User Changelog

Docs home · Installation · Supported status · 简体中文

This page lists user-facing release notes. For support boundaries and early-preview caveats, see Supported Today And Future Work.

Unreleased - main

• Plan review now opens a complete, scrollable workbench instead of truncating the plan inside a

compact status card. All plan steps remain reachable on 32x8 through wide terminals; Esc closes without rejecting, printable input returns to the composer, and actions stay bound to the exact durable plan.

• Revise now asks what should change, keeps the original plan active until a replacement succeeds,

and restores it after every terminal revision failure. Older affected sessions are recovered by a strict read-only compatibility view, and sigil doctor reports ambiguous legacy lineage.

• Agents can ask bounded, typed questions through a durable attention queue. Pending questions

survive TUI exit and resume without a timeout, background-agent questions route to the root session, and an accepted answer resumes exactly one provider attempt without replaying the turn that asked the question. MCP elicitation reuses the form renderer without inheriting durable replay semantics.

v0.0.1-beta.4 - 2026-08-11

This beta stabilizes long-running TUI sessions, makes shell execution auditable in the transcript, and closes durable Task continuation, memory-routing, and final-answer coordination gaps.

• Fixed TUI transcript corruption after a final answer and restored access to older history after

terminal resize or reflow. History inspection now keeps a stable content anchor while new output arrives, and repeated artifact-maintenance failures are shown once per session.

• Busy-turn follow-ups now wait for a safe delivery point without interrupting the active run, and

compact plan, queue, verification, attachment, and agent controls remain aligned and actionable on narrow or short terminals.

• Bash cards now show the policy-safe command during live execution, after reload, and in child-agent

transcripts; running and terminal states no longer collapse into duplicate or misleading cards. Bounded read-only Git metadata checks are recognized without weakening protected .git mutation rules or danger-full-access hard-deny boundaries.

• Continuing a paused Task now selects the exact durable Task, carries the user's continuation

guidance through a reviewed replacement plan, and shows only the current Task in the TUI instead of reviving an obsolete initial plan. Starting an explicit plan review also clears the prior Task from the current-work surface while preserving it as history.

• Explicit durable-memory intent can now be handled in the automatic routing turn. Approved memory

writes settle durably before the selected Chat / PlanReview / Task handoff, while tool results keep the provider's declaration order.

• Automatic-routing reasoning no longer appears as a duplicate Thinking block before the real work

turn. Final-answer coordination now scopes child-agent facts to the active logical run, removes stale snapshots when those facts settle, and bounds repeated blocked-final attempts. A blocked child or Task participant can no longer be reported as completed, and a final candidate rejected by the terminal blocker is no longer left visible as an accepted answer.

• Complete zero-byte tool artifacts are now valid retrievable artifacts, preventing maintenance from

repeatedly deferring otherwise healthy session state.

v0.0.1-beta.3 - 2026-08-06

This beta releases the automatic plan-review and AI orchestration lifecycle, session-scoped scratch with quota/TTL, and the RFC-0062 tool-result settlement closures.

• Automatic Task routing is now the default: ordinary input first runs a host-owned

Chat / PlanReview / Task decision on the review-first baseline, and only an accepted plan can create a durable Task. Direct task execution requires an exact-route qualification manifest; explicit routing_policy = "manual" is the coarse rollback. sigil doctor reports the three automatic-orchestration facts.

• The plan review surface is now a first-class lifecycle: /plan and automatic review share the

same typed plan draft, the same pending plan card (Run / Save / Revise / Reject), and the same HTTP/Desktop/TUI decision commands. Reconnect and reload restore a pending plan.

• Workspace instruction memory and approval-gated durable memory are now enabled by default. Set

[memory].enabled = false or [memory].writable = false to opt out of either capability.

• Plain sigil launches now start a fresh session instead of reopening the most recent conversation.

Explicit resume keeps the portable transcript across safe endpoint corrections, asks before using a changed or unproven destination, and prevents the same session from being opened for writing in two TUI/Desktop/headless owners at once.

• TUI and Desktop now configure an optional context-window size independently for each connection/model pair, including first-run setup. Empty values keep automatic provider metadata and global fallback behavior, so providers without a model catalog are not blocked.
• TUI transcript selections now copy automatically when the mouse is released. Copy attempts both

OSC52 and the system clipboard, preserves a failed selection for Ctrl-C retry, and still excludes the info rail.

• Long TUI runs no longer stop when native scrollback and asynchronous input compete for a cursor

query. Active-run facts no longer look like a host finalization instruction, and typed artifact retrieval remains visible after the initial tool-preview budget is exhausted.

• Interactive tool approvals no longer expire automatically after 300 seconds. `Allow for

session persists across reopening the same session and reuses recognized validation grants across presentation-only tail, head, or grep` changes while retaining exact execution checks.

v0.0.1-beta.2 - 2026-08-03

This beta removes model-catalog discovery from the critical path for first-run setup and allows the frozen TUI packages to ship before the matching Desktop artifacts are ready.

• Quick Setup, /config, Desktop Settings, and the HTTP setup route now treat remote model lists

as optional guidance. Bundled models stay selectable while discovery runs, and an exact model ID can always be entered when a provider has no /models endpoint or discovery fails. The first real generation request remains the authority for credential, protocol, endpoint, and model compatibility.

• Unsupported reasoning parameters are omitted using the exact provider/model capability mapping

instead of being guessed for unknown or non-reasoning models.

• Alpha and beta TUI npm packages can now be published from the frozen release candidate while

the GitHub Release remains a draft. Desktop DMGs, updater archives, Pages updates, Homebrew, and public GitHub Release publication continue later from the same immutable tag.

• macOS Desktop notarization is now asynchronous and resumable. Immutable DMG and app submissions

are recorded in an append-only ledger, status checks are one-shot, and an offline finalizer verifies accepted submissions before the Desktop assets can be uploaded.

v0.0.1-beta.1 - 2026-08-02

This beta publishes the first signed macOS Desktop distribution and the matching cross-platform TUI beta channel.

• Shell permissions now use one immutable, structured plan across policy, approval, audit, and

execution. Compound validation commands are classified per child command; dangerous flags, redirections, dynamic syntax, protected targets, and missing sandbox capabilities continue to fail closed. Exact, containment-bound session grants reduce repeated approval without widening access to unrelated commands. Native PowerShell background work is kept out of the one-shot Shell path, and unresolved shell paths remain conservative instead of aborting the plan.

• Approval decisions now converge from the exact command receipt in both Desktop and TUI instead

of waiting solely for a later live event. Accepted, resolving, execution-started, stale, expired, uncertain, and terminal states are distinct and protected against older receipts.

• Finite commands now use the foreground Shell path exclusively. Persistent and interactive work

uses explicit terminal tasks with readiness and event-driven waits; the runtime no longer uses periodic terminal status discovery or asks the model to poll unchanged logs.

• Release tags now build TUI/npm bytes once and freeze a commit-bound candidate manifest in the

draft. Final publication reuses those admitted tarballs instead of rebuilding, while the release doctor and Desktop upload command bind versions, tag/main/CI, updater keys, signatures, notarization, and both macOS architectures. Publishing automatically starts a bounded public npm/GitHub/Desktop/Pages/Homebrew installation smoke.

• Fixed invalid or non-current configuration blocking every Desktop workspace before Settings was

reachable. Desktop now starts a recovery-capable workspace service and offers an explicit current-format replacement in Settings; TUI Quick Setup offers the same replacement path. Neither surface reuses or migrates values from the invalid file, and both refuse to overwrite a concurrently repaired valid config.

• Published the macOS Desktop beta channel with signed, Apple-notarized Apple Silicon and Intel DMGs plus signed architecture-specific update bundles.
• Added explicit version checks and updates across Desktop Settings, TUI /update, and CLI sigil update; managed installations receive their package-manager command, while standalone updates remain checksum/signature verified and never restart silently.
• Updated the website, README, installation/status docs, and visual tour so Desktop and TUI are coequal entry points with architecture-specific download guidance, a real Desktop capture tour, and the existing TUI real-run demo.
• /compact now treats the command itself as explicit intent: one invocation generates, validates, and atomically activates an admitted recoverable checkpoint without a confirmation modal. Failures keep the current context and remain visible with their exact reason.
• Fixed semantic compaction discarding its own result as stale after its audited provider-attempt and usage records advanced the durable stream.
• Fixed Desktop Enter submissions during active work failing to reach the durable follow-up queue, intermittent loss of live runtime controls, and valid live/durable replay being misclassified as a message conflict.
• Conversation titles can now be generated semantically by the active model after the first turn, while manual and generated renames stay synchronized with the open conversation page; title generation no longer competes with the primary request.
• Aligned the Desktop timeline, activity state, approval card, and composer widths, narrowed the approval presentation, and stopped unknown totals from appearing as stalled percentage progress.
• Added current-source Desktop Gherkin E2E coverage for real approvals, Enter queuing, skill/agent loading, /plan, automatic planning, and parallel Agents, alongside the existing stateful and orchestration TUI PTY acceptance campaigns.

v0.0.1-alpha.6 - 2026-07-30

These changes are included in the packaged v0.0.1-alpha.6 release.

Added

• Added AI-planned Task execution: Sigil can turn one repository goal into a visible, reviewable step plan, run independent steps in parallel, preserve normal tool approvals, and finish against repository-owned verification evidence.
• Added an authenticated, restart-durable historical session catalog to sigil serve for future desktop clients, with bounded pagination, title search, provider/pin/state filters, and explicit stale-cursor recovery. Session logs remain the source of truth, and catalog failures do not stop runs or recording.
• Added a desktop runtime bridge for trusted local clients: durable catalog entries can be reopened after restart, startup and server metadata have one versioned JSON shape, and an opt-in stdin owner pipe triggers graceful shutdown without PID polling.
• Added a source-built desktop dogfood shell with native workspace selection, durable history, conversation runs, exact approval and cancellation controls, and verification evidence over the same authenticated local server used by automation. CI builds short-lived unsigned macOS, Linux, and Windows dogfood artifacts; these are not a public install channel.
• Added exact-route orchestration rollout manifests. Qualified releases may enable auto + proactive for matching new installations, while missing, stale, invalid, or different-route manifests fail closed.
• Added durable, session-scoped tool-output artifacts. Large shell, file, search, terminal, and MCP results keep bounded conversation cards while their policy-safe bytes remain available through typed, size-limited page or literal-search reads in the model, TUI, HTTP, and desktop surfaces.

Changed

• Provider API-key storage now defaults to the owner-only credential file. Existing explicit auto configuration is non-interactive and strictly file-backed; it never queries an older native-system record. Only explicit keyring mode may show a system password prompt.
• Context compaction now uses cache-aware V3 exclusively: stable provider/tool prefixes, source-bound intent continuity across repeated compactions, complete-turn tails, and trusted cache-cost admission. Manual compaction directly requests one recoverable semantic checkpoint; normal semantic compaction makes one extra no-tool LLM request on the current route. Provider-native materialization remains fail-closed until exact-route resume is implemented.
• Reworked the desktop dogfood shell around workspace/session navigation, one conversation task surface, and a verification inspector. It replays bounded saved messages, retains control of runs across navigation while the workspace service stays open, separates final replies from progress and tool output, and provides focused approval, diff, evidence, and draft-aware composer behavior.
• Added one consistent desktop visual system, adaptive wide/two-pane/compact layouts, system light and dark themes, high-contrast and reduced-motion handling, keyboard focus capture/restore, terminal-only streaming announcements, and usable reflow down to 320 CSS pixels.
• Incompatible configurations are rejected rather than migrated. Doctor reports release qualification, and manual + explicit_request_only remains the coarse rollout rollback without deleting Task history.
• New sessions now use the V2 tool-result schema and deterministic tool-output aging before semantic compaction. Development logs written with the pre-V2 tool-result schema are intentionally unsupported: Sigil reports the inline body as Unavailable in a bounded schema diagnostic, leaves the file untouched, and does not backfill, rewrite, or guess missing artifacts.

Fixed

• Fixed non-interactive auto credential checks waiting indefinitely on macOS Keychain. auto no longer contacts native storage, and Doctor falls back to an offline status if an explicit native check does not finish promptly.
• Fixed the packaged desktop app reading Tauri-managed state before its setup lifecycle ran, which previously made the macOS app exit before creating a window.
• Fixed the event-driven TUI input loop stalling after the first key while idle.
• Fixed background tool-artifact maintenance racing with session actions, which could surface a transient lock failure instead of completing the requested action.

v0.0.1-alpha.5 - 2026-07-18

These changes are included in the packaged v0.0.1-alpha.5 release.

Added

• Added explicit OAuth sign-in for remote Streamable HTTP MCP servers, including automatic or manual callback, native credential storage, refresh, sign-out, and specific recovery errors. Every destination still passes the normal network disclosure and destination checks; headless startup never opens a browser.
• Added configurable info-rail visibility, an F2 show/hide shortcut, and a copy command that uses selected transcript text or the latest assistant reply.

Changed

• Windows shell and terminal tools now use PowerShell by default, show the detected shell in Doctor and tool cards, and stop child processes more reliably after a timeout. Local execution remains unconfined.
• Activating or refreshing a remote MCP server now updates its available tools without leaving stale duplicates. Windows also cleans up stopped local MCP process trees more reliably.
• Refreshed the Sigil logo, repository landing page, documentation site, social preview, and launch materials around one consistent product story.

Fixed

• Reply completion, queued work, and session transitions now recover more reliably without duplicating or stranding a final response.
• Long sessions keep timeline tail-index updates bounded, reducing redraw work as histories grow.

v0.0.1-alpha.4 - 2026-07-16

These changes are included in the packaged v0.0.1-alpha.4 release.

Added

• Added default-off, privacy-bounded terminal attention notifications for completed long work, approvals, failures, and user-input requests, with automatic OSC 9/OSC 777/BEL selection.
• Added bounded repository context for Rust, Python, JavaScript/TypeScript, and Go, using available language services with a built-in parser fallback.
• Added TUI image attachments for PNG, JPEG, and WebP through local paths or the system image clipboard, with removable attachment chips and clear provider/model compatibility checks.
• Added sigil doctor --output json, a versioned and redacted local diagnostics format for support requests.
• Added /feedback, which previews included and excluded data before an explicit local-only JSON export; reports are never uploaded automatically.
• Added structured GitHub forms for bugs, feature requests, and documentation issues.

Changed

• Completed the /feedback handoff: exported reports can now be reviewed in the TUI, revealed in the file manager, or paired with an explicitly opened bug form; reports remain local until the user attaches them.

v0.0.1-alpha.3 - 2026-07-15

These changes are included in the packaged v0.0.1-alpha.3 release.

Added

• Added stable sigil run --output json and --output jsonl formats for scripts, plus an advanced bearer-authenticated sigil serve interface that only listens on the local machine.
• Added explicit saved-session actions for safe export, conversation fork, pinning, exact delete review, and retention cleanup preview and confirmation.

Changed

• /compact can now confirm one manual context compaction when the selected model has installed local counting support and the compacted request is proven to fit. Completed long conversations and queued requests may use the same checked path. One pinned official OpenAI Responses model can also recover once from a confirmed pre-output context-limit rejection after separate count and savings checks.

v0.0.1-alpha.2 - 2026-07-15

These changes are included in the packaged v0.0.1-alpha.2 release.

Added

• Added the OpenAI Responses provider through [providers.openai_responses].
• Added stable websearch and supported webfetch routes with separate network controls and visible sources.
• Added a task Verification card, Alt-V focus, recommended checks, and inspectable evidence tied to the reviewed files and changes.
• Added Ctrl-R checkpoint review with controlled restore or conversation fork choices.
• Added a read-only Context Compaction V2 preview through /compact.

Changed

• Expanded local MCP support from stdio servers to include user-root Streamable HTTP servers under the same trust, approval, and secret-egress policy.
• Refreshed the user docs and website navigation around verification, recovery, and context controls.

Current Limitation

• Context Compaction V2 apply, including guarded overflow recovery, remains temporarily frozen while correctness fixes are in progress; /compact is a review-only preview.

v0.0.1-alpha.1 - 2026-07-08

Added

• Published the scoped npm package as @sigil-ai/sigil@alpha.
• Published the Homebrew tap formula as JimmyDaddy/sigil/sigil-ai.
• Documented npm, Homebrew, Cargo git-tag, source, and manual release-archive install paths.
• Added generated GitHub Pages documentation pages for installation, configuration, providers, safety, privacy, MCP, visual tour, troubleshooting, reference, and current status.

Changed

• Clarified that v0.0.1-alpha.1 is an early preview: core TUI workflows are usable, while config, plugin APIs, advanced sandbox behavior, and automation surfaces may still change.
• Made the documentation entrypoints more task-focused: quickstart first, then installation, visual tour, daily workflow, safety, troubleshooting, and reference.
• Updated the user docs to describe the current provider set: DeepSeek, OpenAI-compatible, Anthropic, and Gemini.

Known Limitations

• Self-update is not available.
• Stable plugin API compatibility is not promised for the alpha line.
• Sandbox coverage varies by platform and backend.
• Headless automation cannot show interactive approval modals.

v0.0.1-alpha - 2026-07-07

Added

• First public alpha release for the Sigil TUI.
• TUI entrypoint through the sigil command.
• Quick Setup, /config, sigil doctor, and /doctor.
• Multi-step task and planning flows through /task and /plan.
• Approval-backed file changes, shell execution, MCP usage, and code-intelligence edits.
• Recovery of saved local sessions after a restart.

Known Limitations

• This release was an initial preview and was superseded by v0.0.1-alpha.1.
• Users should prefer the alpha install channel or the latest documented release tag.

Next: Review installation and update paths.