rpotato v0.53.0 - Bounded Runtime Architecture and Autonomous Grounding

Release date: 2026-07-29

This minor release completes a repository-wide responsibility refactor and makes freshness-sensitive questions enter runtime-owned web grounding before a small local model can answer from stale memory. The public CLI and TUI command surface remains compatible.

Included

• Replaces named product, event, and answer-shape exceptions with generic freshness features, typed query plans, evidence coverage, and answer binding.
• Keeps local repository questions and explicit no-web requests offline while routing volatile facts, current versions, and empirical comparisons through bounded WebSearch, WebOpen, and WebFind.
• Preserves runtime source attribution in generated and fallback grounded answers and rejects stale, conflicting, or unrelated evidence.
• Answers current-model identity questions from runtime configuration without intercepting model comparison or recommendation requests.
• Keeps ordinary active conversations tolerant of stale historical source pointers while explicit /resume reports invalid resume artifacts.
• Splits TUI, context, model, backend, install, workflow, patch, collaboration, web, CLI, persistence, terminal, tests, release notes, and retrospectives into bounded owners.

Compatibility and boundaries

• Existing public commands, stored sessions, model artifacts, lazy projectors, and managed llama.cpp behavior remain compatible.
• Web evidence remains bounded, read-only, untrusted input and grants no browser, shell, filesystem-write, credential, or approval authority.
• Supported release targets remain macOS Apple Silicon, macOS Intel, Linux x86_64, Linux ARM64, and Windows x86_64.
• GitHub Releases remains the only supported binary distribution channel.
• This refactor makes no new model-quality, benchmark, memory, or hardware claims.

rpotato v0.52.1 - Managed llama.cpp Structured-Chat Recovery

Release date: 2026-07-29

This patch restores the default TUI conversation path on the managed
llama.cpp backend. The v0.52.0 structured answer schema requested a grammar
repetition larger than the pinned compiler accepts, causing even a simple
greeting to fail with HTTP 400 before generation.

Included

• Removes the grammar-level answer.maxLength repetition while preserving the
  runtime's independent 16 KiB visible-answer validation.
• Rejects unsupported string, array, and object repetition bounds before a
  generation lifecycle starts or a request reaches the backend.
• Traverses only JSON Schema subschema positions so output properties named
  maxLength, minItems, or similar keywords remain valid.
• Covers the verified 1,999/2,000 managed-grammar boundary, property-name
  collisions, the production turn schema, and the native structured TUI
  request.
• Requires a real pinned llama-server and installed supported model smoke
  whenever request-body, response-format, schema, or chat-template contracts
  change.

Compatibility and boundaries

• Existing models, sessions, projectors, web tools, browser restrictions, and
  public commands remain compatible.
• The application still rejects visible structured answers larger than 16 KiB;
  only the incompatible grammar repetition was removed.
• GitHub Releases remains the only supported binary distribution channel.

rpotato v0.52.0 - Structured Small-Model Tool Turns

Release date: 2026-07-28

This minor release replaces free-form web-tool instructions with a
schema-constrained decision and observation loop designed for small local
models. The runtime, rather than model text, now owns the transition from a tool
request through verified evidence to the final visible answer.

Included

• Requires one bounded JSON-schema decision per conversational model turn:
  Answer, WebSearch, WebOpen, WebFind, or ContinueLocal.
• Executes WebSearch, WebOpen, and WebFind through the same
  ToolCall → Observation → Answer lifecycle and never displays a raw tool
  report as the final model answer.
• Generates the final answer only after a separate model call receives the
  runtime-owned, bounded observation.
• Builds follow-up search queries from recent user requests while excluding
  model responses, attachments, credentials, private values, and unrelated
  prior topics.
• Treats malformed structured decisions as a safe local continuation instead
  of exposing private protocol text.
• Distinguishes general analysis questions from repository inspection requests
  unless the request names a repository, project, source, file, or code scope.
• Separates search, open, find, query sanitation, and runtime request support
  into bounded owners protected by architecture and native-terminal tests.
• Prevents PR label and ready-state events from canceling one another while
  retaining same-event supersession for newer candidate commits.

Compatibility and boundaries

• Existing models, sessions, projectors, web grounding, browser restrictions,
  and public commands remain compatible.
• Web observations are read-only, bounded, untrusted evidence. They cannot
  grant browser, shell, filesystem-write, credential, or approval authority.
• GitHub Releases remains the only supported binary distribution channel.

rpotato v0.51.0 - Grounded Conversation Continuity

Release date: 2026-07-28

This minor release keeps conversational memory, context usage, and verified web
evidence coherent across ordinary follow-ups and explicitly resumed sessions.
It replaces several last-request shortcuts with session-derived state while
keeping network access bounded and runtime-owned citations authoritative.

Included

• Calculates the idle TUI context indicator from preserved session dialogue
  instead of the latest isolated model run.
• Retains failed requests and visible runtime errors as bounded conversational
  context so a corrective follow-up does not lose the request that failed.
• Carries bounded prior dialogue into final web synthesis and persists verified
  grounding evidence for relevant follow-ups after /resume.
• Reuses prior web grounding only for referential or topic-overlapping requests;
  unrelated prompts remain local and cannot inherit stale evidence.
• Softly truncates oversized web evidence instead of failing the complete
  research request, while keeping runtime-issued source identifiers intact.
• Replaces uncited or unusable model summaries with concise runtime-grounded
  fallback answers and runtime-owned source links.
• Prevents context-dependent identity questions from being misclassified as
  questions about the agent itself.
• Splits session restoration, web routing, grounded answer binding, and request
  support into bounded owners protected by architecture regression tests.

Compatibility and boundaries

• Existing models, projectors, registries, sessions, web tools, and public
  commands remain compatible.
• Restored web evidence remains bounded and is reused only when the new request
  demonstrates a relevant relationship to it.
• GitHub Releases remains the only supported binary distribution channel.

rpotato v0.50.1

Release date: 2026-07-27

This patch makes the primary conversation TUI reliably preserve user input, present model progress, render bounded Markdown, and keep private runtime protocols out of visible answers. It retains the v0.50.0 web research and restricted browser behavior without widening browser or tool authority.

Included

• Normalizes pasted clipboard image paths into private attachment badges and verifies the real native-terminal paste path.
• Shows an active spinner, elapsed time, and a live context estimate while a local-model request is running.
• Reconciles context metrics across manifest model identifiers and GGUF artifact names.
• Accepts bounded small-model web-tool protocol variants while preventing private WEB TOOL and WEB INPUT instructions from reaching the transcript.
• Keeps short conversational follow-ups local unless current public evidence is actually required.
• Adds terminal-width-aware Markdown, bounded code-fence headers, and history scrolling that preserves unfinished drafts.
• Distinguishes failures before and after request dispatch so uncertain model requests are not submitted again.
• Splits request submission, Markdown, status rendering, and terminal-platform responsibilities into bounded owners with architecture guards.

Compatibility and boundaries

• Existing models, projectors, registries, sessions, web tools, and public commands remain compatible.
• Scrolling never submits a hidden command, and private runtime protocol text is never treated as a public assistant answer.
• GitHub Releases remains the only supported binary distribution channel.

rpotato v0.50.0

Release date: 2026-07-27

This feature release turns one-shot web lookups into bounded, source-linked
research and adds a narrowly scoped Chromium search-form route for explicit
public-site search requests. It keeps model output untrusted and does not grant
general browser automation, authenticated sessions, or arbitrary page actions.

Included

• Adds a bounded Search→Open→Find research loop with query refinement, result
  ranking, main-content extraction, RSS/Atom/Markdown support, and strict
  request, evidence, and elapsed-time budgets.
• Assigns stable runtime source identifiers, rejects invented or swapped
  citations, preserves conflicting evidence, and exposes multi-page source
  navigation through the TUI.
• Hardens direct API-key-free discovery with complementary DuckDuckGo HTML and
  lite parsing, canonical URL deduplication, anti-bot detection, and clear
  bounded failure.
• Adds an isolated installed-Chromium search-form route using a fresh temporary
  profile, accessibility-based field discovery, opaque element handles, and a
  public-HTTPS-only address-pinned loopback CONNECT proxy.
• Blocks private and local targets, existing browser profiles, login, payment,
  posting, upload, download, personal-data entry, project-content submission,
  arbitrary selectors, JavaScript, and unrestricted click or form automation.
• Keeps static WebSearch, WebOpen, WebFind, ordinary local conversation, and
  explicit no-browse behavior available when a supported browser is absent.
• Strengthens native-terminal candidate coverage so long visible commands are
  submitted atomically only after the live composer is ready.

Compatibility and boundaries

• Existing public commands, model files, registries, sessions, and local
  conversations remain compatible.
• Browser support uses an already installed Chromium-family browser and does not
  download or bundle a separate browser.
• GitHub Releases remains the only supported binary distribution channel.

rpotato v0.49.4

v0.49.4 - Reliable macOS Terminal Release Verification

Release date: 2026-07-26

This recovery patch prevents native terminal cleanup from hiding failures behind
an unbounded child-process wait and moves the exact macOS terminal scenario into
the pull-request candidate gate. It supersedes the incomplete v0.49.3 artifact
set while retaining its explicit session resume and fresh-start behavior.

Included

• Bounds Unix PTY cleanup with a SIGTERM grace period followed by SIGKILL,
  so a stopped or unresponsive child cannot hang the test process indefinitely.
• Adds an exact regression test that suspends the PTY child and proves cleanup
  returns within its termination budget.
• Emits native-terminal stage traces on every supported platform and runs the
  candidate and release probes with uncaptured output for actionable failures.
• Adds an exact macOS native-terminal candidate job before merge and tag
  publication, alongside the existing Windows candidate coverage.
• Preserves the v0.49.3 changes that start a fresh conversation by default,
  restore durable history through /resume, and isolate stale historical source
  pointers from unrelated prompts.

Compatibility and boundaries

• Existing public commands, model files, registries, and durable conversations
  remain compatible.
• Existing conversations remain available through /resume.
• GitHub Releases remains the only supported binary distribution channel.

rpotato v0.49.3 (incomplete; superseded by v0.49.4)

⚠️ 이 릴리스의 macOS 검증과 전체 자산 게시가 완료되지 않았습니다. 사용하지 말고 검증이 완료된 v0.49.4를 사용하세요.

대체 릴리스: https://github.com/MCprotein/rolling-potato/releases/tag/v0.49.4

명시적 세션 resume과 신뢰할 수 있는 새 대화

• 기본 rpotato 실행은 최근 transcript를 암묵적으로 불러오지 않고 새로운 durable conversation에서 시작합니다.
• /resume 선택창에서 원하는 저장 대화만 명시적으로 복원하고, /new는 별도의 새 session 경계를 만듭니다.
• 실패한 요청과 내부 오류 보고를 다음 prompt memory에서 제외합니다.
• 변경·삭제된 과거 source pointer는 일반 대화에서 best-effort로 처리하고, 명시적인 reread·patch만 엄격하게 검증합니다.
• runtime ledger guard 자기잠금과 연쇄 timeout을 수정하고 exact regression/preflight 및 Windows native-terminal E2E를 보강했습니다.

기존 durable conversation은 /resume으로 계속 사용할 수 있습니다. 공식 binary 배포는 이 GitHub Release만 사용합니다.

v0.49.2 - Accurate Vision and Model Cache Status

This patch makes the primary TUI report multimodal capability separately from loaded backend readiness and clearly reuses cached model artifacts instead of presenting every switch as a new download.

Highlights:

• Qwen and Gemma show vision on-demand before their projectors are loaded.
• Vision ready requires an exact healthy model, context, and verified projector match.
• First image use prepares the projector automatically; exact model and projector caches are reused.
• Model selection separates base-model cache state from lazy projector state.
• Vision fact replies no longer intercept implementation, debugging, analysis, or test requests.
• Production-shaped regressions cover stopped, stale, mismatched, ready, unsupported, unselected, and cache-hit paths.

An exact cache hit does not download the model again, although SHA-256 verification of a multi-gigabyte artifact can still take time.

GitHub Releases remains the only supported binary distribution channel.

v0.49.1 - Text-Safe Vision Upgrades

This patch keeps existing text-only installations usable after optional vision metadata is added and restores the selected model's manifest context instead of retaining a stale 4,096-token backend.

Highlights:

• Legacy schema-v1 registries remain valid for startup and ordinary text conversation.
• Projector download and binding happen only on first image use.
• Exact projector caches are reused and registry evidence is preserved during atomic v2 migration.
• Projector preparation failures preserve the selected model, text-ready backend, default selection, and registry.
• Model setup no longer eagerly downloads optional multimodal artifacts.
• Candidate preflight now enforces model-upgrade compatibility regressions.

GitHub Releases remains the only supported binary distribution channel.