Releases: lambda-symbolics/autolith
Release list
Autolith 0.41.0
Autolith 0.41.0
Autolith 0.41.0 reduces model context cost, adds authenticated active-image management, and fixes command approval and terminal picker failures.
- Sends capability-gated deferred tool namespaces and hosted tool search to supported Codex models while preserving eager fallback elsewhere.
- Keeps one leading Chat Completions system message and omits completed agenda history from request-local context.
- Adds an opt-in HMAC-authenticated active-image management REPL over private Unix sockets or loopback TCP.
- Makes
once,always, and saved ask-mode approvals explicitly unsandboxed, with fail-closed missing approval UI behavior. - Explains Linux Bubblewrap requirements and actionable installer, launcher, startup, and runtime failures.
- Fixes modal pickers on detached Localgroup sessions by blocking correctly for relayed input.
Autolith 0.40.1
Autolith 0.40.1
Autolith 0.40.1 adds live context-window usage and makes interactive sessions run behind a detachable Localgroup relay.
- Shows provider-reported context usage and the automatic compaction threshold in the terminal modeline.
- Starts ordinary interactive sessions in a detached process so
(detach)releases the relay immediately without stopping active work. - Bounds Localgroup relay-thread cleanup, reaps failed detached starts, and accepts ordinary typed characters during attached authentication.
- Keeps the verified 272K ChatGPT subscription context window for GPT-5.6 models.
Autolith 0.40.0
Autolith 0.40.0
Autolith 0.40.0 makes provider context caching visible and stable across turns.
/statusaggregates token and prompt-cache accounting across primary provider calls and recursive inference requests.- Mutable session state and request-local context now follow durable history instead of rewriting stable request prefixes. Anthropic requests add explicit cache breakpoints for tools, system instructions, and history.
- Provider authentication immediately refreshes dynamic model catalogs, while discovery failures preserve the last catalog and return a visible warning.
- API-key prompts propagate terminal input descriptors correctly and use sanitized semantic styling on capable interactive terminals.
- The context cost report documents the measured request prefix and the cache-stability changes in this release.
- Live OpenRouter verification covered Ox Alpha streaming and a namespaced function-tool call with normalized arguments.
Autolith 0.39.1
Autolith 0.39.1
Autolith 0.39.1 adds official OpenRouter provider support.
- Authenticate with
OPENROUTER_API_KEY,autolith auth openrouter,/auth openrouter, or(auth "openrouter"). - OpenRouter's authenticated catalog is filtered to text models that support function tools, tool choice, and normalized reasoning. Local model identifiers use the collision-safe
openrouter/<vendor>/<model>namespace. - Requests use OpenRouter's normalized reasoning controls, broadly compatible
max_tokensoutput ceiling, and optional application attribution headers. - Live verification covered streaming and persisted turns on GPT-5 Mini, Claude Haiku 4.5, and Gemini 3.5 Flash, plus a complete GPT-5 Mini tool-call round trip.
Autolith 0.39.0
Autolith 0.39.0
Autolith 0.39.0 improves provider resilience, parallel execution, session naming, and live command visibility.
- Provider streams reconnect when they stall without producing events.
- ChatGPT Codex uses the standard Responses request format, supports configurable Fast mode, and can execute independent tool calls concurrently while preserving deterministic transcript order and serializing shared mutable state.
- Sessions receive immediate local titles and optional provider-generated refreshes. Use
/titles on|offor(titles "on"|"off")to control generated titles. - Queued and running primary-agent commands appear in the terminal live region with elapsed time, bounded retention, and shared viewport allocation alongside child activity.
- Reusable Responses request hooks now come from
cl-llm-provider-api; provider-specific transport and execution policy remain in Autolith.
Autolith 0.38.0
Autolith 0.38.0
Autolith 0.38.0 adds Nous Research and Mistral providers while moving reusable provider and Skill protocols into focused libraries.
- Nous Research subscription access uses browser device authentication, process-safe rotating refresh tokens, authenticated model discovery, and automatic Chat Completions or Anthropic Messages routing by model.
- Mistral API-key access discovers chat-capable account models and uses the provider's documented
max_tokensrequest field. - RFC 8628 device authentication is shared between Grok and Nous with strict verification URL handling and issue-time expiry enforcement.
- Provider value types and protocols now come from
cl-llm-provider-api; Skill discovery, parsing, caching, and rendering now come fromcl-skills. - Child activity suppresses stale primary status text while work is still running.
Autolith 0.37.0
Autolith 0.37.0
Autolith 0.37.0 adds standard Agent Skills and keeps active responses visible near the prompt.
- Exact
SKILL.mddiscovery supports validated YAML frontmatter and preserves the Markdown instruction body. - Standard Skills are converted into durable content-hashed native cache entries with source and generated-content integrity checks.
- Successful
skill.loadcalls produce concise request-local transcript markers in compact view, while expanded and failed results remain visible normally. - Streamed assistant output stays immediately above the prompt instead of being displaced by notices and queued-input previews.
- NYAML is locked coherently across Quicklisp and Nix builds.
Autolith 0.36.3
Autolith 0.36.3
Autolith 0.36.3 adds native macOS x86-64 releases and repairs tracked source editing.
- GitHub Actions packages and publishes
x86_64-darwinarchives on themacos-15-intelrunner. - The installer, packaged launcher, runtime builder, and release mirror recognize macOS x86-64 end to end.
- Intel macOS runtime builds use a validated Homebrew SBCL bootstrap and load packaged native
.dyliblibraries. - Workspace tools can edit tracked Autolith launcher and recovery source while installed stable artifacts remain protected.
Autolith 0.36.2
Autolith 0.36.2
Autolith 0.36.2 supersedes 0.36.1 with portable native release packaging fixes.
- Linux x86-64 release builds use Ubuntu 20.04's host SBCL to preserve the archive's glibc 2.31 compatibility floor.
- NetBSD release builds accept the version suffix used by pkgsrc SBCL packages.
- macOS and BSD archive construction correctly represents the absence of a Linux libc identity.
- The release includes the static Linux archives, unified authentication, recursive-inference activity, and runtime efficiency improvements introduced in 0.36.0.
Autolith 0.36.1
Autolith 0.36.1
Autolith 0.36.1 republishes 0.36.0 from unchanged product source after transient release-runner failures prevented the complete native artifact set.
- Release packaging produces genuinely static musl archives for Linux x86-64 and aarch64, preserves platform identity across installation and updates, and validates Linux artifacts in CI.
/authand callable(auth)share one operation with interactive provider selection, explicit-provider bypass, immediate login guidance, hidden Localgroup API-key input, and safe deferred terminal output.rlm.infer,rlm.map, andrlm.completeshow compact live request progress, including nested environment calls, while isolating callback failures and revoking late updates after shutdown.- Streaming, redaction, task progress, image encoding, preview decoding, prompt rendering, and terminal draft handling avoid repeated whole-value copies and unnecessary work.
- Inference trace windows stream from chunk logs, context-object lookup returns verified content, image inspection uses bounded prefix reads, and picker search indexes publish once on close.
- Command authorization accounts for unavailable sandbox isolation, papercut reports show canonical callable forms, and Localgroup status includes the session process identifier.