BeauPi v1.0.2

Fixed

• Fixed DeepSeek and strict OpenAI-compatible providers rejecting privileged_exec by exposing a root object parameter schema while preserving local/terminal branch validation.

BeauPi v1.0.1

Added

• Added an API-first two-file model configuration: global settings.json can now contain models.providers and review.model, while auth.json remains the credential store; optional models.json overrides stay supported with higher precedence. Documented built-in DeepSeek API and review-model setup.

Changed

• BeauPi 1.0.1 npm installs now run a one-time postinstall migration that backs up and replaces ~/.beaupi/agent/settings.json, models.json, and auth.json. Existing API keys and OAuth credentials must be configured again. Default npm install and self-update commands now allow lifecycle scripts.

BeauPi v1.0.0

Breaking Changes

• Changed PrivilegeTerminalControl.start() and PrivilegeCommandSession.start() to stage the sudo command without executing it, and added execute() for the user's explicit terminal release action.
• Renamed the feature-specific terminalOutputReview.model setting to the shared review.model; the old key is no longer read.

Added

• Added current context token usage to the footer and streamed progress for context compaction.
• Added built-in beaupi-dark and beaupi-light themes plus shared ANSI-aware BeauPi visual helpers.
• Added a Session-restorable Task Ledger with phase, Tool/Shell/file/verification facts, repeated git status detection, a responsive Tasks widget, and Footer task status.
• Added the M3 Document Runtime with bounded local Markdown discovery/indexing, structured citations, content-hash invalidation, docs_search, docs_read, docs_resolve_task, and versioned Execution Contract details projected into the existing Task Ledger.
• Added the BeauPi Skill Registry core with deterministic user/project persistence, validation and inventory diagnostics, project-trust gating, conflict-aware precedence, and ResourceLoader reload projection.
• Added local BeauPi Skill Registry management with a searchable /skills UI, local/Claude/Codex directory import, enable/disable, validation, safe removal confirmation, trust gating, collision checks, and hot reload.
• Added remote BeauPi Skill imports and updates from Git, npm, and HTTPS sources with lifecycle-free staging, SHA-256 pins, validation/security review, explicit confirmation, symlink/cache filtering, atomic replacement, structured diagnostics, /skill-update, and controlled ResourceLoader Skill allowlists.
• Added complete /skills registry management details, searchable source/scope/path/diagnostic fields, collision-side display, SKILL.md viewing, and explicit disabled actions for non-updateable sources.
• Added BeauPi M5 in-process sub-agents with isolated AgentSession contexts, shared ModelRuntime/Provider/ResourceLoader lifecycles, AgentProfile Tool/Skill/file boundaries and budgets, bounded concurrency/cancellation/timeouts, non-recursive delegate_task, structured results, and Monitor-ready lifecycle/progress events.
• Added the M6 session-scoped MonitorRuntime and MonitorRegistry with deterministic Process/Tool/Sub-Agent records, fake adapters, the reserved SSH/tmux adapter interface, six monitor_* Tools, cursor/hash incremental logs, lifecycle event deduplication, and Session restoration loss handling.
• Added the interactive /target-server command for creating and updating user/project SSH execution targets, persisting non-secret OpenSSH configuration, and selecting the saved target immediately.
• Added the BeauPi M8 web research loop with a configurable SearXNG provider, web_search and SSRF-safe web_fetch Tools, shared query/URL caching, stable citations, content-hash deduplication, deterministic network budgets, Task Ledger evidence, and controlled sub-agent access.
• Added a systemd user service and installer for starting the local Docker Compose SearXNG instance at boot, with Docker readiness retries and local health verification.
• Added BeauPi M9 ask_user_question with strict bounded schemas, single/multi-select, built-in Other and notes editors, multi-question review, Markdown preview, configurable keybindings, versioned Session/Task Ledger facts, SDK/RPC callbacks, and controlled sub-agent clarification boundaries.
• Added the M10 session-scoped, branch-aware Policy Runtime with conservative Shell/path/error classification, stable redacted signatures, repeated-operation and failure/fallback advisories, Task Ledger facts, and current-branch Session restoration.
• Added bounded sub-agent Monitor activity logs with turn, Tool, target path, outcome, last activity, and virtual monitor_logs output when no file log exists.
• Added BeauPi M11 multi-agent Workflows with a strict versioned YAML/JSON DAG schema, bounded dependency conditions, AgentPool scheduling, shared single-writer coordination, isolated Git Worktrees, five built-in Workflows, workflow_run/status/cancel, Monitor and Task Ledger lifecycle integration, and a responsive DAG renderer.
• Added BeauPi M12 session-scoped background tasks with six background_* Tools, runner-owned local process and existing SSH/tmux Monitor attachment, deterministic triggers, a persistent deduplicated Wake Queue, idle/follow-up AgentSession auto-wake, bounded AgentPool progress reviews, branch-aware recovery, Task Ledger integration, and responsive TUI rendering.
• Added BeauPi M13 per-request controlled sudo execution with privileged_exec, automatic local bash and terminal_bash routing through a session-scoped PrivilegeRuntime, exact single-line or multiline read-only staging with user-controlled Enter execution or Escape cancellation, post-authentication detach with continued work-log capture, interactive-root-shell blocking, secure local-tmux PTY input, one-shot and terminal_send bypass blocking, branch-aware Session/Monitor/Task Ledger integration, responsive TUI rendering, and permission-restricted JSONL audit logs.
• Added the @winbeau/beaupi npm distribution, checksummed GitHub Release standalone binaries, and a versioned install.sh installer for Linux and macOS.

Changed

• Branded the existing coding agent distribution as BeauPi with the beaupi CLI, .beaupi configuration directory, and BEAUPI_CODING_AGENT_* environment variables.
• Changed User, Assistant, and Thinking messages to use compact BeauPi gutters, spacing, and width-safe error states while preserving OSC 133 regions.
• Reworked the interactive TUI with a minimal Tool shell, grouped read operations, streaming Write previews, structured Diff rendering, recent-run Footer statistics, responsive three-line status, and unified Compact/Retry states.
• Changed the streaming working row to show the latest model Thinking summary while preserving explicit extension messages and Tool-associated summaries/results.
• Changed Tool and task completion/failure markers to green/red dots, Diff boundaries to solid lines, and Compact progress to an indented second line.
• Changed ordinary coding sessions to resolve local document constraints through the existing AgentSession/System Prompt lifecycle without injecting full docs/**/*.md contents.
• Changed the default system prompt to use action-first, ADHD-friendly responses and focused implementations, preferring compact commands and scripts without arbitrary code-size limits.
• Changed the default system prompt to keep simple tasks on a short TODO-driven path, prioritize runnable prototypes and near-release fixes, repair root causes in the owning layer, and use risk-based minimal verification without repeated broad test loops.
• Changed Skill updates to reload through an atomic transaction and restore the previous Skill and Registry projection when reload fails.
• Changed Task Todo projection to keep Execution Contracts and requirements in the Task Ledger without rendering separate document-contract or Requirement Todo rows.
• Changed Assistant Thinking summaries to render one item as plain italic text, two as a complete Thought Chain, and longer streams as first/ellipsis/latest without duplicate rows.
• Changed the Tool renderer, Tasks Widget, and Footer to expose live Monitor status, running counts, stalled/failed attention summaries, durations, and complete log paths without exceeding narrow terminal widths.
• Changed SSH/tmux Tool calls to stay on one width-safe line and collapse long command output to a 10-line preview that uses the configurable Tool expansion keybinding.
• Changed Bash Tool calls to stay on one width-safe line with animated in-parentheses truncation, and collapsed output to a line-count summary using the configurable Tool expansion keybinding.
• Changed the built-in sub-agent profiles to use only wall-clock timeouts by default, removing their token and turn caps; custom profiles can still opt into additional budgets.
• Changed all built-in sub-agent wall-clock limits to 8 minutes while retaining optional lower per-profile and per-request budgets.
• Changed the BeauPi roadmap to mark the Policy Engine complete and advance the active milestone to multi-agent Workflow.
• Changed Policy to advisory-only behavior: every managed operation reaches its original executor, former block/confirm/replace/pause conditions become non-sensitive Footer-only advisories, TUI/SDK/RPC confirmation handlers are not invoked, controlled sub-agents omit Policy requests, and Tool/Todo rendering ignores Policy status metadata while legacy Session Policy details remain parseable.
• Changed the BeauPi roadmap to mark multi-agent Workflow complete and advance the active milestone to background task auto-wake.
• Changed the default system prompt so plans for projects expected to exceed roughly 500 implementation lines use a phase-and-milestone main plan with an index to separate per-module sub-plans.
• Changed the default system prompt to trust fast-model-reviewed bash, remote_bash, terminal_bash, and privileged_exec/Sudo Bash results on first inspection, skip full logs after reviewed success, inspect them only after the first reviewed failure, and provide complete sudo installation commands for missing system tools instead of user-local installs or PATH workarounds.
• Changed Terminal Bash and Sudo Bash output handling to return short successful output directly while routing failures, diagnostics, and output over 100 lines through the shared review.model pipeline with deterministic fallback and complete log paths.
• Changed the BeauPi roadmap to define M14 as dynamic Task planning with progress review reusing the shared review.model configuration, while reserving release preparation as the unnumbered M Final so additional feature milestones can be inserted.

Fixed

• Fixed ask_user_question answers ...