dsh-plugin-local-infer 0.1.2 → 0.1.3

package/packages/providers/lib/adapters.js

@@ -1,61 +1,13 @@
 /**
- * The concrete engine adapters: NInfer (docker and local) and Hyper-Qwen
- * (compose). Each composes the provider `ParameterSpec` data into a
- * `LaunchPlan` and drives its deploy, health-check, and stop steps through an
- * injectable `LifecycleHost`, so the tests stub the host and never run a real
- * build, docker, or network in CI. The node implementation
- * (`nodeLifecycleHost`) is what the service wires in on a live host.
+ * The shared engine-adapter helpers: the spec-to-argv and spec-to-env
+ * translators the concrete adapters (the `engines/*Adapter.ts` files) compose
+ * into a `LaunchPlan`, the card's Custom Arguments translators, the
+ * `LifecycleHost` the deploy, health-check, and stop steps drive, and the
+ * health poll. The node implementation (`nodeLifecycleHost`) is what the
+ * service wires in on a live host.
  * @module dsh-plugin-local-infer/providers/adapters
  */
 import { exec, spawn as nodeSpawn } from 'node:child_process';
-import { basename, dirname, join } from 'node:path';
-import { EngineAdapter, formatLaunchCommand } from '../../core/lib/index.js';
-import { syncAndRebuildEngine } from "./acquire.js";
-import { NINFER_PARAMETERS } from "./ninfer.js";
-import { HYPERQWEN_PARAMETERS } from "./hyperqwen.js";
-import { LLAMACPP_MODEL_ARTIFACTS, LLAMACPP_PARAMETERS, llamaCppModelKey, } from "./llamacpp.js";
-import { INFERNIX_MODEL_ARTIFACTS, INFERNIX_PARAMETERS, infernixArtifactFor, infernixDraftModelPath, infernixModelKey, } from "./infernix.js";
-export { INFERNIX_MODEL_ARTIFACTS } from "./infernix.js";
-import { STRATA_PARAMETERS } from "./strata.js";
-/**
- * The NInfer model artifacts, keyed by the model registry id. The 27B entry
- * is the default the DEFAULT_MODELS list names; the 35B-A3B entries are the
- * Qwen3.8 measured artifact (`qwen3-35b`) and the Qwen3.6 default
- * (`qwen36-35b`).
- */
-export const NINFER_MODEL_ARTIFACTS = {
-    'qwen3-27b': {
-        file: 'qwen3_8_27b_nvfp4.ninfer',
-        repo: 'neroued/Qwen3.8-27B-nvfp4-NInfer',
-        modelId: 'Qwen3.8-27B',
-    },
-    // The Swift 1.5 ablated 27B the local launcher bootstraps: the nvfp4-full
-    // DFlash2 artifact the qwen-ninfer script serves. It reuses the 27B model id
-    // the engine advertises, so a chat turn routes to the same alias.
-    'qwen3-27b-swift15': {
-        file: 'qwen3_8_27b_swift15abl_nvfp4full-dflash2.ninfer',
-        repo: 'kaushikvira/Qwen3.8-27B-swift15-uncensored-ajgazin-nvfp4full-dflash2-NInfer-v3',
-        modelId: 'Qwen3.8-27B',
-    },
-    'qwen3-35b': {
-        file: 'qwen3_8_35b_a3b.ninfer',
-        repo: 'Neroued/Qwen3.8-35B-A3B-NInfer',
-        modelId: 'Qwen3.8-35B-A3B',
-    },
-    'qwen36-35b': {
-        file: 'qwen3_6_35b_a3b.ninfer',
-        repo: 'neroued/Qwen3.6-35B-A3B-NInfer',
-        modelId: 'Qwen3.6-35B-A3B',
-    },
-};
-/**
- * artifactFor resolves the artifact a model id deploys. An id no entry names
- * falls back to the 27B artifact, so a bare config always renders a plan.
- */
-function artifactFor(modelId) {
-    const artifact = modelId !== undefined ? NINFER_MODEL_ARTIFACTS[modelId] : undefined;
-    return artifact ?? NINFER_MODEL_ARTIFACTS['qwen3-27b'];
-}
 /**
  * flagsFromSpecs turns a spec surface and a config into the CLI flag list: each
  * spec's `toArgv` runs over the config value (falling back to the spec
@@ -313,7 +265,7 @@ export const nodeLifecycleHost = createNodeLifecycleHost();
  * signals a pid a recycled process has taken over. A handle whose host cannot
  * watch (onExit absent) stays tracked until stop removes it.
  */
-function trackChild(children, key, handle) {
+export function trackChild(children, key, handle) {
     children.set(key, handle);
     handle.onExit?.(() => {
         // Compare identity before deleting: a newer launch that already replaced
@@ -348,839 +300,14 @@ export async function pollHealth(host, plan, options) {
         detail: `health check did not return 200 within ${maxAttempts} attempts`,
     };
 }
-function delay(ms) {
+/**
+ * delay waits `ms` milliseconds. The health polls use it between attempts; a
+ * poll interval of 0 skips the wait, which the tests use to run the loop
+ * without a real sleep.
+ */
+export function delay(ms) {
     return new Promise(resolve => {
         setTimeout(resolve, ms);
     });
 }
-/**
- * The NInfer docker image the adapter pulls and runs. The engineFork platform
- * config overrides the git source for a local build; the image stays the
- * upstream ghcr artifact.
- */
-export const NINFER_DOCKER_IMAGE = 'ghcr.io/iamwavecut/ninfer-all';
-/**
- * NInferAdapter is the canonical both-capable engine. `renderLaunch` renders a
- * local `ninfer-serve` process or a `docker run` of the ghcr image from the
- * same spec surface; `deploy` builds the engine in the engine directory (or
- * pulls the image), downloads the model into the model directory, and starts
- * the engine; `healthCheck` polls the health URL; `stop` signals the tracked
- * process and stops the named container.
- */
-export class NInferAdapter extends EngineAdapter {
-    id = 'ninfer';
-    supports = { docker: true, local: true };
-    host;
-    children = new Map();
-    constructor(host = nodeLifecycleHost) {
-        super();
-        this.host = host;
-    }
-    parameters() {
-        return NINFER_PARAMETERS;
-    }
-    modelRef(config) {
-        // The id the engine advertises on its OpenAI endpoint is the artifact's
-        // --model-id alias, not the VramModel registry key the config names, so
-        // the registrar publishes the alias and a chat turn routes to it.
-        const artifact = artifactFor(typeof config.model === 'string' ? config.model : 'qwen3-27b');
-        return { id: artifact.modelId, name: artifact.modelId };
-    }
-    launchSignature(config = {}) {
-        // The local plan runs the ninfer-serve binary; the docker plan's host
-        // process is docker run of the ghcr image, so the image name is the
-        // command-line marker the reserve probe matches on.
-        const port = this.port(config);
-        if (this.buildMode(config) === 'docker') {
-            return { binary: NINFER_DOCKER_IMAGE, port };
-        }
-        return { binary: 'ninfer-serve', port };
-    }
-    renderLaunch(config, _hw) {
-        const port = this.port(config);
-        const modelId = typeof config.model === 'string' ? config.model : 'qwen3-27b';
-        const artifact = artifactFor(modelId);
-        // A model key that names a local artifact file (a directory-scanned
-        // model) overrides the artifact map's file, so a file the user has on
-        // disk launches as-is.
-        const file = modelId.endsWith('.ninfer') ? modelId : artifact.file;
-        const modelDirectory = typeof config.modelDirectory === 'string' ? config.modelDirectory : '';
-        const flags = flagsFromSpecs(NINFER_PARAMETERS, config);
-        if (this.buildMode(config) === 'docker') {
-            const modelPath = modelDirectory === '' ? file : join(modelDirectory, file);
-            const containerModelPath = `/models/${basename(file)}`;
-            const modelMount = modelDirectory === ''
-                ? `${file}:${containerModelPath}`
-                : `${dirname(modelPath)}:/models`;
-            const containerName = `ninfer-${port}`;
-            const customArgs = customArgFlags(config);
-            return {
-                engineId: this.id,
-                kind: 'docker',
-                argv: [
-                    'docker',
-                    'run',
-                    '--rm',
-                    '--gpus',
-                    'all',
-                    '--name',
-                    containerName,
-                    '-p',
-                    `127.0.0.1:${port}:8080`,
-                    '--ulimit',
-                    'memlock=-1',
-                    '-v',
-                    'ninfer-cache:/cache',
-                    '-v',
-                    modelMount,
-                    NINFER_DOCKER_IMAGE,
-                    'serve',
-                    containerModelPath,
-                    '--model-id',
-                    artifact.modelId,
-                    ...flags,
-                    ...customArgs,
-                ],
-                env: {},
-                containerName,
-                modelId,
-                ...(modelDirectory === '' ? {} : { modelDirectory }),
-                port,
-                healthUrl: `http://127.0.0.1:${port}/health`,
-            };
-        }
-        const modelPath = modelDirectory === '' ? file : join(modelDirectory, file);
-        const customArgs = customArgFlags(config);
-        return {
-            engineId: this.id,
-            kind: 'process',
-            argv: [
-                'ninfer-serve',
-                modelPath,
-                '--host',
-                '127.0.0.1',
-                '--port',
-                String(port),
-                '--model-id',
-                artifact.modelId,
-                ...flags,
-                ...customArgs,
-            ],
-            env: {},
-            modelId,
-            ...(modelDirectory === '' ? {} : { modelDirectory }),
-            port,
-            healthUrl: `http://127.0.0.1:${port}/health`,
-        };
-    }
-    async deploy(plan, platform, onProgress, options) {
-        if (plan.kind === 'process') {
-            const engineDir = platform?.engineDirectory;
-            let buildLogs = '';
-            // Auto Rebuild off (the card's checkbox): the git sync (fetch, rebase,
-            // rebuild) is skipped and the checkout launches as-is.
-            if (engineDir !== undefined && options?.autoRebuild !== false) {
-                // The git sync is the checkout download: the card's button reads
-                // "Downloading..." for it, not "Building...".
-                onProgress?.('download');
-                // The conditional rebuild fetches the checkout, prices it against the
-                // remote branch, and rebuilds only when it is behind (or a fresh
-                // clone). The decision, the build command, and its output are captured
-                // into buildLogs and threaded to the card's Logs section.
-                const sync = await syncAndRebuildEngine(this.host, {
-                    engineId: this.id,
-                    directory: engineDir,
-                    ...(platform?.engineFork !== undefined ? { fork: platform.engineFork } : {}),
-                    ...(plan.argv?.[0] !== undefined ? { binary: plan.argv[0] } : {}),
-                    ...(options?.allowRunDeps !== undefined ? { allowRunDeps: options.allowRunDeps } : {}),
-                }, () => { });
-                buildLogs = sync.logs;
-                if (!sync.ok) {
-                    return { ok: false, detail: sync.detail ?? 'unknown', logs: buildLogs };
-                }
-                // The build step reads as "Building..." only when a build actually
-                // ran (a fresh clone, or a rebase of a checkout that was behind).
-                if (sync.rebuilt) {
-                    onProgress?.('build');
-                }
-            }
-            const withLogs = (result) => buildLogs !== '' ? { ...result, logs: buildLogs } : result;
-            onProgress?.('download');
-            const download = await this.downloadModel(plan);
-            if (!download.ok)
-                return withLogs(download);
-            // A chat template the user picked as an HF repo reference (the combo
-            // control's known choices) is not a path the engine can read, so the
-            // deploy step downloads it into the model directory and rewrites the
-            // --chat-template flag to the local file before the spawn.
-            onProgress?.('download');
-            const templateResult = await this.resolveChatTemplate(plan);
-            if (!templateResult.ok)
-                return withLogs(templateResult);
-            const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {}, engineDir);
-            trackChild(this.children, this.planKey(plan), handle);
-            return withLogs({ ok: true, handle });
-        }
-        if (plan.kind === 'docker') {
-            // The image pull is the docker build step: the card's button reads
-            // "Building..." for it.
-            onProgress?.('build');
-            const pull = await this.host.run(`docker pull ${NINFER_DOCKER_IMAGE}`);
-            if (!pull.ok) {
-                return { ok: false, detail: `docker pull failed: ${pull.detail ?? 'unknown'}` };
-            }
-            const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {});
-            trackChild(this.children, this.planKey(plan), handle);
-            return { ok: true, handle };
-        }
-        // NInfer renders process and docker plans; a compose plan is a no-op.
-        return { ok: true };
-    }
-    async healthCheck(plan, options) {
-        return pollHealth(this.host, plan, options);
-    }
-    async stop(plan) {
-        if (plan.kind === 'docker' && plan.containerName !== undefined) {
-            await this.host.run(`docker stop ${plan.containerName}`);
-        }
-        const handle = this.children.get(this.planKey(plan));
-        if (handle !== undefined) {
-            handle.kill();
-            this.children.delete(this.planKey(plan));
-        }
-    }
-    buildMode(config) {
-        return config.buildMode === 'docker' ? 'docker' : 'local';
-    }
-    port(config) {
-        return typeof config.port === 'number' ? config.port : 18020;
-    }
-    planKey(plan) {
-        return `${plan.engineId}:${plan.port}`;
-    }
-    async downloadModel(plan) {
-        const artifact = plan.modelId !== undefined ? NINFER_MODEL_ARTIFACTS[plan.modelId] : undefined;
-        if (artifact === undefined)
-            return { ok: true };
-        const target = plan.modelDirectory ?? '.';
-        const download = await this.host.run(`hf download ${artifact.repo} ${artifact.file} --local-dir ${target}`);
-        if (!download.ok) {
-            return { ok: false, detail: `model download failed: ${download.detail ?? 'unknown'}` };
-        }
-        return { ok: true };
-    }
-    /**
-     * resolveChatTemplate rewrites a --chat-template flag that carries an HF repo
-     * reference into the local file the engine reads. A value with a slash that
-     * is not an absolute path (the combo control's known choices) is a repo the
-     * deploy step downloads into the model directory; an already-local path is
-     * left as-is. The download is best-effort: a template that is already on
-     * disk is used without a re-fetch, and a missing hf CLI leaves the flag
-     * untouched so the engine reports the missing template itself.
-     */
-    async resolveChatTemplate(plan) {
-        const argv = plan.argv;
-        if (argv === undefined)
-            return { ok: true };
-        const index = argv.indexOf('--chat-template');
-        if (index === -1)
-            return { ok: true };
-        const value = argv[index + 1];
-        if (value === undefined)
-            return { ok: true };
-        // An absolute path (a leading /) is already local; a value with no slash
-        // is a bare file name the engine resolves against its own directory.
-        if (value.startsWith('/') || !value.includes('/'))
-            return { ok: true };
-        // An HF repo reference: download the template file into the model
-        // directory and point the flag at the local copy.
-        const target = plan.modelDirectory ?? '.';
-        const localPath = `${target.endsWith('/') ? '' : `${target}/`}chat_template.jinja`;
-        // A template already on disk skips the re-fetch; the hf CLI is optional,
-        // so a download failure is reported but does not abort the launch.
-        const probe = await this.host.run(`[ -f ${localPath} ]`);
-        if (!probe.ok) {
-            const download = await this.host.run(`hf download ${value} chat_template.jinja --local-dir ${target}`);
-            if (!download.ok) {
-                return {
-                    ok: false,
-                    detail: `chat template download failed: ${download.detail ?? 'unknown'}`,
-                };
-            }
-        }
-        argv[index + 1] = localPath;
-        return { ok: true };
-    }
-}
-/**
- * HyperQwenAdapter is the docker-compose engine. `renderLaunch` renders a
- * compose plan (the compose file, the profile, and the `.env` the spec surface
- * maps to); `deploy` runs `docker compose up`; `healthCheck` polls the health
- * URL; `stop` runs `docker compose down`.
- */
-export class HyperQwenAdapter extends EngineAdapter {
-    id = 'hyperqwen';
-    supports = { docker: true, local: false };
-    host;
-    constructor(host = nodeLifecycleHost) {
-        super();
-        this.host = host;
-    }
-    parameters() {
-        return HYPERQWEN_PARAMETERS;
-    }
-    modelRef(_config) {
-        // vLLM serves the model under the name the compose file passes, and the
-        // plugin tracks no per-model data for Hyper-Qwen yet, so the engine id is
-        // the stable entry the registrar publishes until that data lands.
-        return { id: 'hyperqwen', name: 'Hyper-Qwen' };
-    }
-    launchSignature(config = {}) {
-        // The compose host process is `docker compose ... up`, so the marker is
-        // the command, not an image. The toEnv keys (SPEC, CTX, KV, ...) name the
-        // .env the compose file reads, not the host process env, so they ride
-        // along as the env-key element the matcher checks when a process line
-        // carries env data.
-        const port = this.port(config);
-        const envKeys = Object.keys(envFromSpecs(HYPERQWEN_PARAMETERS, config));
-        return { binary: 'docker compose', port, envKeys };
-    }
-    renderLaunch(config, _hw) {
-        const port = this.port(config);
-        const profile = this.profile(config);
-        // The spec env and the card's Custom Arguments merge into one env map: the
-        // spec keys are the .env the compose file reads, and the custom rows are
-        // the user's own overrides appended after them.
-        const env = { ...envFromSpecs(HYPERQWEN_PARAMETERS, config), ...customArgEnv(config) };
-        const composeFile = typeof config.composeFile === 'string' ? config.composeFile : 'docker-compose.yml';
-        return {
-            engineId: this.id,
-            kind: 'compose',
-            env,
-            composeFile,
-            profile,
-            port,
-            healthUrl: `http://127.0.0.1:${port}/health`,
-        };
-    }
-    async deploy(plan, platform, onProgress, _options) {
-        // The card's docker-tag field carries the compose arguments (the profile
-        // the up step selects); the deploy step applies it to the plan so the up
-        // line and the card's Logs section agree.
-        if (platform?.dockerTag !== undefined) {
-            plan.composeArgs = platform.dockerTag;
-        }
-        // The core formatter is the single source of the compose up line: the
-        // card's Logs section prints the same string this run gets.
-        const command = formatLaunchCommand(plan);
-        if (command === undefined) {
-            return { ok: false, detail: 'compose plan carries no command' };
-        }
-        // The compose up is the docker build step: the card's button reads
-        // "Building..." for it.
-        onProgress?.('build');
-        const up = await this.host.run(command);
-        if (!up.ok) {
-            return { ok: false, detail: `compose up failed: ${up.detail ?? 'unknown'}` };
-        }
-        return { ok: true };
-    }
-    async healthCheck(plan, options) {
-        return pollHealth(this.host, plan, options);
-    }
-    async stop(plan) {
-        // The down step brings down the containers the up step started, so it
-        // runs the same compose arguments (the profile the up selected); an
-        // absent composeArgs falls back to the engine's default profile.
-        const composeArgs = plan.composeArgs?.split(/\s+/).filter((token) => token !== '');
-        const args = composeArgs !== undefined && composeArgs.length > 0
-            ? composeArgs.join(' ')
-            : `--profile ${plan.profile ?? 'single'}`;
-        await this.host.run(`docker compose -f ${plan.composeFile ?? 'docker-compose.yml'} ${args} down`);
-    }
-    port(config) {
-        return typeof config.port === 'number' ? config.port : 18020;
-    }
-    profile(config) {
-        return typeof config.profile === 'string' ? config.profile : 'single';
-    }
-}
-/**
- * The llama.cpp docker image the adapter pulls and runs. The engineFork
- * platform config overrides the git source for a local build; the image stays
- * the upstream ggml artifact.
- */
-export const LLAMACPP_DOCKER_IMAGE = 'ghcr.io/ggml-org/llama.cpp';
-/**
- * llamaCppArtifact resolves the artifact a model key deploys. A key no entry
- * names falls back to the 27B artifact, so a bare config always renders a plan.
- */
-function llamaCppArtifact(modelKey) {
-    const artifact = modelKey !== undefined ? LLAMACPP_MODEL_ARTIFACTS[modelKey] : undefined;
-    return artifact ?? LLAMACPP_MODEL_ARTIFACTS['qwen3-27b'];
-}
-/**
- * LlamaCppAdapter is the llama.cpp engine, both-capable like NInfer.
- * `renderLaunch` renders a local `llama-server` process or a `docker run` of the
- * ggml image from the same `LLAMACPP_PARAMETERS` surface; `deploy` builds the
- * engine in the engine directory (or pulls the image), downloads the GGUF model
- * into the model directory, and starts the engine; `healthCheck` polls the health
- * URL; `stop` signals the tracked process and stops the named container.
- */
-export class LlamaCppAdapter extends EngineAdapter {
-    id = 'llamacpp';
-    supports = { docker: true, local: true };
-    host;
-    children = new Map();
-    constructor(host = nodeLifecycleHost) {
-        super();
-        this.host = host;
-    }
-    parameters() {
-        return LLAMACPP_PARAMETERS;
-    }
-    modelRef(config) {
-        // The engine serves the model under the artifact's advertised id, so the
-        // registrar publishes that alias and a chat turn routes to it.
-        const artifact = llamaCppArtifact(llamaCppModelKey(config));
-        return { id: artifact.modelId, name: artifact.modelId };
-    }
-    launchSignature(config = {}) {
-        // The local plan runs the llama-server binary; the docker plan's host
-        // process is docker run of the ggml image, so the image name is the
-        // command-line marker the reserve probe matches on.
-        const port = this.port(config);
-        if (this.buildMode(config) === 'docker') {
-            return { binary: LLAMACPP_DOCKER_IMAGE, port };
-        }
-        return { binary: 'llama-server', port };
-    }
-    renderLaunch(config, _hw) {
-        const port = this.port(config);
-        const modelId = llamaCppModelKey(config);
-        const artifact = llamaCppArtifact(modelId);
-        // A model key that names a local GGUF file (a directory-scanned model)
-        // overrides the artifact map's file, so a file the user has on disk
-        // launches as-is.
-        const file = modelId.endsWith('.gguf') ? modelId : artifact.file;
-        const modelDirectory = typeof config.modelDirectory === 'string' ? config.modelDirectory : '';
-        const flags = flagsFromSpecs(LLAMACPP_PARAMETERS, config);
-        if (this.buildMode(config) === 'docker') {
-            const modelPath = modelDirectory === '' ? file : join(modelDirectory, file);
-            const containerModelPath = `/models/${basename(file)}`;
-            const modelMount = modelDirectory === ''
-                ? `${file}:${containerModelPath}`
-                : `${dirname(modelPath)}:/models`;
-            const containerName = `llamacpp-${port}`;
-            const customArgs = customArgFlags(config);
-            return {
-                engineId: this.id,
-                kind: 'docker',
-                argv: [
-                    'docker',
-                    'run',
-                    '--rm',
-                    '--gpus',
-                    'all',
-                    '--name',
-                    containerName,
-                    '-p',
-                    `127.0.0.1:${port}:8080`,
-                    '-v',
-                    modelMount,
-                    LLAMACPP_DOCKER_IMAGE,
-                    '--model',
-                    containerModelPath,
-                    ...flags,
-                    ...customArgs,
-                ],
-                env: {},
-                containerName,
-                modelId,
-                ...(modelDirectory === '' ? {} : { modelDirectory }),
-                port,
-                healthUrl: `http://127.0.0.1:${port}/health`,
-            };
-        }
-        const modelPath = modelDirectory === '' ? file : join(modelDirectory, file);
-        const customArgs = customArgFlags(config);
-        return {
-            engineId: this.id,
-            kind: 'process',
-            argv: [
-                'llama-server',
-                '--model',
-                modelPath,
-                '--host',
-                '127.0.0.1',
-                '--port',
-                String(port),
-                ...flags,
-                ...customArgs,
-            ],
-            env: {},
-            modelId,
-            ...(modelDirectory === '' ? {} : { modelDirectory }),
-            port,
-            healthUrl: `http://127.0.0.1:${port}/health`,
-        };
-    }
-    async deploy(plan, platform, onProgress, options) {
-        if (plan.kind === 'process') {
-            const engineDir = platform?.engineDirectory;
-            let buildLogs = '';
-            // Auto Rebuild off (the card's checkbox): the git sync (fetch, rebase,
-            // rebuild) is skipped and the checkout launches as-is.
-            if (engineDir !== undefined && options?.autoRebuild !== false) {
-                // The git sync is the checkout download: the card's button reads
-                // "Downloading..." for it, not "Building...".
-                onProgress?.('download');
-                // The conditional rebuild fetches the checkout, prices it against the
-                // remote branch, and rebuilds only when it is behind (or a fresh
-                // clone). The decision, the build command, and its output are captured
-                // into buildLogs and threaded to the card's Logs section.
-                const sync = await syncAndRebuildEngine(this.host, {
-                    engineId: this.id,
-                    directory: engineDir,
-                    ...(platform?.engineFork !== undefined ? { fork: platform.engineFork } : {}),
-                    ...(plan.argv?.[0] !== undefined ? { binary: plan.argv[0] } : {}),
-                    ...(options?.allowRunDeps !== undefined ? { allowRunDeps: options.allowRunDeps } : {}),
-                }, () => { });
-                buildLogs = sync.logs;
-                if (!sync.ok) {
-                    return { ok: false, detail: sync.detail ?? 'unknown', logs: buildLogs };
-                }
-                // The build step reads as "Building..." only when a build actually
-                // ran (a fresh clone, or a rebase of a checkout that was behind).
-                if (sync.rebuilt) {
-                    onProgress?.('build');
-                }
-            }
-            const withLogs = (result) => buildLogs !== '' ? { ...result, logs: buildLogs } : result;
-            onProgress?.('download');
-            const download = await this.downloadModel(plan);
-            if (!download.ok)
-                return withLogs(download);
-            const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {}, engineDir);
-            trackChild(this.children, this.planKey(plan), handle);
-            return withLogs({ ok: true, handle });
-        }
-        if (plan.kind === 'docker') {
-            // The image pull is the docker build step: the card's button reads
-            // "Building..." for it.
-            onProgress?.('build');
-            const pull = await this.host.run(`docker pull ${LLAMACPP_DOCKER_IMAGE}`);
-            if (!pull.ok) {
-                return { ok: false, detail: `docker pull failed: ${pull.detail ?? 'unknown'}` };
-            }
-            const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {});
-            trackChild(this.children, this.planKey(plan), handle);
-            return { ok: true, handle };
-        }
-        // llama.cpp renders process and docker plans; a compose plan is a no-op.
-        return { ok: true };
-    }
-    async healthCheck(plan, options) {
-        return pollHealth(this.host, plan, options);
-    }
-    async stop(plan) {
-        if (plan.kind === 'docker' && plan.containerName !== undefined) {
-            await this.host.run(`docker stop ${plan.containerName}`);
-        }
-        const handle = this.children.get(this.planKey(plan));
-        if (handle !== undefined) {
-            handle.kill();
-            this.children.delete(this.planKey(plan));
-        }
-    }
-    buildMode(config) {
-        return config.buildMode === 'docker' ? 'docker' : 'local';
-    }
-    port(config) {
-        return typeof config.port === 'number' ? config.port : 8080;
-    }
-    planKey(plan) {
-        return `${plan.engineId}:${plan.port}`;
-    }
-    async downloadModel(plan) {
-        const artifact = plan.modelId !== undefined ? LLAMACPP_MODEL_ARTIFACTS[plan.modelId] : undefined;
-        if (artifact === undefined)
-            return { ok: true };
-        const target = plan.modelDirectory ?? '.';
-        const download = await this.host.run(`hf download ${artifact.repo} ${artifact.file} --local-dir ${target}`);
-        if (!download.ok) {
-            return { ok: false, detail: `model download failed: ${download.detail ?? 'unknown'}` };
-        }
-        return { ok: true };
-    }
-}
-/**
- * The Strata docker image the adapter pulls and runs. Strata is the
- * placeholder engine: the image and the parameter data land when the engine
- * is available, and the adapter renders a bare docker plan until then.
- */
-export const STRATA_DOCKER_IMAGE = 'ghcr.io/strata/strata';
-/**
- * StrataAdapter is the placeholder docker-only engine. `renderLaunch` renders
- * a `docker run` of the Strata image (no spec surface yet, so no flags);
- * `deploy` pulls the image and starts the container; `healthCheck` polls the
- * health URL; `stop` stops the named container. The parameter surface and the
- * per-model data land with the engine, as data edits, never an adapter rewire.
- */
-export class StrataAdapter extends EngineAdapter {
-    id = 'strata';
-    supports = { docker: true, local: false };
-    host;
-    children = new Map();
-    constructor(host = nodeLifecycleHost) {
-        super();
-        this.host = host;
-    }
-    parameters() {
-        return STRATA_PARAMETERS;
-    }
-    modelRef(_config) {
-        // Strata serves the model under the name the container passes, and the
-        // plugin tracks no per-model data for Strata yet, so the engine id is the
-        // stable entry the registrar publishes until that data lands.
-        return { id: 'strata', name: 'Strata' };
-    }
-    launchSignature(config = {}) {
-        // Strata is docker-only, so the signature is the image name the docker
-        // run carries plus the port the plan binds.
-        return { binary: STRATA_DOCKER_IMAGE, port: this.port(config) };
-    }
-    renderLaunch(config, _hw) {
-        const port = this.port(config);
-        const containerName = `strata-${port}`;
-        const customArgs = customArgFlags(config);
-        return {
-            engineId: this.id,
-            kind: 'docker',
-            argv: [
-                'docker',
-                'run',
-                '--rm',
-                '--gpus',
-                'all',
-                '--name',
-                containerName,
-                '-p',
-                `127.0.0.1:${port}:8080`,
-                STRATA_DOCKER_IMAGE,
-                ...customArgs,
-            ],
-            env: {},
-            containerName,
-            port,
-            healthUrl: `http://127.0.0.1:${port}/health`,
-        };
-    }
-    async deploy(plan, _platform, onProgress, _options) {
-        // The image pull is the docker build step: the card's button reads
-        // "Building..." for it.
-        onProgress?.('build');
-        const pull = await this.host.run(`docker pull ${STRATA_DOCKER_IMAGE}`);
-        if (!pull.ok) {
-            return { ok: false, detail: `docker pull failed: ${pull.detail ?? 'unknown'}` };
-        }
-        const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {});
-        trackChild(this.children, this.planKey(plan), handle);
-        return { ok: true, handle };
-    }
-    async healthCheck(plan, options) {
-        return pollHealth(this.host, plan, options);
-    }
-    async stop(plan) {
-        if (plan.containerName !== undefined) {
-            await this.host.run(`docker stop ${plan.containerName}`);
-        }
-        const handle = this.children.get(this.planKey(plan));
-        if (handle !== undefined) {
-            handle.kill();
-            this.children.delete(this.planKey(plan));
-        }
-    }
-    port(config) {
-        return typeof config.port === 'number' ? config.port : 8080;
-    }
-    planKey(plan) {
-        return `${plan.engineId}:${plan.port}`;
-    }
-}
-/**
- * InfernixAdapter is the local-only engine for Infernix, the C++/CUDA fork
- * of NInfer (https://github.com/wallawalla47/infernix) tuned for Qwen models
- * on a single RTX 50-series GPU. `renderLaunch` renders a local
- * `infernix-serve` process from the `INFERNIX_PARAMETERS` surface (the
- * Infernix-only fields, from `--ngram-volume` to the quality routes, spell
- * onto the argv only for this engine); `deploy` builds the engine in the
- * engine directory (the engine's own cmake preset), downloads the model
- * artifact (and the n-gram volume, for the multi-file Flash-Next repo) into
- * the model directory, and starts the engine; `healthCheck` polls the health
- * URL; `stop` signals the tracked process. The engine ships no docker image,
- * so the adapter supports the local build mode only.
- */
-export class InfernixAdapter extends EngineAdapter {
-    id = 'infernix';
-    supports = { docker: false, local: true };
-    host;
-    children = new Map();
-    constructor(host = nodeLifecycleHost) {
-        super();
-        this.host = host;
-    }
-    parameters() {
-        return INFERNIX_PARAMETERS;
-    }
-    modelRef(config) {
-        // The engine serves the model under the artifact's advertised id, so the
-        // registrar publishes that alias and a chat turn routes to it. A model
-        // key that names a local artifact file (a directory-scanned model)
-        // reuses the 27B alias, the same fallback the NInfer adapter applies.
-        const artifact = infernixArtifactFor(infernixModelKey(config));
-        return { id: artifact.modelId, name: artifact.modelId };
-    }
-    launchSignature(config = {}) {
-        // The local plan runs the infernix-serve binary (the .exe on Windows),
-        // so the binary name is the command-line marker the reserve probe
-        // matches on, and the port the plan binds rides along.
-        return { binary: 'infernix-serve', port: this.port(config) };
-    }
-    renderLaunch(config, _hw) {
-        const port = this.port(config);
-        const modelKey = infernixModelKey(config);
-        const artifact = infernixArtifactFor(modelKey);
-        // A model key that names a local artifact file (a directory-scanned
-        // model, a .ninfer or .infernix name) overrides the artifact map's file,
-        // so a file the user has on disk launches as-is.
-        const file = modelKey.endsWith('.ninfer') || modelKey.endsWith('.infernix')
-            ? modelKey
-            : artifact.file;
-        const modelDirectory = typeof config.modelDirectory === 'string' ? config.modelDirectory : '';
-        const modelPath = modelDirectory === '' ? file : join(modelDirectory, file);
-        // The engine binary is the .exe on a Windows host, where the published
-        // release zip ships the server and its DLLs.
-        const binary = process.platform === 'win32' ? 'infernix-serve.exe' : 'infernix-serve';
-        const flags = flagsFromSpecs(INFERNIX_PARAMETERS, config);
-        // The dflash2 backend always loads a drafter through --draft-model, so the
-        // flag is spelled here from the resolved path (the selected DFlash2 preset,
-        // or the Custom Draft model field) rather than by a spec toArgv. A config
-        // that names no drafter path is the misconfiguration the dflash2
-        // guideline names, so the flag is spelled only when a path resolves.
-        if (config.spec === 'dflash2') {
-            const draftPath = infernixDraftModelPath(config);
-            if (draftPath !== null) {
-                flags.push('--draft-model', draftPath);
-            }
-        }
-        const customArgs = customArgFlags(config);
-        return {
-            engineId: this.id,
-            kind: 'process',
-            argv: [
-                binary,
-                modelPath,
-                '--host',
-                '127.0.0.1',
-                '--port',
-                String(port),
-                '--model-id',
-                artifact.modelId,
-                ...flags,
-                ...customArgs,
-            ],
-            env: {},
-            modelId: modelKey,
-            ...(modelDirectory === '' ? {} : { modelDirectory }),
-            port,
-            healthUrl: `http://127.0.0.1:${port}/health`,
-        };
-    }
-    async deploy(plan, platform, onProgress, options) {
-        if (plan.kind !== 'process') {
-            // Infernix renders process plans only; a docker or compose plan is a
-            // no-op.
-            return { ok: true };
-        }
-        const engineDir = platform?.engineDirectory;
-        let buildLogs = '';
-        // Auto Rebuild off (the card's checkbox): the git sync (fetch, rebase,
-        // rebuild) is skipped and the checkout launches as-is.
-        if (engineDir !== undefined && options?.autoRebuild !== false) {
-            // The git sync is the checkout download: the card's button reads
-            // "Downloading..." for it, not "Building...".
-            onProgress?.('download');
-            // The conditional rebuild fetches the checkout, prices it against the
-            // remote branch, and rebuilds only when it is behind (or a fresh
-            // clone). The engine's own quick start (the cmake preset, then the
-            // preset's build) rides the acquisition-data build steps. The decision,
-            // the build command, and its output are captured into buildLogs and
-            // threaded to the card's Logs section.
-            const sync = await syncAndRebuildEngine(this.host, {
-                engineId: this.id,
-                directory: engineDir,
-                ...(platform?.engineFork !== undefined ? { fork: platform.engineFork } : {}),
-                ...(plan.argv?.[0] !== undefined ? { binary: plan.argv[0] } : {}),
-                ...(options?.allowRunDeps !== undefined ? { allowRunDeps: options.allowRunDeps } : {}),
-            }, () => { });
-            buildLogs = sync.logs;
-            if (!sync.ok) {
-                return { ok: false, detail: sync.detail ?? 'unknown', logs: buildLogs };
-            }
-            // The build step reads as "Building..." only when a build actually
-            // ran (a fresh clone, or a rebase of a checkout that was behind).
-            if (sync.rebuilt) {
-                onProgress?.('build');
-            }
-        }
-        const withLogs = (result) => buildLogs !== '' ? { ...result, logs: buildLogs } : result;
-        onProgress?.('download');
-        const download = await this.downloadModel(plan);
-        if (!download.ok)
-            return withLogs(download);
-        const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {}, engineDir);
-        this.children.set(this.planKey(plan), handle);
-        return withLogs({ ok: true, handle });
-    }
-    async healthCheck(plan, options) {
-        return pollHealth(this.host, plan, options);
-    }
-    async stop(plan) {
-        const handle = this.children.get(this.planKey(plan));
-        if (handle !== undefined) {
-            handle.kill();
-            this.children.delete(this.planKey(plan));
-        }
-    }
-    port(config) {
-        // 8080 is the infernix-serve CLI default, so the form seeds it and the
-        // adapter falls back to it when a config names no port.
-        return typeof config.port === 'number' ? config.port : 8080;
-    }
-    planKey(plan) {
-        return `${plan.engineId}:${plan.port}`;
-    }
-    async downloadModel(plan) {
-        const artifact = plan.modelId !== undefined ? INFERNIX_MODEL_ARTIFACTS[plan.modelId] : undefined;
-        // A model key that names a local artifact file carries no artifact entry,
-        // so the file the user has on disk launches as-is without a download.
-        if (artifact === undefined)
-            return { ok: true };
-        const target = plan.modelDirectory ?? '.';
-        // A multi-file artifact (the Flash-Next shards) downloads the whole repo,
-        // which carries the n-gram volume beside the shards, so the
-        // --ngram-volume flag points at a file the same download left on disk.
-        // A single-file artifact downloads the file alone.
-        const repo = artifact.repo;
-        const files = artifact.files ?? [];
-        const download = files.length > 1
-            ? await this.host.run(`hf download ${repo} --local-dir ${target}`)
-            : await this.host.run(`hf download ${repo} ${artifact.file} --local-dir ${target}`);
-        if (!download.ok) {
-            return { ok: false, detail: `model download failed: ${download.detail ?? 'unknown'}` };
-        }
-        return { ok: true };
-    }
-}
 //# sourceMappingURL=adapters.js.map