Docs

• Added md_to_pdf.py to the project tree (it was missing from What ships).
• Corrected the cheatsheet inputs description — errors/log.md is the seed for weakmap and the heart of the cheatsheet's traps section, not its only input (cheatsheet also reads patterns.md, coverage.md, and summary.md).
• Documented the harness environment variables — PAIDEIA_MODEL, PAIDEIA_TIMEOUT, PAIDEIA_ASK_PERMISSIONS, PAIDEIA_PYTHON — and the opencode.json workspace config, in both README and README.ko.

Docs

• README and README.ko note that cheatsheet --pdf requires reportlab (or pandoc).

Added

• Safety timeout on the opencode run. The opencode run invocation is wrapped in a generous timeout (default 30 min, override with PAIDEIA_TIMEOUT in seconds) so a hung or runaway run fails with a clear message instead of blocking the harness indefinitely.

Changed

• Single ingest summary table. The transcription agent no longer prints its own Converted/Skipped/Failed table; the harness prints the authoritative one (it alone knows the skipped/failed counts).

Fixed

• Stale grade scratch. grade now clears any leftover answers/converted/.tmp-<stem>/ from an interrupted run before rasterizing the answer PDF, so a retry starts clean.

Added

• Reliable cheatsheet --pdf. The PDF is now produced deterministically by the harness (assets/scripts/md_to_pdf.py — pandoc when a LaTeX engine is available, otherwise a self-contained reportlab renderer) instead of relying on the agent to run the conversion, which it often skipped. The prompt no longer asks the model to generate the PDF.

Note: cheatsheet --pdf needs reportlab (pip install --user reportlab) or pandoc.

Fixed

• Premature phase detection. Phase and weakmap detection misread the schema example inside the errors/log.md seed comment (which literally contains - problem_id: <id>) as a real graded entry — pushing a brand-new course straight to phase=drill and falsely reporting logged errors. HTML comments are now stripped before matching, so a fresh course correctly reports diag.

First release of PAIDEIA for opencode — a standalone exam-prep harness that drives opencode rather than loading as an editor plugin. The harness owns the study logic and composes a full stage specification per run; opencode is the execution engine (model, tools, auth).

Highlights

• 16 commands + status covering the full formation cycle: ingest → analyze → drill → grade → weakmap → cheatsheet.
• Vision ingest: every course PDF is rendered/resized deterministically in the harness, then transcribed to faithful LaTeX markdown by parallel opencode subagents.
• HW density = exam probability weighting across every drill (🔥🔥 / 🔥 / 🟡 / ⚪ tiers).
• Strategy-based grading of hand-written answer scans; three OCR engines — agent-vision (default), local Qwen3-VL via ollama, and tesseract.
• Drills: quiz, two-phase blind, twin, chain, mock; review: weakmap, error-driven cheatsheet.
• OPTIMETA Exam Radar import (alt) folds the lecture-emphasis signal into coverage.md beside HW density.
• en/ko interface, zero-dependency ESM, --dry-run to inspect the exact opencode command and composed spec.

Install

git clone https://github.com/OPTIMETA/PAIDEIA-opencode && cd PAIDEIA-opencode
npm link && paideia doctor