billion-context-dsh 0.1.9 → 0.2.0

package/README.en.md

@@ -3,7 +3,7 @@
 [English](./README.en.md) | [中文](./README.md)
 
 > **⚠️ Beta notice — not for production use**
-> This project (**v0.1.9**) is a work-in-progress beta. The [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) itself is also in **public beta**. **Do not use either in engineering / production environments** — expect breaking changes and rough edges.
+> This project (**v0.2.0**) is a work-in-progress beta. The [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) itself is also in **public beta**. **Do not use either in engineering / production environments** — expect breaking changes and rough edges.
 
 <p align="center">
 <strong>Built with gratitude on top of these projects</strong> — please give them a ⭐:
@@ -56,6 +56,18 @@ This is the DeepSeek Harness port of [billion-context-pi](https://github.com/ran
 npm install billion-context-dsh
 ```
 
+> 💡 **One-command install via `dsh plugin` (bundle, v0.2.0+)**. The package declares a `dsh.bundle`
+> manifest, so DSH's plugin command installs it into the profile and applies the patch
+> automatically (equivalent to the composition row below):
+
+```bash
+dsh plugin --profile web add billion-context-dsh
+```
+
+Restart `dsh` afterwards (bundle layers are composed at startup). For custom `config`
+(such as `modelContextLimit` / `prompts`), keep the hand-written composition row — the
+bundle patch ([cordis.patch.yml](cordis.patch.yml)) only inserts the default row without `config`.
+
 That's it. Then add a composition row where a compaction backend is expected — two scopes, pick by how wide you want it:
 
 **Global — host plane, every mode** (recommended). In your profile patch (e.g. `~/.dsh/profiles/web/cordis.patch.yml`), add:

package/README.md

@@ -3,7 +3,7 @@
 [中文](./README.md) | [English](./README.en.md)
 
 > **⚠️ 测试版声明——请勿用于生产环境**
-> 本项目（**v0.1.9**）仍处于开发中的测试版。[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 本身也处于**公开测试版**阶段。**请勿将两者用于工程化 / 生产环境**——预期会有破坏性变更与粗糙之处。
+> 本项目（**v0.2.0**）仍处于开发中的测试版。[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 本身也处于**公开测试版**阶段。**请勿将两者用于工程化 / 生产环境**——预期会有破坏性变更与粗糙之处。
 
 <p align="center">
 <strong>衷心感谢以下项目——请给它们一个 ⭐：</strong>
@@ -59,6 +59,17 @@
 npm install billion-context-dsh
 ```
 
+> 💡 **v0.2.0 起支持 `dsh plugin` 一键安装（bundle）**。包已声明 `dsh.bundle`
+> manifest，DSH 的插件命令会把它装进 profile 并自动应用补丁（等价于下面的组合行）：
+
+```bash
+dsh plugin --profile web add billion-context-dsh
+```
+
+装完重启 `dsh`（bundle 层在启动时组合）。需要自定义 `config`（如
+`modelContextLimit` / `prompts`）时仍建议手写组合行——bundle 补丁
+（[cordis.patch.yml](cordis.patch.yml)）只插入无 `config` 的默认行。
+
 就这样。然后在需要压缩后端的位置加组合配置——两种范围，按需选择：
 
 **全局生效（host 平面，所有模式）——推荐**。在你的 profile 补丁（如 `~/.dsh/profiles/web/cordis.patch.yml`）中追加：

package/cordis.patch.yml

@@ -0,0 +1,11 @@
+# dsh bundle patch: inserts the ACP compaction backend into a profile's layer stack.
+#
+# This is what makes the package installable via `dsh plugin --profile web add
+# billion-context-dsh` (declared in package.json as `dsh.bundle.patch`). The row
+# below is exactly the manual composition row documented in docs/INSTALL.md (§2a),
+# minus config: omitting `modelContextLimit` means auto-detection (fallback 128000),
+# and the built-in prompt copy (config.prompts) is used. Users who want custom
+# config can still write the row by hand in their profile patch.
+- insert:
+    - id: compaction-acp
+      name: 'billion-context-dsh'

package/dist/index.d.ts

@@ -30,10 +30,12 @@ import { CompactionEngine, type CompactionAgentContext, type CompactionResult, t
 import { type CompressionCore } from 'acp-kernel';
 import type { Agent } from '@deepseek-ai/dsh-agent';
 import { AcpStateStore } from './state.ts';
+import { type AcpPrompts, type ResolvedPrompts } from './prompts.ts';
 import { type AcpWindow } from './window.ts';
 export { AcpStateStore } from './state.ts';
 export { kernelConfigFor, type KernelConfigInput } from './config.ts';
 export { ACP_SYSTEM_PROMPT, ACP_SYSTEM_PROMPT_ORDER } from './system-prompt.ts';
+export { DEFAULT_PROMPTS, DEFAULT_RESOLVED, renderSystemPrompt, renderTemplate, resolvePrompts, type AcpPrompts, type NudgePrompts, type PromptInput, type PromptOverride, type RangeTablePrompts, type ResolvedPrompts, type ToolPrompts, } from './prompts.ts';
 export { makeTools, type ToolEnvironment } from './tools.ts';
 export { acpCommand } from './commands.ts';
 export { buildNudge, resolveTokenCount, type NudgeEnvironment, type NudgeOutcome } from './nudge.ts';
@@ -87,6 +89,8 @@ export interface AcpConfig {
     readonly autoCommand: boolean;
     /** Inject the nudge into `agent/pre-step` when the kernel recommends it. Default true. */
     readonly autoNudge: boolean;
+    /** Per-stage prompt template overrides (nudge / range table / system prompt / tool descriptions). See docs/configurable-prompts-design.md. */
+    readonly prompts?: AcpPrompts;
 }
 export declare function resolveAcpConfig(config?: Partial<AcpConfig>): AcpConfig;
 /**
@@ -101,6 +105,8 @@ export declare class AcpCompactionEngine extends CompactionEngine {
     readonly store: AcpStateStore;
     /** Resolved engine configuration. */
     readonly config: AcpConfig;
+    /** Resolved prompt templates (validated at construction — fail-fast on template typos). */
+    readonly prompts: ResolvedPrompts;
     private readonly lastNudgeTurn;
     /** Per provider/model route the resolved window (probe failures cached too). */
     private readonly windowCache;

package/dist/index.js

@@ -585,6 +585,111 @@ import {
   defaultCountTokens as defaultCountTokens2
 } from "acp-kernel";
 import { createUserMessage as createUserMessage2 } from "@deepseek-ai/dsh-llm";
+
+// src/prompts.ts
+import { COMPRESS_PHILOSOPHY } from "acp-kernel";
+var NUDGE_ALLOWED = {
+  normal: /* @__PURE__ */ new Set(["pct"]),
+  emergency: /* @__PURE__ */ new Set(["pct"]),
+  guidance: /* @__PURE__ */ new Set(),
+  tier: /* @__PURE__ */ new Set(["tier", "count", "prevTier", "tokens", "seqs"])
+};
+var RANGE_TABLE_ALLOWED = {
+  header: /* @__PURE__ */ new Set(["surface"]),
+  title: /* @__PURE__ */ new Set(["count"]),
+  line: /* @__PURE__ */ new Set(["start", "end", "count", "tokens"]),
+  footer: /* @__PURE__ */ new Set()
+};
+var TOOLS_ALLOWED = {
+  compress: /* @__PURE__ */ new Set(),
+  decompress: /* @__PURE__ */ new Set(),
+  searchContext: /* @__PURE__ */ new Set(),
+  acpStatus: /* @__PURE__ */ new Set()
+};
+var SYSTEM_ALLOWED = /* @__PURE__ */ new Set(["philosophy"]);
+function validateTemplate(template, allowed, path) {
+  const re = /\{([A-Za-z_][A-Za-z0-9_]*)\}/g;
+  let match;
+  while ((match = re.exec(template)) !== null) {
+    const name = match[1];
+    if (!allowed.has(name)) {
+      throw new Error(
+        `${path} contains unknown placeholder {${name}} \u2014 allowed: ${[...allowed].join(", ") || "(none)"}`
+      );
+    }
+  }
+  return template;
+}
+function renderTemplate(template, vars) {
+  return template.replace(/\{([A-Za-z_][A-Za-z0-9_]*)\}/g, (_match, name) => {
+    const value = vars[name];
+    if (value === void 0) {
+      throw new Error(
+        `renderTemplate: missing value for placeholder {${name}} in template "${template.slice(0, 60)}\u2026"`
+      );
+    }
+    return String(value);
+  });
+}
+function mergeGroup(defaults, override, allowed, path) {
+  if (override == null) return defaults;
+  const out = {};
+  for (const key of Object.keys(defaults)) {
+    const value = override[key];
+    out[key] = value === null || value === void 0 ? defaults[key] : validateTemplate(value, allowed[key], `${path}.${String(key)}`);
+  }
+  return out;
+}
+function resolvePrompts(input) {
+  if (input === void 0) return DEFAULT_RESOLVED;
+  return {
+    nudge: mergeGroup(DEFAULT_PROMPTS.nudge, input.nudge, NUDGE_ALLOWED, "prompts.nudge"),
+    rangeTable: mergeGroup(DEFAULT_PROMPTS.rangeTable, input.rangeTable, RANGE_TABLE_ALLOWED, "prompts.rangeTable"),
+    tools: mergeGroup(DEFAULT_PROMPTS.tools, input.tools, TOOLS_ALLOWED, "prompts.tools"),
+    systemPromptTemplate: input.systemPrompt === null || input.systemPrompt === void 0 ? DEFAULT_PROMPTS.systemPromptTemplate : validateTemplate(input.systemPrompt, SYSTEM_ALLOWED, "prompts.systemPrompt")
+  };
+}
+function renderSystemPrompt(prompts) {
+  return renderTemplate(prompts.systemPromptTemplate, { philosophy: COMPRESS_PHILOSOPHY });
+}
+var DEFAULT_PROMPTS = {
+  nudge: {
+    normal: "Context usage is at {pct}%. This is a suggestion, not a requirement \u2014 you decide whether and when to compress.",
+    emergency: "\u26A0\uFE0F Context usage is at {pct}% of the window \u2014 nearly full. Consider compressing consumed ranges soon so working context stays available; the choice and timing are yours.",
+    guidance: "Compress by need, not by percentage: replace only ranges you have genuinely consumed, with dense self-contained summaries.",
+    tier: "Tier {tier}: {count} tier-{prevTier} block(s) distillable ({tokens} tokens) \u2014 compress their summary node(s) [seqs {seqs}] to reclaim the original messages."
+  },
+  rangeTable: {
+    header: "Surface: {surface}",
+    title: "Compressible ranges (suggestions only \u2014 compress any consumed span; refs are surface seqs):",
+    line: "  - seq {start}..{end} \u2014 {count} messages, ~{tokens} tokens",
+    footer: "Compress with: compress({ content: [{ startSeq, endSeq, summary }] }) \u2014 content is an array: batch multiple unrelated segments in one call, each entry its own block. Keep ranges disjoint.\nSnapshot taken at nudge time: the seqs go stale once the surface moves (a later compress shadows them), so re-run acp_status for fresh refs before compressing."
+  },
+  tools: {
+    compress: "Replace older conversation ranges with dense summaries you write. Each message seq is a surface reference. Single range: compress({ content: [{ startSeq, endSeq, summary }] }). Batch multiple unrelated ranges in one call (each content entry becomes its own block); keep ranges disjoint. Never compress content the current step is actively using. Seq refs must come from the CURRENT surface (acp_status or the latest nudge): a span whose edges were shadowed by an earlier compress is auto-remapped to its still-live content, a fully compressed span is reported as already compressed, and invented/other-session seqs fail with guidance.",
+    decompress: "Recover the original content of a compressed block by its blockId (read-only; does not unshadow the range).",
+    searchContext: "Search inside compressed blocks (summaries and original content) for information the model no longer sees in context.",
+    acpStatus: "Report the ACP block ledger: compressed blocks, reclaimed tokens, and current context pressure."
+  },
+  systemPromptTemplate: `Active Context Pruning \u2014 model-driven context management
+
+YOU decide whether and when to compress context. Nothing forces you: the injected "nudge" is a suggestion, not an order, and you may ignore it when compression would not help. Compress only ranges you have genuinely consumed (read tool outputs, finished explorations, superseded steps) that the current work no longer needs verbatim.
+
+{philosophy}
+
+Compression tools (refs are SURFACE SEQS, not ids):
+- compress: replace one or more seq ranges, each with your own dense summary. Single range: compress({ content: [{ startSeq, endSeq, summary }] }). Batch multiple unrelated segments in one call (each entry becomes its own block): compress({ content: [{ startSeq: 1, endSeq: 5, summary: '...' }, { startSeq: 12, endSeq: 18, summary: '...' }] }). Keep ranges disjoint \u2014 overlapping entries in one batch are skipped. Edges are auto-balanced to tool-call/result boundaries; a trailing #callId fragment in a seq is ignored. Seq refs must be on the current surface: seqs from older nudges or earlier compresses go stale as the surface moves, so a stale span is auto-remapped to its still-live remainder (the result reports the adjusted span), a fully compressed span is reported as already compressed, and invented/other-session seqs fail with guidance.
+- decompress: recover a compressed block's original content, read-only. decompress({ blockId }).
+- search_context: find information inside compressed blocks BEFORE decompressing. search_context({ query }).
+- acp_status: current context usage and the live compressible-range list. Run it right before compressing \u2014 the only seqs that never go stale are the ones you just read.
+
+Tiered compression: each compressed block appears on the surface as one summary node. Compressing that node again DISTILLS the block (tier 2): the parent summary folds into your new summary and the original messages are freed. Distilling a tier-2 block yields tier 3. Distill when a summary itself is consumed \u2014 decompress on the tier-2 block recovers the full originals.
+
+When you write a summary, it becomes the ONLY record of that range: keep file paths, signatures, exact values, decisions, and error strings verbatim so a later reader (or you, after decompress) can continue without the original. Never reuse historical seqs \u2014 the surface moves as messages land and compress; verify with acp_status.`
+};
+var DEFAULT_RESOLVED = DEFAULT_PROMPTS;
+
+// src/nudge.ts
 function resolveTokenCount(agent, coreMessages) {
   const projections = agent.ctx?.get?.("sessionProjections");
   const projected = projections?.snapshot?.(agent.session)?.values?.contextPressure?.projectedTokens;
@@ -594,17 +699,24 @@ function resolveTokenCount(agent, coreMessages) {
   if (typeof surface === "number" && surface > 0) return surface;
   return coreMessages.reduce((sum, message) => sum + defaultCountTokens2(message.text ?? ""), 0);
 }
-function rangeTable(session) {
+function rangeTable(session, prompts = DEFAULT_RESOLVED) {
   const ranges = buildCompressibleSeqRanges(session).slice(0, 6);
   if (ranges.length === 0) return "";
-  const lines = ranges.map((range) => `  - seq ${range.start}..${range.end} \u2014 ${range.count} messages, ~${range.tokens} tokens`);
+  const lines = ranges.map(
+    (range) => renderTemplate(prompts.rangeTable.line, {
+      start: range.start,
+      end: range.end,
+      count: range.count,
+      tokens: range.tokens
+    })
+  );
   return [
+    // 前导空串元素产生 nudge 中范围表前的唯一空行(§4:parts 层不再加分隔)。
     "",
-    `Surface: ${surfaceSummary(session)}`,
-    "Compressible ranges (suggestions only \u2014 compress any consumed span; refs are surface seqs):",
+    renderTemplate(prompts.rangeTable.header, { surface: surfaceSummary(session) }),
+    renderTemplate(prompts.rangeTable.title, { count: ranges.length }),
     ...lines,
-    "Compress with: compress({ content: [{ startSeq, endSeq, summary }] }) \u2014 content is an array: batch multiple unrelated segments in one call, each entry its own block. Keep ranges disjoint.",
-    "Snapshot taken at nudge time: the seqs go stale once the surface moves (a later compress shadows them), so re-run acp_status for fresh refs before compressing."
+    prompts.rangeTable.footer
   ].join("\n");
 }
 function measuredTokenCount(agent, coreMessages) {
@@ -626,28 +738,36 @@ function buildNudge(agent, env, lastNudgeTurn) {
   const alreadyShown = !emergency && lastNudgeTurn.get(session.id) === turnNumber;
   if (alreadyShown) return null;
   lastNudgeTurn.set(session.id, turnNumber);
-  const text = buildNudgeText(nudge, emergency, session);
+  const text = buildNudgeText(nudge, emergency, session, env.prompts);
   const message = createUserMessage2({
     content: [{ type: "text", text }],
     source: { kind: "plugin", plugin: "acp-nudge" }
   });
   return { message, emergency };
 }
-function buildNudgeText(nudge, emergency, session) {
+function buildNudgeText(nudge, emergency, session, prompts = DEFAULT_RESOLVED) {
   const pct = Math.round(Math.min(nudge.contextUsage, 1) * 100);
-  const frame = emergency ? `\u26A0\uFE0F Context usage is at ${pct}% of the window \u2014 nearly full. Consider compressing consumed ranges soon so working context stays available; the choice and timing are yours.` : `Context usage is at ${pct}%. This is a suggestion, not a requirement \u2014 you decide whether and when to compress.`;
-  const guidance = "Compress by need, not by percentage: replace only ranges you have genuinely consumed, with dense self-contained summaries.";
-  const parts = [frame, "", guidance];
+  const frame = renderTemplate(
+    emergency ? prompts.nudge.emergency : prompts.nudge.normal,
+    { pct }
+  );
+  const parts = [frame];
+  if (prompts.nudge.guidance !== "") parts.push("", prompts.nudge.guidance);
   if ((nudge.tier === 2 || nudge.tier === 3) && (nudge.tierTargetBlocks?.length ?? 0) > 0) {
     const targets = nudge.tierTargetBlocks;
     const summarySeqs = targets.map((block) => summarySeqOfKernelBlock(session, block.blockId)).filter((seq) => seq !== null);
     const pending = nudge.tier === 2 ? nudge.breakdown?.pendingT2 : nudge.breakdown?.pendingT3;
     const tokens = typeof pending === "number" ? pending : 0;
-    parts.push(
-      `Tier ${nudge.tier}: ${targets.length} tier-${nudge.tier - 1} block(s) distillable (${tokens} tokens) \u2014 compress their summary node(s) [seqs ${summarySeqs.join(", ")}] to reclaim the original messages.`
-    );
+    const tierLine = renderTemplate(prompts.nudge.tier, {
+      tier: nudge.tier,
+      count: targets.length,
+      prevTier: nudge.tier - 1,
+      tokens,
+      seqs: summarySeqs.join(", ")
+    });
+    if (tierLine !== "") parts.push(tierLine);
   }
-  parts.push(rangeTable(session));
+  parts.push(rangeTable(session, prompts));
   return parts.join("\n");
 }
 
@@ -935,10 +1055,11 @@ async function handleStatus(env, _args, exec) {
   return { text: lines.join("\n") };
 }
 function makeTools(env) {
+  const prompts = env.prompts ?? DEFAULT_RESOLVED;
   return [
     defineTool({
       name: "compress",
-      description: "Replace older conversation ranges with dense summaries you write. Each message seq is a surface reference. Single range: compress({ content: [{ startSeq, endSeq, summary }] }). Batch multiple unrelated ranges in one call (each content entry becomes its own block); keep ranges disjoint. Never compress content the current step is actively using. Seq refs must come from the CURRENT surface (acp_status or the latest nudge): a span whose edges were shadowed by an earlier compress is auto-remapped to its still-live content, a fully compressed span is reported as already compressed, and invented/other-session seqs fail with guidance.",
+      description: prompts.tools.compress,
       parameters: compressParameters,
       output: textOutput(),
       async execute(args, exec) {
@@ -947,7 +1068,7 @@ function makeTools(env) {
     }),
     defineTool({
       name: "decompress",
-      description: "Recover the original content of a compressed block by its blockId (read-only; does not unshadow the range).",
+      description: prompts.tools.decompress,
       parameters: decompressParameters,
       output: textOutput(),
       execute(args, exec) {
@@ -956,7 +1077,7 @@ function makeTools(env) {
     }),
     defineTool({
       name: "search_context",
-      description: "Search inside compressed blocks (summaries and original content) for information the model no longer sees in context.",
+      description: prompts.tools.searchContext,
       parameters: searchParameters,
       output: textOutput(),
       execute(args, exec) {
@@ -965,7 +1086,7 @@ function makeTools(env) {
     }),
     defineTool({
       name: "acp_status",
-      description: "Report the ACP block ledger: compressed blocks, reclaimed tokens, and current context pressure.",
+      description: prompts.tools.acpStatus,
       parameters: statusParameters,
       output: textOutput(),
       execute(args, exec) {
@@ -1062,22 +1183,7 @@ function acpCommand(env) {
 }
 
 // src/system-prompt.ts
-import { COMPRESS_PHILOSOPHY } from "acp-kernel";
-var ACP_SYSTEM_PROMPT = `Active Context Pruning \u2014 model-driven context management
-
-YOU decide whether and when to compress context. Nothing forces you: the injected "nudge" is a suggestion, not an order, and you may ignore it when compression would not help. Compress only ranges you have genuinely consumed (read tool outputs, finished explorations, superseded steps) that the current work no longer needs verbatim.
-
-${COMPRESS_PHILOSOPHY}
-
-Compression tools (refs are SURFACE SEQS, not ids):
-- compress: replace one or more seq ranges, each with your own dense summary. Single range: compress({ content: [{ startSeq, endSeq, summary }] }). Batch multiple unrelated segments in one call (each entry becomes its own block): compress({ content: [{ startSeq: 1, endSeq: 5, summary: '...' }, { startSeq: 12, endSeq: 18, summary: '...' }] }). Keep ranges disjoint \u2014 overlapping entries in one batch are skipped. Edges are auto-balanced to tool-call/result boundaries; a trailing #callId fragment in a seq is ignored. Seq refs must be on the current surface: seqs from older nudges or earlier compresses go stale as the surface moves, so a stale span is auto-remapped to its still-live remainder (the result reports the adjusted span), a fully compressed span is reported as already compressed, and invented/other-session seqs fail with guidance.
-- decompress: recover a compressed block's original content, read-only. decompress({ blockId }).
-- search_context: find information inside compressed blocks BEFORE decompressing. search_context({ query }).
-- acp_status: current context usage and the live compressible-range list. Run it right before compressing \u2014 the only seqs that never go stale are the ones you just read.
-
-Tiered compression: each compressed block appears on the surface as one summary node. Compressing that node again DISTILLS the block (tier 2): the parent summary folds into your new summary and the original messages are freed. Distilling a tier-2 block yields tier 3. Distill when a summary itself is consumed \u2014 decompress on the tier-2 block recovers the full originals.
-
-When you write a summary, it becomes the ONLY record of that range: keep file paths, signatures, exact values, decisions, and error strings verbatim so a later reader (or you, after decompress) can continue without the original. Never reuse historical seqs \u2014 the surface moves as messages land and compress; verify with acp_status.`;
+var ACP_SYSTEM_PROMPT = renderSystemPrompt(DEFAULT_PROMPTS);
 var ACP_SYSTEM_PROMPT_ORDER = 150;
 
 // src/index.ts
@@ -1104,12 +1210,15 @@ var AcpCompactionEngine = class extends CompactionEngine {
   store;
   /** Resolved engine configuration. */
   config;
+  /** Resolved prompt templates (validated at construction — fail-fast on template typos). */
+  prompts;
   lastNudgeTurn = /* @__PURE__ */ new Map();
   /** Per provider/model route the resolved window (probe failures cached too). */
   windowCache = /* @__PURE__ */ new Map();
   constructor(ctx, config = {}) {
     super(ctx);
     this.config = resolveAcpConfig(config);
+    this.prompts = resolvePrompts(config.prompts);
     const ports = this.config.countTokens !== void 0 ? { countTokens: this.config.countTokens } : {};
     this.kernel = createCore(ports);
     this.store = new AcpStateStore();
@@ -1122,7 +1231,8 @@ var AcpCompactionEngine = class extends CompactionEngine {
       nudgeMaxContextLimitPct: this.config.nudgeMaxContextLimitPct,
       nudgeEmergencyThresholdPct: this.config.nudgeEmergencyThresholdPct,
       coreOverrides: this.config.coreOverrides,
-      windowFor: (agent) => this.windowFor(agent)
+      windowFor: (agent) => this.windowFor(agent),
+      prompts: this.prompts
     };
     const tools = ctx.get("tools");
     if (tools !== void 0) {
@@ -1171,7 +1281,7 @@ var AcpCompactionEngine = class extends CompactionEngine {
       systemPrompt.section({
         name: "billion-context-dsh",
         order: ACP_SYSTEM_PROMPT_ORDER,
-        text: ACP_SYSTEM_PROMPT
+        text: renderSystemPrompt(this.prompts)
       });
     } else {
       let done = false;
@@ -1183,7 +1293,7 @@ var AcpCompactionEngine = class extends CompactionEngine {
         registry.section({
           name: "billion-context-dsh",
           order: ACP_SYSTEM_PROMPT_ORDER,
-          text: ACP_SYSTEM_PROMPT
+          text: renderSystemPrompt(this.prompts)
         });
       };
       ctx.on("internal/service", (name) => {
@@ -1248,6 +1358,8 @@ export {
   AcpStateStore,
   AlreadyCompressedRangeError,
   DEFAULT_CONTEXT_WINDOW,
+  DEFAULT_PROMPTS,
+  DEFAULT_RESOLVED,
   acpCommand,
   assertNoActiveCompaction,
   blockRefForSummarySeq,
@@ -1264,7 +1376,10 @@ export {
   makeTools,
   projectEvent,
   rebuildBlockLedger,
+  renderSystemPrompt,
+  renderTemplate,
   resolveAcpConfig,
+  resolvePrompts,
   resolveSurfaceRange,
   resolveTokenCount,
   runCompactionTransaction,