@shrkcrft/ai 0.1.0-alpha.13 → 0.1.0-alpha.15

package/dist/ai-request.d.ts

@@ -23,6 +23,14 @@ export interface IAiRequest {
      * progress without the synchronous wait.
      */
     onTokenStream?: (chunk: string) => void;
+    /**
+     * Per-call wall-clock timeout in milliseconds. When set and exceeded, the
+     * provider aborts the in-flight request and returns an `AppError` with code
+     * `TIMEOUT`. Bounds slow local models so a single call can never hang the
+     * command. Takes precedence over the provider's `config.timeoutMs`; when
+     * neither is set, no timeout is applied.
+     */
+    timeoutMs?: number;
 }
 export interface IAiResponse {
     content: string;

package/dist/ai-request.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ai-request.d.ts","sourceRoot":"","sources":["../src/ai-request.ts"],"names":[],"mappings":"AAAA,oBAAY,aAAa;IACvB,MAAM,WAAW;IACjB,IAAI,SAAS;IACb,SAAS,cAAc;CACxB;AAED,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,aAAa,CAAC;IACpB,OAAO,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,UAAU;IACzB,QAAQ,EAAE,SAAS,UAAU,EAAE,CAAC;IAChC,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,cAAc,CAAC,EAAE,iBAAiB,CAAC;IACnC;;;;;;;OAOG;IACH,aAAa,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,CAAC;CACzC;AAED,MAAM,WAAW,WAAW;IAC1B,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IACxD,GAAG,CAAC,EAAE,OAAO,CAAC;CACf;AAED,MAAM,WAAW,iBAAiB;IAChC,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,aAAa,GAAG,aAAa,CAAC;IACpC,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACjC,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB"}
+{"version":3,"file":"ai-request.d.ts","sourceRoot":"","sources":["../src/ai-request.ts"],"names":[],"mappings":"AAAA,oBAAY,aAAa;IACvB,MAAM,WAAW;IACjB,IAAI,SAAS;IACb,SAAS,cAAc;CACxB;AAED,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,aAAa,CAAC;IACpB,OAAO,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,UAAU;IACzB,QAAQ,EAAE,SAAS,UAAU,EAAE,CAAC;IAChC,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,cAAc,CAAC,EAAE,iBAAiB,CAAC;IACnC;;;;;;;OAOG;IACH,aAAa,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,CAAC;IACxC;;;;;;OAMG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,WAAW;IAC1B,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IACxD,GAAG,CAAC,EAAE,OAAO,CAAC;CACf;AAED,MAAM,WAAW,iBAAiB;IAChC,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,aAAa,GAAG,aAAa,CAAC;IACpC,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACjC,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB"}

package/dist/llamacpp/llama-cpp-provider.d.ts

@@ -52,5 +52,5 @@ export declare class LlamaCppProvider extends AbstractAiProvider {
  * loaded. Errors during dispose are swallowed (the alternative is
  * the abort we're trying to prevent).
  */
-export declare function disposeLlamaCppRuntime(): Promise<void>;
+export declare function disposeLlamaCppRuntime(): Promise<boolean>;
 //# sourceMappingURL=llama-cpp-provider.d.ts.map

package/dist/llamacpp/llama-cpp-provider.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llama-cpp-provider.d.ts","sourceRoot":"","sources":["../../src/llamacpp/llama-cpp-provider.ts"],"names":[],"mappings":"AAEA,OAAO,EAAsC,KAAK,QAAQ,EAAE,KAAK,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAChG,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AACvD,OAAO,EAAkC,KAAK,UAAU,EAAE,KAAK,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAKrG;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,qBAAa,gBAAiB,SAAQ,kBAAkB;IACtD,QAAQ,CAAC,EAAE,cAAc;IACzB,QAAQ,CAAC,IAAI,4BAA4B;IAEzC,wDAAwD;IACxD,MAAM,CAAC,iBAAiB,EACpB,CAAC,CAAC,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,MAAM,KAAK,OAAO,CAAC,WAAW,CAAC,CAAC,GAClE,IAAI,CAAQ;IAEhB;;;;OAIG;IACH,MAAM,CAAC,eAAe,IAAI,MAAM,GAAG,IAAI;IAIvC,OAAO,IAAI,OAAO;IAIZ,IAAI,CAAC,OAAO,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,WAAW,EAAE,QAAQ,CAAC,CAAC;YA4IzD,YAAY;CAwB3B;AAWD;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,sBAAsB,IAAI,OAAO,CAAC,IAAI,CAAC,CAU5D"}
+{"version":3,"file":"llama-cpp-provider.d.ts","sourceRoot":"","sources":["../../src/llamacpp/llama-cpp-provider.ts"],"names":[],"mappings":"AAEA,OAAO,EAAsC,KAAK,QAAQ,EAAE,KAAK,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAChG,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AACvD,OAAO,EAAkC,KAAK,UAAU,EAAE,KAAK,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAKrG;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,qBAAa,gBAAiB,SAAQ,kBAAkB;IACtD,QAAQ,CAAC,EAAE,cAAc;IACzB,QAAQ,CAAC,IAAI,4BAA4B;IAEzC,wDAAwD;IACxD,MAAM,CAAC,iBAAiB,EACpB,CAAC,CAAC,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,MAAM,KAAK,OAAO,CAAC,WAAW,CAAC,CAAC,GAClE,IAAI,CAAQ;IAEhB;;;;OAIG;IACH,MAAM,CAAC,eAAe,IAAI,MAAM,GAAG,IAAI;IAIvC,OAAO,IAAI,OAAO;IAIZ,IAAI,CAAC,OAAO,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,WAAW,EAAE,QAAQ,CAAC,CAAC;YAwKzD,YAAY;CAwB3B;AAWD;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,sBAAsB,IAAI,OAAO,CAAC,OAAO,CAAC,CAc/D"}

package/dist/llamacpp/llama-cpp-provider.js

@@ -60,6 +60,9 @@ export class LlamaCppProvider extends AbstractAiProvider {
                 }));
             }
         }
+        let promptAbort;
+        let promptTimer;
+        let promptTimedOut = false;
         try {
             const tf = (await import('node-llama-cpp'));
             const { LlamaChatSession } = tf;
@@ -122,11 +125,23 @@ export class LlamaCppProvider extends AbstractAiProvider {
             }
             const start = Date.now();
             const onChunk = request.onTokenStream;
+            // Per-call wall-clock timeout: abort the decode if it overruns so a
+            // slow model can't hang the command. node-llama-cpp honours an
+            // AbortSignal when `stopOnAbortSignal` is set.
+            const timeoutMs = request.timeoutMs ?? this.config.timeoutMs;
+            if (timeoutMs && timeoutMs > 0) {
+                promptAbort = new AbortController();
+                promptTimer = setTimeout(() => {
+                    promptTimedOut = true;
+                    promptAbort?.abort();
+                }, timeoutMs);
+            }
             const text = await session.prompt(userPrompt, {
                 maxTokens,
                 ...(request.temperature !== undefined ? { temperature: request.temperature } : {}),
                 ...(wantsJson ? { trimWhitespaceSuffix: true } : {}),
                 ...(grammar ? { grammar: grammar } : {}),
+                ...(promptAbort ? { signal: promptAbort.signal, stopOnAbortSignal: true } : {}),
                 ...(onChunk
                     ? {
                         onTextChunk: (chunk) => {
@@ -163,11 +178,20 @@ export class LlamaCppProvider extends AbstractAiProvider {
             });
         }
         catch (e) {
+            if (promptTimedOut) {
+                return err(new AppErrorImpl(ERROR_CODES.TIMEOUT, `node-llama-cpp decode exceeded the per-call timeout and was aborted.`, {
+                    suggestion: 'The model is too slow for the budget. Try a smaller model, fewer --enhance-passes, or raise the budget.',
+                }));
+            }
             return err(new AppErrorImpl(ERROR_CODES.IO_ERROR, `node-llama-cpp call failed: ${e.message}`, {
                 cause: e,
                 suggestion: 'Verify LLAMACPP_MODEL_PATH points to a valid .gguf file readable by llama.cpp.',
             }));
         }
+        finally {
+            if (promptTimer)
+                clearTimeout(promptTimer);
+        }
     }
     async ensureLoaded(modelPath) {
         // Cached at MODULE scope so the disposer can find it on process
@@ -214,13 +238,17 @@ export async function disposeLlamaCppRuntime() {
     const state = sharedLlamaState;
     sharedLlamaState = null;
     if (!state)
-        return;
+        return false;
     // Context first — it holds the sequence pool that depends on the model.
     await callMaybeDispose(state.context);
     // Then the model, which depends on the llama runtime.
     await callMaybeDispose(state.model);
     // Finally the Llama instance itself (releases the Metal device).
     await callMaybeDispose(state.llama);
+    // libggml/Metal was loaded — even after disposing, this Node version still
+    // runs the native static destructor during `exit()` and it can abort with a
+    // GGML backtrace. The caller redirects fd 2 to a log file to contain it.
+    return true;
 }
 async function callMaybeDispose(target) {
     if (!target || typeof target !== 'object')

package/dist/ollama/ollama-provider.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ollama-provider.d.ts","sourceRoot":"","sources":["../../src/ollama/ollama-provider.ts"],"names":[],"mappings":"AAAA,OAAO,EAAsC,KAAK,QAAQ,EAAE,KAAK,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAChG,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AACvD,OAAO,EAAiB,KAAK,UAAU,EAAE,KAAK,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAMpF;;;;;;;;;;;;;;;;;;;GAmBG;AACH,qBAAa,cAAe,SAAQ,kBAAkB;IACpD,QAAQ,CAAC,EAAE,YAAY;IACvB,QAAQ,CAAC,IAAI,yBAAyB;IAEtC,OAAO,IAAI,OAAO;IAIlB;;;;;;;;;;;OAWG;IACG,WAAW,CACf,YAAY,CAAC,EAAE,MAAM,GACpB,OAAO,CAAC,MAAM,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,EAAE,CAAC;QAAC,YAAY,EAAE,OAAO,GAAG,IAAI,CAAA;KAAE,EAAE,QAAQ,CAAC,CAAC;IA+BxF,IAAI,CAAC,OAAO,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,WAAW,EAAE,QAAQ,CAAC,CAAC;CAkExE"}
+{"version":3,"file":"ollama-provider.d.ts","sourceRoot":"","sources":["../../src/ollama/ollama-provider.ts"],"names":[],"mappings":"AAAA,OAAO,EAAsC,KAAK,QAAQ,EAAE,KAAK,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAChG,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AACvD,OAAO,EAAiB,KAAK,UAAU,EAAE,KAAK,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAMpF;;;;;;;;;;;;;;;;;;;GAmBG;AACH,qBAAa,cAAe,SAAQ,kBAAkB;IACpD,QAAQ,CAAC,EAAE,YAAY;IACvB,QAAQ,CAAC,IAAI,yBAAyB;IAEtC,OAAO,IAAI,OAAO;IAIlB;;;;;;;;;;;OAWG;IACG,WAAW,CACf,YAAY,CAAC,EAAE,MAAM,GACpB,OAAO,CAAC,MAAM,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,EAAE,CAAC;QAAC,YAAY,EAAE,OAAO,GAAG,IAAI,CAAA;KAAE,EAAE,QAAQ,CAAC,CAAC;IA+BxF,IAAI,CAAC,OAAO,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,WAAW,EAAE,QAAQ,CAAC,CAAC;CAgGxE"}

package/dist/ollama/ollama-provider.js

@@ -81,11 +81,26 @@ export class OllamaProvider extends AbstractAiProvider {
         const format = formatFor(request.responseFormat);
         if (format !== undefined)
             body.format = format;
+        // Per-call wall-clock timeout. Without this a slow local model (a large
+        // 20B+ model, or one still loading) hangs the request indefinitely — the
+        // root cause of `smart-context` "running too long". Manual controller +
+        // timer (rather than AbortSignal.timeout) so the catch can distinguish a
+        // timeout from an unrelated network error.
+        const timeoutMs = request.timeoutMs ?? this.config.timeoutMs;
+        const controller = timeoutMs && timeoutMs > 0 ? new AbortController() : undefined;
+        let timedOut = false;
+        const timer = controller && timeoutMs
+            ? setTimeout(() => {
+                timedOut = true;
+                controller.abort();
+            }, timeoutMs)
+            : undefined;
         try {
             const res = await fetch(`${baseUrl}/api/chat`, {
                 method: 'POST',
                 headers: { 'content-type': 'application/json' },
                 body: JSON.stringify(body),
+                ...(controller ? { signal: controller.signal } : {}),
             });
             if (!res.ok) {
                 const text = await res.text();
@@ -107,11 +122,20 @@ export class OllamaProvider extends AbstractAiProvider {
             });
         }
         catch (e) {
+            if (timedOut) {
+                return err(new AppErrorImpl(ERROR_CODES.TIMEOUT, `Ollama call exceeded ${timeoutMs}ms and was aborted (model "${model}").`, {
+                    suggestion: `The model is too slow for the budget. Try a smaller --model, fewer --enhance-passes, or raise the budget.`,
+                }));
+            }
             return err(new AppErrorImpl(ERROR_CODES.IO_ERROR, `Failed to call Ollama at ${baseUrl}: ${e.message}`, {
                 cause: e,
                 suggestion: `Is Ollama running? Try \`ollama serve\` or set OLLAMA_HOST to a reachable instance.`,
             }));
         }
+        finally {
+            if (timer)
+                clearTimeout(timer);
+        }
     }
 }
 function roleFor(role) {

package/dist/pipeline/enhancement-pipeline.d.ts

@@ -53,6 +53,19 @@ export interface IEnhancementPipelineOptions {
     temperature?: number;
     /** Override the model selection (forwarded to the provider per call). */
     model?: string;
+    /**
+     * Total wall-clock budget (ms) for the whole pipeline. Before each stage the
+     * elapsed time is checked; once the budget is spent the pipeline stops and
+     * returns the best output so far (degrading to the deterministic seed if not
+     * even the draft finished). Undefined = no budget (legacy behaviour).
+     */
+    budgetMs?: number;
+    /**
+     * Per-call timeout (ms) handed to the provider for each stage. Effective
+     * timeout is `min(perStageTimeoutMs, remaining budget)`. Bounds a single
+     * slow call so it can't blow the whole budget.
+     */
+    perStageTimeoutMs?: number;
     /** Optional progress hook — called once per stage. */
     onStage?: (event: {
         kind: EnhancementStageKind;
@@ -77,6 +90,11 @@ export interface IEnhancementPipelineRun {
      * the deterministic seed unchanged.
      */
     deterministicFallback: boolean;
+    /**
+     * True when the wall-clock `budgetMs` was reached before every planned
+     * stage ran. `finalOutput` still holds the best result produced so far.
+     */
+    budgetExhausted: boolean;
 }
 /**
  * Multi-pass refinement pipeline that turns a deterministic brief into
@@ -120,4 +138,14 @@ export declare class EnhancementPipeline {
  * grounding.
  */
 export declare function buildDefaultEnhancementStages(): IEnhancementStage[];
+/**
+ * The fast default for interactive use: `draft → polish` (2 calls). Skips the
+ * slow critique + refine round-trip (the two passes small/large local models
+ * spend the most wall-clock on) while still applying the polish pass that
+ * gives the agent file:line refs and terse imperative bullets. Materially
+ * better than a single shot, ~half the calls of the full pipeline. Callers who
+ * want maximal density opt into `buildDefaultEnhancementStages()` (the
+ * `--plus` path).
+ */
+export declare function buildFastEnhancementStages(): IEnhancementStage[];
 //# sourceMappingURL=enhancement-pipeline.d.ts.map

package/dist/pipeline/enhancement-pipeline.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"enhancement-pipeline.d.ts","sourceRoot":"","sources":["../../src/pipeline/enhancement-pipeline.ts"],"names":[],"mappings":"AAAA,OAAO,EAAsC,KAAK,QAAQ,EAAE,KAAK,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAChG,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AACrD,OAAO,EAAiB,KAAK,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAElE;;;;;;GAMG;AACH,oBAAY,oBAAoB;IAC9B,KAAK,UAAU;IACf,QAAQ,aAAa;IACrB,MAAM,WAAW;IACjB,MAAM,WAAW;CAClB;AAED,MAAM,WAAW,sBAAsB;IACrC,8DAA8D;IAC9D,eAAe,EAAE,MAAM,CAAC;IACxB,yCAAyC;IACzC,IAAI,EAAE,MAAM,CAAC;IACb,+DAA+D;IAC/D,QAAQ,EAAE,MAAM,CAAC;IACjB,iEAAiE;IACjE,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,oBAAoB,CAAC;IAC3B;;;OAGG;IACH,aAAa,CAAC,KAAK,EAAE,sBAAsB,GAAG,UAAU,EAAE,CAAC;CAC5D;AAED,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,oBAAoB,CAAC;IAC3B,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,uEAAuE;IACvE,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CACzD;AAED,MAAM,WAAW,2BAA2B;IAC1C,6EAA6E;IAC7E,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,4CAA4C;IAC5C,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,iEAAiE;IACjE,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,yEAAyE;IACzE,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,sDAAsD;IACtD,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE;QAAE,IAAI,EAAE,oBAAoB,CAAC;QAAC,EAAE,EAAE,OAAO,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,KAAK,IAAI,CAAC;CACrG;AAED,MAAM,WAAW,uBAAuB;IACtC,uGAAuG;IACvG,WAAW,EAAE,MAAM,CAAC;IACpB,mCAAmC;IACnC,MAAM,EAAE,uBAAuB,EAAE,CAAC;IAClC,4EAA4E;IAC5E,UAAU,EAAE;QAAE,WAAW,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1D;;;;OAIG;IACH,qBAAqB,EAAE,OAAO,CAAC;CAChC;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,qBAAa,mBAAmB;IAC9B,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmC;gBAE9C,MAAM,EAAE,aAAa,CAAC,iBAAiB,CAAC;IAI9C,GAAG,CACP,KAAK,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,eAAe,EAAE,MAAM,CAAA;KAAE,EAChD,QAAQ,EAAE,WAAW,GAAG,IAAI,EAC5B,OAAO,GAAE,2BAAgC,GACxC,OAAO,CAAC,MAAM,CAAC,uBAAuB,EAAE,QAAQ,CAAC,CAAC;CAmFtD;AAED;;;;;;;;GAQG;AACH,wBAAgB,6BAA6B,IAAI,iBAAiB,EAAE,CAOnE"}
+{"version":3,"file":"enhancement-pipeline.d.ts","sourceRoot":"","sources":["../../src/pipeline/enhancement-pipeline.ts"],"names":[],"mappings":"AAAA,OAAO,EAAsC,KAAK,QAAQ,EAAE,KAAK,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAChG,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AACrD,OAAO,EAAiB,KAAK,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAElE;;;;;;GAMG;AACH,oBAAY,oBAAoB;IAC9B,KAAK,UAAU;IACf,QAAQ,aAAa;IACrB,MAAM,WAAW;IACjB,MAAM,WAAW;CAClB;AAED,MAAM,WAAW,sBAAsB;IACrC,8DAA8D;IAC9D,eAAe,EAAE,MAAM,CAAC;IACxB,yCAAyC;IACzC,IAAI,EAAE,MAAM,CAAC;IACb,+DAA+D;IAC/D,QAAQ,EAAE,MAAM,CAAC;IACjB,iEAAiE;IACjE,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,oBAAoB,CAAC;IAC3B;;;OAGG;IACH,aAAa,CAAC,KAAK,EAAE,sBAAsB,GAAG,UAAU,EAAE,CAAC;CAC5D;AAED,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,oBAAoB,CAAC;IAC3B,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,uEAAuE;IACvE,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE;QAAE,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,YAAY,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;CACzD;AAED,MAAM,WAAW,2BAA2B;IAC1C,6EAA6E;IAC7E,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,4CAA4C;IAC5C,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,iEAAiE;IACjE,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,yEAAyE;IACzE,KAAK,CAAC,EAAE,MAAM,CAAC;IACf;;;;;OAKG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;OAIG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,sDAAsD;IACtD,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE;QAAE,IAAI,EAAE,oBAAoB,CAAC;QAAC,EAAE,EAAE,OAAO,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,KAAK,IAAI,CAAC;CACrG;AAED,MAAM,WAAW,uBAAuB;IACtC,uGAAuG;IACvG,WAAW,EAAE,MAAM,CAAC;IACpB,mCAAmC;IACnC,MAAM,EAAE,uBAAuB,EAAE,CAAC;IAClC,4EAA4E;IAC5E,UAAU,EAAE;QAAE,WAAW,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1D;;;;OAIG;IACH,qBAAqB,EAAE,OAAO,CAAC;IAC/B;;;OAGG;IACH,eAAe,EAAE,OAAO,CAAC;CAC1B;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,qBAAa,mBAAmB;IAC9B,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmC;gBAE9C,MAAM,EAAE,aAAa,CAAC,iBAAiB,CAAC;IAI9C,GAAG,CACP,KAAK,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,eAAe,EAAE,MAAM,CAAA;KAAE,EAChD,QAAQ,EAAE,WAAW,GAAG,IAAI,EAC5B,OAAO,GAAE,2BAAgC,GACxC,OAAO,CAAC,MAAM,CAAC,uBAAuB,EAAE,QAAQ,CAAC,CAAC;CAkGtD;AAgBD;;;;;;;;GAQG;AACH,wBAAgB,6BAA6B,IAAI,iBAAiB,EAAE,CAOnE;AAED;;;;;;;;GAQG;AACH,wBAAgB,0BAA0B,IAAI,iBAAiB,EAAE,CAEhE"}

package/dist/pipeline/enhancement-pipeline.js

@@ -50,6 +50,7 @@ export class EnhancementPipeline {
                 stages: [],
                 totalUsage: { inputTokens: 0, outputTokens: 0 },
                 deterministicFallback: true,
+                budgetExhausted: false,
             });
         }
         const cap = options.maxPasses ?? this.stages.length;
@@ -59,7 +60,16 @@ export class EnhancementPipeline {
         let previous = '';
         let lastCritique;
         let lastGood = input.originalContext;
+        const startedAt = Date.now();
+        let budgetExhausted = false;
         for (let i = 0; i < plan.length; i += 1) {
+            // Wall-clock budget guard: stop before starting a stage we have no time
+            // for, and keep the best output produced so far.
+            const remaining = options.budgetMs !== undefined ? options.budgetMs - (Date.now() - startedAt) : undefined;
+            if (remaining !== undefined && remaining <= MIN_STAGE_BUDGET_MS) {
+                budgetExhausted = true;
+                break;
+            }
             const stage = plan[i];
             const messages = stage.buildMessages({
                 originalContext: input.originalContext,
@@ -67,11 +77,14 @@ export class EnhancementPipeline {
                 previous,
                 lastCritique,
             });
+            // Effective per-call timeout = min(configured per-stage, remaining budget).
+            const perStageTimeout = effectiveTimeout(options.perStageTimeoutMs, remaining);
             const stageResult = await callOnceWithRetry(provider, {
                 messages,
                 maxTokens: options.maxTokensPerStage ?? 4096,
                 temperature: options.temperature ?? 0.2,
                 ...(options.model ? { model: options.model } : {}),
+                ...(perStageTimeout !== undefined ? { timeoutMs: perStageTimeout } : {}),
             });
             const onStage = options.onStage;
             if (!stageResult.ok) {
@@ -120,9 +133,23 @@ export class EnhancementPipeline {
             stages: stagesOut,
             totalUsage,
             deterministicFallback: false,
+            budgetExhausted,
         });
     }
 }
+/** Don't start a stage with less than this much budget left (a call needs at
+ * least this long to have any chance of returning). */
+const MIN_STAGE_BUDGET_MS = 250;
+/**
+ * Effective per-call timeout: the tighter of an explicit per-stage cap and the
+ * remaining wall-clock budget. Returns undefined when neither is set.
+ */
+function effectiveTimeout(perStage, remaining) {
+    const candidates = [perStage, remaining].filter((n) => typeof n === 'number' && n > 0);
+    if (candidates.length === 0)
+        return undefined;
+    return Math.min(...candidates);
+}
 /**
  * The default stage set for "make this brief more useful to the Claude
  * agent". Tuned for small local models (Qwen2.5-Coder-3B, Llama-3.1-8B).
@@ -140,6 +167,18 @@ export function buildDefaultEnhancementStages() {
         new PolishStage(),
     ];
 }
+/**
+ * The fast default for interactive use: `draft → polish` (2 calls). Skips the
+ * slow critique + refine round-trip (the two passes small/large local models
+ * spend the most wall-clock on) while still applying the polish pass that
+ * gives the agent file:line refs and terse imperative bullets. Materially
+ * better than a single shot, ~half the calls of the full pipeline. Callers who
+ * want maximal density opt into `buildDefaultEnhancementStages()` (the
+ * `--plus` path).
+ */
+export function buildFastEnhancementStages() {
+    return [new DraftStage(), new PolishStage()];
+}
 class DraftStage {
     kind = EnhancementStageKind.Draft;
     buildMessages(input) {
@@ -285,6 +324,11 @@ async function callOnceWithRetry(provider, request) {
     if (first.ok) {
         return ok({ content: first.value.content, model: first.value.model, usage: first.value.usage });
     }
+    // Don't retry a timeout — the model is too slow for the budget, so a second
+    // attempt just burns another timeout period. Surface the timeout immediately.
+    if (first.error.code === ERROR_CODES.TIMEOUT) {
+        return first;
+    }
     // One retry — small local models routinely 500 on the first request
     // after a daemon start. Idempotent reissue is safe.
     const second = await provider.send(request);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@shrkcrft/ai",
-  "version": "0.1.0-alpha.13",
+  "version": "0.1.0-alpha.15",
   "description": "SharkCraft local LLM provider abstraction: Ollama (HTTP) + llama.cpp (in-process) + multi-pass enhancement pipeline.",
   "license": "MIT",
   "author": "SharkCraft contributors",
@@ -43,8 +43,8 @@
     "typecheck": "tsc --noEmit -p tsconfig.json"
   },
   "dependencies": {
-    "@shrkcrft/core": "^0.1.0-alpha.13",
-    "@shrkcrft/context": "^0.1.0-alpha.13",
+    "@shrkcrft/core": "^0.1.0-alpha.15",
+    "@shrkcrft/context": "^0.1.0-alpha.15",
     "node-llama-cpp": "^3.16.0"
   },
   "publishConfig": {