perf(rectification): evidence turns get Skill §5/§7 only; drop Mastra <available_skills> injection (R4)

T6 of TASK-rectification-grounding-20260927.
- skill-slice.ts: per-action slice rule as a code constant, keyed on heading
  titles (evidence: 「ConversationFocus 与意图承接」「批量证据与日期真实性」, i.e.
  §5/§7 of the 10.0.x layout); other actions and any bound Skill without those
  headings (9.0.0) get the whole body, so historical Cases still run with their
  exact bound Skill (BUG-621).
- The rectification Agent declares providesSkillDiscovery "on-demand"
  (rectificationSkillBoundProcessor): no <available_skills> block with a temp
  path and no "call the skill tool" system message; getSkill still loads the
  bound package.
- Measured on a real Agent + recording model (public AA case, estimate = CJK
  chars + other chars / 4): fixed overhead per call 12,002 → 6,471 tokens
  (step 0: 8,440 → 2,909). Skill text unchanged; no version bump.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
This commit is contained in:
Jesse_Chen
2026-09-27 03:28:27 +08:00
co-authored by Claude Opus 5.5
parent 1ed3e55579
commit c4bb394bbf
4 changed files with 301 additions and 1 deletions
@@ -10,6 +10,7 @@ import type { V9CaseDossier } from "./tool-service";
import { cachedSystemMessage } from "../../agent-generation-settings.ts";
import { OPENING_COLLECT_DOMAINS } from "../user-copy";
import { retryConstraintForAttempt } from "./host-fallback";
import { sliceRectificationSkill } from "./skill-slice";
import type { V9AgentRunOptions } from "./agent-run";
function clockWindow(range: { start_time?: string | null; end_time?: string | null } | null | undefined): string | null {
@@ -54,9 +55,12 @@ export function buildAgentMessages(
const timeContext = options.timeContext
?? `服务端当前时间(权威):${new Date().toISOString()}。涉及“现在、今天、今年、未来几个月”等相对时间时,以此为准。`;
const caseContext = `【服务端 Case ID】${options.caseId}。所有 rectification 工具调用的 caseId 必须原样使用此值。`;
// T6 (TASK-rectification-grounding-20260927): an evidence turn gets only
// the Skill sections it uses; other actions keep the whole bound body.
const boundSkill = sliceRectificationSkill(skillInstructions, options.action).text;
const bootstrapContent = [
"【服务器已绑定当前 Case 的精确 Skill】运行器已在本 attempt 内加载并核验下列指令;不要重复调用 skill。第一步必须调用 rectification-read-case。",
skillInstructions,
boundSkill,
...(attempt > 1 ? [retryConstraintForAttempt(previousErrorCode)] : []),
].join("\n\n");
const bootstrap = cachedSystemMessage(bootstrapContent, options.generationModel)
@@ -0,0 +1,71 @@
/**
* Per-turn slice of the bound rectification Skill (TASK-rectification-
* grounding-20260927 T6).
*
* The runner used to send the whole bound SKILL.md body (about 10.7K
* characters for 10.0.31) on every model call. An evidence turn only uses
* §5 (ConversationFocus and intent) and §7 (batch evidence and date
* truthfulness); the rest describes the opening, case status table, fields
* that only exist in `full_diagnostics` (method_followup_plan,
* CaseConversationSummary, guided_collect_windows …), the candidate language
* of offer turns, and upstream sync. Other actions keep the whole body.
*
* The slice is keyed on `## N. <title>` heading titles, not on numbers:
* the 9.0.0 snapshot numbers its sections differently (its §5 is the
* candidate language). If any wanted title is missing, the whole body is
* sent unchanged, so a historical Case still runs with its exact bound Skill
* (BUG-621). All 10.0.x snapshots share the 10.0.31 headings.
*/
import type { RectificationAgentAction } from "./step-budget";
/** Heading titles per action; evidence = §5 and §7 of the 10.0.x layout. */
export const RECTIFICATION_SKILL_SECTIONS_BY_ACTION: Readonly<
Record<RectificationAgentAction, readonly string[] | "all">
> = {
opening: "all",
read_only: "all",
evidence: ["ConversationFocus 与意图承接", "批量证据与日期真实性"],
rescore: "all",
accept: "all",
confirm: "all",
};
export type RectificationSkillSlice = Readonly<{
text: string;
/** Heading lines sent, or "all" when the whole body went out. */
sections: readonly string[] | "all";
}>;
function headingTitle(line: string): string | null {
const match = /^##\s+(?:\d+\.\s*)?(.+?)\s*$/.exec(line);
return match ? match[1] : null;
}
export function sliceRectificationSkill(
instructions: string,
action: RectificationAgentAction,
): RectificationSkillSlice {
const wanted = RECTIFICATION_SKILL_SECTIONS_BY_ACTION[action];
if (wanted === "all") return { text: instructions, sections: "all" };
const preamble: string[] = [];
const sections: Array<{ title: string; lines: string[] }> = [];
for (const line of instructions.split("\n")) {
const title = line.startsWith("## ") ? headingTitle(line) : null;
if (title !== null) {
sections.push({ title, lines: [line] });
continue;
}
if (sections.length > 0) sections.at(-1)!.lines.push(line);
else preamble.push(line);
}
const picked = wanted.map((title) => sections.find((section) => section.title === title));
if (picked.some((section) => !section)) return { text: instructions, sections: "all" };
const title = preamble.find((line) => line.startsWith("# ")) ?? "";
const headings = picked.map((section) => section!.lines[0]!.trim());
const note = `(本轮是证据轮:以下只附本轮适用的 Skill 章节「${picked.map((section) => section!.title).join("」「")}」;其余章节不适用于本轮,按服务器投影与工具说明执行。)`;
const body = picked.map((section) => section!.lines.join("\n").trim()).join("\n\n");
return {
text: [title, note, body].filter(Boolean).join("\n\n"),
sections: headings,
};
}
@@ -1,4 +1,5 @@
import { Agent } from "@mastra/core/agent";
import type { Processor } from "@mastra/core/processors";
import type { ResolvedLanguageModel } from "./model";
import {
resolveActiveSkillPackage,
@@ -56,8 +57,30 @@ export function getRectificationV9Agent(
model: model.model,
instructions: agenticRectificationInstructions,
skills: [resolveSkillPackageRuntimePath(skillPackage)],
inputProcessors: [rectificationSkillBoundProcessor],
tools: createRectificationV9AgentTools(ctx),
});
}
/**
* R4 (TASK-rectification-grounding-20260927): Mastra's default skills
* processor added an `<available_skills>` block (with a temporary package
* path) and "call the `skill` tool" to every step, contradicting the runner's
* "do not call skill" and costing about 1K characters per call.
* `providesSkillDiscovery: "on-demand"` declares that the caller owns skill
* loading, which is true: the runner loads the exact bound Skill through
* `agent.getSkill` and quotes it (sliced per turn) in the bootstrap message.
* With this marker Mastra adds neither the system block nor the `skill` /
* `skill_search` tools (same mechanism as the consultation agents'
* `jyotishSkillBoundProcessor`).
*/
export const rectificationSkillBoundProcessor: Processor & { processInputStep: NonNullable<Processor["processInputStep"]> } = {
id: "rectification-skill-bound",
name: "Rectification Skill Bound",
providesSkillDiscovery: "on-demand",
processInputStep() {
// Nothing to inject: the bound Skill text rides in the runner's bootstrap message.
},
};
export { RECTIFICATION_V9_SKILL_NAME };