Releases: RutaTang/Clew
Release list
v0.1.12
Ask, sharpened
The Ask agent gets a full hardening pass — smarter exploration, real streaming, and language-server precision.
Language-server navigation for the agent
Ask can now call definition, references, and hover — the same semantic navigation you use, resolved by real language servers that clew manages on the server side. Tracing a call chain no longer drowns in same-named grep hits: the agent jumps straight to the defining line, lists every real usage, and reads type signatures. Servers start lazily per language, reuse across questions, survive crashes, and re-sync files you edit on disk. Nothing is ever auto-installed — provisioning stays a consented action in the app.
The answer streams for real
The agent now signals when it's done exploring and writes the final answer through a streaming request — every token reaches the panel as it's generated, instead of the old arrive-whole-then-chunk imitation. Mid-stream provider failures and dropped connections surface as errors rather than silently truncated "answers".
A tougher loop underneath
- 2.5× the exploration budget (30 steps), so cross-file questions get answered instead of cut off.
- Prompt caching on every step — long explorations now cost a fraction in both money and latency on providers that support it.
- Truncation safety: a step that overflows its token budget retries bigger; half-written tool calls never execute.
- Transient-error resilience: rate limits, overloads, and network blips back off and retry instead of failing the whole question.
- Reasoning-model compatibility: thinking blocks round-trip correctly for providers that emit them.
All of it verified end-to-end against a real rust-analyzer (including symlinked project roots and on-disk edits mid-conversation) and a live provider.
Full Changelog: v0.1.11...v0.1.12
v0.1.11
Jupyter notebooks, natively
.ipynb files now open as a native cell view — no webview, no kernel, nothing executes. clew renders what the notebook saved:
- Markdown cells render as documents, math (
$…$,$$…$$) and mermaid diagrams included. - Code cells get the same tree-sitter highlighting as the editor, with
In [n]badges. - Outputs are collapsed by default — expand per cell, or all at once from the header strip (
6 cells · 4 with output · python · Expand all outputs). Text outputs keep their ANSI colors (red tracebacks look like tracebacks), figures render as images, SVG as SVG. Interactive/HTML-only outputs show a named placeholder. - Everything composes: search greps a clean script projection of the notebook (not raw JSON), the outline panel becomes a cell map (headings + code cells), the Ask agent reads notebooks through the same projection, and jumps land on the target cell with a highlight ring.
Read-only by design, like the rest of clew.
Full Changelog: v0.1.10...v0.1.11
v0.1.10
Better function summaries
- Hover shows the full explanation. The hover peek no longer truncates the cached summary to one sentence — the whole explanation is there, wrapping and scrolling as needed.
- Anywhere on a signature line works. Hovering a function's parameters, parentheses, or keywords shows its summary — no need to land exactly on the name.
- Inline chips are gone. The end-of-line "— This function …" chips crowded the code and were truncated anyway; the hover peek (and the OUTLINE panel, where the toggle still applies as "Outline summaries") is their better home.
- No more empty hover boxes. A hover with nothing to say renders nothing.
Full Changelog: v0.1.9...v0.1.10
v0.1.9
Ask clew is now a real code agent
Ask no longer answers from one-shot retrieval — the server runs an agent turn: the model explores your project with read-only tools (search, read, outline, files, git history, semantic find, cached explanations) until it can answer, then writes an answer that cites real code.
- Step chips: every tool call renders live in the panel (
🔍 search "…" → 6,📄 read src/…) and is clickable — jump straight to the code the agent looked at. - No prerequisites: asking no longer requires building the semantic index first; the agent finds its own way. The retrieval mode remains as an automatic fallback.
- Stop button cancels a turn mid-exploration.
- Read-only by design: clew is a code reader — the agent explores, it never mutates your project.
Richer answers
- Syntax-highlighted code blocks: fenced code in answers (and Explain views) is highlighted by clew's own tree-sitter pipeline — the exact same palette as the editor.
- Clickable citations:
path:linereferences in answers are links that jump to that line in the editor. Unique bare file names (theme.rs:37) resolve too.
Fixes
- Horizontal scrolling now accounts for everything actually drawn — inline function summaries, inlay hints, and blame annotations can all be scrolled into view instead of being cut off at the pane edge (#1).
Full Changelog: v0.1.8...v0.1.9