The Open Framework for Building Agents
fix(eve): eviction-first compaction with full-fidelity summarizer input (#1003)
Rui7df0bf1dc261
diff --git a/.changeset/calm-compaction-checkpoints.md b/.changeset/calm-compaction-checkpoints.mdnew file mode 100644index 000000000..b3617d387--- /dev/null+++ b/.changeset/calm-compaction-checkpoints.md@@ -0,0 +1,5 @@+---+"eve": patch+---++Compaction now reserves room for its checkpoint prompt before reaching the configured threshold. The prompt asks the compaction model to distinguish completed work from remaining work, and later compactions receive the previous checkpoint intact instead of truncating it with ordinary transcript text.diff --git a/.changeset/compaction-eviction-first.md b/.changeset/compaction-eviction-first.mdnew file mode 100644index 000000000..9bbc62bec--- /dev/null+++ b/.changeset/compaction-eviction-first.md@@ -0,0 +1,5 @@+---+"eve": patch+---++Compaction now feeds the summarizer full-fidelity conversation text (tool payloads stay compact) and first tries evicting older tool results before summarizing; the kept recent window retains tool results verbatim. Agents lose less context per compaction and stop re-running completed tools.diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/agent.ts b/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/agent.tsdeleted file mode 100644index 62a0b8da5..000000000--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/agent.ts+++ /dev/null@@ -1,9 +0,0 @@-import { createCompactionRegressionAgent } from "@eve/e2e-compaction-regression-shared/agent";-import type { AgentDefinition } from "eve";--const agent: AgentDefinition = createCompactionRegressionAgent({- compactionModel: "openai/gpt-5.6-sol",- modelFamily: "gpt-5.6",-});--export default agent;diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/advance-checkpoint.ts b/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/advance-checkpoint.tsdeleted file mode 100644index e7c2434b0..000000000--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/advance-checkpoint.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/advance-checkpoint";diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/inspect-repository.ts b/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/inspect-repository.tsdeleted file mode 100644index a90064462..000000000--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/inspect-repository.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/inspect-repository";diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/perform-source-analysis.ts b/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/perform-source-analysis.tsdeleted file mode 100644index 9859f5c2f..000000000--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/tools/perform-source-analysis.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/perform-source-analysis";diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/redundant-tool-calls/gpt-5.6.eval.ts b/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/redundant-tool-calls/gpt-5.6.eval.tsdeleted file mode 100644index fd0aaf09d..000000000--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/redundant-tool-calls/gpt-5.6.eval.ts+++ /dev/null@@ -1,9 +0,0 @@-import { defineEval } from "eve/evals";-import { testRedundantToolCalls } from "@eve/e2e-compaction-regression-shared/evals";--export default defineEval({- description: "GPT-5.6 does not repeat an identical successful call after compaction.",- async test(t) {- await testRedundantToolCalls(t, "gpt-5.6");- },-});diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/stale-todo-work/gpt-5.6.eval.ts b/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/stale-todo-work/gpt-5.6.eval.tsdeleted file mode 100644index 689751f30..000000000--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/stale-todo-work/gpt-5.6.eval.ts+++ /dev/null@@ -1,9 +0,0 @@-import { defineEval } from "eve/evals";-import { testStaleTodoWork } from "@eve/e2e-compaction-regression-shared/evals";--export default defineEval({- description: "GPT-5.6 does not repeat completed work because a todo stayed pending.",- async test(t) {- await testStaleTodoWork(t, "gpt-5.6");- },-});diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/agent.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/agent.tsdeleted file mode 100644index d53a7276b..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/agent.ts+++ /dev/null@@ -1,9 +0,0 @@-import { createCompactionRegressionAgent } from "@eve/e2e-compaction-regression-shared/agent";-import type { AgentDefinition } from "eve";--const agent: AgentDefinition = createCompactionRegressionAgent({- compactionModel: "anthropic/claude-opus-4.8",- modelFamily: "opus-4.8",-});--export default agent;diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/instructions.md b/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/instructions.mddeleted file mode 100644index b296ae3dc..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/instructions.md+++ /dev/null@@ -1,14 +0,0 @@-# Compaction regression fixture--Follow the user's requested regression case exactly.--When a tool reports `completed: true` or returns a `completionMarker`, that work-is complete. Call that tool no more than once. Report the marker without-restarting the work after context compaction.--Completed evidence takes precedence over a pending todo. If a compacted summary-says a work unit is complete but the todo still shows it as pending, do not run-the work again.--If a tool returns `hardStop: true`, call no more tools and report its-`completionMarker` immediately.diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/advance-checkpoint.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/advance-checkpoint.tsdeleted file mode 100644index e7c2434b0..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/advance-checkpoint.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/advance-checkpoint";diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/inspect-repository.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/inspect-repository.tsdeleted file mode 100644index a90064462..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/inspect-repository.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/inspect-repository";diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/perform-source-analysis.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/perform-source-analysis.tsdeleted file mode 100644index 9859f5c2f..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/agent/tools/perform-source-analysis.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/perform-source-analysis";diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/evals.config.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/evals.config.tsdeleted file mode 100644index b159ca9a0..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/evals.config.ts+++ /dev/null@@ -1,6 +0,0 @@-import { defineEvalConfig } from "eve/evals";--export default defineEvalConfig({- maxConcurrency: 1,- timeoutMs: 180_000,-});diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/redundant-tool-calls/opus-4.8.eval.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/redundant-tool-calls/opus-4.8.eval.tsdeleted file mode 100644index 5a5614d7a..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/redundant-tool-calls/opus-4.8.eval.ts+++ /dev/null@@ -1,9 +0,0 @@-import { testRedundantToolCalls } from "@eve/e2e-compaction-regression-shared/evals";-import { defineEval } from "eve/evals";--export default defineEval({- description: "Claude Opus 4.8 does not repeat an identical successful call after compaction.",- async test(t) {- await testRedundantToolCalls(t, "opus-4.8");- },-});diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/stale-todo-work/opus-4.8.eval.ts b/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/stale-todo-work/opus-4.8.eval.tsdeleted file mode 100644index e1493c2ef..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/evals/stale-todo-work/opus-4.8.eval.ts+++ /dev/null@@ -1,9 +0,0 @@-import { testStaleTodoWork } from "@eve/e2e-compaction-regression-shared/evals";-import { defineEval } from "eve/evals";--export default defineEval({- description: "Claude Opus 4.8 does not repeat completed work because a todo stayed pending.",- async test(t) {- await testStaleTodoWork(t, "opus-4.8");- },-});diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/package.json b/e2e/fixtures/agent-compaction-regressions-opus-4-8/package.jsondeleted file mode 100644index bc9fed894..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/package.json+++ /dev/null@@ -1,25 +0,0 @@-{- "name": "agent-compaction-regressions-opus-4-8",- "version": "0.0.0",- "private": true,- "type": "module",- "scripts": {- "build": "eve build",- "dev": "eve dev",- "start": "eve start",- "typecheck": "eve build && tsc",- "test:e2e": "eve eval --strict"- },- "dependencies": {- "@eve/e2e-compaction-regression-shared": "workspace:*",- "@opentelemetry/core": "2.6.1",- "@opentelemetry/sdk-trace-base": "2.6.1",- "@vercel/otel": "2.1.2",- "eve": "workspace:*",- "zod": "catalog:"- },- "devDependencies": {- "@types/node": "catalog:",- "typescript": "catalog:"- }-}diff --git a/e2e/fixtures/agent-compaction-regressions-opus-4-8/tsconfig.json b/e2e/fixtures/agent-compaction-regressions-opus-4-8/tsconfig.jsondeleted file mode 100644index 85e3eceb8..000000000--- a/e2e/fixtures/agent-compaction-regressions-opus-4-8/tsconfig.json+++ /dev/null@@ -1,17 +0,0 @@-{- "compilerOptions": {- "target": "ES2022",- "module": "esnext",- "moduleResolution": "bundler",- "outDir": "dist",- "rootDir": ".",- "strict": true,- "esModuleInterop": true,- "skipLibCheck": true,- "forceConsistentCasingInFileNames": true,- "declaration": true,- "noEmit": true,- "types": ["node"]- },- "include": ["agent/**/*.ts", "evals/**/*.ts"]-}diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/agent.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/agent.tsdeleted file mode 100644index 6841e0910..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/agent.ts+++ /dev/null@@ -1,9 +0,0 @@-import { createCompactionRegressionAgent } from "@eve/e2e-compaction-regression-shared/agent";-import type { AgentDefinition } from "eve";--const agent: AgentDefinition = createCompactionRegressionAgent({- compactionModel: "anthropic/claude-sonnet-5",- modelFamily: "sonnet-5",-});--export default agent;diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/instructions.md b/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/instructions.mddeleted file mode 100644index b296ae3dc..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/instructions.md+++ /dev/null@@ -1,14 +0,0 @@-# Compaction regression fixture--Follow the user's requested regression case exactly.--When a tool reports `completed: true` or returns a `completionMarker`, that work-is complete. Call that tool no more than once. Report the marker without-restarting the work after context compaction.--Completed evidence takes precedence over a pending todo. If a compacted summary-says a work unit is complete but the todo still shows it as pending, do not run-the work again.--If a tool returns `hardStop: true`, call no more tools and report its-`completionMarker` immediately.diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/advance-checkpoint.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/advance-checkpoint.tsdeleted file mode 100644index e7c2434b0..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/advance-checkpoint.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/advance-checkpoint";diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/inspect-repository.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/inspect-repository.tsdeleted file mode 100644index a90064462..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/inspect-repository.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/inspect-repository";diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/perform-source-analysis.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/perform-source-analysis.tsdeleted file mode 100644index 9859f5c2f..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/agent/tools/perform-source-analysis.ts+++ /dev/null@@ -1 +0,0 @@-export { default } from "@eve/e2e-compaction-regression-shared/tools/perform-source-analysis";diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/evals.config.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/evals.config.tsdeleted file mode 100644index b159ca9a0..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/evals.config.ts+++ /dev/null@@ -1,6 +0,0 @@-import { defineEvalConfig } from "eve/evals";--export default defineEvalConfig({- maxConcurrency: 1,- timeoutMs: 180_000,-});diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/redundant-tool-calls/sonnet-5.eval.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/redundant-tool-calls/sonnet-5.eval.tsdeleted file mode 100644index 4bf1c8835..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/redundant-tool-calls/sonnet-5.eval.ts+++ /dev/null@@ -1,9 +0,0 @@-import { testRedundantToolCalls } from "@eve/e2e-compaction-regression-shared/evals";-import { defineEval } from "eve/evals";--export default defineEval({- description: "Claude Sonnet 5 does not repeat an identical successful call after compaction.",- async test(t) {- await testRedundantToolCalls(t, "sonnet-5");- },-});diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/stale-todo-work/sonnet-5.eval.ts b/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/stale-todo-work/sonnet-5.eval.tsdeleted file mode 100644index 3f76c4d02..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/evals/stale-todo-work/sonnet-5.eval.ts+++ /dev/null@@ -1,9 +0,0 @@-import { testStaleTodoWork } from "@eve/e2e-compaction-regression-shared/evals";-import { defineEval } from "eve/evals";--export default defineEval({- description: "Claude Sonnet 5 does not repeat completed work because a todo stayed pending.",- async test(t) {- await testStaleTodoWork(t, "sonnet-5");- },-});diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/package.json b/e2e/fixtures/agent-compaction-regressions-sonnet-5/package.jsondeleted file mode 100644index 2ea9748b5..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/package.json+++ /dev/null@@ -1,25 +0,0 @@-{- "name": "agent-compaction-regressions-sonnet-5",- "version": "0.0.0",- "private": true,- "type": "module",- "scripts": {- "build": "eve build",- "dev": "eve dev",- "start": "eve start",- "typecheck": "eve build && tsc",- "test:e2e": "eve eval --strict"- },- "dependencies": {- "@eve/e2e-compaction-regression-shared": "workspace:*",- "@opentelemetry/core": "2.6.1",- "@opentelemetry/sdk-trace-base": "2.6.1",- "@vercel/otel": "2.1.2",- "eve": "workspace:*",- "zod": "catalog:"- },- "devDependencies": {- "@types/node": "catalog:",- "typescript": "catalog:"- }-}diff --git a/e2e/fixtures/agent-compaction-regressions-sonnet-5/tsconfig.json b/e2e/fixtures/agent-compaction-regressions-sonnet-5/tsconfig.jsondeleted file mode 100644index 85e3eceb8..000000000--- a/e2e/fixtures/agent-compaction-regressions-sonnet-5/tsconfig.json+++ /dev/null@@ -1,17 +0,0 @@-{- "compilerOptions": {- "target": "ES2022",- "module": "esnext",- "moduleResolution": "bundler",- "outDir": "dist",- "rootDir": ".",- "strict": true,- "esModuleInterop": true,- "skipLibCheck": true,- "forceConsistentCasingInFileNames": true,- "declaration": true,- "noEmit": true,- "types": ["node"]- },- "include": ["agent/**/*.ts", "evals/**/*.ts"]-}diff --git a/e2e/fixtures/agent-compaction-regressions/agent/agent.ts b/e2e/fixtures/agent-compaction-regressions/agent/agent.tsnew file mode 100644index 000000000..24eaee9e6--- /dev/null+++ b/e2e/fixtures/agent-compaction-regressions/agent/agent.ts@@ -0,0 +1,181 @@+import { defineAgent } from "eve";+import { mockModel, type MockModelRequest } from "eve/evals";++import {+ COMPACTION_CHECKPOINT_TEXT,+ SECOND_CHECKPOINT_MARKER,+ TASK_PRESERVED_MARKER,+ TASK_TAIL_SENTINEL,+} from "../constants";++const TEST_CONTEXT_WINDOW_TOKENS = 32_000;+const MAX_TOOL_CALLS = 10;++type RegressionCase = "redundant-tool-calls" | "stale-todo-work" | "task-survival";++let activeCase: RegressionCase | undefined;+const checkpointAdvanceCallCounts = new Map<RegressionCase, number>();+const toolCallCounts = new Map<RegressionCase, number>();++let requestCount = 0;++const taskModel = mockModel({+ modelId: "compaction-regression-task-model",+ respond(request) {+ // EVE_E2E_DUMP_CONTEXT=1 prints every request's messages — the context+ // exactly as the model sees it, so compaction, capping, and replay are+ // observable per step while iterating on these evals.+ if (process.env.EVE_E2E_DUMP_CONTEXT) {+ requestCount += 1;+ console.log(`\n=== model request #${requestCount} (${request.messages.length} messages) ===`);+ for (const message of request.messages) {+ const text = message.text.replace(/\s+/g, " ");+ console.log(` [${message.role}] ${text.length} chars | ${text.slice(0, 160)}`);+ }+ }++ const initialCase = findInitialCase(request);+ if (initialCase !== undefined && activeCase !== initialCase) {+ activeCase = initialCase;+ checkpointAdvanceCallCounts.set(initialCase, 0);+ toolCallCounts.set(initialCase, 0);+ }++ if (activeCase === undefined) {+ throw new Error("Compaction regression task model received no case marker.");+ }++ const regressionCase = activeCase;++ if (regressionCase === "task-survival") {+ const compacted = request.messages.some(+ (message) => message.role === "user" && message.text === COMPACTION_CHECKPOINT_TEXT,+ );+ if (compacted) {+ // The harness must hand the model its verbatim task back after+ // compaction — via the kept tail or the resumption replay. Losing it+ // is the trace failure this case pins.+ return request.userMessages.some((text) => text.includes(TASK_TAIL_SENTINEL))+ ? `Task text still visible: ${TASK_PRESERVED_MARKER}`+ : "Task text lost after compaction: TASK_LOST";+ }++ const pressureCalls = toolCallCounts.get(regressionCase) ?? 0;+ if (pressureCalls >= MAX_TOOL_CALLS) {+ return "Hard stop without a compaction: TASK_SURVIVAL_NO_COMPACTION";+ }++ toolCallCounts.set(regressionCase, pressureCalls + 1);+ return {+ toolCalls: [+ {+ id: `inspect-repository-${pressureCalls + 1}`,+ input: { scope: "repository" },+ name: "inspect-repository",+ },+ ],+ };+ }++ const marker = completionMarker(regressionCase);++ // These are fixture markers, not compaction protocol fields. `marker` records the+ // regression work tool; `SECOND_CHECKPOINT_MARKER` records the test-only tool+ // whose output makes the harness cross the compaction threshold a second time.+ // Completion evidence is detected in any assistant message: compaction may+ // leave it as a summarization checkpoint or as an eviction trail line, and+ // the model must not repeat work in either case. User messages are+ // excluded because the eval instructions themselves quote the markers.+ if (assistantEvidenceContains(request.messages, marker)) {+ if (assistantEvidenceContains(request.messages, SECOND_CHECKPOINT_MARKER)) {+ return `Done: ${marker}; ${SECOND_CHECKPOINT_MARKER}`;+ }++ const advanceCalls = checkpointAdvanceCallCounts.get(regressionCase) ?? 0;+ if (advanceCalls >= MAX_TOOL_CALLS) {+ return `Hard stop after ${MAX_TOOL_CALLS} checkpoint advances: ${marker}`;+ }++ checkpointAdvanceCallCounts.set(regressionCase, advanceCalls + 1);+ return {+ toolCalls: [+ {+ id: `advance-checkpoint-${advanceCalls + 1}`,+ input: { regressionCase },+ name: "advance-checkpoint",+ },+ ],+ };+ }++ const completedCalls = toolCallCounts.get(regressionCase) ?? 0;+ if (completedCalls >= MAX_TOOL_CALLS) {+ return `Hard stop after ${MAX_TOOL_CALLS} calls: ${marker}`;+ }++ const attempt = completedCalls + 1;+ toolCallCounts.set(regressionCase, attempt);++ return regressionCase === "redundant-tool-calls"+ ? {+ toolCalls: [+ {+ id: `inspect-repository-${attempt}`,+ input: { scope: "repository" },+ name: "inspect-repository",+ },+ ],+ }+ : {+ toolCalls: [+ {+ id: `perform-source-analysis-${attempt}`,+ input: { approach: `attempt-${attempt}` },+ name: "perform-source-analysis",+ },+ ],+ };+ },+});++export default defineAgent({+ model: taskModel,+ modelContextWindowTokens: TEST_CONTEXT_WINDOW_TOKENS,+ compaction: {+ model: process.env.EVE_E2E_MODEL ?? "openai/gpt-5.6-sol",+ modelContextWindowTokens: TEST_CONTEXT_WINDOW_TOKENS,+ thresholdPercent: 0.02,+ },+ limits: {+ maxInputTokensPerSession: 100_000,+ },+});++function findInitialCase(request: MockModelRequest): RegressionCase | undefined {+ for (const message of request.userMessages) {+ const regressionCase = regressionCaseFromText(message);+ if (regressionCase !== undefined) return regressionCase;+ }++ return undefined;+}++function regressionCaseFromText(text: string): RegressionCase | undefined {+ if (text.includes("[case: redundant-tool-calls]")) return "redundant-tool-calls";+ if (text.includes("[case: stale-todo-work]")) return "stale-todo-work";+ if (text.includes("[case: task-survival]")) return "task-survival";+ return undefined;+}++function completionMarker(regressionCase: Exclude<RegressionCase, "task-survival">): string {+ return regressionCase === "redundant-tool-calls"+ ? "REPOSITORY_INSPECTION_COMPLETE"+ : "SOURCE_ANALYSIS_COMPLETE";+}++function assistantEvidenceContains(+ messages: MockModelRequest["messages"],+ marker: string,+): boolean {+ return messages.some((message) => message.role === "assistant" && message.text.includes(marker));+}diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/instructions.md b/e2e/fixtures/agent-compaction-regressions/agent/instructions.mdsimilarity index 100%rename from e2e/fixtures/agent-compaction-regressions-gpt-5-6/agent/instructions.mdrename to e2e/fixtures/agent-compaction-regressions/agent/instructions.mddiff --git a/e2e/fixtures/compaction-regression-shared/src/advance-checkpoint.ts b/e2e/fixtures/agent-compaction-regressions/agent/tools/advance-checkpoint.tssimilarity index 81%rename from e2e/fixtures/compaction-regression-shared/src/advance-checkpoint.tsrename to e2e/fixtures/agent-compaction-regressions/agent/tools/advance-checkpoint.tsindex d61f1c769..8cdea2bf8 100644--- a/e2e/fixtures/compaction-regression-shared/src/advance-checkpoint.ts+++ b/e2e/fixtures/agent-compaction-regressions/agent/tools/advance-checkpoint.ts@@ -2,16 +2,14 @@ import { defineState } from "eve/context";import { defineTool } from "eve/tools";import { z } from "zod";-import { SECOND_CHECKPOINT_MARKER } from "./constants";+import { SECOND_CHECKPOINT_MARKER } from "../../constants";const invocationCount = defineState("compaction-regression.advance-checkpoint", () => 0);-const modelFamilySchema = z.enum(["gpt-5.6", "opus-4.8", "sonnet-5"]);export default defineTool({description:"Test-only second-compaction trigger tool. Records a completed fixture work unit and adds enough evidence for the harness to cross the compaction threshold again.",inputSchema: z.object({- modelFamily: modelFamilySchema,regressionCase: z.enum(["redundant-tool-calls", "stale-todo-work"]),}),async execute(input) {@@ -21,7 +19,6 @@ export default defineTool({return {checkpointMarker: SECOND_CHECKPOINT_MARKER,completed: true,- modelFamily: input.modelFamily,regressionCase: input.regressionCase,attempt,evidencePadding: "second checkpoint evidence ".repeat(100),diff --git a/e2e/fixtures/compaction-regression-shared/src/inspect-repository.ts b/e2e/fixtures/agent-compaction-regressions/agent/tools/inspect-repository.tssimilarity index 82%rename from e2e/fixtures/compaction-regression-shared/src/inspect-repository.tsrename to e2e/fixtures/agent-compaction-regressions/agent/tools/inspect-repository.tsindex 61a1e072e..a75bb6de0 100644--- a/e2e/fixtures/compaction-regression-shared/src/inspect-repository.ts+++ b/e2e/fixtures/agent-compaction-regressions/agent/tools/inspect-repository.ts@@ -4,21 +4,18 @@ import { z } from "zod";const completionMarker = "REPOSITORY_INSPECTION_COMPLETE";const invocationCount = defineState("compaction-regression.inspect-repository", () => 0);-const modelFamilySchema = z.enum(["gpt-5.6", "opus-4.8", "sonnet-5"]);export default defineTool({description:"Compaction regression tool. Inspect the repository exactly once when the user requests the redundant-tool-calls case.",inputSchema: z.object({- modelFamily: modelFamilySchema,scope: z.literal("repository"),}),- async execute(input) {+ async execute() {const attempt = invocationCount.get() + 1;invocationCount.update(() => attempt);return {- modelFamily: input.modelFamily,completed: true,completionMarker,workUnit: "repository-inspection",diff --git a/e2e/fixtures/compaction-regression-shared/src/perform-source-analysis.ts b/e2e/fixtures/agent-compaction-regressions/agent/tools/perform-source-analysis.tssimilarity index 87%rename from e2e/fixtures/compaction-regression-shared/src/perform-source-analysis.tsrename to e2e/fixtures/agent-compaction-regressions/agent/tools/perform-source-analysis.tsindex 25848bbfe..63b49a988 100644--- a/e2e/fixtures/compaction-regression-shared/src/perform-source-analysis.ts+++ b/e2e/fixtures/agent-compaction-regressions/agent/tools/perform-source-analysis.ts@@ -5,14 +5,12 @@ import { z } from "zod";const completionMarker = "SOURCE_ANALYSIS_COMPLETE";const invocationCount = defineState("compaction-regression.perform-source-analysis", () => 0);-const modelFamilySchema = z.enum(["gpt-5.6", "opus-4.8", "sonnet-5"]);export default defineTool({description:"Compaction regression tool. Complete source analysis exactly once when the user requests the stale-todo-work case.",inputSchema: z.object({approach: z.string().min(1),- modelFamily: modelFamilySchema,}),async execute(input, ctx) {const attempt = invocationCount.get() + 1;@@ -25,7 +23,6 @@ export default defineTool({);return {- modelFamily: input.modelFamily,completed: true,completionMarker,workUnit: "source-analysis",diff --git a/e2e/fixtures/agent-compaction-regressions/constants.ts b/e2e/fixtures/agent-compaction-regressions/constants.tsnew file mode 100644index 000000000..e49abf5b1--- /dev/null+++ b/e2e/fixtures/agent-compaction-regressions/constants.ts@@ -0,0 +1,14 @@+/** Fixture evidence emitted by the test-only second-compaction trigger tool. */+export const SECOND_CHECKPOINT_MARKER = "SECOND_CHECKPOINT_READY";++/** Trailing sentinel of the task-survival case's long task message. */+export const TASK_TAIL_SENTINEL = "TASK_TAIL_SENTINEL_AFTER_PADDING";++/**+ * Reported by the mock task model only when, after at least one compaction,+ * the verbatim task sentinel is still visible in a user message.+ */+export const TASK_PRESERVED_MARKER = "TASK_PRESERVED_AFTER_COMPACTION";++/** Checkpoint marker the harness inserts when compaction summarizes. */+export const COMPACTION_CHECKPOINT_TEXT = "Summary of our conversation so far:";diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/evals.config.ts b/e2e/fixtures/agent-compaction-regressions/evals/evals.config.tssimilarity index 100%rename from e2e/fixtures/agent-compaction-regressions-gpt-5-6/evals/evals.config.tsrename to e2e/fixtures/agent-compaction-regressions/evals/evals.config.tsdiff --git a/e2e/fixtures/agent-compaction-regressions/evals/redundant-tool-calls.eval.ts b/e2e/fixtures/agent-compaction-regressions/evals/redundant-tool-calls.eval.tsnew file mode 100644index 000000000..eeb9b43f8--- /dev/null+++ b/e2e/fixtures/agent-compaction-regressions/evals/redundant-tool-calls.eval.ts@@ -0,0 +1,31 @@+import { defineEval } from "eve/evals";++import { SECOND_CHECKPOINT_MARKER } from "../constants";++export default defineEval({+ description: "The model does not repeat an identical successful call after compaction.",+ async test(t) {+ const turn = await t.send(+ [+ "[case: redundant-tool-calls]",+ "Call inspect-repository exactly once with scope repository.",+ "After it succeeds, report REPOSITORY_INSPECTION_COMPLETE and call no more tools.",+ ].join("\n"),+ );++ turn.expectOk();+ t.succeeded();+ t.calledTool("inspect-repository", {+ count: 1,+ input: { scope: "repository" },+ output: { completed: true, completionMarker: "REPOSITORY_INSPECTION_COMPLETE" },+ });+ t.calledTool("advance-checkpoint", {+ count: 1,+ output: { checkpointMarker: SECOND_CHECKPOINT_MARKER, completed: true },+ });+ t.event("compaction.completed", { count: 2 });+ t.messageIncludes("REPOSITORY_INSPECTION_COMPLETE");+ t.messageIncludes(SECOND_CHECKPOINT_MARKER);+ },+});diff --git a/e2e/fixtures/agent-compaction-regressions/evals/stale-todo-work.eval.ts b/e2e/fixtures/agent-compaction-regressions/evals/stale-todo-work.eval.tsnew file mode 100644index 000000000..71deb8508--- /dev/null+++ b/e2e/fixtures/agent-compaction-regressions/evals/stale-todo-work.eval.ts@@ -0,0 +1,31 @@+import { defineEval } from "eve/evals";++import { SECOND_CHECKPOINT_MARKER } from "../constants";++export default defineEval({+ description: "The model does not redo completed work because a stale todo stayed pending.",+ async test(t) {+ const turn = await t.send(+ [+ "[case: stale-todo-work]",+ "Call perform-source-analysis exactly once with approach initial.",+ "The tool deliberately leaves its completed work in a pending todo.",+ "After it succeeds, report SOURCE_ANALYSIS_COMPLETE and call no more tools.",+ ].join("\n"),+ );++ turn.expectOk();+ t.succeeded();+ t.calledTool("perform-source-analysis", {+ count: 1,+ output: { completed: true, workUnit: "source-analysis" },+ });+ t.calledTool("advance-checkpoint", {+ count: 1,+ output: { checkpointMarker: SECOND_CHECKPOINT_MARKER, completed: true },+ });+ t.event("compaction.completed", { count: 2 });+ t.messageIncludes("SOURCE_ANALYSIS_COMPLETE");+ t.messageIncludes(SECOND_CHECKPOINT_MARKER);+ },+});diff --git a/e2e/fixtures/agent-compaction-regressions/evals/task-survival.eval.ts b/e2e/fixtures/agent-compaction-regressions/evals/task-survival.eval.tsnew file mode 100644index 000000000..95d88abf2--- /dev/null+++ b/e2e/fixtures/agent-compaction-regressions/evals/task-survival.eval.ts@@ -0,0 +1,26 @@+import { defineEval } from "eve/evals";++import { TASK_PRESERVED_MARKER, TASK_TAIL_SENTINEL } from "../constants";++export default defineEval({+ description: "The verbatim task text survives compaction and reaches the model.",+ async test(t) {+ const turn = await t.send(+ [+ "[case: task-survival]",+ "Call inspect-repository once to generate context pressure, then confirm the task.",+ // Sized to a narrow window: long enough that a 280-char summarizer+ // input cap would destroy the tail sentinel, but short enough that+ // the task alone cannot cross the fixture's ~640-token threshold —+ // otherwise compaction fires before the first model call and the+ // mock never generates pressure.+ `Requirements: ${"handle the edge case exactly. ".repeat(12)}${TASK_TAIL_SENTINEL}`,+ ].join("\n"),+ );++ turn.expectOk();+ t.succeeded();+ t.calledTool("inspect-repository", { count: 1 });+ t.messageIncludes(TASK_PRESERVED_MARKER);+ },+});diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/package.json b/e2e/fixtures/agent-compaction-regressions/package.jsonsimilarity index 82%rename from e2e/fixtures/agent-compaction-regressions-gpt-5-6/package.jsonrename to e2e/fixtures/agent-compaction-regressions/package.jsonindex 84b2f40e2..13b9c3b24 100644--- a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/package.json+++ b/e2e/fixtures/agent-compaction-regressions/package.json@@ -1,5 +1,5 @@{- "name": "agent-compaction-regressions-gpt-5-6",+ "name": "agent-compaction-regressions","version": "0.0.0","private": true,"type": "module",@@ -11,7 +11,6 @@"test:e2e": "eve eval --strict"},"dependencies": {- "@eve/e2e-compaction-regression-shared": "workspace:*","@opentelemetry/core": "2.6.1","@opentelemetry/sdk-trace-base": "2.6.1","@vercel/otel": "2.1.2",diff --git a/e2e/fixtures/agent-compaction-regressions-gpt-5-6/tsconfig.json b/e2e/fixtures/agent-compaction-regressions/tsconfig.jsonsimilarity index 100%rename from e2e/fixtures/agent-compaction-regressions-gpt-5-6/tsconfig.jsonrename to e2e/fixtures/agent-compaction-regressions/tsconfig.jsondiff --git a/e2e/fixtures/compaction-regression-shared/package.json b/e2e/fixtures/compaction-regression-shared/package.jsondeleted file mode 100644index 7c3e6963c..000000000--- a/e2e/fixtures/compaction-regression-shared/package.json+++ /dev/null@@ -1,17 +0,0 @@-{- "name": "@eve/e2e-compaction-regression-shared",- "version": "0.0.0",- "private": true,- "type": "module",- "exports": {- "./agent": "./src/agent.ts",- "./evals": "./src/evals.ts",- "./tools/advance-checkpoint": "./src/advance-checkpoint.ts",- "./tools/inspect-repository": "./src/inspect-repository.ts",- "./tools/perform-source-analysis": "./src/perform-source-analysis.ts"- },- "dependencies": {- "eve": "workspace:*",- "zod": "catalog:"- }-}diff --git a/e2e/fixtures/compaction-regression-shared/src/agent.ts b/e2e/fixtures/compaction-regression-shared/src/agent.tsdeleted file mode 100644index b951f5020..000000000--- a/e2e/fixtures/compaction-regression-shared/src/agent.ts+++ /dev/null@@ -1,150 +0,0 @@-import { defineAgent, type AgentDefinition } from "eve";-import { mockModel, type MockModelRequest } from "eve/evals";--import { SECOND_CHECKPOINT_MARKER } from "./constants";--const TEST_CONTEXT_WINDOW_TOKENS = 32_000;-const MAX_TOOL_CALLS = 10;--export type ModelFamily = "gpt-5.6" | "opus-4.8" | "sonnet-5";-type RegressionCase = "redundant-tool-calls" | "stale-todo-work";--interface ActiveRegression {- readonly regressionCase: RegressionCase;-}--export function createCompactionRegressionAgent(input: {- readonly compactionModel: string;- readonly modelFamily: ModelFamily;-}): AgentDefinition {- let activeRegression: ActiveRegression | undefined;- const checkpointAdvanceCallCounts = new Map<RegressionCase, number>();- const toolCallCounts = new Map<RegressionCase, number>();- const taskModel = mockModel({- modelId: `compaction-regression-task-model-${input.modelFamily}`,- respond(request) {- const initialRegression = findInitialRegression(request, input.modelFamily);- if (- initialRegression !== undefined &&- activeRegression?.regressionCase !== initialRegression.regressionCase- ) {- activeRegression = initialRegression;- checkpointAdvanceCallCounts.set(initialRegression.regressionCase, 0);- toolCallCounts.set(initialRegression.regressionCase, 0);- }-- if (activeRegression === undefined) {- throw new Error("Compaction regression task model received no case marker.");- }-- const regression = activeRegression;- const marker = completionMarker(regression.regressionCase);-- // These are fixture markers, not compaction protocol fields. `marker` records the- // regression work tool; `SECOND_CHECKPOINT_MARKER` records the test-only tool- // whose output makes the harness cross the compaction threshold a second time.- if (checkpointContains(request.messages, marker)) {- if (checkpointContains(request.messages, SECOND_CHECKPOINT_MARKER)) {- return `Done: ${marker}; ${SECOND_CHECKPOINT_MARKER}`;- }-- const advanceCalls = checkpointAdvanceCallCounts.get(regression.regressionCase) ?? 0;- if (advanceCalls >= MAX_TOOL_CALLS) {- return `Hard stop after ${MAX_TOOL_CALLS} checkpoint advances: ${marker}`;- }-- checkpointAdvanceCallCounts.set(regression.regressionCase, advanceCalls + 1);- return {- toolCalls: [- {- id: `advance-checkpoint-${advanceCalls + 1}`,- input: {- modelFamily: input.modelFamily,- regressionCase: regression.regressionCase,- },- name: "advance-checkpoint",- },- ],- };- }-- const completedCalls = toolCallCounts.get(regression.regressionCase) ?? 0;- if (completedCalls >= MAX_TOOL_CALLS) {- return `Hard stop after ${MAX_TOOL_CALLS} calls: ${marker}`;- }-- const attempt = completedCalls + 1;- toolCallCounts.set(regression.regressionCase, attempt);-- return regression.regressionCase === "redundant-tool-calls"- ? {- toolCalls: [- {- id: `inspect-repository-${attempt}`,- input: { modelFamily: input.modelFamily, scope: "repository" },- name: "inspect-repository",- },- ],- }- : {- toolCalls: [- {- id: `perform-source-analysis-${attempt}`,- input: { approach: `attempt-${attempt}`, modelFamily: input.modelFamily },- name: "perform-source-analysis",- },- ],- };- },- });-- return defineAgent({- model: taskModel,- modelContextWindowTokens: TEST_CONTEXT_WINDOW_TOKENS,- compaction: {- model: input.compactionModel,- modelContextWindowTokens: TEST_CONTEXT_WINDOW_TOKENS,- thresholdPercent: 0.02,- },- limits: {- maxInputTokensPerSession: 100_000,- },- });-}--function findInitialRegression(- request: MockModelRequest,- modelFamily: ModelFamily,-): ActiveRegression | undefined {- for (const message of request.userMessages) {- if (!message.includes(`[model: ${modelFamily}]`)) continue;-- const regressionCase = regressionCaseFromText(message);- if (regressionCase !== undefined) return { regressionCase };- }-- return undefined;-}--function regressionCaseFromText(text: string): RegressionCase | undefined {- if (text.includes("[case: redundant-tool-calls]")) return "redundant-tool-calls";- if (text.includes("[case: stale-todo-work]")) return "stale-todo-work";- return undefined;-}--function completionMarker(regressionCase: RegressionCase): string {- return regressionCase === "redundant-tool-calls"- ? "REPOSITORY_INSPECTION_COMPLETE"- : "SOURCE_ANALYSIS_COMPLETE";-}--function checkpointContains(messages: MockModelRequest["messages"], marker: string): boolean {- return messages.some((message, index) => {- if (message.role !== "user" || message.text !== "Summary of our conversation so far:") {- return false;- }-- const checkpoint = messages[index + 1];- return checkpoint?.role === "assistant" && checkpoint.text.includes(marker);- });-}diff --git a/e2e/fixtures/compaction-regression-shared/src/constants.ts b/e2e/fixtures/compaction-regression-shared/src/constants.tsdeleted file mode 100644index 9773bae5d..000000000--- a/e2e/fixtures/compaction-regression-shared/src/constants.ts+++ /dev/null@@ -1,2 +0,0 @@-/** Fixture evidence emitted by the test-only second-compaction trigger tool. */-export const SECOND_CHECKPOINT_MARKER = "SECOND_CHECKPOINT_READY";diff --git a/e2e/fixtures/compaction-regression-shared/src/evals.ts b/e2e/fixtures/compaction-regression-shared/src/evals.tsdeleted file mode 100644index a2d670472..000000000--- a/e2e/fixtures/compaction-regression-shared/src/evals.ts+++ /dev/null@@ -1,62 +0,0 @@-import type { EveEvalContext } from "eve/evals";--import type { ModelFamily } from "./agent";-import { SECOND_CHECKPOINT_MARKER } from "./constants";--export async function testRedundantToolCalls(- t: EveEvalContext,- modelFamily: ModelFamily,-): Promise<void> {- const turn = await t.send(- [- `[model: ${modelFamily}]`,- "[case: redundant-tool-calls]",- "Call inspect-repository exactly once with scope repository.",- "After it succeeds, report REPOSITORY_INSPECTION_COMPLETE and call no more tools.",- ].join("\n"),- );-- turn.expectOk();- t.succeeded();- t.calledTool("inspect-repository", {- count: 1,- input: { scope: "repository" },- output: { completed: true, completionMarker: "REPOSITORY_INSPECTION_COMPLETE" },- });- t.calledTool("advance-checkpoint", {- count: 1,- output: { checkpointMarker: SECOND_CHECKPOINT_MARKER, completed: true },- });- t.event("compaction.completed", { count: 2 });- t.messageIncludes("REPOSITORY_INSPECTION_COMPLETE");- t.messageIncludes(SECOND_CHECKPOINT_MARKER);-}--export async function testStaleTodoWork(- t: EveEvalContext,- modelFamily: ModelFamily,-): Promise<void> {- const turn = await t.send(- [- `[model: ${modelFamily}]`,- "[case: stale-todo-work]",- "Call perform-source-analysis exactly once with approach initial.",- "The tool deliberately leaves its completed work in a pending todo.",- "After it succeeds, report SOURCE_ANALYSIS_COMPLETE and call no more tools.",- ].join("\n"),- );-- turn.expectOk();- t.succeeded();- t.calledTool("perform-source-analysis", {- count: 1,- output: { completed: true, workUnit: "source-analysis" },- });- t.calledTool("advance-checkpoint", {- count: 1,- output: { checkpointMarker: SECOND_CHECKPOINT_MARKER, completed: true },- });- t.event("compaction.completed", { count: 2 });- t.messageIncludes("SOURCE_ANALYSIS_COMPLETE");- t.messageIncludes(SECOND_CHECKPOINT_MARKER);-}diff --git a/packages/eve/src/harness/compaction-prompt.test.ts b/packages/eve/src/harness/compaction-prompt.test.tsindex 293c47af9..d9fd455e6 100644--- a/packages/eve/src/harness/compaction-prompt.test.ts+++ b/packages/eve/src/harness/compaction-prompt.test.ts@@ -18,7 +18,7 @@ describe("createCompactionPrompt", () => {expect(result.prompt).toContain(markerAfterTextLimit);});- it("summarizes structured tool messages without dumping raw JSON", () => {+ it("passes tool payloads to the summarizer raw so it can judge what matters", () => {const messages: ModelMessage[] = [{content: [@@ -51,11 +51,70 @@ describe("createCompactionPrompt", () => {expect(result.prompt).toContain("Conversation transcript:");expect(result.prompt).toContain("### assistant");- expect(result.prompt).toContain("Called search with object(query=debug)");- expect(result.prompt).toContain(- "Tool search returned object(type=json, value=array(4: alpha, beta, gamma, …))",- );- expect(result.prompt).not.toContain('{"query"');- expect(result.prompt).not.toContain('{"items"');+ // Raw clipped JSON, not a pre-digested object(…) skeleton: the checkpoint+ // model is the one deciding which parts of a payload matter.+ expect(result.prompt).toContain('Called search with {"query":"debug"}');+ expect(result.prompt).toContain('"alpha"');+ expect(result.prompt).toContain('"delta"');+ });++ it("clips oversized tool payloads instead of reproducing them", () => {+ const big = "match line ".repeat(1_000);+ const messages: ModelMessage[] = [+ {+ content: [+ {+ output: { type: "json", value: { content: big } },+ toolCallId: "call-1",+ toolName: "grep",+ type: "tool-result",+ },+ ],+ role: "tool",+ },+ ];++ const result = createCompactionPrompt({ messages, previousCheckpoint: undefined });++ expect(result.prompt).toContain("Tool grep returned");+ expect(result.prompt).toContain("match line");+ expect(result.prompt).not.toContain(big);+ expect(result.prompt).toContain("…");+ });++ it("renders conversational text verbatim regardless of length", () => {+ // A delegated task message destroyed here is unrecoverable after the first+ // compaction, so user/assistant text must reach the summarizer whole.+ const taskTail = "CRITICAL_REQUIREMENT_AFTER_280_CHARACTERS";+ const task = `${"do the following work item. ".repeat(30)}${taskTail}`;++ const result = createCompactionPrompt({+ messages: [+ { content: task, role: "user" },+ { content: [{ text: task, type: "text" }], role: "assistant" },+ ],+ previousCheckpoint: undefined,+ });++ expect(result.prompt).toContain(taskTail);+ expect(result.prompt.split(taskTail)).toHaveLength(3);+ });++ it("degrades the oldest conversational text first under budget pressure", () => {+ const oldest = `${"oldest message padding. ".repeat(400)}OLDEST_TAIL_MARKER`;+ const newest = `${"newest message padding. ".repeat(400)}NEWEST_TAIL_MARKER`;++ const result = createCompactionPrompt({+ messages: [+ { content: oldest, role: "user" },+ { content: newest, role: "user" },+ ],+ // Fits one full entry plus a degraded one, but not both full.+ transcriptBudgetTokens: 3_500,+ previousCheckpoint: undefined,+ });++ expect(result.prompt).not.toContain("OLDEST_TAIL_MARKER");+ expect(result.prompt).toContain("NEWEST_TAIL_MARKER");});});diff --git a/packages/eve/src/harness/compaction-prompt.ts b/packages/eve/src/harness/compaction-prompt.tsindex 9f3ce0773..86f511dde 100644--- a/packages/eve/src/harness/compaction-prompt.ts+++ b/packages/eve/src/harness/compaction-prompt.ts@@ -1,7 +1,20 @@import type { ModelMessage } from "ai";+import { estimateTokens } from "#harness/token-estimate.js";+export const COMPACTION_CHECKPOINT_MARKER = "Summary of our conversation so far:";+/** Synthetic resumption prompt used when no real user message can be replayed. */+export const COMPACTION_RESUMPTION_MESSAGE = "Continue.";++/**+ * Label line of the framework-injected todo preservation message. Owned here+ * so compaction can recognize the message as synthetic when picking a user+ * message to replay after compaction.+ */+export const TODO_COMPACTION_PRESERVATION_LABEL =+ "[Your task list was preserved across context compaction]";+const COMPACTION_SYSTEM_PROMPT = `You are performing a CONTEXT CHECKPOINT COMPACTION. Create a handoff summary for another LLM that will resume the task.Include:@@ -15,12 +28,15 @@ Be concise, structured, and focused on helping the next LLM seamlessly continueconst COMPACTION_CHECKPOINT_PROMPT = `Update the previous checkpoint with the newer information in the conversation. If there is no previous checkpoint, create one from the conversation.-Make completed work explicit so the next model does not repeat it. Keep completed work separate from current and remaining work, and do not describe completed work as pending unless later messages show it must be redone. Preserve exact file paths, function names, commands, error messages, identifiers, and measured values when they are needed to continue.`;+Make completed work explicit so the next model does not repeat it. Keep completed work separate from current and remaining work, and do not describe completed work as pending unless later messages show it must be redone. Preserve exact file paths, function names, commands, error messages, identifiers, and measured values when they are needed to continue.-const COMPACTION_TEXT_LIMIT = 280;-const COMPACTION_COLLECTION_LIMIT = 3;+Large tool outputs are the main thing to compress: reduce each to the findings the next model needs — what was searched or read, what it established, and the exact identifiers involved — rather than reproducing the output. The next model cannot see the originals, so nothing it would need to act on may be lost.`;-interface CompactionTranscriptMessage {+// Fallback cap for conversational text, applied oldest-first only when the+// rendered transcript exceeds the caller's token budget.+const DEGRADED_TEXT_LIMIT = 2_000;++interface CompactionTranscriptEntry {readonly content: string;readonly role: ModelMessage["role"];}@@ -36,25 +52,79 @@ export const COMPACTION_PROMPT_ENVELOPE = {system: COMPACTION_SYSTEM_PROMPT,} satisfies CompactionPrompt;-/** Builds the compaction model input from framework-owned checkpoint state and older messages. */+/**+ * Builds the compaction model input from framework-owned checkpoint state and+ * older messages.+ *+ * Conversational text is rendered verbatim. When `transcriptBudgetTokens` is+ * set and the rendered prompt exceeds it, conversational text is capped at+ * {@link DEGRADED_TEXT_LIMIT} starting from the oldest entries until the+ * prompt fits; the previous checkpoint is never truncated.+ */export function createCompactionPrompt(input: {readonly messages: readonly ModelMessage[];readonly previousCheckpoint: string | undefined;+ readonly transcriptBudgetTokens?: number;}): CompactionPrompt {- const transcript = input.messages.map((message) => ({- content: summarizeCompactionMessageContent(message),+ const entries = input.messages.map((message) => ({+ content: renderCompactionMessageContent(message),role: message.role,}));+ degradeOversizedTranscript(input, entries);+return {prompt: formatCompactionPrompt({previousCheckpoint: input.previousCheckpoint?.trim() ?? "(none)",- transcript: formatCompactionTranscript(transcript),+ transcript: formatCompactionTranscript(entries),}),system: COMPACTION_SYSTEM_PROMPT,};}+/**+ * Re-renders the oldest entries with degraded (capped) conversational text+ * until the estimated prompt fits the budget. Mutates `entries` in place.+ * Savings are tracked per entry instead of re-estimating the whole prompt per+ * iteration; the char-length delta divided by 4 matches the+ * {@link estimateTokens} ruler closely enough for a soft budget.+ */+function degradeOversizedTranscript(+ input: {+ readonly messages: readonly ModelMessage[];+ readonly previousCheckpoint: string | undefined;+ readonly transcriptBudgetTokens?: number;+ },+ entries: { content: string; role: ModelMessage["role"] }[],+): void {+ const budget = input.transcriptBudgetTokens;+ if (budget === undefined) {+ return;+ }++ const fullPrompt = formatCompactionPrompt({+ previousCheckpoint: input.previousCheckpoint?.trim() ?? "(none)",+ transcript: formatCompactionTranscript(entries),+ });+ let excessTokens = estimateTokens(fullPrompt) - budget;++ for (let index = 0; index < entries.length && excessTokens > 0; index += 1) {+ const entry = entries[index];+ const message = input.messages[index];+ if (entry === undefined || message === undefined) {+ continue;+ }++ const degraded = renderCompactionMessageContent(message, DEGRADED_TEXT_LIMIT);+ if (degraded.length >= entry.content.length) {+ continue;+ }++ excessTokens -= (entry.content.length - degraded.length) / 4;+ entries[index] = { content: degraded, role: entry.role };+ }+}+function formatCompactionPrompt(input: {readonly previousCheckpoint: string;readonly transcript: string;@@ -71,7 +141,7 @@ ${input.transcript}${COMPACTION_CHECKPOINT_PROMPT}`;}-function formatCompactionTranscript(messages: readonly CompactionTranscriptMessage[]): string {+function formatCompactionTranscript(messages: readonly CompactionTranscriptEntry[]): string {const sections = messages.filter((message) => message.content.trim().length > 0).map((message) => `### ${message.role}\n${message.content.trim()}`);@@ -79,13 +149,16 @@ function formatCompactionTranscript(messages: readonly CompactionTranscriptMessareturn sections.length === 0 ? "(empty)" : sections.join("\n\n");}-function summarizeCompactionMessageContent(message: ModelMessage): string {+function renderCompactionMessageContent(+ message: ModelMessage,+ conversationTextLimit?: number,+): string {if (typeof message.content === "string") {- return summarizeText(message.content);+ return renderConversationText(message.content, conversationTextLimit);}return message.content- .map((part) => summarizeCompactionContentPart(part))+ .map((part) => renderCompactionContentPart(part, conversationTextLimit)).filter((summary) => summary.length > 0).join("\n").trim();@@ -93,10 +166,13 @@ function summarizeCompactionMessageContent(message: ModelMessage): string {type ModelMessageContentPart = Exclude<ModelMessage["content"], string>[number];-function summarizeCompactionContentPart(part: ModelMessageContentPart): string {+function renderCompactionContentPart(+ part: ModelMessageContentPart,+ conversationTextLimit?: number,+): string {switch (part.type) {case "text":- return summarizeText(part.text);+ return renderConversationText(part.text, conversationTextLimit);case "reasoning":return "";case "file":@@ -104,78 +180,61 @@ function summarizeCompactionContentPart(part: ModelMessageContentPart): string {? `Attached file ${part.filename} (${part.mediaType})`: `Attached file attachment (${part.mediaType})`;case "tool-call":- return summarizeToolCallPart(part);+ return renderTranscriptToolCall(part, conversationTextLimit);case "tool-result":- return summarizeToolResultPart(part);+ return renderTranscriptToolResult(part, conversationTextLimit);default:return "";}}-function summarizeToolCallPart(part: { toolName: string; input?: unknown }): string {- const input = part.input !== undefined ? summarizeCompactValue(part.input) : "";- return input ? `Called ${part.toolName} with ${input}` : `Called ${part.toolName}`;+// Raw tool payloads reach the summarizer clipped, not pre-summarized: the+// checkpoint model decides what matters in a grep result or file read. The+// transcript limit keeps enough content to judge a payload without letting+// one result dominate the prompt; the compact limit applies where a one-line+// rendering is the point — budget-degraded transcript entries.+export const TRANSCRIPT_PAYLOAD_LIMIT = 2_000;+const COMPACT_PAYLOAD_LIMIT = 280;++function renderTranscriptToolCall(+ part: { toolName: string; input?: unknown },+ conversationTextLimit?: number,+): string {+ const limit =+ conversationTextLimit === undefined ? TRANSCRIPT_PAYLOAD_LIMIT : COMPACT_PAYLOAD_LIMIT;+ return renderToolCall(part, limit);}-function summarizeToolResultPart(part: {- toolName: string;- output?: unknown;- isError?: boolean;-}): string {- const output = part.output !== undefined ? summarizeCompactValue(part.output) : "";+function renderTranscriptToolResult(+ part: { toolName: string; output?: unknown; isError?: boolean },+ conversationTextLimit?: number,+): string {+ const limit =+ conversationTextLimit === undefined ? TRANSCRIPT_PAYLOAD_LIMIT : COMPACT_PAYLOAD_LIMIT;const status = part.isError ? "errored" : "returned";+ const output = renderPayload(part.output, limit);return output ? `Tool ${part.toolName} ${status} ${output}` : `Tool ${part.toolName} ${status}`;}-function summarizeCompactValue(value: unknown, depth = 0): string {- if (value === null) return "null";- if (value === undefined) return "";- if (typeof value === "string") return summarizeText(value);- if (typeof value === "number" || typeof value === "boolean" || typeof value === "bigint") {- return String(value);- }-- if (Array.isArray(value)) {- if (value.length === 0) {- return "array(0)";- }-- if (depth >= 2) {- return `array(${value.length})`;- }-- const entries = value- .slice(0, COMPACTION_COLLECTION_LIMIT)- .map((item) => summarizeCompactValue(item, depth + 1));- const suffix = value.length > COMPACTION_COLLECTION_LIMIT ? ", …" : "";- return `array(${value.length}: ${entries.join(", ")}${suffix})`;- }-- if (typeof value === "object") {- const entries = Object.entries(value as Record<string, unknown>);- if (entries.length === 0) {- return "object(0)";- }-- if (depth >= 2) {- return `object(${entries.length} keys)`;- }+function renderToolCall(part: { toolName: string; input?: unknown }, limit: number): string {+ const input = renderPayload(part.input, limit);+ return input ? `Called ${part.toolName} with ${input}` : `Called ${part.toolName}`;+}- const rendered = entries- .slice(0, COMPACTION_COLLECTION_LIMIT)- .map(([key, nested]) => `${key}=${summarizeCompactValue(nested, depth + 1)}`);- const suffix = entries.length > COMPACTION_COLLECTION_LIMIT ? ", …" : "";- return `object(${rendered.join(", ")}${suffix})`;- }+function renderPayload(value: unknown, limit: number): string {+ if (value === undefined) return "";+ return capText(JSON.stringify(value) ?? "", limit);+}- return "";+function renderConversationText(value: string, limit?: number): string {+ return limit === undefined ? value.trim() : capText(value, limit);}-function summarizeText(value: string): string {+function capText(value: string, limit: number): string {const normalized = value.replace(/\s+/g, " ").trim();- if (normalized.length <= COMPACTION_TEXT_LIMIT) {+ if (normalized.length <= limit) {return normalized;}- return `${normalized.slice(0, COMPACTION_TEXT_LIMIT).trimEnd()}…`;+ return `${normalized.slice(0, limit).trimEnd()}…`;}diff --git a/packages/eve/src/harness/compaction.test.ts b/packages/eve/src/harness/compaction.test.tsindex abde90235..0b484cd17 100644--- a/packages/eve/src/harness/compaction.test.ts+++ b/packages/eve/src/harness/compaction.test.ts@@ -4,11 +4,11 @@ import { afterEach, describe, expect, it, vi } from "vitest";import { COMPACTION_PROMPT_ENVELOPE } from "#harness/compaction-prompt.js";import {compactMessages,- estimateTokens,getInputTokenCount,resolveCompactionModel,shouldCompact,} from "#harness/compaction.js";+import { estimateTokens } from "#harness/token-estimate.js";import type { CompactionConfig } from "#harness/types.js";vi.mock("ai", () => ({@@ -252,354 +252,384 @@ describe("resolveCompactionModel", () => {});});-describe("compactMessages", () => {- it("carries a prior checkpoint into the next compaction without truncating it", async () => {- const { generateText } = await import("ai");- const markerAfterTextLimit = "CRITICAL_STATE_AFTER_280_CHARACTERS";- const previousCheckpoint = `${"completed work ".repeat(24)}${markerAfterTextLimit}`;-- vi.mocked(generateText).mockResolvedValue({- text: "Updated checkpoint",- } as Awaited<ReturnType<typeof generateText>>);+// --- compactMessages ---------------------------------------------------+//+// compactMessages escalates through heuristics before summarizing, and which+// strategy a test hits is pure threshold arithmetic:+//+// - The tool-result cap heuristic is accepted only when the capped history+// PLUS the fixed compaction prompt envelope fits the threshold.+// `HEURISTICS_FORBIDDEN` sits below the envelope estimate, so no heuristic+// can be accepted and the summarization fallback always runs.+// - `ROOMY` accepts capping whenever the history's bulk is tool output.+//+// Every result is checked against `expectWellFormedCompaction`: no orphaned+// tool_result, never trailing on an assistant message, and — unless the+// window was exhausted — within the threshold on shouldCompact's own ruler,+// so the result cannot immediately re-trigger compaction.++const ENVELOPE_TOKENS = estimateTokens([+ { content: COMPACTION_PROMPT_ENVELOPE.system, role: "system" },+ { content: COMPACTION_PROMPT_ENVELOPE.prompt, role: "user" },+] satisfies ModelMessage[]);+const HEURISTICS_FORBIDDEN = Math.floor(ENVELOPE_TOKENS);+const ROOMY = 100_000;++const CHECKPOINT_MARKER = "Summary of our conversation so far:";++function user(text: string): ModelMessage {+ return { content: text, role: "user" };+}++function assistant(text: string): ModelMessage {+ return { content: text, role: "assistant" };+}++/** An assistant tool-call message and its paired tool-result message. */+function toolExchange(input: {+ readonly callId: string;+ readonly payloadChars: number;+ readonly prose?: string;+}): [ModelMessage, ModelMessage] {+ return [+ {+ content: [+ ...(input.prose === undefined ? [] : [{ text: input.prose, type: "text" as const }]),+ {+ input: { pattern: "todo" },+ toolCallId: input.callId,+ toolName: "grep",+ type: "tool-call" as const,+ },+ ],+ role: "assistant",+ },+ {+ content: [+ {+ output: { type: "json" as const, value: { content: "x".repeat(input.payloadChars) } },+ toolCallId: input.callId,+ toolName: "grep",+ type: "tool-result" as const,+ },+ ],+ role: "tool",+ },+ ];+}++function checkpointHead(text: string): ModelMessage[] {+ return [user(CHECKPOINT_MARKER), assistant(text)];+}++function expectWellFormedCompaction(result: ModelMessage[], threshold: number): void {+ const seenCallIds = new Set<string>();+ for (const message of result) {+ if (typeof message.content === "string") continue;+ for (const part of message.content) {+ if (part.type === "tool-call") seenCallIds.add(part.toolCallId);+ if (part.type === "tool-result") {+ expect(seenCallIds.has(part.toolCallId), "orphaned tool_result").toBe(true);+ }+ }+ }++ expect(result.at(-1)?.role, "history must not trail on assistant content").not.toBe("assistant");++ if (result.length > 2) {+ expect(estimateTokens(result)).toBeLessThanOrEqual(threshold);+ }+}++async function compact(+ messages: ModelMessage[],+ overrides: Partial<CompactionConfig> & { readonly summary?: string } = {},+): Promise<{ result: ModelMessage[]; summarizer: ReturnType<typeof vi.mocked<never>> }> {+ const { generateText } = await import("ai");+ const summarizer = vi.mocked(generateText);+ summarizer.mockResolvedValue({+ text: overrides.summary ?? "checkpoint text",+ } as Awaited<ReturnType<typeof generateText>>);++ const compactionConfig: CompactionConfig = {+ recentWindowSize: overrides.recentWindowSize ?? 4,+ threshold: overrides.threshold ?? ROOMY,+ };+ const result = await compactMessages(+ messages,+ {} as Parameters<typeof compactMessages>[1],+ compactionConfig,+ );++ expectWellFormedCompaction(result, compactionConfig.threshold);+ return { result, summarizer: summarizer as ReturnType<typeof vi.mocked<never>> };+}++describe("compactMessages: tool-result cap heuristic", () => {+ it("caps oversized older tool results in place without calling the summarizer", async () => {+ const [call, resultMsg] = toolExchange({+ callId: "call-0",+ payloadChars: 4_000,+ prose: "Searching first.",+ });+ const messages = [user("investigate the bug"), call, resultMsg, user("what did you find?")];++ const { result, summarizer } = await compact(messages, { recentWindowSize: 1 });++ expect(summarizer).not.toHaveBeenCalled();+ // Structure survives untouched: the assistant message keeps its prose and+ // its tool-call part, and the result stays a tool message.+ expect(result[0]).toEqual(user("investigate the bug"));+ expect(result[1]).toBe(call);+ const cappedPart = Array.isArray(result[2]?.content) ? result[2].content[0] : undefined;+ expect(cappedPart?.type).toBe("tool-result");+ const output = cappedPart?.type === "tool-result" ? cappedPart.output : undefined;+ const value =+ typeof output === "object" && output !== null && "value" in output+ ? String(output.value)+ : "";+ // Annotation leads; a real content prefix follows; the 4k bulk is gone.+ expect(value).toContain("Truncated by eve");+ expect(value).toContain("xxx");+ expect(value.length).toBeLessThan(2_400);+ expect(result.at(-1)).toEqual(user("what did you find?"));+ // The capped result cannot immediately re-trigger compaction.+ expect(shouldCompact(result, { recentWindowSize: 1, threshold: ROOMY })).toBe(false);+ });- const messages: ModelMessage[] = [- { content: "Summary of our conversation so far:", role: "user" },- { content: previousCheckpoint, role: "assistant" },- { content: "new older context", role: "user" },- { content: "new older response", role: "assistant" },- { content: "recent question", role: "user" },- { content: "recent answer", role: "assistant" },+ it("keeps the recent tail verbatim, tool results included", async () => {+ const [olderCall, olderResult] = toolExchange({ callId: "call-0", payloadChars: 4_000 });+ const [recentCall, recentResult] = toolExchange({ callId: "call-1", payloadChars: 50 });+ const messages = [+ user("older question"),+ olderCall,+ olderResult,+ user("do the thing"),+ recentCall,+ recentResult,];- const model = {} as Parameters<typeof compactMessages>[1];- await compactMessages(messages, model, {- recentWindowSize: 2,- threshold: 100_000,- });+ const { result, summarizer } = await compact(messages, { recentWindowSize: 3 });- const call = vi.mocked(generateText).mock.calls[0]?.[0];- expect(call?.prompt).toContain(previousCheckpoint);- expect(call?.prompt).toContain(markerAfterTextLimit);+ expect(summarizer).not.toHaveBeenCalled();+ // A tool-ending tail gets no "Continue." guard, so the exchange is last.+ expect(result.slice(-2)).toEqual([recentCall, recentResult]);});- it("replaces a prior checkpoint instead of retaining it in the recent window", async () => {- const { generateText } = await import("ai");- const previousCheckpoint = "Previous checkpoint";-- vi.mocked(generateText).mockResolvedValue({- text: "Updated checkpoint",- } as Awaited<ReturnType<typeof generateText>>);-- const messages: ModelMessage[] = [- { content: "Summary of our conversation so far:", role: "user" },- { content: previousCheckpoint, role: "assistant" },- { content: "new evidence", role: "user" },- { content: "latest response", role: "assistant" },+ it("retains the prior checkpoint pair so chaining survives cap-only cycles", async () => {+ const [call, resultMsg] = toolExchange({ callId: "call-0", payloadChars: 4_000 });+ const messages = [+ ...checkpointHead("Previous checkpoint"),+ user("new evidence"),+ call,+ resultMsg,+ user("recent question"),];- const model = {} as Parameters<typeof compactMessages>[1];- const result = await compactMessages(messages, model, {- recentWindowSize: 10,- threshold: 100_000,- });+ const { result, summarizer } = await compact(messages, { recentWindowSize: 1 });- expect(result.filter((message) => message.content === previousCheckpoint)).toHaveLength(0);- expect(result.filter((message) => message.content === "Updated checkpoint")).toHaveLength(1);+ expect(summarizer).not.toHaveBeenCalled();+ expect(result[0]).toEqual(user(CHECKPOINT_MARKER));+ expect(result[1]).toEqual(assistant("Previous checkpoint"));});- it("summarizes older messages and keeps recent window", async () => {- const { generateText } = await import("ai");+ it("keeps a tool result paired with its call when the window would split them", async () => {+ const [call, resultMsg] = toolExchange({ callId: "call-1", payloadChars: 40 });+ // recentWindowSize 2 splits between call and result; the snap pulls the+ // result into the older region, and capping leaves both untouched.+ const messages = [user("old context"), call, resultMsg, user("next question")];- vi.mocked(generateText).mockResolvedValue({- text: "Summary of prior context",- } as Awaited<ReturnType<typeof generateText>>);+ const { result, summarizer } = await compact(messages, { recentWindowSize: 2 });- const messages: ModelMessage[] = [- { content: "old message 1", role: "user" },- { content: "old message 2", role: "assistant" },- { content: "recent 1", role: "user" },- { content: "recent 2", role: "assistant" },- ];+ expect(summarizer).not.toHaveBeenCalled();+ expect(result).toContainEqual(call);+ expect(result).toContainEqual(resultMsg);+ expect(result.indexOf(call)).toBeLessThan(result.indexOf(resultMsg));+ });+});- const model = {} as Parameters<typeof compactMessages>[1];- const result = await compactMessages(messages, model, config);+describe("compactMessages: summarization fallback", () => {+ it("summarizes when capping cannot free enough space", async () => {+ // All bulk is conversational prose — capping removes nothing — and the+ // threshold sits below the prompt envelope, so no heuristic can be+ // accepted regardless.+ const messages = [user("old context to fold away"), assistant("old reply"), user("continue")];- expect(result).toHaveLength(5);- expect(result[0]).toEqual({- content: "Summary of our conversation so far:",- role: "user",- });- expect(result[1]).toEqual({- content: "Summary of prior context",- role: "assistant",+ const { result, summarizer } = await compact(messages, {+ recentWindowSize: 1,+ summary: "Distilled story",+ threshold: HEURISTICS_FORBIDDEN,});- expect(result[2]).toEqual({ content: "recent 1", role: "user" });- expect(result[3]).toEqual({ content: "recent 2", role: "assistant" });- expect(result[4]).toEqual({ content: "Continue.", role: "user" });- });-- it("forwards provider options to the compaction model call", async () => {- const { generateText } = await import("ai");- vi.mocked(generateText).mockResolvedValue({- text: "Summary of prior context",- } as Awaited<ReturnType<typeof generateText>>);+ expect(summarizer).toHaveBeenCalledTimes(1);+ expect(summarizer.mock.calls[0]?.[0]?.prompt).toContain("old context to fold away");+ expect(result[0]).toEqual(user(CHECKPOINT_MARKER));+ expect(result[1]).toEqual(assistant("Distilled story"));+ expect(result.some((m) => m.content === "old context to fold away")).toBe(false);+ });- const messages: ModelMessage[] = [- { content: "old message 1", role: "user" },- { content: "old message 2", role: "assistant" },- { content: "recent 1", role: "user" },- { content: "recent 2", role: "assistant" },+ it("feeds the previous checkpoint to the summarizer untruncated and replaces it", async () => {+ const markerPast280 = "CRITICAL_STATE_AFTER_280_CHARACTERS";+ const previousCheckpoint = `${"completed work ".repeat(24)}${markerPast280}`;+ const messages = [+ ...checkpointHead(previousCheckpoint),+ user("new evidence"),+ assistant("latest response"),];- const model = {} as Parameters<typeof compactMessages>[1];- const providerOptions = {- anthropic: {- thinking: {- budget_tokens: 128,- },- },- };-- await compactMessages(messages, model, config, providerOptions);+ const { result, summarizer } = await compact(messages, {+ summary: "Updated checkpoint",+ threshold: HEURISTICS_FORBIDDEN,+ });- expect(generateText).toHaveBeenCalledWith(- expect.objectContaining({- model,- providerOptions,- }),- );+ expect(summarizer.mock.calls[0]?.[0]?.prompt).toContain(previousCheckpoint);+ expect(summarizer.mock.calls[0]?.[0]?.prompt).toContain(markerPast280);+ expect(result.filter((m) => m.content === previousCheckpoint)).toHaveLength(0);+ expect(result.filter((m) => m.content === "Updated checkpoint")).toHaveLength(1);});- it("forwards the turn abort signal to the compaction model call", async () => {- const { generateText } = await import("ai");-- vi.mocked(generateText).mockResolvedValue({- text: "Summary of prior context",- } as Awaited<ReturnType<typeof generateText>>);+ it("keeps the recent window verbatim after summarizing when it fits", async () => {+ // Prose bulk forces summarization; the threshold has room for the tail.+ const oldProse = user("investigation notes ".repeat(2_000));+ const [recentCall, recentResult] = toolExchange({ callId: "call-1", payloadChars: 100 });+ const messages = [oldProse, assistant("done reading"), recentCall, recentResult];- const messages: ModelMessage[] = [- { content: "old message 1", role: "user" },- { content: "old message 2", role: "assistant" },- { content: "recent 1", role: "user" },- { content: "recent 2", role: "assistant" },- ];+ const { result, summarizer } = await compact(messages, {+ recentWindowSize: 2,+ threshold: 5_000,+ });- const model = {} as Parameters<typeof compactMessages>[1];- const abortController = new AbortController();+ expect(summarizer).toHaveBeenCalledTimes(1);+ expect(result.slice(2, 4)).toEqual([recentCall, recentResult]);+ });- await compactMessages(- messages,- model,- config,- undefined,- undefined,- undefined,- abortController.signal,- );+ it("strips tool activity from the tail when verbatim does not fit but text does", async () => {+ const oldProse = user("investigation notes ".repeat(2_000));+ const [recentCall, recentResult] = toolExchange({+ callId: "call-1",+ payloadChars: 1_400,+ prose: "Running the tool.",+ });+ const tail = [user("do the thing"), recentCall, recentResult];+ const messages = [oldProse, ...tail];++ // Derive a threshold between the stripped and verbatim tail sizes so the+ // regime is explicit rather than encoded in magic numbers. The summary is+ // sized to exceed the window-selection reserve, which is what makes the+ // verbatim tail overshoot after the summary head is added.+ const summary = "s".repeat(2_400);+ const summaryHead = [user(CHECKPOINT_MARKER), assistant(summary)];+ const verbatimSize = estimateTokens([...summaryHead, ...tail]);+ const strippedSize = estimateTokens([+ ...summaryHead,+ user("do the thing"),+ assistant("Running the tool."),+ ]);+ const threshold = Math.floor((verbatimSize + strippedSize) / 2);+ expect(strippedSize).toBeLessThan(threshold);+ expect(verbatimSize).toBeGreaterThan(threshold);++ const { result, summarizer } = await compact(messages, {+ recentWindowSize: 3,+ summary,+ threshold,+ });- expect(generateText).toHaveBeenCalledWith(- expect.objectContaining({- abortSignal: abortController.signal,- model,- }),- );+ expect(summarizer).toHaveBeenCalledTimes(1);+ expect(result).toContainEqual(user("do the thing"));+ expect(result).toContainEqual(assistant("Running the tool."));+ expect(result.some((m) => m.role === "tool")).toBe(false);});- it("folds oversized recent tool results into the summary when the raw tail does not fit", async () => {- const { generateText } = await import("ai");-- vi.mocked(generateText).mockResolvedValue({- text: "Summary of the large SQL result",- } as Awaited<ReturnType<typeof generateText>>);+ it("folds everything into the summary when even the stripped tail cannot fit", async () => {+ const [call, resultMsg] = toolExchange({ callId: "call-1", payloadChars: 2_000 });+ const messages = [user("Find the relevant rows."), call, resultMsg];- const messages: ModelMessage[] = [- { content: "Find the relevant rows.", role: "user" },- {- content: [- {- input: { sql: "select * from events" },- toolCallId: "call-1",- toolName: "execute_sql",- type: "tool-call",- },- ],- role: "assistant",- },- {- content: [- {- output: {- type: "json",- value: {- rows: Array.from({ length: 50 }, (_, index) => ({- id: index,- payload: "x".repeat(200),- })),- },- },- toolCallId: "call-1",- toolName: "execute_sql",- type: "tool-result",- },- ],- role: "tool",- },- ];-- const model = {} as Parameters<typeof compactMessages>[1];- const result = await compactMessages(messages, model, {+ const { result } = await compact(messages, {recentWindowSize: 10,- threshold: 150,+ summary: "Summary of the large SQL result",+ threshold: HEURISTICS_FORBIDDEN,});+ // The folded-away user prompt is replayed as the live turn, so the model+ // resumes against its actual instruction rather than a bare "Continue.".expect(result).toEqual([- {- content: "Summary of our conversation so far:",- role: "user",- },- {- content: "Summary of the large SQL result",- role: "assistant",- },- { content: "Continue.", role: "user" },+ user(CHECKPOINT_MARKER),+ assistant("Summary of the large SQL result"),+ user("Find the relevant rows."),]);- expect(vi.mocked(generateText)).toHaveBeenCalledTimes(1);});- it("drops tool results and strips assistant tool calls from the kept tail", async () => {- const { generateText } = await import("ai");-- vi.mocked(generateText).mockResolvedValue({- text: "summary",- } as Awaited<ReturnType<typeof generateText>>);-- const messages: ModelMessage[] = [- { content: "old context to summarize", role: "user" },- { content: "older reply", role: "assistant" },- // Recent window (last 4):- { content: "do the thing", role: "user" },- {- content: [- { text: "Running the tool.", type: "text" },- { input: { x: 1 }, toolCallId: "call-1", toolName: "run", type: "tool-call" },- ],- role: "assistant",- },- {- content: [- {- output: { type: "json", value: { ok: true } },- toolCallId: "call-1",- toolName: "run",- type: "tool-result",- },- ],- role: "tool",- },- { content: "All done.", role: "assistant" },+ it("replays the folded-away user prompt when the tail would trail on assistant content", async () => {+ const messages = [+ user("please fix the flaky test"),+ assistant("working on it"),+ assistant("still going"),];- const model = {} as Parameters<typeof compactMessages>[1];- const result = await compactMessages(messages, model, {- recentWindowSize: 4,- threshold: 100_000,+ const { result } = await compact(messages, {+ recentWindowSize: 1,+ threshold: HEURISTICS_FORBIDDEN,});- // No tool-result message and no tool_use/tool_result part survives anywhere,- // so the rebuilt history has no dangling tool calls.- for (const message of result) {- expect(message.role).not.toBe("tool");- if (Array.isArray(message.content)) {- for (const part of message.content) {- expect(part.type).not.toBe("tool-call");- expect(part.type).not.toBe("tool-result");- }- }- }-- // User messages are kept; the assistant tool-call turn is reduced to its text.- expect(result).toEqual([- { content: "Summary of our conversation so far:", role: "user" },- { content: "summary", role: "assistant" },- { content: "do the thing", role: "user" },- { content: "Running the tool.", role: "assistant" },- { content: "All done.", role: "assistant" },- { content: "Continue.", role: "user" },- ]);+ expect(result.at(-1)).toEqual(user("please fix the flaky test"));});- it("forwards headers to the generateText call", async () => {- const { generateText } = await import("ai");-- vi.mocked(generateText).mockResolvedValue({- text: "Summary of prior context",- } as Awaited<ReturnType<typeof generateText>>);-- const messages: ModelMessage[] = [- { content: "old message 1", role: "user" },- { content: "old message 2", role: "assistant" },- { content: "recent 1", role: "user" },- { content: "recent 2", role: "assistant" },- ];-- const model = {} as Parameters<typeof compactMessages>[1];- const headers = {- "x-title": "My Agent",- "http-referer": "https://my-agent.vercel.app",- };+ it("falls back to a synthetic resumption when the real user prompt survives in the tail", async () => {+ const messages = [user("old context"), user("latest question"), assistant("answering")];- await compactMessages(messages, model, config, undefined, undefined, headers);+ const { result } = await compact(messages, {+ recentWindowSize: 2,+ threshold: HEURISTICS_FORBIDDEN,+ });- expect(generateText).toHaveBeenCalledWith(- expect.objectContaining({- headers,- model,- }),- );+ // "latest question" is already in the kept tail — replaying it would ask+ // the model to answer again instead of continuing.+ expect(result.at(-1)).toEqual(user("Continue."));+ expect(result.filter((m) => m.content === "latest question")).toHaveLength(1);});- it("appends synthetic user message when recent window trails with assistant", async () => {- const { generateText } = await import("ai");-- vi.mocked(generateText).mockResolvedValue({- text: "Summary of prior context",- } as Awaited<ReturnType<typeof generateText>>);-- const messages: ModelMessage[] = [- { content: "old message", role: "user" },- { content: "old reply", role: "assistant" },- { content: "assistant trailing", role: "assistant" },- ];+ it("does not append any resumption when the tail already ends on a user turn", async () => {+ const messages = [user("old"), assistant("old reply"), user("latest question")];- const model = {} as Parameters<typeof compactMessages>[1];- const result = await compactMessages(messages, model, {+ const { result } = await compact(messages, {recentWindowSize: 1,- threshold: 100,+ threshold: HEURISTICS_FORBIDDEN,});- expect(result.at(-1)).toEqual({ content: "Continue.", role: "user" });+ expect(result.at(-1)).toEqual(user("latest question"));+ expect(result.filter((m) => m.content === "Continue.")).toHaveLength(0);});- it("does not append synthetic user message when recent window ends with user or tool", async () => {+ it("forwards model options to the summarization call", async () => {const { generateText } = await import("ai");-vi.mocked(generateText).mockResolvedValue({- text: "Summary",+ text: "summary",} as Awaited<ReturnType<typeof generateText>>);- const messages: ModelMessage[] = [- { content: "old", role: "user" },- { content: "old reply", role: "assistant" },- { content: "latest question", role: "user" },- ];-+ const messages = [user("old message"), assistant("old reply"), user("continue")];const model = {} as Parameters<typeof compactMessages>[1];- const result = await compactMessages(messages, model, {- recentWindowSize: 1,- threshold: 100,- });+ const providerOptions = { anthropic: { thinking: { budget_tokens: 128 } } };+ const headers = { "x-title": "My Agent" };+ const abortController = new AbortController();++ await compactMessages(+ messages,+ model,+ { recentWindowSize: 1, threshold: HEURISTICS_FORBIDDEN },+ providerOptions,+ undefined,+ headers,+ abortController.signal,+ );- expect(result.at(-1)).toEqual({ content: "latest question", role: "user" });+ expect(generateText).toHaveBeenCalledWith(+ expect.objectContaining({+ abortSignal: abortController.signal,+ headers,+ model,+ providerOptions,+ }),+ );});});diff --git a/packages/eve/src/harness/compaction.ts b/packages/eve/src/harness/compaction.tsindex d37044a81..2d0d6e752 100644--- a/packages/eve/src/harness/compaction.ts+++ b/packages/eve/src/harness/compaction.ts@@ -3,8 +3,12 @@ import { generateText, type LanguageModel, type ModelMessage, type TelemetryOptiimport {COMPACTION_CHECKPOINT_MARKER,COMPACTION_PROMPT_ENVELOPE,+ COMPACTION_RESUMPTION_MESSAGE,createCompactionPrompt,+ TODO_COMPACTION_PRESERVATION_LABEL,+ TRANSCRIPT_PAYLOAD_LIMIT,} from "#harness/compaction-prompt.js";+import { estimateTokens } from "#harness/token-estimate.js";import type { RuntimeModelReference } from "#runtime/agent/bootstrap.js";import type { CompactionConfig, ToolLoopHarnessConfig } from "#harness/types.js";@@ -15,18 +19,9 @@ const COMPACTION_SUMMARY_RESERVE_TOKENS = 2_048;*/type ModelMessageContentPart = Exclude<ModelMessage["content"], string>[number];-/**- * Rough token estimate: serialized JSON length / 4. Good enough for- * deciding whether compaction is needed; the real token count comes back- * from the model each step via {@link CompactionConfig.lastKnownInputTokens}.- *- * Accepts any JSON-serializable value so callers can apply the same heuristic- * to whole message arrays or individual content parts on one consistent ruler.- */-export function estimateTokens(value: unknown): number {- return JSON.stringify(value).length / 4;-}-+// Static envelope estimate stays valid because createCompactionPrompt bounds+// its transcript to the caller's threshold budget, so the summarization call+// itself never grows past threshold + envelope + checkpoint.const COMPACTION_PROMPT_OVERHEAD_TOKENS = estimateTokens([{ content: COMPACTION_PROMPT_ENVELOPE.system, role: "system" },{ content: COMPACTION_PROMPT_ENVELOPE.prompt, role: "user" },@@ -98,9 +93,87 @@ export async function resolveCompactionModel(input: {};}+/** Conversation regions and config handed to each compaction heuristic. */+interface CompactionHeuristicInput {+ readonly config: CompactionConfig;+ readonly conversation: readonly ModelMessage[];+ readonly older: readonly ModelMessage[];+ readonly previousCheckpoint: string | undefined;+ readonly recent: readonly ModelMessage[];+}++/**+ * A heuristic either produces a complete replacement history that fits the+ * threshold, or signals that the next strategy must run. Heuristics have no+ * failure channel — anything exceptional throws.+ */+type CompactionHeuristicOutcome =+ | { readonly messages: ModelMessage[]; readonly type: "within-limit" }+ | { readonly type: "insufficient" };++type CompactionHeuristic = (input: CompactionHeuristicInput) => CompactionHeuristicOutcome;++/**+ * Model-free heuristics tried in order before falling back to LLM+ * summarization. Each is a pure transformation of the split conversation;+ * composing a new strategy (pruning variants, protect-lists) means adding an+ * entry here.+ */+const COMPACTION_HEURISTICS: readonly CompactionHeuristic[] = [toolResultCapHeuristic];++/**+ * Caps oversized tool-result outputs in the older region while every message+ * — including its tool calls and the recent tail — stays structurally+ * verbatim. Most history bulk is a handful of large tool outputs, so this+ * usually suffices with no model call and no rewriting.+ */+function toolResultCapHeuristic(input: CompactionHeuristicInput): CompactionHeuristicOutcome {+ const checkpointHead: ModelMessage[] =+ input.previousCheckpoint === undefined+ ? []+ : [+ { content: COMPACTION_CHECKPOINT_MARKER, role: "user" },+ { content: input.previousCheckpoint, role: "assistant" },+ ];+ const capped = withResumptionGuard(+ [...checkpointHead, ...capToolResults(input.older), ...input.recent],+ input.conversation,+ );++ // Evaluate on the same ruler shouldCompact uses (envelope included):+ // capping can be a near no-op when the older region holds few large+ // results, and accepting one on a looser ruler would let shouldCompact+ // re-fire every step without compaction ever making progress.+ const evaluation = evaluateThreshold(capped, input.config, "should-compact");+ return evaluation.type === "within-limit"+ ? { messages: capped, type: "within-limit" }+ : { type: "insufficient" };+}++/**+ * Measures a candidate history against the compaction threshold.+ * The "should-compact" ruler includes the fixed prompt-envelope overhead that+ * {@link shouldCompact} adds, so a history accepted on it cannot immediately+ * re-trigger compaction; "estimate" is the bare history size.+ */+function evaluateThreshold(+ messages: readonly ModelMessage[],+ config: CompactionConfig,+ ruler: "estimate" | "should-compact",+): { readonly estimatedTokens: number; readonly type: "over-limit" | "within-limit" } {+ const overhead = ruler === "should-compact" ? COMPACTION_PROMPT_OVERHEAD_TOKENS : 0;+ const estimatedTokens = estimateTokens(messages) + overhead;+ return {+ estimatedTokens,+ type: estimatedTokens <= config.threshold ? "within-limit" : "over-limit",+ };+}+/**- * Compacts messages by summarizing older history and keeping only the most- * recent messages.+ * Compacts messages by escalation: try each {@link CompactionHeuristic} in+ * order, then fall back to summarizing the older region with the compaction+ * model — keeping the recent tail verbatim when it fits, degrading it to+ * text-only, then shrinking the window.*/export async function compactMessages(messages: ModelMessage[],@@ -114,13 +187,28 @@ export async function compactMessages(const { conversation, previousCheckpoint } = extractPreviousCheckpoint(messages);let keep = selectRecentWindowSize(conversation, config);- while (true) {+ {const { older, recent } = splitMessagesForCompaction(conversation, keep);if (older.length === 0 && previousCheckpoint === undefined) {return keepNonToolResultMessages(recent);}- const summaryPrompt = createCompactionPrompt({ messages: older, previousCheckpoint });+ for (const heuristic of COMPACTION_HEURISTICS) {+ const outcome = heuristic({ config, conversation, older, previousCheckpoint, recent });+ if (outcome.type === "within-limit") {+ return outcome.messages;+ }+ }+ }++ while (true) {+ const { older, recent } = splitMessagesForCompaction(conversation, keep);++ const summaryPrompt = createCompactionPrompt({+ messages: older,+ previousCheckpoint,+ transcriptBudgetTokens: config.threshold,+ });const result = await generateText({abortSignal,@@ -133,37 +221,127 @@ export async function compactMessages(temperature: 0,});- // Keep recent context as plain conversation: tool results are dropped (the- // summary above already captures the older ones) and assistant tool calls- // are stripped, so no tool_use survives without its result. The summarized- // older region is the durable record of tool activity.- const keptTail = keepNonToolResultMessages(recent);-- // The kept tail may be empty or trail with an assistant message; the summary- // assistant message also precedes it. Providers that don't support assistant- // prefill reject a request that ends on assistant content, so append a- // synthetic user message to resume from a user turn.- const lastKeptRole = keptTail.at(-1)?.role;- const trailingAssistantGuard: ModelMessage[] =- lastKeptRole === undefined || lastKeptRole === "assistant"- ? [{ role: "user", content: "Continue." }]- : [];-- const compacted: ModelMessage[] = [+ const summaryHead: ModelMessage[] = [{ content: COMPACTION_CHECKPOINT_MARKER, role: "user" },{ content: result.text, role: "assistant" },- ...keptTail,- ...trailingAssistantGuard,];- if (estimateTokens(compacted) <= config.threshold || keep === 0) {- return compacted;+ // Prefer keeping the recent tail verbatim — surviving tool results are the+ // model's evidence that work already ran. Degrade to text-only, then to a+ // smaller window, only under threshold pressure.+ const verbatim = withResumptionGuard([...summaryHead, ...recent], conversation);+ if (evaluateThreshold(verbatim, config, "estimate").type === "within-limit") {+ return verbatim;+ }++ const stripped = withResumptionGuard(+ [...summaryHead, ...keepNonToolResultMessages(recent)],+ conversation,+ );+ if (evaluateThreshold(stripped, config, "estimate").type === "within-limit" || keep === 0) {+ return stripped;}keep -= 1;}}+const CAPPED_RESULT_ANNOTATION =+ "[Truncated by eve: tool result reduced during context compaction. Re-run the tool if you need the full output.]";++/**+ * Caps oversized tool-result outputs in place, keeping message structure and+ * tool_use/tool_result pairing untouched. The cap matches the transcript+ * payload clip, so a later summarization sees the same content the model+ * kept — capping never destroys material the summarizer would need, unlike+ * dropping results outright. The annotation leads the value so it survives+ * prefix-capped renderings.+ */+function capToolResults(messages: readonly ModelMessage[]): ModelMessage[] {+ return messages.map((message) => {+ if (message.role !== "tool" || typeof message.content === "string") {+ return message;+ }++ let changed = false;+ const content = message.content.map((part) => {+ if (part.type !== "tool-result") {+ return part;+ }+ const serialized = JSON.stringify(part.output) ?? "";+ if (serialized.length <= TRANSCRIPT_PAYLOAD_LIMIT) {+ return part;+ }++ changed = true;+ return {+ ...part,+ output: {+ type: "text" as const,+ value: `${CAPPED_RESULT_ANNOTATION}\n\n${serialized.slice(0, TRANSCRIPT_PAYLOAD_LIMIT)}`,+ },+ };+ });++ return changed ? { ...message, content } : message;+ });+}++/**+ * Providers that don't support assistant prefill reject a request that ends on+ * assistant content, so compaction must resume from a user turn. Rather than+ * a contentless synthetic prompt, replay the conversation's last real user+ * message when compaction folded it away — the model resumes against its+ * actual instruction, with the checkpoint as background. Falls back to+ * "Continue." when the last real user message still survives in the kept+ * messages (the model was mid-work on it) or none exists.+ */+function withResumptionGuard(+ messages: ModelMessage[],+ conversation: readonly ModelMessage[],+): ModelMessage[] {+ const lastRole = messages.at(-1)?.role;+ if (lastRole !== undefined && lastRole !== "assistant") {+ return messages;+ }++ const replay = findLastRealUserMessage(conversation);+ const alreadyKept =+ replay !== undefined &&+ messages.some((message) => message.role === "user" && message.content === replay.content);++ return [+ ...messages,+ replay !== undefined && !alreadyKept+ ? replay+ : { content: COMPACTION_RESUMPTION_MESSAGE, role: "user" },+ ];+}++/**+ * Latest user message authored by the user rather than synthesized by the+ * framework (resumption prompts, checkpoint markers, and todo preservation+ * messages are all `role: "user"` but carry no user intent).+ */+function findLastRealUserMessage(conversation: readonly ModelMessage[]): ModelMessage | undefined {+ for (let index = conversation.length - 1; index >= 0; index -= 1) {+ const message = conversation[index];+ if (message?.role !== "user" || typeof message.content !== "string") {+ continue;+ }+ if (+ message.content === COMPACTION_RESUMPTION_MESSAGE ||+ message.content === COMPACTION_CHECKPOINT_MARKER ||+ message.content.startsWith(TODO_COMPACTION_PRESERVATION_LABEL)+ ) {+ continue;+ }+ return message;+ }++ return undefined;+}+function extractPreviousCheckpoint(messages: readonly ModelMessage[]): {readonly conversation: ModelMessage[];readonly previousCheckpoint: string | undefined;@@ -280,8 +458,16 @@ function splitMessagesForCompaction(};}+ // The recent tail survives verbatim, so it must not open with tool results+ // whose tool calls fall in the older region — providers reject a tool_result+ // without its preceding tool_use. Snap such messages into the older region.+ let split = messages.length - keep;+ while (split < messages.length && messages[split]?.role === "tool") {+ split += 1;+ }+return {- older: messages.slice(0, -keep),- recent: messages.slice(-keep),+ older: messages.slice(0, split),+ recent: messages.slice(split),};}diff --git a/packages/eve/src/harness/token-estimate.ts b/packages/eve/src/harness/token-estimate.tsnew file mode 100644index 000000000..94ab84b21--- /dev/null+++ b/packages/eve/src/harness/token-estimate.ts@@ -0,0 +1,11 @@+/**+ * Rough token estimate: serialized JSON length / 4. Good enough for+ * deciding whether compaction is needed; the real token count comes back+ * from the model each step via `CompactionConfig.lastKnownInputTokens`.+ *+ * Accepts any JSON-serializable value so callers can apply the same heuristic+ * to whole message arrays or individual content parts on one consistent ruler.+ */+export function estimateTokens(value: unknown): number {+ return JSON.stringify(value).length / 4;+}diff --git a/packages/eve/src/runtime/framework-tools/todo.ts b/packages/eve/src/runtime/framework-tools/todo.tsindex b1acf6550..056e1b218 100644--- a/packages/eve/src/runtime/framework-tools/todo.ts+++ b/packages/eve/src/runtime/framework-tools/todo.ts@@ -3,6 +3,7 @@ import { z } from "#compiled/zod/index.js";import { loadContext } from "#context/container.js";import { ContextKey } from "#context/key.js";+import { TODO_COMPACTION_PRESERVATION_LABEL } from "#harness/compaction-prompt.js";import type { ResolvedToolDefinition } from "#runtime/types.js";// ---------------------------------------------------------------------------@@ -39,7 +40,7 @@ function formatTodoSummary(state: TodoState): string | undefined {return `- [${check}] [${item.priority}] ${item.content}`;});- return `[Your task list was preserved across context compaction]\n${lines.join("\n")}`;+ return `${TODO_COMPACTION_PRESERVATION_LABEL}\n${lines.join("\n")}`;}/**diff --git a/pnpm-lock.yaml b/pnpm-lock.yamlindex e5f050c41..534060907 100644--- a/pnpm-lock.yaml+++ b/pnpm-lock.yaml@@ -655,67 +655,8 @@ importers:specifier: 'catalog:'version: 7.0.2- e2e/fixtures/agent-compaction-regressions-gpt-5-6:+ e2e/fixtures/agent-compaction-regressions:dependencies:- '@eve/e2e-compaction-regression-shared':- specifier: workspace:*- version: link:../compaction-regression-shared- '@opentelemetry/core':- specifier: 2.6.1- version: 2.6.1(@opentelemetry/[email protected])- '@opentelemetry/sdk-trace-base':- specifier: 2.6.1- version: 2.6.1(@opentelemetry/[email protected])- '@vercel/otel':- specifier: 2.1.2- version: 2.1.2(@opentelemetry/[email protected])(@opentelemetry/[email protected])(@opentelemetry/[email protected](@opentelemetry/[email protected])([email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))- eve:- specifier: workspace:*- version: link:../../../packages/eve- zod:- specifier: 'catalog:'- version: 4.4.3- devDependencies:- '@types/node':- specifier: 'catalog:'- version: 25.9.1- typescript:- specifier: 'catalog:'- version: 7.0.2-- e2e/fixtures/agent-compaction-regressions-opus-4-8:- dependencies:- '@eve/e2e-compaction-regression-shared':- specifier: workspace:*- version: link:../compaction-regression-shared- '@opentelemetry/core':- specifier: 2.6.1- version: 2.6.1(@opentelemetry/[email protected])- '@opentelemetry/sdk-trace-base':- specifier: 2.6.1- version: 2.6.1(@opentelemetry/[email protected])- '@vercel/otel':- specifier: 2.1.2- version: 2.1.2(@opentelemetry/[email protected])(@opentelemetry/[email protected])(@opentelemetry/[email protected](@opentelemetry/[email protected])([email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))(@opentelemetry/[email protected](@opentelemetry/[email protected]))- eve:- specifier: workspace:*- version: link:../../../packages/eve- zod:- specifier: 'catalog:'- version: 4.4.3- devDependencies:- '@types/node':- specifier: 'catalog:'- version: 25.9.1- typescript:- specifier: 'catalog:'- version: 7.0.2-- e2e/fixtures/agent-compaction-regressions-sonnet-5:- dependencies:- '@eve/e2e-compaction-regression-shared':- specifier: workspace:*- version: link:../compaction-regression-shared'@opentelemetry/core':specifier: 2.6.1version: 2.6.1(@opentelemetry/[email protected])@@ -1037,15 +978,6 @@ importers:specifier: 'catalog:'version: 7.0.2- e2e/fixtures/compaction-regression-shared:- dependencies:- eve:- specifier: workspace:*- version: link:../../../packages/eve- zod:- specifier: 'catalog:'- version: 4.4.3-e2e/fixtures/dist-extensions:dependencies:eve: