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

fix: preserve source data and recover tool math at presentation

Browse Source
This commit is contained in:
amosblomqvist
2026-10-26 04:49:47 +02:00
parent b277ee3e0f
commit d3c8709bd2
21 changed files with 1134 additions and 652 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/tool-parts.test.ts
+30 -32
View File
@@ -12,7 +12,7 @@ import assert from "node:assert/strict";
import { describe, it } from "node:test";
import { quizInputProblems } from "../src/core/question-core.ts";
import { Value } from "typebox/value";
import { restoreMathEscapes } from "../src/core/math-delims.ts";
import { repairMathForDisplay, restoreMathEscapes } from "../src/core/math-delims.ts";
import {
parseAskInput,
parseProbeDoneInput,
@@ -170,7 +170,7 @@ describe("LaTeX-escape restoration (the JSON-debris fix, owner ruling 2026-09-26
// The exact live shapes from the modal-logic session: single-backslash
// LaTeX inside JSON args — JSON escape resolution ate the command's
// first letter (\f → FF ate 'f' of \forall, \v → VT ate 'v' of \varphi).
it("parseQuizInput restores escape debris inside math spans (question + options + explanation)", () => {
it("presentation restores debris from parsed source fields (question + options + explanation)", () => {
const parsed = parseQuizInput({
question: "What does \\(\\forall x\\, \\varphi(x)\\) mean?",
options: [
@@ -182,18 +182,18 @@ describe("LaTeX-escape restoration (the JSON-debris fix, owner ruling 2026-09-26
});
assert.ok(parsed);
assert.equal(parsed.question, "What does \\(\\forall x\\, \\varphi(x)\\) mean?");
assert.equal(parsed.options[0]!.label, "\\(\\varphi\\) is true when every element satisfies it");
assert.equal(parsed.options[1]!.label, "\\(\\forall\\) and \\(\\exists\\) are just shorthand");
assert.equal(parsed.explanation, "\\(\\forall\\) quantifies over the domain.");
assert.equal(repairMathForDisplay(parsed.options[0]!.label), "\\(\\varphi\\) is true when every element satisfies it");
assert.equal(repairMathForDisplay(parsed.options[1]!.label), "\\(\\forall\\) and \\(\\exists\\) are just shorthand");
assert.equal(repairMathForDisplay(parsed.explanation), "\\(\\forall\\) quantifies over the domain.");
});
it("parseAskInput restores debris in ask option labels too", () => {
it("presentation restores debris from parsed ask option labels too", () => {
const parsed = parseAskInput({
question: "Goal?",
options: [{ label: "\\(\x0corall\\) rigor" }, { label: "Intuition first" }],
});
assert.ok(parsed);
assert.equal(parsed.options![0]!.label, "\\(\\forall\\) rigor");
assert.equal(repairMathForDisplay(parsed.options![0]!.label), "\\(\\forall\\) rigor");
});
it("restores TAB debris (the \\to case) inside spans only", () => {
@@ -204,7 +204,7 @@ describe("LaTeX-escape restoration (the JSON-debris fix, owner ruling 2026-09-26
explanation: "",
});
assert.ok(parsed);
assert.equal(parsed.question, "Is \\(p\\to q\\) valid?");
assert.equal(repairMathForDisplay(parsed.question), "Is \\(p\\to q\\) valid?");
});
it("control chars OUTSIDE math spans are left alone (real prose/pasted text)", () => {
@@ -253,7 +253,7 @@ describe("display-math newline rule (Option A, owner ruling 2026-09-27)", () =>
explanation: "",
});
assert.ok(parsed);
assert.equal(parsed.question, "Evaluate \\[\\left( a \\right)\\] carefully.");
assert.equal(repairMathForDisplay(parsed.question), "Evaluate \\[\\left( a \\right)\\] carefully.");
});
it("sloppy \\neq in display math stays broken — the accepted cost (newline kept, byte-identical to a real line break)", () => {
@@ -276,7 +276,7 @@ describe("display-math newline rule (Option A, owner ruling 2026-09-27)", () =>
explanation: "",
});
assert.ok(parsed);
assert.equal(parsed.question, "Is \\(p \\neq q\\) valid?");
assert.equal(repairMathForDisplay(parsed.question), "Is \\(p \\neq q\\) valid?");
});
});
@@ -289,9 +289,9 @@ describe("dollar-delimiter recognition — existing escape restoration rules", (
explanation: "Here $p\to q$ is an implication.",
});
assert.ok(parsed);
assert.equal(parsed.question, "What does $\\forall x$ mean?");
assert.equal(parsed.options[0]!.label, "$\\varphi$ holds");
assert.equal(parsed.explanation, "Here $p\\to q$ is an implication.");
assert.equal(repairMathForDisplay(parsed.question), "What does $\\forall x$ mean?");
assert.equal(repairMathForDisplay(parsed.options[0]!.label), "$\\varphi$ holds");
assert.equal(repairMathForDisplay(parsed.explanation), "Here $p\\to q$ is an implication.");
});
it("keeps the same six-character inline recovery map", () => {
@@ -320,21 +320,21 @@ describe("dollar-delimiter recognition — existing escape restoration rules", (
}
});
it("preserves legacy replacement frequency when a span has multiple debris characters", () => {
it("restores all debris in one pass consistently across delimiter syntaxes", () => {
const body = "p\to q, \rho";
const legacy = restoreMathEscapes(`\\(${body}\\)`).slice(2, -2);
assert.equal(restoreMathEscapes(`$${body}$`), `$${legacy}$`);
assert.equal(legacy, "p\\to q, \rho", "the existing first-match-only behavior is unchanged");
assert.equal(legacy, "p\\to q, \\rho", "recovery converges instead of requiring a second render");
});
it("shares dollar recovery with ask and free-recall fields", () => {
const ask = parseAskInput({ question: "Goal?", options: [{ label: "$\x0corall$ rigor" }, { label: "Intuition" }] });
assert.ok(ask);
assert.equal(ask.options![0]!.label, "$\\forall$ rigor");
assert.equal(repairMathForDisplay(ask.options![0]!.label), "$\\forall$ rigor");
const explain = parseExplainInput({ question: "Explain $p\to q$", correctAnswer: "$p\neq q$ is different." });
assert.ok(explain);
assert.equal(explain.question, "Explain $p\\to q$");
assert.equal(explain.correctAnswer, "$p\\neq q$ is different.");
assert.equal(repairMathForDisplay(explain.question), "Explain $p\\to q$");
assert.equal(repairMathForDisplay(explain.correctAnswer), "$p\\neq q$ is different.");
});
it("leaves non-math debris, escaped opening dollars, and code examples untouched", () => {
@@ -401,13 +401,9 @@ describe("parseExplainInput: the dock's display seam for explain calls", () => {
});
});
describe("over-escaped command heal at the parse seam (owner ruling 2026-10-15)", () => {
// xiaomi/mimo-v2.6-pro and deepseek-v4.1-flash stochastically DOUBLE
// LaTeX command backslashes inside JSON tool arguments — per-call and
// per-field (gradient-descent msg 3: question doubled, explanation and
// options in the SAME call clean). restoreMathEscapesDeep now rides
// healOverescapedCommands after the debris restore, so doubled spans
// heal at every display parse: live dock, settled cards, old records.
describe("over-escaped command recovery after shape parsing", () => {
// Keep these real authoring shapes, but recover only for presentation.
// Parsed source is also used to build durable answers and identities.
const CLEAN =
"Let $f : \\mathbb{R}^n \\to \\mathbb{R}$ be differentiable at $x$, with $\\nabla f(x) \\neq 0$. Among all unit directions $\\|u\\|_2 = 1$, which makes $D_u f(x)$ most negative?";
const DOUBLE = (s: string): string => s.replace(/\\/g, "\\\\");
@@ -420,19 +416,21 @@ describe("over-escaped command heal at the parse seam (owner ruling 2026-10-15)"
explanation: "By Cauchy\u2013Schwarz: $$u^\\star = -\\frac{\\nabla f(x)}{\\|\\nabla f(x)\\|_2}$$",
};
it("parseQuizInput heals a doubled question field (the gradient-descent msg-3 shape)", () => {
it("presentation heals a doubled question field (the gradient-descent msg-3 shape)", () => {
const parsed = parseQuizInput({ ...BASE, question: DOUBLE(CLEAN) });
assert.ok(parsed, "parses");
assert.equal(parsed.question, CLEAN);
assert.equal(parsed.question, DOUBLE(CLEAN), "shape parsing retains source");
assert.equal(repairMathForDisplay(parsed.question), CLEAN);
// The CLEAN sibling fields of the same call stay byte-identical.
assert.equal(parsed.options[0]!.label, BASE.options[0]!.label);
assert.equal(parsed.explanation, BASE.explanation);
});
it("parseQuizInput heals a quadrupled field (the agent's over-correction, msg 6)", () => {
it("presentation heals a quadrupled field (the agent's over-correction, msg 6)", () => {
const parsed = parseQuizInput({ ...BASE, question: DOUBLE(DOUBLE(CLEAN)) });
assert.ok(parsed, "parses");
assert.equal(parsed.question, CLEAN);
assert.equal(parsed.question, DOUBLE(DOUBLE(CLEAN)), "shape parsing retains source");
assert.equal(repairMathForDisplay(parsed.question), CLEAN);
});
it("doubled option labels and explanations heal too (the inference-runtime msg-7 shape)", () => {
@@ -448,10 +446,10 @@ describe("over-escaped command heal at the parse seam (owner ruling 2026-10-15)"
assert.ok(parsed, "parses");
assert.equal(parsed.options[0]!.label, "$mkn$");
assert.equal(parsed.options[1]!.label, "$2mkn$");
assert.equal(parsed.explanation, "$$c_{ij} = \\sum_{\\ell=1}^{k} a_{i\\ell}\\, b_{\\ell j}$$");
assert.equal(repairMathForDisplay(parsed.explanation), "$$c_{ij} = \\sum_{\\ell=1}^{k} a_{i\\ell}\\, b_{\\ell j}$$");
});
it("parseAskInput heals doubled ask-option descriptions the same way", () => {
it("presentation heals doubled ask-option descriptions the same way", () => {
const parsed = parseAskInput({
question: "Depth?",
options: [
@@ -460,7 +458,7 @@ describe("over-escaped command heal at the parse seam (owner ruling 2026-10-15)"
],
});
assert.ok(parsed, "parses");
assert.equal(parsed.options![0]!.description, "why $-\\nabla f$ is steepest descent");
assert.equal(repairMathForDisplay(parsed.options![0]!.description!), "why $-\\nabla f$ is steepest descent");
});
it("clean records pass through byte-identical (no over-heal)", () => {