@poncho-ai/harness 0.52.2 → 0.55.0

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.52.2 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.55.0 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
 > node scripts/embed-docs.js && tsup src/index.ts --format esm --dts
 
 [embed-docs] Generated poncho-docs.ts with 4 topics
@@ -8,9 +8,9 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js            540.60 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 204ms
+ESM dist/index.js            545.48 KB
+ESM ⚡️ Build success in 258ms
 DTS Build start
-DTS ⚡️ Build success in 7242ms
-DTS dist/index.d.ts 93.60 KB
+DTS ⚡️ Build success in 7628ms
+DTS dist/index.d.ts 94.38 KB

package/CHANGELOG.md

@@ -1,5 +1,78 @@
 # @poncho-ai/harness
 
+## 0.55.0
+
+### Minor Changes
+
+- [#149](https://github.com/cesr/poncho-ai/pull/149) [`f5a8260`](https://github.com/cesr/poncho-ai/commit/f5a8260d0515038afc1797d00507908c334115ff) Thanks [@cesr](https://github.com/cesr)! - compaction: preserve subagent context and prior summaries, harden the split
+
+  Three improvements to context compaction (fires at ~75% context):
+  - **Split safety**: `findSafeSplitPoint` now refuses a split whose compacted
+    side would end on an assistant message with unanswered `tool_calls` (its
+    answering `role:"tool"` result having moved to the preserved side), walking
+    earlier to the next clean `user` boundary. Prevents orphaning a tool-call
+    relationship inside the summary boundary. Still returns `-1` when no safe
+    point exists.
+  - **Subagent ledger**: while compacting, scans for subagent-callback records
+    (metadata `_subagentCallback`/`subagentCallback`, or text starting with
+    `[Subagent Result]`) and any `## Subagents` block embedded in a prior
+    compaction summary, then renders a combined, deduped (by `subagentId`)
+    ledger that is appended VERBATIM after the LLM summary text — so the model
+    can never paraphrase or truncate subagent results away. Cumulative across
+    successive compactions.
+  - **Cumulative summary**: when the first compacted message is itself a prior
+    compaction summary, it is passed to the summarizer in full (not truncated
+    to 1200 chars) and the prompt instructs the model to merge-and-update the
+    prior working state rather than re-summarize it from scratch. All other
+    messages keep the 1200-char truncation.
+
+## 0.54.0
+
+### Minor Changes
+
+- [#147](https://github.com/cesr/poncho-ai/pull/147) [`a3eed14`](https://github.com/cesr/poncho-ai/commit/a3eed142832318b6397cd73819d3296c79d6eff0) Thanks [@cesr](https://github.com/cesr)! - storage: add append-only conversation-entry substrate (unused groundwork)
+
+  Pure entry types + rebuild functions (`buildLlmContext`,
+  `buildDisplaySnapshot`, `getPendingSubagentResults`) for the eventual
+  append-only conversation model that removes the mutable-blob clobber race
+  (the root cause behind lost subagent results). No storage-engine wiring
+  and no live callers yet — additive, deploys nothing behavioral. The
+  rebuild logic (compaction overlay, amendment folding, callback-consumption)
+  is covered by unit tests so the design is proven before the bigger
+  dual-write / migration / cutover PRs.
+
+## 0.53.0
+
+### Minor Changes
+
+- [#145](https://github.com/cesr/poncho-ai/pull/145) [`bfa4976`](https://github.com/cesr/poncho-ai/commit/bfa4976ac8b05a300e22271e23c3bae4aadae2a8) Thanks [@cesr](https://github.com/cesr)! - events: add stable identity so streaming clients match instead of guess
+
+  Additive fields that let a streaming client reconstruct view-state by
+  identity rather than inferring structure from event order (the source of a
+  class of reconnect/subagent rendering bugs):
+  - `tool:started` / `tool:completed` / `tool:error` now carry `toolCallId`
+    (already in scope as `call.id` / `result.callId`). Clients match tool
+    pills by id instead of by tool name.
+  - `subagent:spawned|completed|error|stopped` now carry `parentToolCallId`
+    (the `spawn_subagent` tool call's id) and `task`; `completed`/`error`
+    also carry `resultText`. Clients attach subagent state to the spawning
+    tool's pill and render the result inline — no header-regex or
+    sequential-cursor pairing needed.
+  - `ToolContext` gains `toolCallId` so the `spawn_subagent` handler can
+    record which call produced the subagent (plumbed: tool-dispatcher →
+    spawn handler → `SubagentSpawnOptions.parentToolCallId` →
+    `subagentMeta.parentToolCallId` → the events above).
+  - `run:started` gains an optional `cause` field in the type
+    (`user|continuation|subagent_callback|approval_resume`); emission is
+    deferred to a later pass.
+
+  All fields are additive; older clients ignore them.
+
+### Patch Changes
+
+- Updated dependencies [[`bfa4976`](https://github.com/cesr/poncho-ai/commit/bfa4976ac8b05a300e22271e23c3bae4aadae2a8)]:
+  - @poncho-ai/sdk@1.15.0
+
 ## 0.52.2
 
 ### Patch Changes

package/dist/index.d.ts

@@ -96,6 +96,12 @@ declare const estimateTotalTokens: (systemPrompt: string, messages: Message[], t
  * and everything from it onward is preserved. The split always lands just
  * before a `user` message to avoid breaking assistant+tool pairs.
  *
+ * Defensive guard: even at a `user` boundary, refuse a split whose compacted
+ * side would END on an assistant message with unanswered tool_calls (its
+ * `tool` result having moved to the preserved side). Such a split would
+ * orphan the tool_calls inside the summary boundary. When that happens we
+ * walk earlier to the next safe `user` boundary.
+ *
  * Returns -1 if no valid split point is found.
  */
 declare const findSafeSplitPoint: (messages: Message[], keepRecentMessages: number) => number;
@@ -199,6 +205,9 @@ interface Conversation {
          *  subagent's runs emit no telemetry (e.g. spawned from an incognito
          *  turn). Read by the orchestrator's runSubagent / continuation. */
         suppressTelemetry?: boolean;
+        /** The parent's `spawn_subagent` tool call id — echoed onto subagent:*
+         *  events so a client can attach subagent state to that tool's pill. */
+        parentToolCallId?: string;
     };
     channelMeta?: {
         platform: string;
@@ -1190,6 +1199,10 @@ interface SubagentManager {
         /** Inherit the parent run's telemetry choice — when true, the subagent
          *  run (and its re-runs) emit no telemetry. */
         suppressTelemetry?: boolean;
+        /** The id of the `spawn_subagent` tool call that produced this subagent,
+         *  so its events can carry `parentToolCallId` and a client can attach
+         *  subagent state to the spawning tool's pill. */
+        parentToolCallId?: string;
     }): Promise<SubagentSpawnResult>;
     sendMessage(subagentId: string, message: string): Promise<SubagentSpawnResult>;
     stop(subagentId: string): Promise<void>;

package/dist/index.js

@@ -374,6 +374,9 @@ var SUMMARIZATION_PROMPT = `Summarize the following conversation into a structur
 
 Be concise but preserve all information needed to continue the task.
 Omit any section that has no relevant content.`;
+var CUMULATIVE_SUMMARY_PROMPT = `The FIRST message below (tagged [prior-summary]) is an existing working-state summary produced by an earlier compaction. Treat it as the authoritative prior working state: MERGE AND UPDATE it with the newer messages that follow it, carrying forward all still-relevant detail. Do NOT discard or re-compress information from the prior summary just because it is older \u2014 only drop it if the newer messages explicitly supersede it.`;
+var SUBAGENT_DIGEST_CHARS = 500;
+var SUBAGENT_LEDGER_HEADING = "## Subagents";
 var resolveCompactionConfig = (explicit) => {
   if (!explicit) return { ...DEFAULT_COMPACTION_CONFIG };
   return {
@@ -398,32 +401,57 @@ var estimateTotalTokens = (systemPrompt, messages, toolDefinitionsJson) => {
   }
   return tokens;
 };
+var assistantHasToolCalls = (msg) => {
+  if (msg.role !== "assistant") return false;
+  if (typeof msg.content !== "string") return false;
+  if (!msg.content.includes('"tool_calls"')) return false;
+  try {
+    const parsed = JSON.parse(msg.content);
+    return Array.isArray(parsed.tool_calls) && parsed.tool_calls.length > 0;
+  } catch {
+    return false;
+  }
+};
+var splitOrphansToolCalls = (messages, idx) => {
+  if (idx <= 0 || idx >= messages.length) return false;
+  const lastCompacted = messages[idx - 1];
+  return assistantHasToolCalls(lastCompacted);
+};
 var findSafeSplitPoint = (messages, keepRecentMessages) => {
   const candidateIdx = messages.length - keepRecentMessages;
   if (candidateIdx < MIN_COMPACTABLE_MESSAGES) return -1;
   for (let i = candidateIdx; i >= MIN_COMPACTABLE_MESSAGES; i--) {
-    if (messages[i].role === "user") {
+    if (messages[i].role === "user" && !splitOrphansToolCalls(messages, i)) {
       return i;
     }
   }
   for (let i = candidateIdx + 1; i < messages.length - 1; i++) {
-    if (messages[i].role === "user") {
+    if (messages[i].role === "user" && !splitOrphansToolCalls(messages, i)) {
       if (i < MIN_COMPACTABLE_MESSAGES) return -1;
       return i;
     }
   }
   return -1;
 };
+var isCompactionSummary = (msg) => msg.metadata?.isCompactionSummary === true;
 var buildSummarizationMessages = (messagesToCompact, instructions) => {
+  const hasPriorSummary = messagesToCompact.length > 0 && isCompactionSummary(messagesToCompact[0]);
   const conversationLines = [];
-  for (const msg of messagesToCompact) {
+  for (let i = 0; i < messagesToCompact.length; i++) {
+    const msg = messagesToCompact[i];
     const text = getTextContent(msg);
-    const truncated = text.length > SUMMARIZATION_MESSAGE_TRUNCATION_CHARS ? text.slice(0, SUMMARIZATION_MESSAGE_TRUNCATION_CHARS) + "\n...[truncated]" : text;
-    conversationLines.push(`[${msg.role}]: ${truncated}`);
+    const isPrior = i === 0 && hasPriorSummary;
+    const rendered = isPrior || text.length <= SUMMARIZATION_MESSAGE_TRUNCATION_CHARS ? text : text.slice(0, SUMMARIZATION_MESSAGE_TRUNCATION_CHARS) + "\n...[truncated]";
+    const tag = isPrior ? "prior-summary" : msg.role;
+    conversationLines.push(`[${tag}]: ${rendered}`);
   }
-  const prompt = instructions ? `${SUMMARIZATION_PROMPT}
+  let prompt = SUMMARIZATION_PROMPT;
+  if (hasPriorSummary) prompt = `${prompt}
 
-Additional focus: ${instructions}` : SUMMARIZATION_PROMPT;
+${CUMULATIVE_SUMMARY_PROMPT}`;
+  if (instructions) prompt = `${prompt}
+
+Additional focus: ${instructions}`;
   return [
     {
       role: "user",
@@ -435,6 +463,67 @@ ${conversationLines.join("\n\n")}`
     }
   ];
 };
+var SUBAGENT_RESULT_HEADER = /^\[Subagent Result\] Subagent "([^"]*)" \(([^)]*)\) (\S+):/;
+var parseSubagentCallback = (msg) => {
+  if (msg.role !== "user") return null;
+  const meta = msg.metadata ?? {};
+  const text = getTextContent(msg);
+  const hasMetaFlag = meta._subagentCallback === true || meta.subagentCallback === true;
+  const hasTextMarker = text.startsWith("[Subagent Result]");
+  if (!hasMetaFlag && !hasTextMarker) return null;
+  const headerMatch = text.match(SUBAGENT_RESULT_HEADER);
+  const subagentId = typeof meta.subagentId === "string" && meta.subagentId ? meta.subagentId : headerMatch?.[2] ?? "";
+  if (!subagentId) return null;
+  const task = typeof meta.task === "string" && meta.task ? meta.task : headerMatch?.[1] ?? "";
+  const status = headerMatch?.[3] ?? "completed";
+  const bodyStart = text.indexOf("\n\n");
+  const body = bodyStart >= 0 ? text.slice(bodyStart + 2) : text;
+  const digest = body.length > SUBAGENT_DIGEST_CHARS ? body.slice(0, SUBAGENT_DIGEST_CHARS) + "\u2026" : body;
+  return { subagentId, task, status, digest };
+};
+var parsePriorLedger = (summaryText) => {
+  const headingIdx = summaryText.indexOf(SUBAGENT_LEDGER_HEADING);
+  if (headingIdx < 0) return [];
+  const block = summaryText.slice(headingIdx + SUBAGENT_LEDGER_HEADING.length);
+  const entries = [];
+  const entryRe = /^- \*\*(.*?)\*\* \((.+?)\) — (\S+)\n {2}(.*)$/gm;
+  let m;
+  while ((m = entryRe.exec(block)) !== null) {
+    entries.push({
+      task: m[1],
+      subagentId: m[2],
+      status: m[3],
+      digest: m[4]
+    });
+  }
+  return entries;
+};
+var collectSubagentLedger = (messagesToCompact) => {
+  const byId = /* @__PURE__ */ new Map();
+  const order = [];
+  const upsert = (entry) => {
+    if (!byId.has(entry.subagentId)) order.push(entry.subagentId);
+    byId.set(entry.subagentId, entry);
+  };
+  for (const msg of messagesToCompact) {
+    if (isCompactionSummary(msg)) {
+      for (const prior of parsePriorLedger(getTextContent(msg))) upsert(prior);
+      continue;
+    }
+    const entry = parseSubagentCallback(msg);
+    if (entry) upsert(entry);
+  }
+  return order.map((id) => byId.get(id));
+};
+var renderSubagentLedger = (entries) => {
+  if (entries.length === 0) return "";
+  const lines = entries.map(
+    (e) => `- **${e.task}** (${e.subagentId}) \u2014 ${e.status}
+  ${e.digest.replace(/\n/g, " ")}`
+  );
+  return `${SUBAGENT_LEDGER_HEADING}
+${lines.join("\n")}`;
+};
 var buildContinuationMessage = (summary) => ({
   role: "user",
   content: `[CONTEXT COMPACTION] This conversation was automatically compacted. The summary below covers earlier messages.
@@ -483,7 +572,11 @@ var compactMessages = async (model, messages, config, options) => {
         warning: "Summarization returned empty result"
       };
     }
-    const continuationMessage = buildContinuationMessage(summary);
+    const ledger = renderSubagentLedger(collectSubagentLedger(toCompact));
+    const summaryWithLedger = ledger ? `${summary}
+
+${ledger}` : summary;
+    const continuationMessage = buildContinuationMessage(summaryWithLedger);
     const compactedMessages = [continuationMessage, ...toPreserve];
     return {
       compacted: true,
@@ -8323,7 +8416,8 @@ var createSubagentTools = (manager) => [
         parentConversationId: conversationId,
         ownerId,
         tenantId: context.tenantId,
-        suppressTelemetry: context.suppressTelemetry
+        suppressTelemetry: context.suppressTelemetry,
+        parentToolCallId: context.toolCallId
       });
       return { subagentId, status: "running" };
     }
@@ -8585,7 +8679,7 @@ var ToolDispatcher = class {
       };
     }
     try {
-      const output = await definition.handler(call.input, context);
+      const output = await definition.handler(call.input, { ...context, toolCallId: call.id });
       if (context.abortSignal?.aborted) {
         return {
           callId: call.id,
@@ -11170,7 +11264,7 @@ ${textContent}` };
               return;
             }
             const runtimeToolName = exposedToolNames.get(call.name) ?? call.name;
-            yield pushEvent({ type: "tool:started", tool: runtimeToolName, input: call.input });
+            yield pushEvent({ type: "tool:started", tool: runtimeToolName, toolCallId: call.id, input: call.input });
             if (this.requiresApprovalForToolCall(runtimeToolName, call.input)) {
               approvalNeeded.push({
                 approvalId: `approval_${randomUUID5()}`,
@@ -11363,6 +11457,7 @@ ${textContent}` };
               yield pushEvent({
                 type: "tool:error",
                 tool: result2.tool,
+                toolCallId: result2.callId,
                 error: result2.error,
                 recoverable: true
               });
@@ -11404,6 +11499,7 @@ ${textContent}` };
               yield pushEvent({
                 type: "tool:completed",
                 tool: result2.tool,
+                toolCallId: result2.callId,
                 input: callInputMap.get(result2.callId),
                 output: result2.output,
                 duration: now() - batchStart,
@@ -12826,11 +12922,14 @@ var AgentOrchestrator = class {
       result: { status: "completed", response: responseText, steps: 0, tokens: { input: 0, output: 0, cached: 0 }, duration: 0 },
       timestamp: Date.now()
     };
-    await this.conversationStore.appendSubagentResult(conv.parentConversationId, pendingResult);
+    await this.appendSubagentResultReliable(conv.parentConversationId, pendingResult);
     await this.eventSink(conv.parentConversationId, {
       type: "subagent:completed",
       subagentId,
-      conversationId: subagentId
+      conversationId: subagentId,
+      task: conv.subagentMeta?.task ?? conv.title,
+      parentToolCallId: conv.subagentMeta?.parentToolCallId,
+      resultText: responseText
     });
     await this.triggerParentCallback(conv.parentConversationId);
   }
@@ -12906,9 +13005,11 @@ var AgentOrchestrator = class {
     let latestRunId = "";
     let runResult;
     let runError;
+    let parentToolCallId;
     try {
       const conversation = await this.conversationStore.getWithArchive(childConversationId);
       if (!conversation) throw new Error("Subagent conversation not found");
+      parentToolCallId = conversation.subagentMeta?.parentToolCallId;
       if (conversation.subagentMeta?.status === "stopped") return;
       conversation.lastActivityAt = Date.now();
       await this.conversationStore.update(conversation);
@@ -13126,7 +13227,10 @@ var AgentOrchestrator = class {
       await this.eventSink(parentConversationId, {
         type: "subagent:completed",
         subagentId: childConversationId,
-        conversationId: childConversationId
+        conversationId: childConversationId,
+        task,
+        parentToolCallId,
+        resultText: subagentResponse
       });
       this.triggerParentCallback(parentConversationId).catch(
         (err) => console.error(`[poncho][subagent] Parent callback failed:`, err instanceof Error ? err.message : err)
@@ -13157,7 +13261,9 @@ var AgentOrchestrator = class {
         type: "subagent:error",
         subagentId: childConversationId,
         conversationId: childConversationId,
-        error: errMsg
+        error: errMsg,
+        task,
+        parentToolCallId
       });
       this.triggerParentCallback(parentConversationId).catch(
         (err2) => console.error(`[poncho][subagent] Parent callback failed:`, err2 instanceof Error ? err2.message : err2)
@@ -13486,7 +13592,10 @@ ${resultBody}`,
       await this.eventSink(parentConversationId, {
         type: "subagent:completed",
         subagentId: conversationId,
-        conversationId
+        conversationId,
+        task,
+        parentToolCallId: conversation.subagentMeta?.parentToolCallId,
+        resultText: subagentResponse
       });
       if (parentConv) {
         if (this.isServerless) {
@@ -13530,7 +13639,9 @@ ${resultBody}`,
       await this.eventSink(conversation.parentConversationId, {
         type: "subagent:completed",
         subagentId: conversationId,
-        conversationId
+        conversationId,
+        task,
+        parentToolCallId: conversation.subagentMeta?.parentToolCallId
       });
       if (parentConv) {
         if (this.isServerless) {
@@ -13560,7 +13671,7 @@ ${resultBody}`,
           opts.tenantId ?? null,
           {
             parentConversationId: opts.parentConversationId,
-            subagentMeta: { task: opts.task, status: "running", suppressTelemetry: opts.suppressTelemetry },
+            subagentMeta: { task: opts.task, status: "running", suppressTelemetry: opts.suppressTelemetry, parentToolCallId: opts.parentToolCallId },
             messages: [{ role: "user", content: opts.task }]
           }
         );
@@ -13572,7 +13683,8 @@ ${resultBody}`,
           type: "subagent:spawned",
           subagentId: conversation.conversationId,
           conversationId: conversation.conversationId,
-          task: opts.task
+          task: opts.task,
+          parentToolCallId: opts.parentToolCallId
         });
         if (this.isServerless) {
           this.hooks.dispatchBackground("subagent-run", conversation.conversationId);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.52.2",
+  "version": "0.55.0",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",
@@ -34,7 +34,7 @@
     "mustache": "^4.2.0",
     "yaml": "^2.4.0",
     "zod": "^3.22.0",
-    "@poncho-ai/sdk": "1.14.0"
+    "@poncho-ai/sdk": "1.15.0"
   },
   "peerDependencies": {
     "esbuild": ">=0.17.0",