Skip to content

CHANGELOG.md

Latest commit

 

History

History
339 lines (224 loc) · 20.8 KB

CHANGELOG.md

File metadata and controls

339 lines (224 loc) · 20.8 KB

Changelog

All notable changes to GoAgent will be documented here.

This project follows semantic versioning once public releases begin.

0.4.12 - Local KataGo Default Hotfix

Fixed

  • Changed analysis defaults back to the local computer: upgraded users who previously saved a remote Zhizi/iKataGo mode are migrated once to the local auto mode.
  • Stopped auto mode from silently falling back to Zhizi cloud when local KataGo is missing or fails. Remote compute is now used only when explicitly enabled.
  • Updated the settings copy so users can clearly tell that GoAgent is using local KataGo unless they manually choose Zhizi cloud.

Verified

  • node --test tests/local-analysis-default-contract.test.mjs tests/zhizi-cloud-engine-contract.test.mjs tests/real-teaching-engine-pool-v2-contract.test.mjs

0.4.11 - Try Move Trial Branches

Added

  • Added a temporary try-move mode for exploring variations without modifying the original SGF.
  • Added trial branch state, legal move handling, undo/clear/exit interactions, and distinct trial-stone rendering.
  • Added trial KataGo analysis through a dedicated IPC path so candidate points, PV and teacher evidence can follow the temporary branch without polluting mainline cache or winrate data.
  • Added teacher context and board-image handling so current-move teaching and freeform questions can explicitly distinguish trial branches from real game moves.

Verified

  • pnpm test
  • pnpm typecheck
  • pnpm build
  • pnpm check:teacher-quality

0.4.3 - Adaptive Analysis Runtime and Review Evidence

Added

  • Wired the adaptive analysis runtime into KataGo position, streaming position, and quick game analysis paths.
  • Added main-side analysis cache metadata, adaptive analysis profiles, teaching readiness gates, and runtime evidence so teacher output can distinguish cache hits, lower-quality evidence, and cases that should be deepened.
  • Added move classification 2.0, PV confidence, evidence bundles, timeline review models, and board variation playback helpers as the foundation for sharper review interactions.

Maintained

  • Keeps the v0.4.2 Tool-first Agent flow, Vision Evidence Chain, KataGo Trace Translator, local knowledge tools, TTS provider policy, and release artifact smoke checks.

0.4.1 - Hotfix: freeform chat unblocked + KataGo timeout buffer

Fixed

  • Manual prompts to the teacher no longer raise 棋盘图证据不完整 ("teacher task requires a board image") when the question happens to match a keyword like 为什么这里, 这里不好, or any coordinate / current-move phrase that the backend classifier interpreted as a vision-required intent. The vision requirement is now driven solely by the renderer's explicit request.mode (current-move / move-range); freeform chats route as freeform regardless of the inferred intent, and the agent gets the same boardImageAttached=false context so it falls back to SGF / KataGo tools instead of hallucinating an image.
  • KataGo whole-game analysis no longer times out partway through long games. The per-query budget is raised from 2.5 s to 5 s (and the floor from 120 s to 180 s), giving the zhizi b28 network and other heavy networks enough room to finish a 200-move sweep without aborting halfway through and leaving the winrate curve stuck on the opening moves.

0.4.0 - Lizzie-clean Winrate Timeline + zhizi b28 Bundled

Added

  • Bundled the official zhizi b28 KataGo model (kata1-zhizi-b28c512nbt-muonfd2.bin.gz) into the release so users no longer need to download it from the settings panel after installing the app.
  • Added pnpm prepare:zhizi-b28 and pnpm dist:local:mac|win|linux scripts so the release can be packaged locally without depending on GitHub Actions credits.
  • Added --extra-model support to scripts/prepare_katago_assets.mjs and a bundledModels array in data/katago/manifest.json so any extra preset can be shipped alongside the default model.

Changed

  • Redesigned the winrate timeline header and chart to a Lizzie-clean light theme: a single-line KPI bar (move counter, black winrate with delta, score lead, severity legend, range chip), a porcelain canvas surface aligned with the rest of the light theme, and severity markers (blunder / mistake / inaccuracy) now drawn directly on the curve instead of only appearing inside the hover tooltip.
  • LLM model picker no longer hardcodes provider-specific defaults like gpt-5.5 / claude-3-5-sonnet-latest. When a Base URL and API key are configured, the picker fetches the model list from the user's proxy automatically; otherwise it shows an explicit "fill in Base URL and API key" hint.
  • LLM settings auto-save on every edit (Base URL and selects on change, API key on blur), so the Test and Refresh-Models buttons immediately operate against the latest configuration without requiring a separate Save click. The Save button is replaced by an inline "Auto-saved" status indicator.

Fixed

  • The settings panel's "Apply selected weight" button no longer triggers a redundant download or surfaces a confusing error when the selected preset is already bundled with the installer. Bundled models are now detected up front and reused, and only non-bundled presets fall through to the network download path.

0.3.16 - Strict Offline Kokoro TTS

Added

  • Added a strict selected-provider TTS system with bundled Kokoro zh-CN offline synthesis as the default provider.
  • Added teacher answer playback controls and a TTS settings panel.
  • Added custom OpenAI-compatible, HTTP JSON, and local-service TTS providers that only run when explicitly selected.
  • Added Kokoro asset preparation, validation, provider-policy checks, and real strict offline synthesis smoke testing.

Changed

  • Updated release packaging so GitHub Release builds prepare Kokoro ONNX assets before packaging installers.
  • Kept the no-fallback policy: no system voice, no Web Speech, no provider chain, and no automatic provider switching.

0.3.15 - GoAgent Brand Identity

Changed

  • Renamed the product, package metadata, app id, preload API, data directory, release assets, documentation, and workflow references to GoAgent / goagent.
  • Updated the public README set to present GoAgent as an agentic AI teacher for Go / Weiqi / Baduk and the Chinese product name as 围棋智能体.
  • Added a brand contract test so old product identity strings cannot reappear in current source, docs, scripts, or tests.

Maintained

  • Preserved the v0.3.14 interactive teacher review surface: clickable move references, clickable board coordinates, and polished coordinate flash.
  • Kept the existing teacher-quality, release-quality, Real Eval, persistent KataGo engine pool, multilingual UI, and release artifact checks.

0.3.14-dev.2 - Real Eval Prompt Grounding Hotfix

Fixed

  • Tightened the real teaching eval prompt so strict runs only allow coordinates already present in the KataGo evidence coordinate set.
  • Verified strict real eval locally with bundled KataGo v1.16.4 Metal, b18 recommended model, CLIProxyAPI, and gpt-5.5.

0.3.14-dev.1 - Real Teaching Eval and Persistent KataGo Engine

Added

  • Added real local teaching evaluation that can run true KataGo analysis plus a true LLM teacher response when GOAGENT_REAL_EVAL=1 and local credentials/assets are configured.
  • Added opt-in persistent KataGo analysis engine reuse through GOAGENT_KATAGO_ENGINE_POOL=1, while keeping the existing spawn-per-batch fallback as the default.
  • Added a deep teacher-quality command that includes strict real teaching evaluation without forcing it into default CI.

Notes

  • pnpm eval:real-teaching skips cleanly when real KataGo or LLM configuration is missing.
  • pnpm eval:real-teaching:strict is intended for local/release machines with real KataGo and LLM credentials.
  • Development releases are marked as prerelease and are not promoted over the latest stable GitHub release.

0.3.10 - Windows OpenCL Runtime Bundle Hotfix

Fixed

  • Fixed the standard Windows release package so it restores the full official KataGo OpenCL runtime directory from wimi321/lizzieyzy-next instead of packaging only a bare katago.exe.
  • The standard Windows installer and portable ZIP now keep the KataGo OpenCL adjacent runtime files, including *.dll files such as compression, OpenSSL, and OpenCL loader dependencies supplied by the upstream bundle.
  • Release checks now guard the OpenCL asset source and runtime-directory packaging path.

Notes

  • GPU vendor OpenCL drivers still come from the user's graphics driver. GoAgent bundles the KataGo OpenCL runtime files, not a replacement for NVIDIA/AMD/Intel display drivers.

0.3.9 - Real Eval Gates and Teacher Sessions

Added

  • Added Real Eval / Engine Silver fixture coverage with a high-visits KataGo silver-oracle schema gate.
  • Added KataGo engine pool telemetry for task wait, run, success, error, and timeout timing, preparing the codebase for a persistent engine queue.
  • Added release artifact smoke checks and wired them into the release quality gate.
  • Added default student level, student age range, and teacher style settings, with an explicit evidence boundary so persona controls only change expression and pacing.
  • Added right-side teacher sessions with new, close/archive, list, restore, and history support instead of only clearing the chat.

Improved

  • Selectively absorbed layiku's PR #6 by keeping move-range progression analysis and board text rendering while preserving the shared move-range parser boundary.
  • Move-range review remains key-move-only for long ranges and avoids reintroducing repeated renderer runtime imports from main-only moveRange paths.
  • Release quality now carries forward grounded shape recognition engine, local pattern matcher, knowledge source-policy gates, optimized move-range review, and eval gates.

Thanks

  • Thanks to layiku for PR #6's move-range progression and board text rendering ideas, now aligned with GoAgent's evidence and quality-gate system.
  • Thanks to layiku for PR #5, PR #4, and PR #3, and thanks to wimi321 for PR #1 / PR #2.

0.3.8 - NVIDIA Edition Runtime Detection Hotfix

Fixed

  • Fixed first-launch runtime detection for the Windows NVIDIA edition when the bundled model is preserved as models/default.bin.gz.
  • Updated KataGo asset preparation so release packages rewrite manifest.json to the actual bundled model path and checksums.
  • Added runtime and settings-panel fallback detection for edition.json metadata and compatible bundled .bin.gz models.

Maintained

  • Keeps the v0.3.7 NVIDIA packaging, multilingual release page, grounded shape recognition engine, local pattern matcher, knowledge source-policy gates, optimized move-range review, and quality checks and eval gates.

0.3.7 - Windows NVIDIA Edition and Multilingual Release Page

Added

  • Dedicated Windows NVIDIA release pipeline that restores a real NVIDIA KataGo runtime directory and bundled model from wimi321/lizzieyzy-next assets.
  • NVIDIA-specific release artifacts: GoAgent-0.3.7-win-x64-nvidia.exe and GoAgent-0.3.7-win-x64-nvidia-portable.zip.
  • Multilingual release notes in Chinese, Traditional Chinese, English, Japanese, Korean, Thai, and Vietnamese.
  • Release checks for NVIDIA asset wiring and multilingual release-note completeness.
  • KataGo asset preparation can now scan extracted archives, copy a full runtime directory, preserve model names, and write package edition metadata.

Improved

  • Release workflow now uses a handwritten multilingual release body instead of relying only on generated GitHub notes.
  • KataGo release checks accept a compatible bundled model when a package flavor intentionally uses a model name different from the default manifest model.
  • The v0.3.6 grounded shape recognition engine, local pattern matcher, knowledge source-policy gates, optimized move-range review from PR #5, and quality checks and eval gates remain part of the release quality baseline.

0.3.6 - Grounded Shape Recognition and Move Range Review

Added

  • Grounded shape recognition engine with KataGo-derived shape features and local pattern matching.
  • Knowledge cards v6-v11, source registry coverage, and source-policy gates for local teaching evidence.
  • Optimized move-range review from PR #5 with Alt+drag timeline selection, shared multilingual range parser, and key-move screenshots.
  • Teacher quality checks for claim verification, structured output gating, knowledge coverage, shape recognition, and move-range contracts.

Improved

  • Move-range teaching now starts from range trends and then focuses on top-loss key moves instead of逐手 expensive long-range analysis.
  • The teacher prompt now asks each key move to cite KataGo evidence, analysis quality, and shape or tactical signals.
  • Renderer and main process share src/shared/moveRange.ts, keeping renderer runtime away from main-only imports.
  • Japanese learning intent detection keeps 強くな as a training signal while adding move-range intent detection.

Thanks

  • Thanks to layiku for PR #3 and PR #4 on global arrow-key navigation; PR #4 is merged and improves review操作 flow.
  • Thanks to layiku for PR #5's move-range review direction, Alt+drag interaction, and multilingual parser, now integrated with quality gates and shape evidence.
  • Thanks to wimi321 for PR #1 and PR #2, which established the P0 beta, v5 teaching knowledge, joseki data, and evidence chain foundation.

0.3.5 - Keyboard Move Navigation

Added

  • Global Left/Right arrow navigation for stepping backward and forward through moves.
  • Home/End shortcuts for jumping to the first and final board position.

Improved

  • Rapid keyboard move stepping now debounces live KataGo analysis while keeping the board responsive.
  • Keyboard navigation stays out of editable fields, selection controls, buttons, and modified key combinations.
  • Cancelled live KataGo analysis can preserve a usable partial result instead of surfacing a cancellation error.

Fixed

  • LLM response helper tests now transpile the TypeScript source before importing it in Node's test runner.

0.3.4 - Settings Paste and Desktop Polish

Added

  • Native Electron edit menu and editable-field context menu so Base URL, model name, and API Key fields support normal copy/paste behavior.
  • Settings can reveal the saved LLM API Key on demand for user verification without exposing it in the public dashboard payload.
  • Contract coverage for LLM settings paste-friendly inputs and Electron native paste controls.

Improved

  • Reworked the desktop settings header into the light GoAgent visual system with compact KataGo/LLM readiness badges.
  • LLM settings inputs now disable browser-style autocorrect/autocapitalization and use monospace text for easier API configuration checks.
  • API Key helper copy was removed from the settings row so the expected workflow is clear: paste from the provider dashboard, optionally reveal to verify, then save.

0.3.3 - Teacher Pacing Control

Added

  • Current-move teacher analysis now includes internal pacing advice so common joseki can be explained briefly while middle-game fights receive deeper human-style commentary.
  • Added teaching density modes: minimal, branch, detailed, and caution.
  • Added variation teaching hints that tell the LLM when to explain purpose, expected reply, PV continuation, and practical result.

Improved

  • Current-move prompt now explicitly asks the teacher to control explanation length: say less for routine joseki, show key branches for joseki variations, and explain purpose/reply/follow-up for middle-game fighting.
  • KataGo and knowledge tool results now carry teachingPacing to the agent without changing the visible UI into a report.
  • Real LLM smoke now validates the agent runtime output instead of expecting the removed legacy llm.multimodalTeacher log.

0.3.2 - Agent Runtime and Analysis Polish

Added

  • Teacher runtime now follows a Claude Code-style tool loop: tool calls are executed, returned to the model, and the model continues until the final answer.
  • Current-move teacher prompt now lightly requires board-image reading, KataGo evidence, and local knowledge matching without forcing a fixed report template.
  • Teacher replies can be selected and copied, with a lightweight copy button on each assistant response.
  • Tool calls in the teacher thread now show clean step titles with running-state animation instead of verbose engineering details.
  • KataGo analysis runs can now be cancelled when the user changes game, move, or analysis mode.

Improved

  • Fast winrate graph generation uses a KaTrain-style low-visit sweep and refines suspected mistakes without blocking the first curve.
  • KataGo candidate loss and issue ranking use first-choice versus played-move winrate loss from the player-to-move perspective.
  • Board markers now distinguish the current move from the previous move with a subtler professional marker.
  • Candidate overlays and variation preview behavior are more stable during hover.
  • LLM provider handling now accepts tool-call turns and streamed tool-call deltas instead of treating empty text plus tools as an error.

Fixed

  • Teacher analysis no longer falls back to deterministic pseudo-explanations when the LLM fails.
  • Teacher tool trace no longer exposes long result summaries, shell output, or implementation detail in the main chat.
  • Right-side assistant output keeps auto-scroll behavior and supports copying during normal streamed answers.

0.3.1 - Teacher Smoke Fixes

Fixed

  • Teacher structured-result parsing now handles JSON followed by evidence verification notes.
  • Pure JSON teacher responses are converted into readable markdown instead of being shown as raw JSON.
  • Natural markdown teacher responses now populate a structured headline from the first meaningful line.
  • Teacher runtime now supplies fallback training problem recommendations from weak or joseki-linked matches when strong tactical matches do not provide drills.
  • Teacher LLM smoke now validates the evidence verifier happy path and no longer depends on removed legacy prompt wording.
  • UI Gallery dark panel headers now have sufficient contrast for visual QA screenshots.

0.3.0 - Teaching Knowledge v5

Added

  • Built-in joseki database bundle with source manifest and licensing notes.
  • Motif and joseki recognition services for stronger knowledge matching.
  • Teacher evidence validation so LLM explanations stay tied to KataGo candidates, board state, and matched knowledge.
  • Multilingual UI language option for Chinese, English, Japanese, Korean, Thai, and Vietnamese.
  • Additional elite pattern and joseki knowledge cards for professional teaching explanations.
  • Joseki bundle inspection script for release checks.

Improved

  • Current-move teacher analysis now includes recognized motifs, teaching evidence, and verification metadata in saved reports.
  • Teacher prompts are shorter, more human, and grounded by evidence instead of rigid templates.
  • Local knowledge source registry now records bundled joseki data sources and source-risk decisions.

Known Issues

  • Bundled KataGo binaries and models are still distributed through release assets/build preparation, not normal Git files.
  • macOS packages may still require manual trust if unsigned/not notarized.
  • Windows packages may still trigger SmartScreen when unsigned.

0.2.0-beta.1 - P0 Beta Candidate

Added

  • Three-column desktop workbench with library, board, winrate graph, and teacher chat.
  • Fox public game sync by nickname or UID.
  • SGF upload and mainline parsing.
  • KTrain/Lizzie-inspired board with coordinates, stone assets, last-move marker, and candidate marks.
  • Automatic low-visit full-game winrate graph on game load.
  • KataGo runtime resolver with bundled-runtime and local fallback paths.
  • Official KataGo model presets in settings.
  • OpenAI-compatible multimodal LLM settings.
  • Current-move multimodal teacher analysis.
  • Full-game and recent-10-game teacher quick actions.
  • Local knowledge search and long-term student profile storage.
  • Markdown and JSON report output.
  • Cross-platform CI for macOS, Windows, and Linux.
  • GitHub Release workflow for macOS, Windows, and Linux artifacts.
  • P0 release readiness checks for automation, assets, installers, signing, Windows smoke, and visual QA.
  • Local release evidence collection under release-evidence/.

Fixed

  • GPT/reasoning model response parsing when no plain content field is returned.
  • Fox-style SGF komi values such as KM[375].
  • SGF parser incorrectly reading comments and variations as mainline moves.
  • Board and winrate graph layout overlap in the center workspace.

Known Issues

  • Windows ARM64 is not supported in the P0 beta because the bundled KataGo manifest only supports Windows x64.
  • macOS public distribution requires Developer ID signing and notarization before tagging.
  • Windows public distribution should use an EV/OV certificate or Microsoft Trusted Signing; unsigned installers are internal beta only.
  • Windows 11 x64 real-machine smoke and visual QA evidence are required before creating v0.2.0-beta.1.