lifecycle logic for pi-managed prism llama.cpp models (Bonsai-27B first), shared between the
+ * `/models add` install pipeline (modes/interactive/local-model-commands.ts) and the router's
+ * readiness gate (local-runtime-controller.ts). ONE place owns "ensure both GGUF files exist, then
+ * serve" so a turn routed to a cold Bonsai-27B and a fresh `/models add` both go through the
+ * identical self-healing path instead of two parallel, potentially-diverging implementations.
+ *
+ * Lives in core/models/ (not modes/) so the headless readiness gate — used by print/RPC/interactive
+ * sessions and isolated/background lanes alike — never depends on the interactive-only UI layer;
+ * mirrors the existing local-runtime.ts (mechanics) / local-runtime-controller.ts (gate) split.
+ */
+import { type PrismLlamaCppRuntime, type PrismModelDescriptor } from "./llamacpp-runtime.ts";
+/**
+ * Curated prism llama.cpp descriptors, keyed by repo id — matches the modelId model-ref.ts's
+ * curated `PRISM_LLAMACPP_MODEL_IDS` route returns, and the id `registerPrismLlamaCppModel` writes
+ * as the model's id in models.json. Add future curated prism-ml models to both places together.
+ *
+ * Also the pi-managed-vs-user-owned discriminator (see {@link isPiManagedPrismLlamaCppModel}): a
+ * `llama-cpp` provider model is gated/self-healed by pi ONLY when its id is a key here — a user's
+ * own hand-configured entry (e.g. the built-in `llama-cpp/local` catalog model on port 8080,
+ * pointing at a server the user runs themselves) is never touched.
+ */
+export declare const PRISM_LLAMACPP_DESCRIPTORS: Record<string, PrismModelDescriptor>;
+/**
+ * True only for a `llama-cpp` provider model pi itself registered (id present in
+ * {@link PRISM_LLAMACPP_DESCRIPTORS}) — never true for a user's own hand-configured llama-cpp
+ * model, including the built-in `llama-cpp/local` catalog entry. This is the gate that must run
+ * before ANY self-heal/serve action so pi never touches a server it doesn't own.
+ */
+export declare function isPiManagedPrismLlamaCppModel(model: {
+    provider: string;
+    id: string;
+}): boolean;
+/**
+ * Fixed local port for pi's own managed prism llama.cpp server. Deliberately NOT the generic
+ * built-in `llama-cpp/local` catalog model's conventional port 8080 (models.generated.ts) — that
+ * port is the convention for a user's OWN manually started llama-server pi merely points at; pi's
+ * managed instance must never collide with it.
+ */
+export declare const PRISM_LLAMACPP_SERVE_PORT = 8090;
+/**
+ * Conservative RAM-only context cap for the pi-managed prism llama.cpp server, used at FIRST
+ * install time only (`/models add`) — a later self-heal re-serve reuses the model's already
+ * registered `contextWindow` instead (see the readiness gate in local-runtime-controller.ts), never
+ * re-deriving from current RAM, so a served context size can't drift from what pi already told the
+ * rest of the session (compaction, etc.) to expect. PrismLlamaCppRuntime has no `/api/show`-
+ * equivalent GGUF metadata endpoint (unlike Ollama, which context-sizing.ts's
+ * deriveLocalContextSizing depends on), so this is a coarse rung table mirroring
+ * context-sizing.ts's CONTEXT_RUNGS, hard-capped at 32768 regardless of headroom: Bonsai-27B
+ * advertises a 262K-class context, but that is not a realistic KV budget for a CPU-served model on
+ * consumer hardware, and there is no larger validated rung to grow into yet.
+ */
+export declare function derivePrismLlamaCppNumCtx(totalMemBytes: number): number;
+export type EnsurePrismModelServedResult = {
+    ok: true;
+    baseUrl: string;
+} | {
+    ok: false;
+    stage: "model-download" | "mmproj-download" | "serve";
+    error: string;
+};
+/**
+ * The SOLE path allowed to call `runtime.serve()` for a prism llama.cpp model — no caller may spawn
+ * llama-server without first passing through here (both `addPrismLlamaCppModel`'s install pipeline
+ * and the readiness gate's self-heal path call this, never `runtime.serve()` directly).
+ * Unconditionally re-verifies BOTH GGUF files via `runtime.downloadModel()` immediately before every
+ * serve — that call is idempotent (skips a re-download when the local file already matches the
+ * remote size, so the happy path costs one stat) and self-heals a file that went missing from disk
+ * (deleted, a partial download from a prior crash, host cleanup, etc.) by re-fetching it instead of
+ * ever starting llama-server without it. The vision projector is mandatory whenever the descriptor
+ * declares one (Bonsai-27B always does — see BONSAI_27B.mmprojFile): a download failure on EITHER
+ * file means llama-server is never spawned. There is no text-only fallback — Bonsai-27B never
+ * silently degrades to a vision-less serve because the projector went missing.
+ */
+export declare function ensurePrismModelFilesThenServe(runtime: PrismLlamaCppRuntime, descriptor: PrismModelDescriptor, args: {
+    port: number;
+    numCtx: number;
+}, onProgress: (message: string) => void): Promise<EnsurePrismModelServedResult>;
+/**
+ * Direct HTTP reachability probe against a prism llama.cpp server root (NOT the `/v1` OpenAI-compat
+ * suffix models.json registers — same root/health split as llama-server's own health endpoint used
+ * internally by PrismLlamaCppRuntime#serve). PrismLlamaCppRuntime has no public method for this (its
+ * own equivalent, `_healthUp`, is private and scoped to a `serve()` call in progress), so the
+ * readiness gate needs its own — this is deliberately independent of any single runtime instance:
+ * it answers "is a server already alive at this URL", true even for a server this session never
+ * spawned (started earlier, or by a different process).
+ */
+export declare function isPrismLlamaCppServerHealthy(serverUrl: string, fetchFn?: typeof fetch, timeoutMs?: number): Promise<boolean>;
+//# sourceMappingURL=prism-llamacpp-lifecycle.d.ts.map

package/dist/core/models/prism-llamacpp-lifecycle.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"prism-llamacpp-lifecycle.d.ts","sourceRoot":"","sources":["../../../src/core/models/prism-llamacpp-lifecycle.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAc,KAAK,oBAAoB,EAAE,KAAK,oBAAoB,EAAE,MAAM,uBAAuB,CAAC;AAGzG;;;;;;;;;GASG;AACH,eAAO,MAAM,0BAA0B,EAAE,MAAM,CAAC,MAAM,EAAE,oBAAoB,CAE3E,CAAC;AAEF;;;;;GAKG;AACH,wBAAgB,6BAA6B,CAAC,KAAK,EAAE;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAAG,OAAO,CAE9F;AAED;;;;;GAKG;AACH,eAAO,MAAM,yBAAyB,OAAO,CAAC;AAE9C;;;;;;;;;;;GAWG;AACH,wBAAgB,yBAAyB,CAAC,aAAa,EAAE,MAAM,GAAG,MAAM,CAMvE;AAED,MAAM,MAAM,4BAA4B,GACrC;IAAE,EAAE,EAAE,IAAI,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAC7B;IAAE,EAAE,EAAE,KAAK,CAAC;IAAC,KAAK,EAAE,gBAAgB,GAAG,iBAAiB,GAAG,OAAO,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC;AAEvF;;;;;;;;;;;;GAYG;AACH,wBAAsB,8BAA8B,CACnD,OAAO,EAAE,oBAAoB,EAC7B,UAAU,EAAE,oBAAoB,EAChC,IAAI,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,EACtC,UAAU,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,IAAI,GACnC,OAAO,CAAC,4BAA4B,CAAC,CA6BvC;AAID;;;;;;;;GAQG;AACH,wBAAsB,4BAA4B,CACjD,SAAS,EAAE,MAAM,EACjB,OAAO,GAAE,OAAO,KAAa,EAC7B,SAAS,GAAE,MAAwC,GACjD,OAAO,CAAC,OAAO,CAAC,CAOlB","sourcesContent":["/**\n * Core lifecycle logic for pi-managed prism llama.cpp models (Bonsai-27B first), shared between the\n * `/models add` install pipeline (modes/interactive/local-model-commands.ts) and the router's\n * readiness gate (local-runtime-controller.ts). ONE place owns \"ensure both GGUF files exist, then\n * serve\" so a turn routed to a cold Bonsai-27B and a fresh `/models add` both go through the\n * identical self-healing path instead of two parallel, potentially-diverging implementations.\n *\n * Lives in core/models/ (not modes/) so the headless readiness gate — used by print/RPC/interactive\n * sessions and isolated/background lanes alike — never depends on the interactive-only UI layer;\n * mirrors the existing local-runtime.ts (mechanics) / local-runtime-controller.ts (gate) split.\n */\n\nimport { BONSAI_27B, type PrismLlamaCppRuntime, type PrismModelDescriptor } from \"./llamacpp-runtime.ts\";\nimport { PRISM_LLAMACPP_PROVIDER } from \"./local-registration.ts\";\n\n/**\n * Curated prism llama.cpp descriptors, keyed by repo id — matches the modelId model-ref.ts's\n * curated `PRISM_LLAMACPP_MODEL_IDS` route returns, and the id `registerPrismLlamaCppModel` writes\n * as the model's id in models.json. Add future curated prism-ml models to both places together.\n *\n * Also the pi-managed-vs-user-owned discriminator (see {@link isPiManagedPrismLlamaCppModel}): a\n * `llama-cpp` provider model is gated/self-healed by pi ONLY when its id is a key here — a user's\n * own hand-configured entry (e.g. the built-in `llama-cpp/local` catalog model on port 8080,\n * pointing at a server the user runs themselves) is never touched.\n */\nexport const PRISM_LLAMACPP_DESCRIPTORS: Record<string, PrismModelDescriptor> = {\n\t[BONSAI_27B.repo]: BONSAI_27B,\n};\n\n/**\n * True only for a `llama-cpp` provider model pi itself registered (id present in\n * {@link PRISM_LLAMACPP_DESCRIPTORS}) — never true for a user's own hand-configured llama-cpp\n * model, including the built-in `llama-cpp/local` catalog entry. This is the gate that must run\n * before ANY self-heal/serve action so pi never touches a server it doesn't own.\n */\nexport function isPiManagedPrismLlamaCppModel(model: { provider: string; id: string }): boolean {\n\treturn model.provider === PRISM_LLAMACPP_PROVIDER && Object.hasOwn(PRISM_LLAMACPP_DESCRIPTORS, model.id);\n}\n\n/**\n * Fixed local port for pi's own managed prism llama.cpp server. Deliberately NOT the generic\n * built-in `llama-cpp/local` catalog model's conventional port 8080 (models.generated.ts) — that\n * port is the convention for a user's OWN manually started llama-server pi merely points at; pi's\n * managed instance must never collide with it.\n */\nexport const PRISM_LLAMACPP_SERVE_PORT = 8090;\n\n/**\n * Conservative RAM-only context cap for the pi-managed prism llama.cpp server, used at FIRST\n * install time only (`/models add`) — a later self-heal re-serve reuses the model's already\n * registered `contextWindow` instead (see the readiness gate in local-runtime-controller.ts), never\n * re-deriving from current RAM, so a served context size can't drift from what pi already told the\n * rest of the session (compaction, etc.) to expect. PrismLlamaCppRuntime has no `/api/show`-\n * equivalent GGUF metadata endpoint (unlike Ollama, which context-sizing.ts's\n * deriveLocalContextSizing depends on), so this is a coarse rung table mirroring\n * context-sizing.ts's CONTEXT_RUNGS, hard-capped at 32768 regardless of headroom: Bonsai-27B\n * advertises a 262K-class context, but that is not a realistic KV budget for a CPU-served model on\n * consumer hardware, and there is no larger validated rung to grow into yet.\n */\nexport function derivePrismLlamaCppNumCtx(totalMemBytes: number): number {\n\tconst totalGb = totalMemBytes / 1e9;\n\tif (totalGb >= 64) return 32_768;\n\tif (totalGb >= 32) return 16_384;\n\tif (totalGb >= 16) return 8_192;\n\treturn 4_096;\n}\n\nexport type EnsurePrismModelServedResult =\n\t| { ok: true; baseUrl: string }\n\t| { ok: false; stage: \"model-download\" | \"mmproj-download\" | \"serve\"; error: string };\n\n/**\n * The SOLE path allowed to call `runtime.serve()` for a prism llama.cpp model — no caller may spawn\n * llama-server without first passing through here (both `addPrismLlamaCppModel`'s install pipeline\n * and the readiness gate's self-heal path call this, never `runtime.serve()` directly).\n * Unconditionally re-verifies BOTH GGUF files via `runtime.downloadModel()` immediately before every\n * serve — that call is idempotent (skips a re-download when the local file already matches the\n * remote size, so the happy path costs one stat) and self-heals a file that went missing from disk\n * (deleted, a partial download from a prior crash, host cleanup, etc.) by re-fetching it instead of\n * ever starting llama-server without it. The vision projector is mandatory whenever the descriptor\n * declares one (Bonsai-27B always does — see BONSAI_27B.mmprojFile): a download failure on EITHER\n * file means llama-server is never spawned. There is no text-only fallback — Bonsai-27B never\n * silently degrades to a vision-less serve because the projector went missing.\n */\nexport async function ensurePrismModelFilesThenServe(\n\truntime: PrismLlamaCppRuntime,\n\tdescriptor: PrismModelDescriptor,\n\targs: { port: number; numCtx: number },\n\tonProgress: (message: string) => void,\n): Promise<EnsurePrismModelServedResult> {\n\tconst downloadedModel = await runtime.downloadModel({ repo: descriptor.repo, file: descriptor.file }, (progress) =>\n\t\tonProgress(`  ${descriptor.file}: ${progress}`),\n\t);\n\tif (!downloadedModel.ok || !downloadedModel.path) {\n\t\treturn { ok: false, stage: \"model-download\", error: downloadedModel.error ?? \"unknown error\" };\n\t}\n\n\tlet mmprojPath: string | undefined;\n\tif (descriptor.mmprojFile) {\n\t\tconst downloadedMmproj = await runtime.downloadModel(\n\t\t\t{ repo: descriptor.repo, file: descriptor.mmprojFile },\n\t\t\t(progress) => onProgress(`  ${descriptor.mmprojFile}: ${progress}`),\n\t\t);\n\t\tif (!downloadedMmproj.ok || !downloadedMmproj.path) {\n\t\t\treturn { ok: false, stage: \"mmproj-download\", error: downloadedMmproj.error ?? \"unknown error\" };\n\t\t}\n\t\tmmprojPath = downloadedMmproj.path;\n\t}\n\n\tonProgress(`Starting ${descriptor.displayName} on 127.0.0.1:${args.port} (context ${args.numCtx})…`);\n\tconst served = await runtime.serve({\n\t\tmodelPath: downloadedModel.path,\n\t\tmmprojPath,\n\t\tport: args.port,\n\t\tnumCtx: args.numCtx,\n\t});\n\tif (!served.ok) return { ok: false, stage: \"serve\", error: served.error };\n\treturn { ok: true, baseUrl: served.baseUrl };\n}\n\nconst DEFAULT_HEALTH_CHECK_TIMEOUT_MS = 2_000;\n\n/**\n * Direct HTTP reachability probe against a prism llama.cpp server root (NOT the `/v1` OpenAI-compat\n * suffix models.json registers — same root/health split as llama-server's own health endpoint used\n * internally by PrismLlamaCppRuntime#serve). PrismLlamaCppRuntime has no public method for this (its\n * own equivalent, `_healthUp`, is private and scoped to a `serve()` call in progress), so the\n * readiness gate needs its own — this is deliberately independent of any single runtime instance:\n * it answers \"is a server already alive at this URL\", true even for a server this session never\n * spawned (started earlier, or by a different process).\n */\nexport async function isPrismLlamaCppServerHealthy(\n\tserverUrl: string,\n\tfetchFn: typeof fetch = fetch,\n\ttimeoutMs: number = DEFAULT_HEALTH_CHECK_TIMEOUT_MS,\n): Promise<boolean> {\n\ttry {\n\t\tconst response = await fetchFn(`${serverUrl}/health`, { signal: AbortSignal.timeout(timeoutMs) });\n\t\treturn response.ok;\n\t} catch {\n\t\treturn false;\n\t}\n}\n"]}

package/dist/core/models/prism-llamacpp-lifecycle.js

@@ -0,0 +1,121 @@
+/**
+ * Core lifecycle logic for pi-managed prism llama.cpp models (Bonsai-27B first), shared between the
+ * `/models add` install pipeline (modes/interactive/local-model-commands.ts) and the router's
+ * readiness gate (local-runtime-controller.ts). ONE place owns "ensure both GGUF files exist, then
+ * serve" so a turn routed to a cold Bonsai-27B and a fresh `/models add` both go through the
+ * identical self-healing path instead of two parallel, potentially-diverging implementations.
+ *
+ * Lives in core/models/ (not modes/) so the headless readiness gate — used by print/RPC/interactive
+ * sessions and isolated/background lanes alike — never depends on the interactive-only UI layer;
+ * mirrors the existing local-runtime.ts (mechanics) / local-runtime-controller.ts (gate) split.
+ */
+import { BONSAI_27B } from "./llamacpp-runtime.js";
+import { PRISM_LLAMACPP_PROVIDER } from "./local-registration.js";
+/**
+ * Curated prism llama.cpp descriptors, keyed by repo id — matches the modelId model-ref.ts's
+ * curated `PRISM_LLAMACPP_MODEL_IDS` route returns, and the id `registerPrismLlamaCppModel` writes
+ * as the model's id in models.json. Add future curated prism-ml models to both places together.
+ *
+ * Also the pi-managed-vs-user-owned discriminator (see {@link isPiManagedPrismLlamaCppModel}): a
+ * `llama-cpp` provider model is gated/self-healed by pi ONLY when its id is a key here — a user's
+ * own hand-configured entry (e.g. the built-in `llama-cpp/local` catalog model on port 8080,
+ * pointing at a server the user runs themselves) is never touched.
+ */
+export const PRISM_LLAMACPP_DESCRIPTORS = {
+    [BONSAI_27B.repo]: BONSAI_27B,
+};
+/**
+ * True only for a `llama-cpp` provider model pi itself registered (id present in
+ * {@link PRISM_LLAMACPP_DESCRIPTORS}) — never true for a user's own hand-configured llama-cpp
+ * model, including the built-in `llama-cpp/local` catalog entry. This is the gate that must run
+ * before ANY self-heal/serve action so pi never touches a server it doesn't own.
+ */
+export function isPiManagedPrismLlamaCppModel(model) {
+    return model.provider === PRISM_LLAMACPP_PROVIDER && Object.hasOwn(PRISM_LLAMACPP_DESCRIPTORS, model.id);
+}
+/**
+ * Fixed local port for pi's own managed prism llama.cpp server. Deliberately NOT the generic
+ * built-in `llama-cpp/local` catalog model's conventional port 8080 (models.generated.ts) — that
+ * port is the convention for a user's OWN manually started llama-server pi merely points at; pi's
+ * managed instance must never collide with it.
+ */
+export const PRISM_LLAMACPP_SERVE_PORT = 8090;
+/**
+ * Conservative RAM-only context cap for the pi-managed prism llama.cpp server, used at FIRST
+ * install time only (`/models add`) — a later self-heal re-serve reuses the model's already
+ * registered `contextWindow` instead (see the readiness gate in local-runtime-controller.ts), never
+ * re-deriving from current RAM, so a served context size can't drift from what pi already told the
+ * rest of the session (compaction, etc.) to expect. PrismLlamaCppRuntime has no `/api/show`-
+ * equivalent GGUF metadata endpoint (unlike Ollama, which context-sizing.ts's
+ * deriveLocalContextSizing depends on), so this is a coarse rung table mirroring
+ * context-sizing.ts's CONTEXT_RUNGS, hard-capped at 32768 regardless of headroom: Bonsai-27B
+ * advertises a 262K-class context, but that is not a realistic KV budget for a CPU-served model on
+ * consumer hardware, and there is no larger validated rung to grow into yet.
+ */
+export function derivePrismLlamaCppNumCtx(totalMemBytes) {
+    const totalGb = totalMemBytes / 1e9;
+    if (totalGb >= 64)
+        return 32_768;
+    if (totalGb >= 32)
+        return 16_384;
+    if (totalGb >= 16)
+        return 8_192;
+    return 4_096;
+}
+/**
+ * The SOLE path allowed to call `runtime.serve()` for a prism llama.cpp model — no caller may spawn
+ * llama-server without first passing through here (both `addPrismLlamaCppModel`'s install pipeline
+ * and the readiness gate's self-heal path call this, never `runtime.serve()` directly).
+ * Unconditionally re-verifies BOTH GGUF files via `runtime.downloadModel()` immediately before every
+ * serve — that call is idempotent (skips a re-download when the local file already matches the
+ * remote size, so the happy path costs one stat) and self-heals a file that went missing from disk
+ * (deleted, a partial download from a prior crash, host cleanup, etc.) by re-fetching it instead of
+ * ever starting llama-server without it. The vision projector is mandatory whenever the descriptor
+ * declares one (Bonsai-27B always does — see BONSAI_27B.mmprojFile): a download failure on EITHER
+ * file means llama-server is never spawned. There is no text-only fallback — Bonsai-27B never
+ * silently degrades to a vision-less serve because the projector went missing.
+ */
+export async function ensurePrismModelFilesThenServe(runtime, descriptor, args, onProgress) {
+    const downloadedModel = await runtime.downloadModel({ repo: descriptor.repo, file: descriptor.file }, (progress) => onProgress(`  ${descriptor.file}: ${progress}`));
+    if (!downloadedModel.ok || !downloadedModel.path) {
+        return { ok: false, stage: "model-download", error: downloadedModel.error ?? "unknown error" };
+    }
+    let mmprojPath;
+    if (descriptor.mmprojFile) {
+        const downloadedMmproj = await runtime.downloadModel({ repo: descriptor.repo, file: descriptor.mmprojFile }, (progress) => onProgress(`  ${descriptor.mmprojFile}: ${progress}`));
+        if (!downloadedMmproj.ok || !downloadedMmproj.path) {
+            return { ok: false, stage: "mmproj-download", error: downloadedMmproj.error ?? "unknown error" };
+        }
+        mmprojPath = downloadedMmproj.path;
+    }
+    onProgress(`Starting ${descriptor.displayName} on 127.0.0.1:${args.port} (context ${args.numCtx})…`);
+    const served = await runtime.serve({
+        modelPath: downloadedModel.path,
+        mmprojPath,
+        port: args.port,
+        numCtx: args.numCtx,
+    });
+    if (!served.ok)
+        return { ok: false, stage: "serve", error: served.error };
+    return { ok: true, baseUrl: served.baseUrl };
+}
+const DEFAULT_HEALTH_CHECK_TIMEOUT_MS = 2_000;
+/**
+ * Direct HTTP reachability probe against a prism llama.cpp server root (NOT the `/v1` OpenAI-compat
+ * suffix models.json registers — same root/health split as llama-server's own health endpoint used
+ * internally by PrismLlamaCppRuntime#serve). PrismLlamaCppRuntime has no public method for this (its
+ * own equivalent, `_healthUp`, is private and scoped to a `serve()` call in progress), so the
+ * readiness gate needs its own — this is deliberately independent of any single runtime instance:
+ * it answers "is a server already alive at this URL", true even for a server this session never
+ * spawned (started earlier, or by a different process).
+ */
+export async function isPrismLlamaCppServerHealthy(serverUrl, fetchFn = fetch, timeoutMs = DEFAULT_HEALTH_CHECK_TIMEOUT_MS) {
+    try {
+        const response = await fetchFn(`${serverUrl}/health`, { signal: AbortSignal.timeout(timeoutMs) });
+        return response.ok;
+    }
+    catch {
+        return false;
+    }
+}
+//# sourceMappingURL=prism-llamacpp-lifecycle.js.map

package/dist/core/models/prism-llamacpp-lifecycle.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"prism-llamacpp-lifecycle.js","sourceRoot":"","sources":["../../../src/core/models/prism-llamacpp-lifecycle.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,UAAU,EAAwD,MAAM,uBAAuB,CAAC;AACzG,OAAO,EAAE,uBAAuB,EAAE,MAAM,yBAAyB,CAAC;AAElE;;;;;;;;;GASG;AACH,MAAM,CAAC,MAAM,0BAA0B,GAAyC;IAC/E,CAAC,UAAU,CAAC,IAAI,CAAC,EAAE,UAAU;CAC7B,CAAC;AAEF;;;;;GAKG;AACH,MAAM,UAAU,6BAA6B,CAAC,KAAuC,EAAW;IAC/F,OAAO,KAAK,CAAC,QAAQ,KAAK,uBAAuB,IAAI,MAAM,CAAC,MAAM,CAAC,0BAA0B,EAAE,KAAK,CAAC,EAAE,CAAC,CAAC;AAAA,CACzG;AAED;;;;;GAKG;AACH,MAAM,CAAC,MAAM,yBAAyB,GAAG,IAAI,CAAC;AAE9C;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,yBAAyB,CAAC,aAAqB,EAAU;IACxE,MAAM,OAAO,GAAG,aAAa,GAAG,GAAG,CAAC;IACpC,IAAI,OAAO,IAAI,EAAE;QAAE,OAAO,MAAM,CAAC;IACjC,IAAI,OAAO,IAAI,EAAE;QAAE,OAAO,MAAM,CAAC;IACjC,IAAI,OAAO,IAAI,EAAE;QAAE,OAAO,KAAK,CAAC;IAChC,OAAO,KAAK,CAAC;AAAA,CACb;AAMD;;;;;;;;;;;;GAYG;AACH,MAAM,CAAC,KAAK,UAAU,8BAA8B,CACnD,OAA6B,EAC7B,UAAgC,EAChC,IAAsC,EACtC,UAAqC,EACG;IACxC,MAAM,eAAe,GAAG,MAAM,OAAO,CAAC,aAAa,CAAC,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,EAAE,CAAC,QAAQ,EAAE,EAAE,CAClH,UAAU,CAAC,KAAK,UAAU,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC,CAC/C,CAAC;IACF,IAAI,CAAC,eAAe,CAAC,EAAE,IAAI,CAAC,eAAe,CAAC,IAAI,EAAE,CAAC;QAClD,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,KAAK,EAAE,eAAe,CAAC,KAAK,IAAI,eAAe,EAAE,CAAC;IAChG,CAAC;IAED,IAAI,UAA8B,CAAC;IACnC,IAAI,UAAU,CAAC,UAAU,EAAE,CAAC;QAC3B,MAAM,gBAAgB,GAAG,MAAM,OAAO,CAAC,aAAa,CACnD,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,IAAI,EAAE,UAAU,CAAC,UAAU,EAAE,EACtD,CAAC,QAAQ,EAAE,EAAE,CAAC,UAAU,CAAC,KAAK,UAAU,CAAC,UAAU,KAAK,QAAQ,EAAE,CAAC,CACnE,CAAC;QACF,IAAI,CAAC,gBAAgB,CAAC,EAAE,IAAI,CAAC,gBAAgB,CAAC,IAAI,EAAE,CAAC;YACpD,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,iBAAiB,EAAE,KAAK,EAAE,gBAAgB,CAAC,KAAK,IAAI,eAAe,EAAE,CAAC;QAClG,CAAC;QACD,UAAU,GAAG,gBAAgB,CAAC,IAAI,CAAC;IACpC,CAAC;IAED,UAAU,CAAC,YAAY,UAAU,CAAC,WAAW,iBAAiB,IAAI,CAAC,IAAI,aAAa,IAAI,CAAC,MAAM,MAAI,CAAC,CAAC;IACrG,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,KAAK,CAAC;QAClC,SAAS,EAAE,eAAe,CAAC,IAAI;QAC/B,UAAU;QACV,IAAI,EAAE,IAAI,CAAC,IAAI;QACf,MAAM,EAAE,IAAI,CAAC,MAAM;KACnB,CAAC,CAAC;IACH,IAAI,CAAC,MAAM,CAAC,EAAE;QAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,CAAC;IAC1E,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC;AAAA,CAC7C;AAED,MAAM,+BAA+B,GAAG,KAAK,CAAC;AAE9C;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,4BAA4B,CACjD,SAAiB,EACjB,OAAO,GAAiB,KAAK,EAC7B,SAAS,GAAW,+BAA+B,EAChC;IACnB,IAAI,CAAC;QACJ,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,GAAG,SAAS,SAAS,EAAE,EAAE,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC;QAClG,OAAO,QAAQ,CAAC,EAAE,CAAC;IACpB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,KAAK,CAAC;IACd,CAAC;AAAA,CACD","sourcesContent":["/**\n * Core lifecycle logic for pi-managed prism llama.cpp models (Bonsai-27B first), shared between the\n * `/models add` install pipeline (modes/interactive/local-model-commands.ts) and the router's\n * readiness gate (local-runtime-controller.ts). ONE place owns \"ensure both GGUF files exist, then\n * serve\" so a turn routed to a cold Bonsai-27B and a fresh `/models add` both go through the\n * identical self-healing path instead of two parallel, potentially-diverging implementations.\n *\n * Lives in core/models/ (not modes/) so the headless readiness gate — used by print/RPC/interactive\n * sessions and isolated/background lanes alike — never depends on the interactive-only UI layer;\n * mirrors the existing local-runtime.ts (mechanics) / local-runtime-controller.ts (gate) split.\n */\n\nimport { BONSAI_27B, type PrismLlamaCppRuntime, type PrismModelDescriptor } from \"./llamacpp-runtime.ts\";\nimport { PRISM_LLAMACPP_PROVIDER } from \"./local-registration.ts\";\n\n/**\n * Curated prism llama.cpp descriptors, keyed by repo id — matches the modelId model-ref.ts's\n * curated `PRISM_LLAMACPP_MODEL_IDS` route returns, and the id `registerPrismLlamaCppModel` writes\n * as the model's id in models.json. Add future curated prism-ml models to both places together.\n *\n * Also the pi-managed-vs-user-owned discriminator (see {@link isPiManagedPrismLlamaCppModel}): a\n * `llama-cpp` provider model is gated/self-healed by pi ONLY when its id is a key here — a user's\n * own hand-configured entry (e.g. the built-in `llama-cpp/local` catalog model on port 8080,\n * pointing at a server the user runs themselves) is never touched.\n */\nexport const PRISM_LLAMACPP_DESCRIPTORS: Record<string, PrismModelDescriptor> = {\n\t[BONSAI_27B.repo]: BONSAI_27B,\n};\n\n/**\n * True only for a `llama-cpp` provider model pi itself registered (id present in\n * {@link PRISM_LLAMACPP_DESCRIPTORS}) — never true for a user's own hand-configured llama-cpp\n * model, including the built-in `llama-cpp/local` catalog entry. This is the gate that must run\n * before ANY self-heal/serve action so pi never touches a server it doesn't own.\n */\nexport function isPiManagedPrismLlamaCppModel(model: { provider: string; id: string }): boolean {\n\treturn model.provider === PRISM_LLAMACPP_PROVIDER && Object.hasOwn(PRISM_LLAMACPP_DESCRIPTORS, model.id);\n}\n\n/**\n * Fixed local port for pi's own managed prism llama.cpp server. Deliberately NOT the generic\n * built-in `llama-cpp/local` catalog model's conventional port 8080 (models.generated.ts) — that\n * port is the convention for a user's OWN manually started llama-server pi merely points at; pi's\n * managed instance must never collide with it.\n */\nexport const PRISM_LLAMACPP_SERVE_PORT = 8090;\n\n/**\n * Conservative RAM-only context cap for the pi-managed prism llama.cpp server, used at FIRST\n * install time only (`/models add`) — a later self-heal re-serve reuses the model's already\n * registered `contextWindow` instead (see the readiness gate in local-runtime-controller.ts), never\n * re-deriving from current RAM, so a served context size can't drift from what pi already told the\n * rest of the session (compaction, etc.) to expect. PrismLlamaCppRuntime has no `/api/show`-\n * equivalent GGUF metadata endpoint (unlike Ollama, which context-sizing.ts's\n * deriveLocalContextSizing depends on), so this is a coarse rung table mirroring\n * context-sizing.ts's CONTEXT_RUNGS, hard-capped at 32768 regardless of headroom: Bonsai-27B\n * advertises a 262K-class context, but that is not a realistic KV budget for a CPU-served model on\n * consumer hardware, and there is no larger validated rung to grow into yet.\n */\nexport function derivePrismLlamaCppNumCtx(totalMemBytes: number): number {\n\tconst totalGb = totalMemBytes / 1e9;\n\tif (totalGb >= 64) return 32_768;\n\tif (totalGb >= 32) return 16_384;\n\tif (totalGb >= 16) return 8_192;\n\treturn 4_096;\n}\n\nexport type EnsurePrismModelServedResult =\n\t| { ok: true; baseUrl: string }\n\t| { ok: false; stage: \"model-download\" | \"mmproj-download\" | \"serve\"; error: string };\n\n/**\n * The SOLE path allowed to call `runtime.serve()` for a prism llama.cpp model — no caller may spawn\n * llama-server without first passing through here (both `addPrismLlamaCppModel`'s install pipeline\n * and the readiness gate's self-heal path call this, never `runtime.serve()` directly).\n * Unconditionally re-verifies BOTH GGUF files via `runtime.downloadModel()` immediately before every\n * serve — that call is idempotent (skips a re-download when the local file already matches the\n * remote size, so the happy path costs one stat) and self-heals a file that went missing from disk\n * (deleted, a partial download from a prior crash, host cleanup, etc.) by re-fetching it instead of\n * ever starting llama-server without it. The vision projector is mandatory whenever the descriptor\n * declares one (Bonsai-27B always does — see BONSAI_27B.mmprojFile): a download failure on EITHER\n * file means llama-server is never spawned. There is no text-only fallback — Bonsai-27B never\n * silently degrades to a vision-less serve because the projector went missing.\n */\nexport async function ensurePrismModelFilesThenServe(\n\truntime: PrismLlamaCppRuntime,\n\tdescriptor: PrismModelDescriptor,\n\targs: { port: number; numCtx: number },\n\tonProgress: (message: string) => void,\n): Promise<EnsurePrismModelServedResult> {\n\tconst downloadedModel = await runtime.downloadModel({ repo: descriptor.repo, file: descriptor.file }, (progress) =>\n\t\tonProgress(`  ${descriptor.file}: ${progress}`),\n\t);\n\tif (!downloadedModel.ok || !downloadedModel.path) {\n\t\treturn { ok: false, stage: \"model-download\", error: downloadedModel.error ?? \"unknown error\" };\n\t}\n\n\tlet mmprojPath: string | undefined;\n\tif (descriptor.mmprojFile) {\n\t\tconst downloadedMmproj = await runtime.downloadModel(\n\t\t\t{ repo: descriptor.repo, file: descriptor.mmprojFile },\n\t\t\t(progress) => onProgress(`  ${descriptor.mmprojFile}: ${progress}`),\n\t\t);\n\t\tif (!downloadedMmproj.ok || !downloadedMmproj.path) {\n\t\t\treturn { ok: false, stage: \"mmproj-download\", error: downloadedMmproj.error ?? \"unknown error\" };\n\t\t}\n\t\tmmprojPath = downloadedMmproj.path;\n\t}\n\n\tonProgress(`Starting ${descriptor.displayName} on 127.0.0.1:${args.port} (context ${args.numCtx})…`);\n\tconst served = await runtime.serve({\n\t\tmodelPath: downloadedModel.path,\n\t\tmmprojPath,\n\t\tport: args.port,\n\t\tnumCtx: args.numCtx,\n\t});\n\tif (!served.ok) return { ok: false, stage: \"serve\", error: served.error };\n\treturn { ok: true, baseUrl: served.baseUrl };\n}\n\nconst DEFAULT_HEALTH_CHECK_TIMEOUT_MS = 2_000;\n\n/**\n * Direct HTTP reachability probe against a prism llama.cpp server root (NOT the `/v1` OpenAI-compat\n * suffix models.json registers — same root/health split as llama-server's own health endpoint used\n * internally by PrismLlamaCppRuntime#serve). PrismLlamaCppRuntime has no public method for this (its\n * own equivalent, `_healthUp`, is private and scoped to a `serve()` call in progress), so the\n * readiness gate needs its own — this is deliberately independent of any single runtime instance:\n * it answers \"is a server already alive at this URL\", true even for a server this session never\n * spawned (started earlier, or by a different process).\n */\nexport async function isPrismLlamaCppServerHealthy(\n\tserverUrl: string,\n\tfetchFn: typeof fetch = fetch,\n\ttimeoutMs: number = DEFAULT_HEALTH_CHECK_TIMEOUT_MS,\n): Promise<boolean> {\n\ttry {\n\t\tconst response = await fetchFn(`${serverUrl}/health`, { signal: AbortSignal.timeout(timeoutMs) });\n\t\treturn response.ok;\n\t} catch {\n\t\treturn false;\n\t}\n}\n"]}