@hackage/shikumi-tools

Upgraded the baikai dependency to the 0.4 series (>=0.4 && <0.5). Dependency bounds only — no changes to the exported API.

ReAct exposes Proposal, renderStepLine, and summaryStep; native ReAct turns can dispatch multiple tool calls in order before the next model turn.
CodeActConfig now includes trajectory compaction settings, matching the ReAct context-window recovery path.
Tooling exposes infrastructure-error classification, fetch-policy controls, capped HTTP response reading, and symlink information in directory listings.

BREAKING ProtocolImpl.parsePropose now returns Proposal instead of Action, and trajectories can contain the new Summarized action injected by compaction.
BREAKING CodeActConfig record construction requires the new compaction field.
Tool bodies now rethrow budget and context-window exhaustion instead of feeding them back to the model as recoverable observations.
Built-in filesystem, shell, and web tools are hardened: execution timeouts are clamped, default fetch policy rejects common local/private targets, and fetches cap bytes while streaming.
Refreshed the internal shikumi bound for the 0.3 series.

Built-in filesystem, shell, and web tool modules, plus helpers for assembling built-in tool registries.
Shikumi.Tool.Env, an execution-environment record for filesystem and process backed tools.
Shikumi.Tool.Web, a swappable HTTP/search client layer for web tools.

ReActConfig now includes a compaction field, and ReAct agents compact older trajectory steps when usage approaches the model context window.
Refreshed the internal shikumi bound for the 0.2 series.

Constrained the baikai dependency to the 0.2 series and refreshed internal shikumi bounds.

Initial Hackage release of typed tool registries and ReAct-style agents for shikumi programs.
CodeAct, program-of-thought helpers, prompt utilities, and a restricted deterministic code interpreter for offline tests.