sorokin.oleg.edu/learnharness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

tool-result diet: question answers return only learner-added truth (compact JSON for the model, full record in details); dead-code sweep

Browse Source
This commit is contained in:
amosblomqvist
2026-10-02 04:04:04 +03:00
parent 9a2e434748
commit a85087528e
8 changed files with 131 additions and 212 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/core/native-messages.ts
+7 -4
View File
@@ -37,6 +37,7 @@ import type { AgentMessage } from "@earendil-works/pi-agent-core";
import type { AssistantMessage, Message, ToolResultMessage, UserMessage } from "@earendil-works/pi-ai";
import { QUESTION_TOOLS } from "./pending.ts";
import { HANDOFF_TEXT } from "./contracts/chat.ts";
import { compactAnswerForModel, type QuestionAnswerData } from "./question-core.ts";
/** The hard-invariant wording for an aborted turn's unanswered calls —
* carried over verbatim from the pre-native era so the model's view
@@ -103,9 +104,11 @@ export function newUserMessage(text: string): LhMessage & UserMessage {
}
/** The learner's answer — the question call's tool result (fork 1).
* `answer` is the structured QuestionAnswerData the client graded; the
* MODEL sees its JSON encoding (the pre-native wire shape),
* the UI sees the structured payload via `details`. */
* TWO AUDIENCES (owner, tool-result diet 2026-10-02): the MODEL sees the
* compact learner-only record (compactAnswerForModel — the call's own
* arguments already carry the question/options/explanation/commitment;
* restating them doubled every answer's tokens), while `details` keeps
* the FULL graded QuestionAnswerData for the record and the UI. */
export function newAnswerResult(
toolCallId: string,
toolName: string,
@@ -116,7 +119,7 @@ export function newAnswerResult(
role: "toolResult",
toolCallId,
toolName,
content: [{ type: "text", text: JSON.stringify(answer) }],
content: [{ type: "text", text: JSON.stringify(compactAnswerForModel(answer as QuestionAnswerData)) }],
// 0.87.1 narrowed ToolResultMessage.details to JsonRepresentation —
// our payloads are JSON-safe answer structures, so the cast is honest.
details: answer as ToolResultMessage["details"],
src/core/question-core.ts
+27 -86
View File
@@ -339,9 +339,10 @@ export interface QuestionData {
createdAt: string;
}
/** The learner's answer — full structured record, stored in `lh_question_answered`. */
/** The learner's answer — full structured record, stored in `lh_question_answered`.
* NO questionId (owner, tool-result diet 2026-10-02): it was the
* toolCallId restated — the toolResult message already carries it. */
export interface QuestionAnswerData {
questionId: string;
kind: QuestionKind;
questionText: string;
options?: { index: number; label: string }[]; // snapshot for rendering
@@ -425,9 +426,6 @@ export function toggleOther(mode: SelectMode, sel: DockSelection): DockSelection
return { selected: new Set(), otherSelected: true, dontKnow: false };
}
export const QUESTION_ENTRY = "lh_question";
export const QUESTION_ANSWERED_ENTRY = "lh_question_answered";
// ─── options ────────────────────────────────────────────────────────────────
export function normalizeOptions(
@@ -546,10 +544,6 @@ export function isCorrect(selectedIndices: number[], correctIndices: number[]):
// ─── arming + answering ─────────────────────────────────────────────────────
export function makeId(): string {
return `q_${Date.now().toString(36)}_${Math.random().toString(36).slice(2, 10)}`;
}
export function toClientQuestion(qd: QuestionData): ClientQuestion {
return {
id: qd.id,
@@ -572,7 +566,6 @@ function formatOptionRef(options: QuizOption[], index: number): string {
*/
export function buildAnswer(qd: QuestionData, req: AnswerRequest): QuestionAnswerData {
const base: QuestionAnswerData = {
questionId: qd.id,
kind: qd.kind,
questionText: qd.question,
options: qd.options.map((o, i) => ({ index: i + 1, label: o.label })),
@@ -613,82 +606,30 @@ export function buildAnswer(qd: QuestionData, req: AnswerRequest): QuestionAnswe
return base;
}
/** The user message that carries the answer into the agent's context. */
export function formatAnswerMessage(ad: QuestionAnswerData): string {
const label = ad.kind === "quiz" ? "Quiz answer" : ad.kind === "explain" ? "Explain answer" : "Answer to your question";
const head = `[${label}]`;
const lines: string[] = [head];
if (ad.dismissed) {
lines.push("The learner dismissed this question without answering.");
return lines.join("\n");
}
if (ad.kind === "quiz") {
const correctStr = (ad.correctIndices ?? []).map((i) => formatOptionRefSnapshot(ad, i)).join(", ");
if (ad.dontKnow) {
lines.push(
`The learner selected "I don't know" — they did NOT attempt an answer (a genuine knowledge gap, not a wrong guess).`,
);
} else {
const verdict = ad.correct ? "CORRECT" : "INCORRECT";
const selectedStr = ad.answers.map((a) => `${a.index}. ${a.label}`).join(", ");
lines.push(`Selected: ${selectedStr || "(none)"}`);
lines.push(`Result: ${verdict}`);
if (ad.confidence !== undefined) {
lines.push(`Confidence: ${ad.confidence}% — the learner's stated probability that his selected answer is correct.`);
}
}
lines.push(`Correct: ${correctStr}`);
} else if (ad.kind === "explain") {
// No opt-out row on explain (owner ruling 2026-09-27): if he doesn't
// know, he says so in his own words — his text is all there is.
lines.push(`The learner answered: ${ad.text ?? "(no answer)"}`);
if (ad.committed) {
lines.push(`Your committed correct answer (immutable — grade against it, never revise it): ${ad.committed}`);
}
lines.push("Grade his answer strictly against your committed answer in your reply: name what he got and what he missed.");
} else {
if (ad.text) {
lines.push(`The learner answered: ${ad.text}`);
} else if (ad.other) {
lines.push(`The learner answered (Other): ${ad.other}`);
} else if (ad.answers.length > 0) {
const selectedStr = ad.answers.map((a) => `${a.index}. ${a.label}`).join(", ");
lines.push(`Selected: ${selectedStr}`);
} else {
lines.push("(no answer)");
}
}
if (ad.note) lines.push(`Note: ${ad.note}`);
return lines.join("\n");
}
function formatOptionRefSnapshot(ad: QuestionAnswerData, index: number): string {
const opt = (ad.options ?? []).find((o) => o.index === index);
return `${index}. ${opt ? opt.label : "(unknown)"}`;
}
/**
* Derive the pending (oldest unanswered) question from a session's entries.
* Used on resume/backfill — pending state lives in the session file itself.
* The MODEL-VISIBLE answer record (owner, tool-result diet 2026-10-02):
* the tool call already carries everything the model itself authored —
* the question, every option, the correct answer, the explanation, the
* explain commitment — so the result returns ONLY what the learner added:
* his selection or words, the graded verdict, his stated confidence, his
* note. The full graded QuestionAnswerData rides `details` for the record
* and the UI; this is the JSON the agent actually reads.
*/
export function findPendingQuestion(
entries: Array<{ type: string; customType?: string; data?: unknown }>,
): ClientQuestion | null {
const answered = new Set<string>();
for (const e of entries) {
if (e.type === "custom" && e.customType === QUESTION_ANSWERED_ENTRY) {
const d = e.data as QuestionAnswerData | undefined;
if (d?.questionId) answered.add(d.questionId);
}
}
for (const e of entries) {
if (e.type === "custom" && e.customType === QUESTION_ENTRY) {
const d = e.data as QuestionData | undefined;
if (d && !answered.has(d.id)) return toClientQuestion(d);
}
}
return null;
export function compactAnswerForModel(ad: QuestionAnswerData): Record<string, unknown> {
if (ad.dismissed) return { kind: ad.kind, dismissed: true };
const out: Record<string, unknown> = { kind: ad.kind };
if (ad.dontKnow) out.dontKnow = true;
// Array.isArray: the seam is called with `unknown` answers (test
// fixtures, foreign records) — degrade to omission, never crash.
if (Array.isArray(ad.answers) && ad.answers.length > 0)
out.answers = ad.answers.map((a) => ({ index: a.index, label: a.label }));
if (ad.text) out.text = ad.text;
if (ad.other) out.other = ad.other;
// The verdict + stated confidence are the learner's PERFORMANCE — new
// information the model needs for its next move. (dontKnow implies
// incorrect and carries no confidence: both stay implicit.)
if (ad.correct !== undefined && !ad.dontKnow) out.correct = ad.correct;
if (ad.confidence !== undefined) out.confidence = ad.confidence;
if (ad.note) out.note = ad.note;
return out;
}