AI 程式代理
Autolith 0.38 Splits Provider and Skills Protocols into Standalone Libraries While the Agent Remains Live-Rewritable
Common Lisp programming agent Autolith released versions 0.36 through 0.38 within two days, adding the standard Agent Skills format, Nous and Mistral backends, and cross-platform distribution packages. Its recoverable, replayable self-modification design improves inspectability, but generated code still runs with the user’s privileges and must not be treated as a secure sandbox.
Lambda Symbolics released Autolith 0.38.0 on August 23, adding support for Nous Research subscriptions and the Mistral API to its persistent, terminal-based programming agent implemented in Common Lisp. The Nous integration uses RFC 8628 device authorization, refresh-token rotation that works across processes, and post-authentication model discovery. It automatically selects either the Chat Completions or Anthropic Messages format according to the model. The Mistral backend queries the account for chat-capable models and uses Mistral’s specified `max_tokens` field.
Of greater interest to engineering teams is the restructuring of internal boundaries. Provider data types and protocols have moved to `cl-llm-provider-api`, while Skills discovery, parsing, caching, and rendering have been extracted into `cl-skills`. The preceding 0.37 release added support for the standard `SKILL.md` format: it validates YAML frontmatter, preserves the Markdown instruction body, and converts it into a persistent cache identified by a content hash, while checking the integrity of both source and generated content. This means skills are no longer merely an Autolith-specific internal format, and it reduces the extent to which the agent core must be modified when adding a new model provider.
The same release cycle also added static musl packages for Linux aarch64 and x86-64, an Intel macOS build, and builds for several BSD variants. The 0.36 series improved live progress reporting for recursive inference, streaming, and large-content handling, avoiding repeated copying of complete values. Its `rlm.complete` stores corpora that exceed the model’s context window as content-addressed objects and gives the root model only their labels, sizes, and hashes. A budget-constrained Lisp environment can then slice and search the content or launch sub-inferences.
Autolith’s distinguishing feature remains its modifiable runtime: the agent can redefine functions or classes while they are running, record changes in an append-only journal, and, after testing, save them as a private image commit containing a complete replay script. If the active image crashes, a separately built clean recovery image can restore a known-good generation. The official demonstration does document recovery after deliberately breaking a function, but it is not an independent reliability benchmark.
The largest risk is equally direct: model-generated code runs with the user’s privileges. Process separation is primarily intended to support recovery, not to defend against malicious code. Teams should initially test Autolith in an isolated account or virtual machine, pin versions, and audit Skills, provider authorization, and replay artifacts. Several releases arriving in quick succession also indicate that interface stability remains uncertain.