Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
fix(mcp-host): measure rehydrated history like the context manager (#739)
Alfredo Lopez28bad7276b37
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.jsonindex b131b02f9..0a6553daf 100644--- a/mcp-host/package-lock.json+++ b/mcp-host/package-lock.json@@ -1,12 +1,12 @@{"name": "clerum-mcp-host",- "version": "0.1.413",+ "version": "0.1.414","lockfileVersion": 3,"requires": true,"packages": {"": {"name": "clerum-mcp-host",- "version": "0.1.413",+ "version": "0.1.414","license": "SEE LICENSE IN LICENSE","dependencies": {"@anthropic-ai/sdk": "^0.32.0",diff --git a/mcp-host/package.json b/mcp-host/package.jsonindex 2440156cd..c84fa6ae0 100644--- a/mcp-host/package.json+++ b/mcp-host/package.json@@ -1,6 +1,6 @@{"name": "clerum-mcp-host",- "version": "0.1.413",+ "version": "0.1.414","description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude","main": "dist/main.js","engines": {diff --git a/mcp-host/src/agent/taskExecutor.ts b/mcp-host/src/agent/taskExecutor.tsindex d9ab0fe70..fb75ab158 100644--- a/mcp-host/src/agent/taskExecutor.ts+++ b/mcp-host/src/agent/taskExecutor.ts@@ -23,6 +23,7 @@ import type { ApprovalConfig } from '../core/extensions/approvalTypes'import {COMPACTION_PRESSURE_THRESHOLD,PressureContextManager,+ tierDecisionTokens,} from '../core/extensions/contextManager'import {UnifiedApprovalGateController,@@ -1045,13 +1046,15 @@ export class TaskExecutor {const historyStart = Date.now()let messages = this.deps.conversationManager.buildMessageHistory(this.conversation!)// #731: the history is compacted at the same share of the model's window- // at which the context manager starts compacting, not at a fixed count.+ // at which the context manager starts compacting, not at a fixed count,+ // and measured the way the manager measures it for that decision (#739).const compactionThreshold = Math.floor(COMPACTION_PRESSURE_THRESHOLD * this.contextMaxTokens())- messages = compactConversation(+ const tokenCounter = this.getOrCreateTokenCounter()+ messages = await compactConversation(messages,undefined,compactionThreshold,- this.getOrCreateTokenCounter(),+ msgs => tierDecisionTokens(msgs, [], tokenCounter, appConfig.tokenizerDryrun),{enabled: appConfig.compactionPrePruneEnabled,options: {diff --git a/mcp-host/src/core/conversation/__tests__/compaction.test.ts b/mcp-host/src/core/conversation/__tests__/compaction.test.tsindex 5a8ac51b4..d34f8cde4 100644--- a/mcp-host/src/core/conversation/__tests__/compaction.test.ts+++ b/mcp-host/src/core/conversation/__tests__/compaction.test.ts@@ -30,7 +30,7 @@ describe('estimateTokens', () => {})describe('compactConversation', () => {- it('should keep system message + last N turns (Risk 5.5)', () => {+ it('should keep system message + last N turns (Risk 5.5)', async () => {const messages: ChatMessage[] = [{ role: 'system', content: 'You are an assistant.' },{ role: 'user', content: 'Turn 1' },@@ -44,7 +44,7 @@ describe('compactConversation', () => {]// Force compaction by setting threshold to 0- const compacted = compactConversation(messages, 2, 0)+ const compacted = await compactConversation(messages, 2, 0)// System message preservedexpect(compacted[0]).toEqual({@@ -60,7 +60,7 @@ describe('compactConversation', () => {expect(compacted[4].content).toBe('Response 4')})- it('should not compact when under threshold', () => {+ it('should not compact when under threshold', async () => {const messages: ChatMessage[] = [{ role: 'system', content: 'System' },{ role: 'user', content: 'Hello' },@@ -68,11 +68,11 @@ describe('compactConversation', () => {]// Default threshold is 80000 — these messages are tiny- const result = compactConversation(messages)+ const result = await compactConversation(messages)expect(result).toEqual(messages)})- it("T1.3: never archives the last user message ('Active Task' anchor)", () => {+ it("T1.3: never archives the last user message ('Active Task' anchor)", async () => {const messages: ChatMessage[] = [{ role: 'system', content: 'System' },{ role: 'user', content: 'turn 1' },@@ -83,13 +83,13 @@ describe('compactConversation', () => {{ role: 'assistant', content: 'working' },]- const compacted = compactConversation(messages, 1, 0)+ const compacted = await compactConversation(messages, 1, 0)// Even with maxTurns=1 + threshold=0 (force aggressive compaction), the// anchor pulls the cut back so the active user task survives.expect(compacted.some(m => m.content === 'ACTIVE TASK')).toBe(true)})- it('T1.3: kept set never violates tool linkages after compaction', () => {+ it('T1.3: kept set never violates tool linkages after compaction', async () => {const messages: ChatMessage[] = [{ role: 'system', content: 'System' },{ role: 'user', content: 'old' },@@ -104,7 +104,7 @@ describe('compactConversation', () => {{ role: 'assistant', content: 'done' },]- const compacted = compactConversation(messages, 1, 0)+ const compacted = await compactConversation(messages, 1, 0)expect(() => validateToolLinkages(compacted)).not.toThrow()})})@@ -134,16 +134,16 @@ describe('compactConversation — pre-prune is gated at the threshold (#731)', (}const prePruneOn = { enabled: true }- it('T-R2-4a returns the input untouched below the threshold even with pre-prune on', () => {+ it('T-R2-4a returns the input untouched below the threshold even with pre-prune on', async () => {const messages = prunableHistory()- expect(compactConversation(messages, undefined, 1_000_000, undefined, prePruneOn)).toBe(+ expect(await compactConversation(messages, undefined, 1_000_000, undefined, prePruneOn)).toBe(messages)})- it('T-R2-4a witness: the same history is pre-pruned above the threshold', () => {+ it('T-R2-4a witness: the same history is pre-pruned above the threshold', async () => {const messages = prunableHistory()- const compacted = compactConversation(messages, undefined, 1_000, undefined, prePruneOn)+ const compacted = await compactConversation(messages, undefined, 1_000, undefined, prePruneOn)const oldResult = compacted.find(m => m.role === 'tool')expect(oldResult?.content).not.toBe(OLD_RESULT)expect(oldResult!.content.length).toBeLessThan(OLD_RESULT.length / 10)diff --git a/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.ts b/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.tsindex 71c399815..a078cade7 100644--- a/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.ts+++ b/mcp-host/src/core/conversation/__tests__/compactionTranscriptInvariant.test.ts@@ -51,7 +51,7 @@ describe('compaction ↔ durable transcript invariant (FU2, spec §8 "does not aexpect(messages).toHaveLength(12) // 6 turns × (user + assistant)// Force compaction (threshold 0) keeping only the last 2 turns.- const compacted = compactConversation(messages, 2, 0)+ const compacted = await compactConversation(messages, 2, 0)// Compaction actually ran on the context window.expect(compacted.length).toBeLessThan(messages.length)@@ -66,7 +66,7 @@ describe('compaction ↔ durable transcript invariant (FU2, spec §8 "does not aconst { manager, conv } = await buildConversationWithTurns(6)const fullBefore = manager.buildMessageHistory(conv)- compactConversation(fullBefore, 2, 0)+ await compactConversation(fullBefore, 2, 0)// buildMessageHistory re-derives from conversation.turns each call, so the// post-compaction rebuild is identical — proving compaction did not erode@@ -81,7 +81,7 @@ describe('compaction ↔ durable transcript invariant (FU2, spec §8 "does not aconst messages = manager.buildMessageHistory(conv)const inputSnapshot = messages.map(m => ({ ...m }))- const compacted = compactConversation(messages, 2, 0)+ const compacted = await compactConversation(messages, 2, 0)expect(compacted).not.toBe(messages)expect(messages).toEqual(inputSnapshot) // input array left intactdiff --git a/mcp-host/src/core/conversation/compaction.ts b/mcp-host/src/core/conversation/compaction.tsindex 83431829f..b6ae581f2 100644--- a/mcp-host/src/core/conversation/compaction.ts+++ b/mcp-host/src/core/conversation/compaction.ts@@ -1,6 +1,5 @@import { type PrePruneOptions, prePrune } from '../extensions/prePrune'import { heuristicCount } from '../tokenizer/heuristic'-import type { TokenCounter } from '../tokenizer/tokenCounter'import { ChatMessage } from '../types'/**@@ -226,25 +225,23 @@ export function applyAlignedCut(* 3. Apply the T1.3 boundary aligner so cuts never split tool pairs and* never archive the last user message.*/-export function compactConversation(+export async function compactConversation(messages: ChatMessage[],maxTurns: number = 5,threshold: number = 80000,- counter?: TokenCounter,+ measure: (msgs: ChatMessage[]) => Promise<number> = async msgs => heuristicCount(msgs),prePruneOpts?: { enabled: boolean; options?: PrePruneOptions }-): ChatMessage[] {- // P.2: prefer the provider-aware counter when available. `countSync` is a- // best-effort path that never makes a network call — for Anthropic it- // returns the heuristic upper bound, for OpenAI/tiktoken it's exact after- // warmup. The caller (`runAgentLoop`) issues warmup at executor start.- const count = (msgs: ChatMessage[]) => (counter ? counter.countSync(msgs) : heuristicCount(msgs))+): Promise<ChatMessage[]> {+ // #739: the caller (`runAgentLoop`) passes the measure behind+ // `PressureContextManager`'s tier decision (`tierDecisionTokens`), so the+ // threshold it scales from that manager's means the same history here.// #731: under the threshold the history is returned untouched, by// reference. `CLERUM_COMPACTION_PRE_PRUNE` defaults to `true`, so running// pre-prune first would rewrite every history on every task, whatever the// window. This mirrors `PressureContextManager.manage()`, which passes// through below its threshold before pre-pruning.- if (count(messages) < threshold) {+ if ((await measure(messages)) < threshold) {return messages}@@ -252,7 +249,7 @@ export function compactConversation(let working = messagesif (prePruneOpts?.enabled) {working = prePrune(messages, prePruneOpts.options).messages- if (count(working) < threshold) {+ if ((await measure(working)) < threshold) {return working}}diff --git a/mcp-host/src/core/extensions/contextManager.ts b/mcp-host/src/core/extensions/contextManager.tsindex 191f85202..452d58719 100644--- a/mcp-host/src/core/extensions/contextManager.ts+++ b/mcp-host/src/core/extensions/contextManager.ts@@ -145,6 +145,29 @@ export interface PressureContextManagerOptions {*/export const COMPACTION_PRESSURE_THRESHOLD = 0.8+/**+ * The token count behind `PressureContextManager`'s tier decision. Without a+ * counter, or in the dry run, the byte heuristic decides; otherwise the+ * counter does. `taskExecutor` measures the rehydrated history with this same+ * function, so a history the manager passes through is never compacted at+ * rehydration (#739).+ *+ * The `lastObservedInputTokens` shortcut (Hermes `update_from_response`) is+ * intentionally NOT applied: the manager calls this once per loop iteration,+ * so the per-decision count is affordable, and skipping `count()` would risk+ * under-counting messages added since the last response. T2.2 (prompt cache)+ * revisits this with a per-iteration shape diff.+ */+export async function tierDecisionTokens(+ messages: ChatMessage[],+ tools: ToolDefinition[],+ tokenCounter: TokenCounter | undefined,+ dryRun: boolean+): Promise<number> {+ if (!tokenCounter || dryRun) return estimateTokens(messages) + heuristicCountTools(tools)+ return tokenCounter.count(messages, tools)+}+type PressureTier = 'passthrough' | 'workspace' | 'summarize' | 'truncate'function tierFor(pressure: number): PressureTier {@@ -507,12 +530,12 @@ export class PressureContextManager implements ContextManager {}/**- * Resolve the token pressure ratio according to the dry-run gate. Three branches:+ * Resolve the token pressure ratio from `tierDecisionTokens`. Three branches:* - No counter: legacy heuristic (preserves pre-P.2 behavior bit-for-bit).- * - dryRun=true: compute BOTH numbers; the heuristic decides the tier; the- * delta and any tier mismatch are emitted as metrics for the bake-week.- * - dryRun=false: the counter (with `lastObservedInputTokens` shortcut)- * drives the decision directly.+ * - dryRun=true: the heuristic decides the tier; the counter's number is+ * computed too, and the delta and any tier mismatch are emitted as+ * metrics for the bake-week.+ * - dryRun=false: the counter drives the decision directly.* Every branch counts `tools` with the messages: both travel in the request* the provider caps (#731).*/@@ -520,44 +543,37 @@ export class PressureContextManager implements ContextManager {messages: ChatMessage[],tools: ToolDefinition[]): Promise<number> {- if (!this.tokenCounter) {- return (estimateTokens(messages) + heuristicCountTools(tools)) / this.maxTokens+ const decided = await tierDecisionTokens(messages, tools, this.tokenCounter, this.dryRun)+ if (this.tokenCounter && this.dryRun) {+ await this.observeDryrunDelta(this.tokenCounter, messages, tools, decided)}- if (this.dryRun) {- const heuristic = estimateTokens(messages) + heuristicCountTools(tools)- let real: number- try {- real = await this.measureWithCounter(messages, tools)- } catch (err) {- logger.warn({ component: 'ContextManager', err }, 'dryrun counter failed; using heuristic')- return heuristic / this.maxTokens- }- const heuristicTier = tierFor(heuristic / this.maxTokens)- const realTier = tierFor(real / this.maxTokens)- const delta = heuristic > 0 ? real / heuristic : 0- tokenizerDryrunDelta.observe(- { provider: this.tokenCounter.providerName, tier_chosen: heuristicTier },- delta- )- if (heuristicTier !== realTier) {- tokenizerDryrunTierMismatchTotal.inc({ from: heuristicTier, to: realTier })- }- return heuristic / this.maxTokens- }- return (await this.measureWithCounter(messages, tools)) / this.maxTokens+ return decided / this.maxTokens}- private async measureWithCounter(+ /** Dry-run metrics: how far the counter's number is from the heuristic's. */+ private async observeDryrunDelta(+ tokenCounter: TokenCounter,messages: ChatMessage[],- tools: ToolDefinition[]- ): Promise<number> {- // The `lastObservedInputTokens` shortcut (Hermes `update_from_response`)- // is intentionally NOT applied here in the first PR: the call site runs- // once per loop iteration so the per-decision call is affordable, and- // skipping `count()` would risk under-counting messages added since the- // last response. T2.2 (prompt cache) revisits this with a per-iteration- // shape diff.- return this.tokenCounter!.count(messages, tools)+ tools: ToolDefinition[],+ heuristic: number+ ): Promise<void> {+ let real: number+ try {+ real = await tokenCounter.count(messages, tools)+ } catch (err) {+ logger.warn({ component: 'ContextManager', err }, 'dryrun counter failed; using heuristic')+ return+ }+ const heuristicTier = tierFor(heuristic / this.maxTokens)+ const realTier = tierFor(real / this.maxTokens)+ const delta = heuristic > 0 ? real / heuristic : 0+ tokenizerDryrunDelta.observe(+ { provider: tokenCounter.providerName, tier_chosen: heuristicTier },+ delta+ )+ if (heuristicTier !== realTier) {+ tokenizerDryrunTierMismatchTotal.inc({ from: heuristicTier, to: realTier })+ }}/**