@womp/spark-agent 5.8.0 → 5.9.0

package/.env.example

@@ -0,0 +1,6 @@
+OPENAI_API_KEY=
+ANTHROPIC_API_KEY=
+GEMINI_API_KEY=
+KAKAPO_URL=ws://127.0.0.1:5502
+PORT=8787
+SPARK_STORAGE_PATH=.spark/spark.sqlite

package/dist/context-budget.d.ts

@@ -12,6 +12,15 @@ export declare const TRIM_TRIGGER_RATIO = 0.9;
 export declare const TRIM_TARGET_RATIO = 0.75;
 /** A tool result at or below this is left whole: not worth the detail it loses. */
 export declare const SMALL_TOOL_OUTPUT_TOKENS = 500;
+/**
+ * Steps of the CURRENT turn kept whole, however long that turn runs.
+ *
+ * The model is mid-task and reasons from what it just did, so the recent steps
+ * have to survive; the ones before them are history like any other and are worth
+ * more as window than as detail. Five is small enough that a long run always
+ * recovers room and large enough to hold a read-edit-verify cycle.
+ */
+export declare const PROTECTED_CURRENT_STEPS = 5;
 export declare const OMITTED_HISTORY_TEXT = "[Earlier messages in this conversation were left out of this request to fit the model's context window.]";
 export declare const OMITTED_HISTORY_IMAGE_TEXT = "[An earlier image was left out of this request to fit the model's context window.]";
 export declare const OMITTED_TOOL_OUTPUT_TEXT = "[This earlier tool result was left out of this request to fit the model's context window.]";

package/dist/context-budget.js

@@ -22,9 +22,22 @@
 //     still paired with its result. A note in their place tells the model that
 //     earlier messages were left out.
 //
-// The current turn -- the latest user message and everything after it -- is never
-// changed. When it alone is over the window the request goes out as it is and the
-// provider's own error says so; nothing here can make that turn smaller.
+// THE CURRENT TURN is the latest user message and everything after it. The
+// message itself -- the question and any image the user just attached -- is never
+// changed, and neither are the last PROTECTED_CURRENT_STEPS steps after it: the
+// model has to be able to see what it just did.
+//
+// Its EARLIER steps are trimmed like any other history, by stages 1 and 2 above.
+// They were exempt until 2026-10-01, on the assumption that a turn is small. An
+// agentic run's turn is not: every step it takes lands after that user message,
+// and `maxSteps` is 100. dev0 thread 142417 reached the window that way -- ten
+// minutes of work, earlier history already trimmed to nothing, and what was left
+// over the limit was the run's own tool results, which nothing was allowed to
+// touch. Stage 3 still stops at the current turn, because dropping whole messages
+// inside it cannot keep calls paired with their results.
+//
+// When the protected part alone is over the window the request still goes out as
+// it is, and the provider's own error says so.
 //
 // ESTIMATES, NOT COUNTS. Counting needs the provider's tokenizer or a request.
 // The estimate is deliberately high (CHARS_PER_TOKEN is under the ~3.5-4 of prose)
@@ -50,6 +63,15 @@ export const TRIM_TRIGGER_RATIO = 0.9;
 export const TRIM_TARGET_RATIO = 0.75;
 /** A tool result at or below this is left whole: not worth the detail it loses. */
 export const SMALL_TOOL_OUTPUT_TOKENS = 500;
+/**
+ * Steps of the CURRENT turn kept whole, however long that turn runs.
+ *
+ * The model is mid-task and reasons from what it just did, so the recent steps
+ * have to survive; the ones before them are history like any other and are worth
+ * more as window than as detail. Five is small enough that a long run always
+ * recovers room and large enough to hold a read-edit-verify cycle.
+ */
+export const PROTECTED_CURRENT_STEPS = 5;
 export const OMITTED_HISTORY_TEXT = "[Earlier messages in this conversation were left out of this request to fit the model's context window.]";
 export const OMITTED_HISTORY_IMAGE_TEXT = "[An earlier image was left out of this request to fit the model's context window.]";
 export const OMITTED_TOOL_OUTPUT_TEXT = "[This earlier tool result was left out of this request to fit the model's context window.]";
@@ -76,6 +98,25 @@ export function contextWindowFor(modelId) {
         return 1_048_576;
     return undefined;
 }
+/**
+ * Where the current turn's protected tail begins: the Nth-from-last step.
+ *
+ * Counted by ASSISTANT messages, because that is what starts a step -- the tool
+ * messages that answer it follow and are protected with it, so no call is ever
+ * separated from its result. A turn with fewer steps than that protects all of
+ * them, which is what this did for every turn before it could trim one.
+ */
+function protectedTailStart(messages, current) {
+    let steps = 0;
+    for (let i = messages.length - 1; i > current; i -= 1) {
+        if (messages[i]?.role !== "assistant")
+            continue;
+        steps += 1;
+        if (steps >= PROTECTED_CURRENT_STEPS)
+            return i;
+    }
+    return current + 1;
+}
 export function fitMessagesToContext(messages, options) {
     if (options.contextWindow === undefined)
         return { messages };
@@ -96,10 +137,15 @@ export function fitMessagesToContext(messages, options) {
     };
     if (originalTokens <= budgetTokens * TRIM_TRIGGER_RATIO)
         return { messages, report };
-    // The current turn starts at the latest user message and is never changed.
+    // The current turn starts at the latest user message. With no user message at
+    // all there is no turn to reason about and nothing is touched.
     const current = lastIndex(messages, ({ role }) => role === "user");
-    if (current <= 0)
+    if (current < 0)
         return { messages, report };
+    // Stages 1 and 2 rewrite everything before this: all earlier turns, and the
+    // current turn's own earlier steps. `current` itself is skipped -- the user's
+    // question and the image they just attached are not history.
+    const protectedFrom = protectedTailStart(messages, current);
     const target = budgetTokens * TRIM_TARGET_RATIO;
     const out = messages.slice();
     let total = originalTokens;
@@ -110,7 +156,9 @@ export function fitMessagesToContext(messages, options) {
      * message, which is then measured again.
      */
     const rewriteEarlier = (rewrite) => {
-        for (let i = 0; i < current && total > target; i += 1) {
+        for (let i = 0; i < protectedFrom && total > target; i += 1) {
+            if (i === current)
+                continue;
             const message = out[i];
             if (message === undefined || !Array.isArray(message.content))
                 continue;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@womp/spark-agent",
-  "version": "5.8.0",
+  "version": "5.9.0",
   "description": "Portable skills-based AI agent with local and durable chat transports for Womp",
   "type": "module",
   "main": "./dist/index.js",