On this page
Supported Today And Future Work
Docs home · Installation · Changelog · 简体中文
This page separates what users can rely on today from what is experimental, limited, or future packaging work. The current alpha line is an early preview, not a stable API or plugin compatibility promise. Release versions and install commands are maintained in Installation and the Changelog.
Version boundary: this page and the GitHub Pages site track main. The packaged alpha is v0.0.1-alpha.4 and can lag behind the capabilities below; check Unreleased before relying on a newer feature.
Supported Today
| Area | Status |
|---|---|
| TUI entrypoint | sigil opens the TUI and is the primary product surface. |
| Distribution | npm alpha, Homebrew tap, Cargo git-tag, source, and release-archive paths are available; use Installation for current commands and channel details. |
| Quick Setup | First-run setup can create a usable local config. |
| Doctor | sigil doctor and /doctor report config, auth, workspace, MCP, code intelligence, and terminal readiness. |
| Feedback support | sigil doctor --output json produces a versioned redacted report; /feedback previews its privacy boundary before an explicit local-only export. |
| Automation output | sigil run --output json emits one versioned result, while --output jsonl emits ordered versioned events followed by one terminal result or error. |
| Chat workflow | Users can work through the composer with visible tool activity. |
| Image input | An idle Build composer accepts bounded PNG, JPEG, and WebP paths or image-clipboard input, shows removable metadata chips, and sends only to explicitly recognized image-capable OpenAI Responses, Anthropic, or Gemini model IDs. |
| Tool approvals | File writes, shell execution, external paths, and external tools can be reviewed before execution. |
| Session recovery | Session and control records are append-only and restore visible state after restart for current V2 session logs. Older raw session logs are reported as unsupported and left unchanged. |
| Checkpoint recovery | Ctrl-R previews an evidence-bound checkpoint and offers controlled file restore or a conversation fork that leaves files unchanged. |
| Planning | /plan runs read-only planning prompts and can explicitly hand an accepted plan to durable /task execution; /task <task> creates durable multi-step task state directly, and /task continue resumes the latest task. |
| Task verification | The Verification card exposes readiness, recommended checks, and inspectable snapshot and changeset evidence; Alt-V focuses it. |
| Context controls | Context pressure stays visible. Manual, fully idle hard-threshold, and queued pre-turn portable apply require exact local admission. The narrow OpenAI Responses overflow path requires audited server counts and exact economics. Owned preparation and source/queue/rejection CAS prevent stale dispatch. |
| DeepSeek provider | DeepSeek is the default Quick Setup path. |
| OpenAI-compatible provider | Supported through [providers.openai_compat] for compatible Chat Completions endpoints. |
| OpenAI Responses provider | Supported through [providers.openai_responses] for Responses streaming endpoints. The official pinned snapshot can perform one guarded, non-recursive overflow recovery after an exact output-free context rejection. |
| Anthropic provider | Supported through [providers.anthropic] for Anthropic Messages streaming. Its native-compaction beta driver records encrypted candidates only; it is not a user action and does not automatically change context. |
| Gemini provider | Supported through [providers.gemini] for Gemini streamGenerateContent streaming. |
| Web data tools | Stable websearch and capability-backed webfetch routes use separate network policy, durable egress disclosure, and external-source provenance. |
| MCP servers | Local stdio and user-root Streamable HTTP servers are supported through [[mcp_servers]] with trust, approval, and secret-egress policy. |
| Code intelligence | Optional, disabled by default, with LSP discovery and Rust fallback behavior. |
| Terminal controls | Mouse capture, OSC52 copy, scroll sensitivity, diagnostics, and default-off privacy-bounded attention notifications are documented and configurable. |
Limited Or Advanced
| Area | Current expectation |
|---|---|
| Release archives | Available on tagged GitHub releases for manual installs; package-manager installs are preferred. |
| Package-manager channels | Packaging names and availability may evolve during alpha; use Installation as the current source of truth. |
| OpenAI-compatible differences | The provider intentionally omits DeepSeek-only prefix/FIM/beta behaviors. |
| Provider-specific options | Each provider page explains its available setup and options; normal tool approvals, privacy, and session behavior stay consistent. |
| Code intelligence | Depends on installed language servers and local environment; normal chat does not require it. |
| MCP lazy startup | Lazy servers are configured but do not register fake tools until activated. |
| External directories | Disabled by default and should stay narrow and approval-backed. |
| Headless automation | sigil run is useful for scripts but cannot show interactive approval modals. |
| Local server | sigil serve starts an advanced local HTTP/SSE service with retained event replay and graceful shutdown. V1 is loopback-only, requires bearer auth for every route except health, and is not a remote or multi-user service. |
| Execution sandbox | macOS, Linux, Docker, PTY, MCP stdio, and trusted plugin-hook paths have core coverage and receipts where supported, but coverage is not equivalent across all platforms and remote/container daemon scenarios. |
| Context help | New requests can include a bounded, request-local set of relevant Rust, Python, JavaScript/TypeScript, or Go snippets. Warm relevant LSP data is preferred; bundled Tree-sitter fallback remains ignore-, secret-, size-, and budget-limited. This is not a persistent semantic graph or full-repository index. |
| Model quality reports | Internal automated checks exist, but repeatable end-user model comparisons and release trends are not a supported product feature yet. |
Future Work
These are not the current supported path unless a later release says otherwise:
- self-update;
- desktop shell;
- hosted documentation search;
- richer generated release notes;
- broader provider-specific setup assistants;
- full semantic repo graph or vector retrieval by default;
- transparent in-flight crash resume for running shell/agent processes;
- parallel write-agent worktree isolation as a default workflow;
- stable plugin API compatibility;
- all-platform equivalent OS sandbox behavior;
- built-in end-user model quality benchmarking;
- release-to-release visual regression review for generated terminal screenshots.
How To Read The Docs
User docs describe current behavior unless a section explicitly says "future work", "limited", or "advanced". Treat the alpha line as usable for trials, not as a stable compatibility promise.