@ashlr/hub 2.2.0 → 3.0.1

package/dist/core/run/model-profile.d.ts

@@ -0,0 +1,67 @@
+/**
+ * model-profile.ts — M41 model-adaptive capability profiles.
+ *
+ * Local models vary wildly: a 1.5B chat model and a 32B coder need very
+ * different scaffolding. A ModelProfile bundles the knobs that should adapt to
+ * the model actually serving a task — prompt verbosity, tool-call format,
+ * step cap, sampling temperature, and a prompt/context size budget.
+ *
+ * Resolution is by model-NAME pattern only (zero I/O, fully deterministic and
+ * testable), reusing the exact size/coder vocabulary already used by
+ * provider-client.pickModel and router.pickBestLocalModel so detection stays
+ * consistent across the codebase.
+ *
+ * EVERYTHING here is additive and gated: profiles only take effect when
+ * adaptive prompts are enabled (adaptivePromptsEnabled). With the flag off the
+ * harness behaves exactly as before.
+ */
+import type { AshlrConfig } from '../types.js';
+export type PromptVerbosity = 'terse' | 'standard' | 'rich';
+/**
+ * How a model best receives tool intent.
+ *  - 'native': provider tool API (tool_calls) — capable models.
+ *  - 'json':   ask for a JSON tool-call block in content — weak/small models
+ *              that fumble native tool calling but can emit structured text.
+ *  - 'none':   no tools (no-tool contract only).
+ */
+export type ToolFormat = 'native' | 'json' | 'none';
+export interface ModelProfile {
+    /** Stable profile id (e.g. 'coder', 'general', 'small', 'default'). */
+    id: string;
+    /** Prompt verbosity tier — selects the layer variant during assembly. */
+    verbosity: PromptVerbosity;
+    /** Preferred tool-call format. */
+    toolFormat: ToolFormat;
+    /** Optional one-liner appended to the tool layer (e.g. JSON-block hint). */
+    toolFormatHint?: string;
+    /** Per-task ReAct step ceiling (overrides agent-loop TASK_STEP_CAP). */
+    stepCap: number;
+    /** Suggested sampling temperature (lower = more disciplined). */
+    temperature: number;
+    /** Hard character ceiling for the assembled system prompt. */
+    promptCharCap: number;
+    /** Approx context window in tokens (informs the prompt/memory budget). */
+    contextTokens: number;
+}
+/** Fallback when the model name is unknown — behaves like GENERAL. */
+export declare const DEFAULT_PROFILE: ModelProfile;
+/**
+ * Resolve a ModelProfile from a model name. Deterministic, no I/O.
+ *
+ * Bands (most specific first):
+ *  - small  : labeled ≤3B or an explicitly-small family.
+ *  - coder  : coder-specialized name, OR a large (≥20B) capable model.
+ *  - general: any other recognizable chat model.
+ *  - default: unknown / undefined name.
+ *
+ * `overrides` allows per-id shallow tuning (e.g. cfg.models.profiles['coder']).
+ */
+export declare function resolveModelProfile(modelName?: string, overrides?: Record<string, Partial<ModelProfile>>): ModelProfile;
+/**
+ * Whether the M41 adaptive-prompt suite is active. Default OFF.
+ *
+ * Precedence: ASHLR_ADAPTIVE_PROMPTS env (1/true | 0/false) overrides config;
+ * otherwise cfg.models.adaptivePrompts === true. Absent everywhere → false, so
+ * the harness keeps its legacy prompts/step-cap untouched.
+ */
+export declare function adaptivePromptsEnabled(cfg?: AshlrConfig): boolean;

package/dist/core/run/model-profile.js

@@ -0,0 +1,132 @@
+/**
+ * model-profile.ts — M41 model-adaptive capability profiles.
+ *
+ * Local models vary wildly: a 1.5B chat model and a 32B coder need very
+ * different scaffolding. A ModelProfile bundles the knobs that should adapt to
+ * the model actually serving a task — prompt verbosity, tool-call format,
+ * step cap, sampling temperature, and a prompt/context size budget.
+ *
+ * Resolution is by model-NAME pattern only (zero I/O, fully deterministic and
+ * testable), reusing the exact size/coder vocabulary already used by
+ * provider-client.pickModel and router.pickBestLocalModel so detection stays
+ * consistent across the codebase.
+ *
+ * EVERYTHING here is additive and gated: profiles only take effect when
+ * adaptive prompts are enabled (adaptivePromptsEnabled). With the flag off the
+ * harness behaves exactly as before.
+ */
+// ---------------------------------------------------------------------------
+// Name-pattern vocabulary (kept in sync with provider-client / router)
+// ---------------------------------------------------------------------------
+/** Parse a parameter count in billions from a model name ("3b" / "7b" / "72b"). */
+function sizeOf(name) {
+    const b = name.match(/(\d+(?:\.\d+)?)\s*b\b/i);
+    if (b)
+        return parseFloat(b[1]);
+    if (/mini|small|tiny|nano|phi/i.test(name))
+        return 3; // unlabeled small models
+    return -1; // unknown size
+}
+/** Coder/code-specialized models that reward richer engineering discipline. */
+function isCoderModel(name) {
+    return /coder|codellama|code-?llama|codegemma|codestral|starcoder|deepseek-?coder|code-/i.test(name);
+}
+/** Explicitly small/weak families regardless of a labeled size. */
+function isSmallName(name) {
+    return /tinyllama|tiny|nano|mini|gemma:?2b|phi-?[123]|qwen2?\.?5?:?(0\.5|1\.5|1|3)b/i.test(name);
+}
+// ---------------------------------------------------------------------------
+// Named profiles
+// ---------------------------------------------------------------------------
+const SMALL_HINT = 'If native tool calls fail, emit ONE JSON object: {"tool":"<name>","arguments":{...}} and nothing else.';
+/** Weak/small models (≤3B): tight prompt, lower step budget, JSON tool hint. */
+const SMALL = {
+    id: 'small',
+    verbosity: 'terse',
+    toolFormat: 'json',
+    toolFormatHint: SMALL_HINT,
+    stepCap: 12,
+    temperature: 0.2,
+    promptCharCap: 1200,
+    contextTokens: 4096,
+};
+/** Capable general chat models (≈4–14B). Balanced defaults. */
+const GENERAL = {
+    id: 'general',
+    verbosity: 'standard',
+    toolFormat: 'native',
+    stepCap: 20,
+    temperature: 0.3,
+    promptCharCap: 2200,
+    contextTokens: 8192,
+};
+/** Coder-specialized or large (≥20B) models: full discipline, more headroom. */
+const CODER = {
+    id: 'coder',
+    verbosity: 'rich',
+    toolFormat: 'native',
+    stepCap: 28,
+    temperature: 0.2,
+    promptCharCap: 3200,
+    contextTokens: 16384,
+};
+/** Fallback when the model name is unknown — behaves like GENERAL. */
+export const DEFAULT_PROFILE = { ...GENERAL, id: 'default' };
+// ---------------------------------------------------------------------------
+// Resolution
+// ---------------------------------------------------------------------------
+/**
+ * Resolve a ModelProfile from a model name. Deterministic, no I/O.
+ *
+ * Bands (most specific first):
+ *  - small  : labeled ≤3B or an explicitly-small family.
+ *  - coder  : coder-specialized name, OR a large (≥20B) capable model.
+ *  - general: any other recognizable chat model.
+ *  - default: unknown / undefined name.
+ *
+ * `overrides` allows per-id shallow tuning (e.g. cfg.models.profiles['coder']).
+ */
+export function resolveModelProfile(modelName, overrides) {
+    if (!modelName || modelName.trim().length === 0) {
+        return applyOverride(DEFAULT_PROFILE, overrides);
+    }
+    const name = modelName.trim();
+    const size = sizeOf(name);
+    const coder = isCoderModel(name);
+    let base;
+    if (isSmallName(name) || (size > 0 && size <= 3)) {
+        base = SMALL;
+    }
+    else if (coder || (size >= 20 && size < 900)) {
+        base = CODER;
+    }
+    else {
+        // Recognizable mid-size chat model, or unknown size that still names a
+        // known family → general. Truly unknown strings fall here too (safe).
+        base = GENERAL;
+    }
+    return applyOverride(base, overrides);
+}
+function applyOverride(profile, overrides) {
+    const o = overrides?.[profile.id];
+    return o ? { ...profile, ...o } : { ...profile };
+}
+// ---------------------------------------------------------------------------
+// Feature gate
+// ---------------------------------------------------------------------------
+/**
+ * Whether the M41 adaptive-prompt suite is active. Default OFF.
+ *
+ * Precedence: ASHLR_ADAPTIVE_PROMPTS env (1/true | 0/false) overrides config;
+ * otherwise cfg.models.adaptivePrompts === true. Absent everywhere → false, so
+ * the harness keeps its legacy prompts/step-cap untouched.
+ */
+export function adaptivePromptsEnabled(cfg) {
+    const env = process.env.ASHLR_ADAPTIVE_PROMPTS?.trim().toLowerCase();
+    if (env === '1' || env === 'true' || env === 'on')
+        return true;
+    if (env === '0' || env === 'false' || env === 'off')
+        return false;
+    return cfg?.models?.adaptivePrompts === true;
+}
+//# sourceMappingURL=model-profile.js.map

package/dist/core/run/model-profile.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-profile.js","sourceRoot":"","sources":["../../../src/core/run/model-profile.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAkCH,8EAA8E;AAC9E,uEAAuE;AACvE,8EAA8E;AAE9E,mFAAmF;AACnF,SAAS,MAAM,CAAC,IAAY;IAC1B,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,wBAAwB,CAAC,CAAC;IAC/C,IAAI,CAAC;QAAE,OAAO,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC/B,IAAI,2BAA2B,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,CAAC,CAAC,CAAC,yBAAyB;IAC/E,OAAO,CAAC,CAAC,CAAC,CAAC,eAAe;AAC5B,CAAC;AAED,+EAA+E;AAC/E,SAAS,YAAY,CAAC,IAAY;IAChC,OAAO,kFAAkF,CAAC,IAAI,CAC5F,IAAI,CACL,CAAC;AACJ,CAAC;AAED,mEAAmE;AACnE,SAAS,WAAW,CAAC,IAAY;IAC/B,OAAO,8EAA8E,CAAC,IAAI,CACxF,IAAI,CACL,CAAC;AACJ,CAAC;AAED,8EAA8E;AAC9E,iBAAiB;AACjB,8EAA8E;AAE9E,MAAM,UAAU,GACd,wGAAwG,CAAC;AAE3G,gFAAgF;AAChF,MAAM,KAAK,GAAiB;IAC1B,EAAE,EAAE,OAAO;IACX,SAAS,EAAE,OAAO;IAClB,UAAU,EAAE,MAAM;IAClB,cAAc,EAAE,UAAU;IAC1B,OAAO,EAAE,EAAE;IACX,WAAW,EAAE,GAAG;IAChB,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,IAAI;CACpB,CAAC;AAEF,+DAA+D;AAC/D,MAAM,OAAO,GAAiB;IAC5B,EAAE,EAAE,SAAS;IACb,SAAS,EAAE,UAAU;IACrB,UAAU,EAAE,QAAQ;IACpB,OAAO,EAAE,EAAE;IACX,WAAW,EAAE,GAAG;IAChB,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,IAAI;CACpB,CAAC;AAEF,gFAAgF;AAChF,MAAM,KAAK,GAAiB;IAC1B,EAAE,EAAE,OAAO;IACX,SAAS,EAAE,MAAM;IACjB,UAAU,EAAE,QAAQ;IACpB,OAAO,EAAE,EAAE;IACX,WAAW,EAAE,GAAG;IAChB,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,KAAK;CACrB,CAAC;AAEF,sEAAsE;AACtE,MAAM,CAAC,MAAM,eAAe,GAAiB,EAAE,GAAG,OAAO,EAAE,EAAE,EAAE,SAAS,EAAE,CAAC;AAE3E,8EAA8E;AAC9E,aAAa;AACb,8EAA8E;AAE9E;;;;;;;;;;GAUG;AACH,MAAM,UAAU,mBAAmB,CACjC,SAAkB,EAClB,SAAiD;IAEjD,IAAI,CAAC,SAAS,IAAI,SAAS,CAAC,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAChD,OAAO,aAAa,CAAC,eAAe,EAAE,SAAS,CAAC,CAAC;IACnD,CAAC;IACD,MAAM,IAAI,GAAG,SAAS,CAAC,IAAI,EAAE,CAAC;IAC9B,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC;IAC1B,MAAM,KAAK,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC;IAEjC,IAAI,IAAkB,CAAC;IACvB,IAAI,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,IAAI,IAAI,IAAI,CAAC,CAAC,EAAE,CAAC;QACjD,IAAI,GAAG,KAAK,CAAC;IACf,CAAC;SAAM,IAAI,KAAK,IAAI,CAAC,IAAI,IAAI,EAAE,IAAI,IAAI,GAAG,GAAG,CAAC,EAAE,CAAC;QAC/C,IAAI,GAAG,KAAK,CAAC;IACf,CAAC;SAAM,CAAC;QACN,uEAAuE;QACvE,sEAAsE;QACtE,IAAI,GAAG,OAAO,CAAC;IACjB,CAAC;IACD,OAAO,aAAa,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;AACxC,CAAC;AAED,SAAS,aAAa,CACpB,OAAqB,EACrB,SAAiD;IAEjD,MAAM,CAAC,GAAG,SAAS,EAAE,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC;IAClC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,OAAO,EAAE,CAAC;AACnD,CAAC;AAED,8EAA8E;AAC9E,eAAe;AACf,8EAA8E;AAE9E;;;;;;GAMG;AACH,MAAM,UAAU,sBAAsB,CAAC,GAAiB;IACtD,MAAM,GAAG,GAAG,OAAO,CAAC,GAAG,CAAC,sBAAsB,EAAE,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACrE,IAAI,GAAG,KAAK,GAAG,IAAI,GAAG,KAAK,MAAM,IAAI,GAAG,KAAK,IAAI;QAAE,OAAO,IAAI,CAAC;IAC/D,IAAI,GAAG,KAAK,GAAG,IAAI,GAAG,KAAK,OAAO,IAAI,GAAG,KAAK,KAAK;QAAE,OAAO,KAAK,CAAC;IAClE,OAAO,GAAG,EAAE,MAAM,EAAE,eAAe,KAAK,IAAI,CAAC;AAC/C,CAAC"}

package/dist/core/run/orchestrator.d.ts

@@ -87,7 +87,7 @@ export declare function saveRun(s: RunState): void;
 export declare function planGoal(goal: string, client: ProviderClient, onUsage?: (usage: {
     tokensIn: number;
     tokensOut: number;
-}) => void, memoryContext?: string): Promise<RunTask[]>;
+}) => void, memoryContext?: string, adaptive?: boolean): Promise<RunTask[]>;
 /**
  * Top-level orchestrator driver. Builds/loads RunState, plans (unless resuming),
  * executes the DAG with parallelism up to opts.parallel, enforces HARD budget,