Release notes

Stella follows semantic versioning. Every merge to main cuts a release and publishes prebuilt binaries plus a Homebrew formula, so brew upgrade stella always tracks the latest build. This page records the notable, user-facing changes; the full commit-level changelog for any tag lives on its GitHub Release.

0.5.0

The first minor release of the 0.5.x line. It gathers the stabilization and feature work that accumulated across the 0.4.x patch series into one coherent step: a smarter edit path, graph-aware tooling, a durable context engine, and fleet-scale telemetry.

Editing & tools

• read→edit drift oracle. edit_file now keeps a per-file record of the content the model last saw (via a read, or its own successful write). When an edit's old_string no longer matches, Stella distinguishes an out-of-band change to the file from a genuine not-found, echoes the current content, and lets the model recover — instead of blindly retrying a stale edit.
• apply_edits — transactional multi-file edits. A new tool applies a batch of edits across many files atomically, with a dry-run preflight that reports exactly what would change before anything is written. Either the whole batch applies or none of it does.
• run_tests scope:"impacted". Test selection is now graph-driven: Stella walks the tree-sitter code graph from your changes to the tests that actually exercise them, so a focused change runs a focused suite.
• graph-derived planner context. The planner localizes a goal to the relevant files deterministically from the code graph, giving each turn a tighter, more relevant working set.

Context engine

• Adaptive-context baseline (Phase 0 & 1). The foundation of Stella's durable memory: a typed record taxonomy with scope and temporal axes, and canonical JCS hashing for stable content identity. Context receipts now emit typed decision events (loop-detection, budget, retry, and policy decisions) as parseable twins of the prose the agent already narrates, bridged into the policy-plane journal.
• compaction preserves the prompt-cache prefix. When compaction dedupes repeated blocks it now keeps the earliest copy, so the stable prefix the provider's prompt cache keys on survives — fewer cache misses, lower bills.

Fleet & telemetry

• Observatory: org / hub telemetry dashboard. A new view aggregates telemetry across an organization's agents and workspaces, with hub-and-spoke scoping for org / workspace / repo.
• Benchmark harnesses. loop-bench (a cheap turn-loop + context-query correctness harness) and the stella arena arena-bench adapter, which records runs as ContextGraph traces.

Terminal UI

• Syntax highlighting in edit mode. Markdown and TOML are now highlighted when editing skills and agents in the TUI.
• Interactive Command Deck. The docs landing page gained an interactive Command Deck with animated terminals.

Fixes

• Loop detection now sees through read_file's session-tally footer, so legitimate re-reads are no longer misread as a loop.
• The conversational fast path no longer authors a witness test for a bare greeting like hi.
• Exhaustive TUI event matches handle SpeculationDiscarded cleanly, unbreaking the clippy gate.

Install & upgrade

# Homebrew (macOS / Linux)
brew install macanderson/tap/stella
brew upgrade stella

# From a release tarball — pick your target
# https://github.com/macanderson/stella/releases/latest

Prebuilt binaries are published for aarch64/x86_64 macOS and Linux. Verify a download against the SHA256SUMS attached to the release.