mocode-ai 1.1.6 → 1.1.8

package/dist/agent/index.js

@@ -15,7 +15,6 @@ import { createPetHooks } from '../pet/state.js';
 import { t } from '../i18n/index.js';
 import { isToolErrorOutput } from '../tools/result.js';
 import { appendCurrentSessionTraceEvent } from '../session/index.js';
-import { buildActiveNotesPlanReminder } from '../session/notes-plan.js';
 /** 当前 turn 的 batch id(runAgent 内闭包变量;一条 turn 一轮 tool batch 结束即清空)。 */
 let currentBatchId = null;
 let turnFileChanges = [];
@@ -202,6 +201,8 @@ onContextUpdate) {
         },
         onStepStart: () => spinner.start(t('agent.thinking')),
         onChatDone: () => spinner.stop(),
+        // 流式实时用量 → 底栏 context 进度条左侧 chip;轮末由 repl 清空。
+        onLiveUsage: (u) => layout.setLiveUsage(u),
         onTextEnd: () => {
             if (lastChar && lastChar !== '\n') {
                 layout.contentWrite('\n');
@@ -240,24 +241,6 @@ onContextUpdate) {
             flushToolBatch();
             layout.contentWrite(`${ui.dim}${t('agent.aborted')}${ui.reset}\n`);
         },
-        onValidationStart: (command) => {
-            flushToolBatch();
-            spinner.start(t('agent.validating', { command }));
-        },
-        onValidationResult: (validation) => {
-            spinner.stop();
-            const color = validation.status === 'passed'
-                ? ui.green
-                : validation.status === 'failed'
-                    ? ui.red
-                    : ui.yellow;
-            const command = validation.command ?? t('agent.validationNoCommand');
-            const detail = validation.status === 'skipped' && validation.skipReason
-                ? `${validation.status}: ${validation.skipReason}`
-                : validation.status;
-            const symbol = validation.status === 'passed' ? '●' : validation.status === 'failed' ? '×' : '!';
-            layout.contentWrite(`  ${color}${symbol}${ui.reset} ${t('agent.validationResult', { command, status: detail })}\n\n`);
-        },
         onDone: (elapsedMs, usage) => {
             flushToolBatch();
             writeChangeOverview();
@@ -281,9 +264,7 @@ onContextUpdate) {
             userInput,
             signal,
             onContextUpdate,
-            dynamicSystemSuffix: buildActiveNotesPlanReminder,
             hooks: combinedHooks,
-            autoValidate: config.autoValidate,
             onTraceEvent: appendCurrentSessionTraceEvent,
         });
     }

package/dist/agent/spawn.js

@@ -31,8 +31,8 @@ const SUBAGENT_ROLE = `## Sub-agent execution
 You are executing one delegated sub-task with the same engineering standards and capabilities as mocode.
 - Treat Task context as authoritative facts already established by the main agent; do not rediscover them without evidence they are stale.
 - Focus on the delegated scope, but continue until it is genuinely complete. Do not stop to save tokens.
-- Do not recursively call sub-agent. A write task runs in an isolated overlay; the coordinator merges and performs final unified verification.
-- Return concise findings, changes, verification evidence, and blockers to the coordinator.`;
+- Do not recursively call sub-agent. A write task runs in an isolated overlay; the coordinator merges it safely.
+- Return concise findings, changes, checks you chose to run, and blockers to the coordinator.`;
 /**
  * 派生一个子 agent 执行独立子任务。
  *
@@ -51,7 +51,7 @@ export async function spawnAgent(opts) {
             summary: null,
             completed: false,
             transcript: 'Sub-agent execution is disabled. Enable it with /subagent on.',
-            status: 'failed', findings: [], readSet: [], changeSet: null, verification: null,
+            status: 'failed', findings: [], readSet: [], changeSet: null,
             usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0, cachedTokens: 0, reasoningTokens: 0 },
         };
     }
@@ -135,7 +135,6 @@ export async function spawnAgent(opts) {
         maxSteps,
         toolsOverride,
         contextState: localContextState,
-        autoValidate: false,
         onToolOutcome: (tool, args) => {
             if (tool === 'read_file' && typeof args.path === 'string')
                 readSet.add(args.path);
@@ -173,7 +172,6 @@ export async function spawnAgent(opts) {
         findings: result.finalText ? [result.finalText] : [],
         readSet: [...readSet].sort(),
         changeSet,
-        verification: null, // 主 Agent 在所有 coordinator merge 完成后统一验证
         usage: {
             promptTokens: result.usage?.promptTokens ?? 0,
             completionTokens: result.usage?.completionTokens ?? 0,

package/dist/agent/work-discipline.js

@@ -1,18 +1,5 @@
-// PROMPT-01: Build-and-Self-Verify working discipline.
-//
-// 把"完成必须验证"作为 system prompt 的一等公民,而不是事后选项。注入一段
-// 4 阶段纪律(Plan & Discover → Build → Verify → Fix),并提供 per-model
-// 措辞(anthropic / openai / qwen)让 base model 拿到最适合自己的表述。
-//
-// 关键约束:
-// - 纯函数,无副作用,无 config 依赖 → 不踩 TDZ,易测,易回滚。
-// - 段标题在 buildMocodeCorePrompt 之外,不会被 `## Project context` 索引
-//   切片误伤;且 buildBasePrompt 注入位置在 ## Workflow 之前,确保 LLM
-//   先看到纪律再看工具/平台细节。
-// - per-model 措辞是"轻量"差异:3 个家族共享 4 阶段结构,只在首句
-//   上贴近该家族的指令遵从习惯;真正的 prompt 反演化交给 AHE。
-// - 语种统一英文:4 份都用同一份核心纪律文本,避免多语种漂移;用户语言
-//   偏好由现有 i18n 段(assistant.languageInstruction)负责。
+// Lightweight, advisory working guidance. The agent decides how much discovery and
+// validation each task needs; the framework does not enforce a completion gate.
 /**
  * 从 config.model 字符串里嗅探 model family。匹配规则尽量宽松,够用即可。
  * 未来 AHE 闭环后可以换成 config.modelFamily 字段。
@@ -33,73 +20,32 @@ export function inferModelFamily(model) {
  * 4 阶段核心纪律(英文)。4 个 model family 共用此文本,只在首句与标题
  * 标签上做轻量变体。保持短小，详细的完成检查由动态 checklist 按需注入。
  */
-const CORE_SECTION = `## Working discipline — coding tasks (Build-and-Self-Verify)
-
-Treat "verification" as a first-class part of the task, not an afterthought. Use the smallest evidence-driven loop below.
-
-### Phase 1 — Plan & Discover
-- Open with a one-sentence restatement of your interpretation of the request; if a materially different reading exists, name it briefly before proceeding. This catches misunderstanding before any work is wasted.
-- State the goal and a concrete acceptance signal, then inspect the relevant code before changing it.
-- Ask only when an unresolved choice is high-impact or user-owned; otherwise follow repository evidence and proceed.
-
-### Phase 2 — Build
-- Make the smallest coherent change; avoid unrelated refactors.
-- Add or update a focused test when behavior changes and the project has an applicable test suite.
-- Re-read only when a dependent edit needs fresh exact content or state may be stale.
-
-### Phase 3 — Verify
-- Run the smallest executable check that proves the requested behavior, then read its complete result.
-- Compare evidence with the user's request, not merely with the diff.
-
-### Phase 4 — Fix
-- Diagnose the root cause, make a focused correction, and rerun the relevant check.
-- After two identical failures, change the approach instead of repeating the same call.
-
-**Hard rule (non-negotiable):** "I read the code and it looks right" is not a completion signal. Report the verification performed, or state clearly why it could not be run.
-
-**Hard rule (non-negotiable):** Never invent file paths, APIs, config keys, flags, or behavior. Every claim about the codebase must trace to tool output in this conversation; explicitly label anything you have not verified as an assumption.`;
-/**
- * 把核心段适配到指定 model family:只替换首行(语序 / 强动词),段标题
- * 保持原样。Phase 内容保持原样,4 份共享同一份结构化文本。
- * 注意:不再往标题注入 "[model: X]" 标签——它对模型是无意义噪声,
- * 还可能引发自我指涉,反而干扰遵从。
- */
-function adapt(_model, opener) {
-    return CORE_SECTION.replace('Treat "verification" as a first-class part of the task, not an afterthought.', opener);
-}
+const CORE_SECTION = `## Working discipline — coding tasks
+
+Use your judgment to choose the shortest reliable path from the request to a useful result.
+
+- Inspect only the code and context needed for the next decision.
+- Make the smallest coherent change and avoid unrelated refactors.
+- Decide whether validation is useful based on risk, scope, available commands, and the user's request. Validation is optional, not a completion gate.
+- When validation is useful, choose the smallest relevant check yourself; do not run broad test/build suites by default.
+- Re-read or rerun only when evidence is stale or the next edit depends on exact current content.
+- On failure, diagnose before retrying; after repeated identical failures, change approach.
+- Report honestly what you changed, what you checked, and anything left uncertain.
+
+Never invent file paths, APIs, config keys, flags, or behavior. Distinguish repository evidence from assumptions.`;
 /** ASK-01: only user-owned, high-impact choices should interrupt autonomous execution. */
-const ASK_WHITELIST_SECTION = `## When to ask instead of guess
-
-Call \`ask_human\` before coding only when repository evidence cannot resolve a user-owned, high-impact choice:
-1. irreversible deletion, migration, security, permission, or external side effect;
-2. public API compatibility (keep, deprecate, rename, or remove);
-3. multiple reasonable options that materially change product behavior;
-4. the request itself admits two or more materially different readings that lead to different deliverables (do not silently pick one and guess).
-
-For naming, implementation detail, and verification commands, follow repository precedent and choose the safest reversible default. Disclose any consequential assumption.
-
+const ASK_WHITELIST_SECTION = `## When to ask instead of guess
+
+Call \`ask_human\` before coding only when repository evidence cannot resolve a user-owned, high-impact choice:
+1. irreversible deletion, migration, security, permission, or external side effect;
+2. public API compatibility (keep, deprecate, rename, or remove);
+3. multiple reasonable options that materially change product behavior;
+4. the request itself admits two or more materially different readings that lead to different deliverables (do not silently pick one and guess).
+
+For naming and implementation details, follow repository precedent and choose the safest reversible default. Disclose any consequential assumption.
+
 Budget: at most 2 \`ask_human\` calls per turn. Beyond that, use the safest reversible default and disclose it in the final reply.`;
-/**
- * 拼出纪律段 + ASK-01 卡点白名单。返回完整段(两段用 \`\\n\\n\` 隔开);
- * 工厂之前只返回纪律段,ASK-01 落地后变成纪律 + 白名单两段;
- * ASK-01 段是固定英文,不参与 per-model 适配(避免 4 份变体维护成本)。
- */
-export function buildWorkDisciplineSection(modelFamily) {
-    let section;
-    switch (modelFamily) {
-        case 'anthropic':
-            section = adapt('anthropic', 'Verification is a hard prerequisite for completion, not a courtesy.');
-            break;
-        case 'openai':
-            section = adapt('openai', 'Every coding task MUST complete these four phases in order. Skipping or merging phases is treated as a failure. For trivial or read-only requests, phases may collapse.');
-            break;
-        case 'qwen':
-            section = adapt('qwen', 'Verification is a hard prerequisite for completion; "I wrote the code" is not evidence the code works.');
-            break;
-        case 'other':
-        case undefined:
-        default:
-            section = CORE_SECTION;
-    }
-    return `${section}\n\n${ASK_WHITELIST_SECTION}`;
+/** Advisory guidance shared by main and sub-agents. */
+export function buildWorkDisciplineSection(_modelFamily) {
+    return `${CORE_SECTION}\n\n${ASK_WHITELIST_SECTION}`;
 }

package/dist/config/index.js

@@ -41,6 +41,7 @@ export const languageFromShell = process.env.MOCODE_LANGUAGE !== undefined;
 // 仿 themeFromShell 模式:shell export 的环境变量在 loadEnvFiles 中不被回填(优先级最高),
 // 故 /model 写入 ~/.mocode/config 的同名键下次启动会被 shell 值覆盖——据此给 dim 警告。
 const LLM_ENV_KEYS = ['LLM_BASE_URL', 'LLM_API_KEY', 'LLM_MODEL', 'CONTEXT_WINDOW_TOKENS'];
+export const DEFAULT_CONTEXT_WINDOW_TOKENS = 256000;
 const llmKeysFromShell = LLM_ENV_KEYS.filter((k) => process.env[k] !== undefined);
 loadEnvFiles();
 setLanguage(detectLanguage(process.env.MOCODE_LANGUAGE));
@@ -195,11 +196,12 @@ ${buildWorkDisciplineSection(inferModelFamily(config.model))}
 
 ## Workflow
 - Use existing conversation and tool evidence before gathering more. Inspect only what supports the next decision; do not guess.
-- Keep changes focused. After modifications, run the smallest relevant executable verification that actually exercises the requested behavior, and report its result. A command exiting 0 is not proof the task is done — confirm the specific behavior the user asked for is observed, not merely that the diff applied.
+- Keep changes focused. Decide for yourself whether a check is worth running; prefer the smallest relevant check and avoid broad test/build suites unless the task or risk justifies them.
 - Use web search only when freshness materially affects the answer.
 ${buildCodegraphSection()}
 
 ## Tool use
+- During tool-calling turns, stay silent unless something important enough must reach the user — otherwise just call the tool and let it run.
 - Go directly to a known path or symbol; use discovery tools only when the location is unknown.
 - Edit against a FRESH read: before any edit_file/write_file, call read_file on the exact path and copy both its latest hash and the exact target text. Never reconstruct old_string from a grep/summary/diff — those lose whitespace and indentation and cause edit failures.
 - A read_file hash from before a compaction, session resume, edit conflict, or external change is STALE and will be rejected — re-read rather than reuse an old hash.
@@ -297,16 +299,14 @@ export const config = {
     get systemPrompt() {
         return buildBasePrompt();
     },
-    contextWindowTokens: Number(process.env.CONTEXT_WINDOW_TOKENS) || 128000,
-    compactThreshold: Number(process.env.COMPACT_THRESHOLD) || 0.85,
+    contextWindowTokens: Number(process.env.CONTEXT_WINDOW_TOKENS) || DEFAULT_CONTEXT_WINDOW_TOKENS,
     includeUsage: process.env.LLM_STREAM_USAGE !== 'false',
     autoCompact: process.env.AUTO_COMPACT !== 'false',
-    autoValidate: process.env.MOCODE_AUTO_VALIDATE !== 'false',
-    contextOptimize: process.env.MOCODE_CONTEXT_OPTIMIZE !== 'false',
-    contextRelprune: process.env.MOCODE_CONTEXT_RELPRUNE !== 'false',
+    contextOptimize: process.env.MOCODE_CONTEXT_OPTIMIZE === 'true',
+    contextRelprune: process.env.MOCODE_CONTEXT_RELPRUNE === 'true',
     contextLifecycle: process.env.MOCODE_LIFECYCLE !== 'false',
     contextBudget: process.env.MOCODE_BUDGET_SCHEDULER !== 'false',
-    autoReflect: process.env.AUTO_REFLECT !== 'false',
+    autoReflect: process.env.AUTO_REFLECT === 'true',
     memoryEnabled: process.env.MEMORY_ENABLED === 'true',
     reflectEveryN: Number(process.env.REFLECT_EVERY_N) || 5,
     maxSteps: Number(process.env.MAX_STEPS) || 1000,

package/dist/context/age-aware.js

@@ -1,12 +1,8 @@
-// Age-aware tool-result encoding coordinator.
-// Initial pushes stay conservative; old Cold results are re-encoded before chat.
-import { TOOL_OLD_AGE } from './budget.js';
+// Pressure-only tool-result encoding coordinator.
+// Normal pushes are raw apart from the per-result hard safety cap.
 import { optimizeToolResult } from './pipeline.js';
 import { canonicalizePath, extractPath, isToolResultSuccess, toText, } from './utils.js';
-/**
- * Tracks successful first reads and tool-result age without coupling encoders to
- * lifecycle's mutable history indexes. All methods are fail-safe and idempotent.
- */
+/** Rebuilds records from current history for each pressure pass. */
 export class AgeAwareEncodingState {
     pushOrdinal = 0;
     records = new Map();
@@ -14,36 +10,16 @@ export class AgeAwareEncodingState {
     constructor(history = []) {
         this.rehydrate(history);
     }
-    /** Build the conservative context for a newly completed tool result. */
-    preparePush(tc, succeeded) {
-        const path = tc.name === 'read_file'
-            ? canonicalizePath(extractPath(tc.arguments))
-            : null;
-        const isFirstRead = path ? !this.seenReadPaths.has(path) : undefined;
-        this.records.set(tc.id, {
-            toolCallId: tc.id,
-            toolName: tc.name,
-            argsRaw: tc.arguments,
-            pushOrdinal: this.pushOrdinal,
-            succeeded,
-            isFirstRead,
-            agedEncoded: false,
-        });
-        this.pushOrdinal++;
-        // Failed reads must not consume the "first successful read" privilege.
-        if (succeeded && path)
-            this.seenReadPaths.add(path);
-        return {
-            age: 0,
-            isCold: false,
-            isFirstRead,
-            phase: 'push',
-        };
-    }
-    /** Re-encode eligible tool messages in the Cold prefix in place. */
-    sweep(history, hotBoundary) {
+    /**
+     * Pressure-only, progressive encoding of Cold logs and retrievable searches.
+     * It never touches code reads, skills, human decisions, or sub-agent output;
+     * repeated pressure passes may further reduce content only when strictly shorter.
+     */
+    sweepPressure(history, hotBoundary) {
         try {
+            const pressureEncodable = new Set(['run_command', 'grep', 'glob', 'web_search', 'web_fetch']);
             const end = Math.min(Math.max(hotBoundary, 1), history.length);
+            let encodedCount = 0;
             for (let idx = 1; idx < end; idx++) {
                 const message = history[idx];
                 if (message.role !== 'tool')
@@ -51,32 +27,27 @@ export class AgeAwareEncodingState {
                 const toolMessage = message;
                 const id = toolMessage.tool_call_id;
                 const record = id ? this.records.get(id) : undefined;
-                if (!record || !record.succeeded || record.agedEncoded)
+                if (!record || !record.succeeded || !pressureEncodable.has(record.toolName))
                     continue;
                 const content = toText(toolMessage.content);
-                if (!content || content.startsWith('⌦[')) {
-                    record.agedEncoded = true;
+                if (!content || content.startsWith('⌦['))
                     continue;
-                }
-                // Exclude the result's own push: immediately after insertion its age is 0.
                 const age = Math.max(0, this.pushOrdinal - record.pushOrdinal - 1);
-                if (age < TOOL_OLD_AGE)
-                    continue;
                 const encoded = optimizeToolResult(record.toolName, content, record.argsRaw, {
                     age,
                     isCold: true,
                     isFirstRead: record.isFirstRead,
                     phase: 'sweep',
                 });
-                // Aged encoding is a degradation step: never replace content with a
-                // representation that is equal-sized or larger.
-                if (encoded.length < content.length)
+                if (encoded.length < content.length) {
                     toolMessage.content = encoded;
-                record.agedEncoded = true;
+                    encodedCount++;
+                }
             }
+            return encodedCount;
         }
         catch {
-            // Context optimization must never block an agent request.
+            return 0;
         }
     }
     /** Rebuild stable state after resume or structural history compaction. */
@@ -119,7 +90,6 @@ export class AgeAwareEncodingState {
                     pushOrdinal: this.pushOrdinal,
                     succeeded,
                     isFirstRead,
-                    agedEncoded: false,
                 });
                 this.pushOrdinal++;
                 if (succeeded && path)

package/dist/context/artifacts.js

@@ -132,10 +132,17 @@ export function recordArtifact(state, history, idx, output, succeeded) {
     updateStats(state, stateFor(state));
 }
 function affected(artifact, changed) {
-    return artifact.dependencies.some((dependency) => dependency.path === '*' || changed.has(dependency.path));
+    // '*' 依赖(无法解析出具体文件路径的诊断/搜索结果)不与任何具体写操作关联:
+    // 任何文件写入都会作废全部 '*' artifact,等于每次 mutation 都销毁
+    // git/测试/构建等历史证据,模型被迫反复 re-run,轮次爆炸。只失效路径明确命中的。
+    return artifact.dependencies.some((dependency) => dependency.path !== '*' && changed.has(dependency.path));
 }
-/** Mark and immediately stub stale facts; this is stronger than waiting for budget pressure. */
-export function invalidateArtifacts(state, history, changedFiles) {
+/**
+ * Mark precise, file-backed facts stale without changing the evidence in history.
+ * A stale result can still explain a later edit or failure; pressure compression is
+ * the only path allowed to replace its content with a compact marker.
+ */
+export function invalidateArtifacts(state, _history, changedFiles) {
     const changed = new Set(changedFiles.map(canonicalizePath).filter((item) => !!item));
     if (changed.size === 0)
         return 0;
@@ -145,15 +152,6 @@ export function invalidateArtifacts(state, history, changedFiles) {
         if (artifact.freshness !== 'fresh' || !affected(artifact, changed))
             continue;
         artifact.freshness = 'stale';
-        const message = history[artifact.messageIndex];
-        if (message?.role === 'tool') {
-            const original = toText(message.content);
-            const paths = artifact.dependencies.map((item) => item.path).join(', ');
-            const stub = `${STALE_PREFIX}${artifact.source.tool}] source=${artifact.id} dependencies=${paths} ` +
-                `invalidated-by=${[...changed].join(', ')}; re-run ${artifact.source.tool} before using this fact.`;
-            message.content = stub;
-            artifact.tokenCount = estimateTokens(stub);
-        }
         count++;
     }
     updateStats(state, artifactState);
@@ -193,7 +191,7 @@ export function rehydrateArtifacts(state, history) {
     }
     updateStats(state, artifactState);
 }
-/** Compare captured dependency versions before each model step to detect external edits. */
+/** Compare captured dependency versions before each model step and mark stale metadata only. */
 export function refreshArtifactFreshness(state, history) {
     const changed = new Set();
     for (const artifact of stateFor(state).artifacts.values()) {
@@ -208,19 +206,23 @@ export function refreshArtifactFreshness(state, history) {
     }
     return changed.size > 0 ? invalidateArtifacts(state, history, [...changed]) : 0;
 }
-/** Scheduler entry point: stale artifacts are already stubs; normalize any resumed stale message first. */
-export function pruneStaleArtifacts(state, history) {
+/**
+ * Pressure-only stage: replace stale evidence in the Cold prefix with a compact
+ * marker. Recent/current work stays intact, and normal mutation handling never
+ * calls this function.
+ */
+export function pruneStaleArtifacts(state, history, coldBoundary) {
     const artifactState = stateFor(state);
     let pruned = 0;
     for (const artifact of artifactState.artifacts.values()) {
-        if (artifact.freshness !== 'stale')
+        if (artifact.freshness !== 'stale' || artifact.messageIndex >= coldBoundary)
             continue;
         const message = history[artifact.messageIndex];
         if (message?.role !== 'tool')
             continue;
         const content = toText(message.content);
         if (!content.startsWith(STALE_PREFIX)) {
-            message.content = `${STALE_PREFIX}${artifact.source.tool}] source=${artifact.id}; re-run before use.`;
+            message.content = `${STALE_PREFIX}${artifact.source.tool}] source=${artifact.id}; source changed, so this content may be outdated.`;
             artifact.tokenCount = estimateTokens(String(message.content));
             pruned++;
         }

package/dist/context/budget.js

@@ -1,12 +1,8 @@
-// 五区 Context Budget Scheduler。
+// 五区 Context Budget accounting and the shared pressure threshold.
 //
-// 目的:把当前请求拆成 System / History / Tool-Recent / Tool-Old / Summary + Reserve，
-//      统一报告各区占用，并只调度执行层能够真正落地的 warn / compact_history。
-//
-// push-time cap、pipeline、relevance、lifecycle 与 age-aware sweep 负责工具结果优化；
-// scheduler 在这些处理完成后评估，不重复生成 Cold/Hot tool 压缩动作。
-// 本文件保持叶子级，只依赖 ChatMessage / token estimate，具体执行由 session/scheduler.ts 完成。
-// contextBudget 开关关闭时，agent/core.ts 退化为直接调用 maybeCompact。
+// This module only estimates and reports. session/scheduler.ts owns the sole
+// automatic rewrite sequence and starts all pressure cleanup plus compact_history
+// when corrected or raw request occupancy reaches 80%.
 import { chatTools, estimateMessagesTokens, estimateToolSchemaTokens, messageTokens, } from '../llm/index.js';
 import { toText } from './utils.js';
 /** 五区分账(占比对齐 CONTEXT_WINDOW)。顺序固定,便于遍历。 */
@@ -28,9 +24,8 @@ export const DEFAULT_BUDGET_POLICY = {
         reserve: 0.05,
     },
     hotTurnWindow: 4,
-    toolOldAge: 2,
     compactKeepRatio: 0.40,
-    totalTriggerRatio: 0.82,
+    pressureTriggerRatio: 0.80,
     schedulerTargetRatio: 0.80,
     estimateSafetyFactor: 1.05,
     compactHeadroomTokens: 1500,
@@ -38,7 +33,6 @@ export const DEFAULT_BUDGET_POLICY = {
 /** 兼容既有调用方的只读别名；配置只在 DEFAULT_BUDGET_POLICY 中维护。 */
 export const BUDGET_RATIO = DEFAULT_BUDGET_POLICY.ratios;
 export const HOT_TURN_WINDOW = DEFAULT_BUDGET_POLICY.hotTurnWindow;
-export const TOOL_OLD_AGE = DEFAULT_BUDGET_POLICY.toolOldAge;
 function msgTokens(m) {
     // 与请求预估复用同一实现，避免角色结构开销、多模态和 tool_calls 在两个预算路径中漂移。
     return messageTokens(m);
@@ -69,13 +63,17 @@ export function evaluateBudget(history, window, step = 0, correction = 1, active
     }
     // 校正后的 token 数:raw * correction,最小 1(raw > 0 时)。
     const adj = (raw) => (raw > 0 ? Math.max(1, Math.round(raw * correction)) : 0);
+    // 裸总量(不乘 correction):硬闸用它判断,防止 correction 折扣否决真实溢出。
+    let rawTotal = 0;
     const sysMsg = history[0];
     const systemCosts = {
         prompt: sysMsg ? msgTokens(sysMsg) : 0,
         toolSchemas: estimateToolSchemaTokens(activeTools),
     };
     // 工具 schema 与 system prompt 同属请求固定开销；必须计入总量才能可靠触发压缩。
-    layers.system.actual = adj(systemCosts.prompt + systemCosts.toolSchemas);
+    const systemRaw = systemCosts.prompt + systemCosts.toolSchemas;
+    layers.system.actual = adj(systemRaw);
+    rawTotal += systemRaw;
     // Summary 检测:role:'system' 且不是 history[0] 的,视为摘要(compact.ts 摘要插 index 1)。
     // 简单启发:若 history[1]?.role === 'system' 且 content 含「# 会话摘要」特征串,计入 summary。
     // 命中时循环跳过 i=1;不命中时当作普通 message(罕见,落到下方 user/assistant 分支)。
@@ -83,7 +81,9 @@ export function evaluateBudget(history, window, step = 0, correction = 1, active
     if (history.length > 1 && history[1].role === 'system') {
         const c1 = toText(history[1].content);
         if (c1.startsWith('# 会话摘要') || c1.includes('会话摘要')) {
-            layers.summary.actual = adj(msgTokens(history[1]));
+            const summaryRaw = msgTokens(history[1]);
+            layers.summary.actual = adj(summaryRaw);
+            rawTotal += summaryRaw;
             summaryHit = true;
         }
     }
@@ -94,16 +94,18 @@ export function evaluateBudget(history, window, step = 0, correction = 1, active
         const m = history[i];
         if (i === 1 && summaryHit)
             continue; // summary 已单独算过
+        const raw = msgTokens(m);
         if (m.role === 'tool') {
-            const t = adj(msgTokens(m));
             if (i >= hotStart)
-                layers.toolRecent.actual += t;
+                layers.toolRecent.actual += adj(raw);
             else
-                layers.toolOld.actual += t;
+                layers.toolOld.actual += adj(raw);
+            rawTotal += raw;
         }
         else if (m.role !== 'system') {
             // user / assistant 全部计入 history(对话轨迹)
-            layers.history.actual += adj(msgTokens(m));
+            layers.history.actual += adj(raw);
+            rawTotal += raw;
         }
         // 其它 system(几乎不存在)跳过
     }
@@ -120,10 +122,11 @@ export function evaluateBudget(history, window, step = 0, correction = 1, active
     // 按 overRatio 降序
     triggers.sort((a, b) => layers[b].overRatio - layers[a].overRatio);
     const total = BUDGET_LAYERS.reduce((s, k) => s + (k === 'reserve' ? 0 : layers[k].actual), 0);
-    const totalOver = total >= DEFAULT_BUDGET_POLICY.totalTriggerRatio * window;
+    const totalOver = total >= DEFAULT_BUDGET_POLICY.pressureTriggerRatio * window;
     return {
         step,
         total,
+        rawTotal,
         window,
         layers,
         systemCosts,
@@ -134,15 +137,11 @@ export function evaluateBudget(history, window, step = 0, correction = 1, active
     };
 }
 /** 根据 BudgetReport 生成可执行动作。
- * push-time cap、relevance、lifecycle 与 age-aware sweep 已在评估前完成，
- * 因此这里不再生成无法执行的 Cold/Hot tool action。
- * History 或总量超预算时才考虑昂贵的 LLM 摘要。 */
+ * History compaction is the final fallback after pressure-only tool stages.
+ * Per-layer overages are diagnostics, never independent rewrite triggers. */
 export function scheduleActions(report) {
     const actions = [];
-    const { layers, totalOver, total } = report;
-    const policy = DEFAULT_BUDGET_POLICY;
-    const headroom = policy.schedulerTargetRatio * report.window
-        - total * policy.estimateSafetyFactor;
+    const { layers } = report;
     if (layers.system.overBudget) {
         const { prompt, toolSchemas } = report.systemCosts;
         const { actual, budget } = layers.system;
@@ -155,8 +154,8 @@ export function scheduleActions(report) {
                 + `提示 ${prompt} + 工具 ${toolSchemas}，×${report.correction.toFixed(2)}。`,
         });
     }
-    if ((layers.history.overBudget || totalOver)
-        && headroom < -policy.compactHeadroomTokens) {
+    const pressureLine = DEFAULT_BUDGET_POLICY.pressureTriggerRatio * report.window;
+    if (Math.max(report.rawTotal, report.total) >= pressureLine) {
         actions.push({ kind: 'compact_history' });
     }
     return actions;
@@ -164,7 +163,7 @@ export function scheduleActions(report) {
 /** 拍平成人类可读(供 /context 命令与 check-budget 脚本用)。 */
 export function formatReport(report) {
     const lines = [];
-    lines.push(`step ${report.step}  total ${report.total}/${report.window}  (${((report.total / report.window) * 100).toFixed(1)}%)`);
+    lines.push(`step ${report.step}  total ${report.total}/${report.window}  (${((report.total / report.window) * 100).toFixed(1)}%)  raw ${report.rawTotal}`);
     for (const k of BUDGET_LAYERS) {
         const lb = report.layers[k];
         const pct = lb.budget > 0 ? ((lb.actual / lb.budget) * 100).toFixed(0) : '-';

package/dist/context/classifier.js

@@ -29,7 +29,6 @@ const BY_NAME = {
     edit_file: 'status',
     write_file: 'status',
     ask_human: 'status',
-    drop_context: 'status',
     memory_save: 'status',
     memory_update: 'status',
     memory_forget: 'status',

package/dist/context/encoders/index.js

@@ -1,14 +1,7 @@
-// 内置 encoder 清单。启动期 pipeline 首次调用时经 registerAll 注册到 registry。
-//
-// Phase 1:passthrough(identity 兜底)→ 全链路零行为变化。
-// Phase 2:tree / search / log / table / memory(高价值低风险)。
-// Phase 2.5:code / graph / doc / summary(覆盖剩余有 encoder 的 kind)。
-//   - code(read_file):仅折叠 ≥3 连续空行,行号保真(edit_file 依赖)。
-//   - graph(保留 kind,暂无 builtin 工具直接命中)/ doc(web_fetch, use_skill)/ summary(task):
-//     去 ANSI + 折叠空行,保守不重构结构。
-//   - status(edit/write/ask_human/mem 增删改)无 encoder:本就是一行,无需编码。
-//
-// 加 encoder:新建 encoders/xxx.ts 导出 ContextEncoder,在此数组加一行。无需动 agent / llm / core。
+// 内置 encoder 清单。仅由 opt-in pressure stage 首次调用时懒注册。
+// Normal tool pushes do not pass through these encoders.
+// Available transforms:tree / search / log / table / memory / code / doc / summary.
+// status 类工具本就是短结果，无专用 encoder。
 import { passthroughEncoder } from './passthrough.js';
 import { treeEncoder } from './tree.js';
 import { searchEncoder } from './search.js';

package/dist/context/index.js

@@ -1,12 +1,9 @@
-// context/ barrel:Context Optimization Pipeline + 五区 Budget Scheduler。
-//
-// 单一入口 optimizeToolResult(agent/core.ts pushToolResult 调)接管"工具结果进 LLM 前"的表示。
-// 单一入口 runScheduler(agent/core.ts 步前调)接管"何时调用哪一闸"的调度。
-// 不调 LLM、不碰 Tool Calling schema / executeTool / tool_call_id 配对 / TUI 渲染
-// (叶子级:仅 stdlib + tools/constants + session/compact 的 capToolResultForHistory 兜底 + config 开关)。
+// context/ barrel: metadata tracking, optional pressure encoders, and budget reporting.
+// Normal tool pushes stay raw apart from the hard per-result cap. The session
+// scheduler is the only automatic rewrite coordinator at real pressure.
 export { optimizeToolResult } from './pipeline.js';
 export { classify, knownToolKinds } from './classifier.js';
 export { recordArtifact, invalidateArtifacts, rehydrateArtifacts, refreshArtifactFreshness, pruneStaleArtifacts, collectArtifactRefs, formatArtifactTokenSources, } from './artifacts.js';
 export { registerEncoder, registerAll, getEncoder, registeredKinds, } from './registry.js';
 // ── Context Budget Scheduler ───────────────────────────────────────────────
-export { evaluateBudget, scheduleActions, formatReport, quickEstimate, userTurnBoundary, BUDGET_LAYERS, DEFAULT_BUDGET_POLICY, BUDGET_RATIO, HOT_TURN_WINDOW, TOOL_OLD_AGE, } from './budget.js';
+export { evaluateBudget, scheduleActions, formatReport, quickEstimate, userTurnBoundary, BUDGET_LAYERS, DEFAULT_BUDGET_POLICY, BUDGET_RATIO, HOT_TURN_WINDOW, } from './budget.js';