Changelog

v0.5.0 - 2026-09-08

Modules and coordination

• Added five public module operations: patterns/install, catalog, source, update, and call. Modules are editable run-local definitions, discovered from project files, user files, or packaged resources. Explicitly saved definitions can be loaded in a fresh run. See the migration guide.
• Breaking change: Removed the previous named pattern wrappers. The bundled library now contains relay and mailing-list; check-result, ralph, fix-loop, team, and clean-prompt are retired. Callers compose shell work and configured agents explicitly.
• Added immutable module installer ownership, explicit acknowledgment for another agent's edits, and dogfood journals for successful module installations and updates.
• Exposed direct mailing-list operations for subscriptions, posting, notification, message retrieval, digest, and acknowledgment. Initialization and list creation subscribe their caller atomically; child onboarding examples establish subscriptions before generation.

Tool results and context

• Breaking change: External tools return raw {:ok … :out … :err …} envelopes, with applicable process/HTTP status. Automatic shortening and the :truncated flag belong to serialization. Requested line ranges, character windows, and pagination select the result. See bounded results.
• Replaced stored-output representations with ordinary bounded snapshots. Complete raw values remain available to computation; retain needed evidence explicitly before pruning context. Updated recovery and context guidance explains ephemeral bindings and durable observations.

Runtime and performance

• Prevented duplicate completion notifications when rejoining a retained future after an unrelated wake. One subscription remains active through receipt; explicitly awaiting again after receipt delivers the completed value without repeating the computation.

• Preserved queued messages through context compaction, corrected dormant-agent startup from computations, and reset consecutive-error limits after successful model turns.

• Made Ctrl+C exit interactive runs and restore terminal state. Development workflow runners clean up their child processes on interruption and output failures.

• Added a configurable five-minute deadline for complete Codex HTTP responses, including stalled response bodies. Timeout does not automatically retry the request.

• Reduced trace-export allocation, removed verbose-only presentation delays, detached selected file lines from full-file backing storage, and avoided unnecessary parser sanitizer allocation. Added reproducible runtime benchmark workloads and measurements.

• Fixed multiline-string corruption during comment normalization, character-literal and reader-comment balancing, and regex escape preservation.

• Changed the CLI default provider to codex-tc, using local Codex authentication. The default model remains GPT-6 Astra with medium reasoning.

v0.4.0 - 2026-09-06

Communication and context

• Each run now owns an atomic coordinator for agent lifecycles, mailboxes, requests, and wakeups. Concurrent runs keep their coordination state separate.
• Added immediate agents/ask and agents/spawn-ask: request results from one or more agents, continue working, then use agents/!wait or agents/!sleep. Multi-target requests collect one result per target. Replies identify their request, and unrelated messages can awaken a caller while its collections remain pending. See multi-agent communication.
• Enforced an ordering rule for communication waits to prevent coordinator-wait deadlock, assuming fair scheduling and eventual external progress. Refused waits raise recoverable errors so agents can handle outstanding requests. Added atomic capacity admission through :coordinator {:max-edges 10000}.
• Added tracked terminal requests through /ask, /requests, and /cancel. Users can collect results from several agents while continuing to exchange messages and answer clarification requests.
• Made message receipt explicit for raw !llm-self calls: messages remain queued unless {:receive? true} is supplied. Convenience wrappers receive automatically. Added (receive completion) to accept messages into a completed program without evaluating it. Waits and wakeups retain the latest resumable context, and recovery preserves the failing call's receipt policy.
• Replaced the future-only blocking/completion-promise helper with atomic blocking/request tokens. Requests from futures participate in the coordinator's dependency tracking.
• Unified tool, agent-message, and MCP result rendering under :context-max-chars, defaulting to 10,000 characters per contribution. Oversized values remain complete in run-local storage and can be inspected in smaller pieces. Negative local display limits now use the run cap.
• Improved agent guidance for retaining continuations, checking which communication actions executed, recovering from refused waits, and answering outstanding requests before returning.

Models and providers

• Changed the CLI, OpenAI, and Codex model-profile defaults to GPT-6 Astra with medium reasoning. Added Astra aliases, model limits, and standard-tier cost tracking. Higher pricing above 272K input tokens is documented but not applied automatically.
• Fixed Codex streaming responses that deliver completed tool or message items before an empty final output list, while still rejecting failed or incomplete responses.
• Added opt-in Kimi K3 aliases and Fireworks pricing.
• Fixed retries for incomplete OpenAI Responses output when a response status is a string, preserving reported usage.

Documentation

• Added the VitePress documentation site, with guides, API reference, local search, and a production-build check. The README links to the documentation, and both landing pages introduce the communication model.

v0.3.0

• Changed the CLI and OpenAI model-profile default to GPT-5.6 Sol with medium reasoning, while retaining explicit provider and reasoning overrides.
• Removed the CLI launcher's unnecessary dependency on rlwrap by invoking clojure directly.
• Added a new-spec-only MCP 2026-07-28 client with generated permissioned namespaces, Streamable HTTP and stdio transports, tools, resources, prompts, completion, subscriptions, and an explorer-style spell mcp CLI.
• Added a runnable MCP Everything-style example backed by the official Python SDK, including retained GPT-5.6 Sol real-model validation.
• Integrated MCP server-profile, permission, CLI, security, and supported-surface documentation into the public API and configuration reference.
• Added --trace-dir DIR for durable CLI traces and automatic agent/trace context on feedback entries.
• Added --dogfood to expose the feedback namespace to the main agent and its workers only for explicit self-improvement runs.
• Added --agents-md to prepend the current working directory's AGENTS.md, capped at 32 KiB, to natural-language CLI tasks.
• Added Agent Skills discovery with on-demand SKILL.md disclosure via a prompt-only skills namespace, removed reminders, renamed bundled skills to spell-* names where applicable, and added SnakeYAML for safe skill metadata parsing.
• Changed error recovery so proven trailing-expression failures reopen in place, while other evaluation and reader failures use one-turn inert recovery context that is pruned before subsequent turns; added a public recovery guide.
• Added Claude Fable 5.1 model aliases, pricing, adaptive thinking support, and automatic tool choice.

v0.2.0

First public-facing software release.

Notable release-surface changes relative to v0.1.0:

• Added the public spell.api/run entry point with explicit :model-profile and :agent-profile inputs.
• Renamed checked-in runtime configuration to model profiles under config/model-profiles/ and agent profiles under config/agent-profiles/.
• Curated the public examples to hello-world, coin-flip, twenty-questions, telephone, auction, and chat.
• Consolidated public documentation around README.md, AGENTS.md, and docs/api.md.
• Removed the experimental React namespace/profile from the public release surface.
• Added repo-local Spell skills for setup, configuration, and source navigation.

v0.1.0 - 2026-05-06

Paper/reproducibility-oriented baseline release.