v0.6.1

What's Changed

• Fix Claude Agent SDK edit compatibility by accepting deprecated legacy top-level replace args by @dszmaj in #16
• refactor: migrate npm scope to @earendil-works, switch bun→npm+vitest by @RimuruW in #19
• fix(hashline): align rendered line numbers by @XYenon in #18

New Contributors

• @dszmaj made their first contribution in #16
• @XYenon made their first contribution in #18

Full Changelog: v0.6.0...v0.6.1

v0.6.0

Added

• Add first-class replace_text edits.
• Add full and ranges return modes for edit previews.
• Add protocol metadata for snapshot validation, outlines, stale refresh anchors, and edit metrics.

Fixed

• Preserve hard links during atomic writes.
• Canonicalize mutation targets across aliases.
• Render applied edit diffs in the UI only while keeping model-visible responses compact.
• Preserve legacy payload compatibility and fenced result sections.
• Tighten range and snapshot metadata handling.

Changed

• Slim read/edit prompt guidance and edit response text for token efficiency.
• Drop snapshot ID rejection and silent autocorrection behavior from the protocol.
• Improve text-like MIME handling, including XML candidates.

Tests

• Expand file-kind coverage for XML read guard paths and harden related test coverage.

v0.5.4

Highlights

This release fixes #12: If text-file sample for file-kind ends containing a partial utf-8 character, file is incorrectly determined to be binary.

Fixed

• Preserve rendered diff previews when the edit tool returns results.
• Preserve @ signs and Unicode spaces when normalizing relative paths.
• Remove edit-application autocorrection heuristics to keep strict hashline semantics.
• Tighten UTF-8 classification for full-window reads while tolerating incomplete truncated input. (#12)
• Prevent unbounded _hasherCache memory growth.

Changed

• Share file reads between file-kind detection and the read/edit tools to reduce duplicate I/O.

Performance

• Cache the XXH32 hasher instance in hashline processing.

Documentation

• Sync README.md, AGENTS.md, and prompts with current tool behavior.
• Fix broken links in README.md.

Tests

• Add coverage for stale-position compound edits, EACCES/EPERM paths, line-ending/BOM helpers, and Windows-specific permission guards.

v0.5.3

What's Changed

Features

• Return updated anchors in edit results for chained editing workflows
• Shared file-mutation-queue bridge with tighter schema and prepareArguments hook
• Add hashline-specific prompt metadata (snippets & guidelines) for read and edit tools

Fixes

• Fix legacy edit line-range and multi-line delete tracking
• Fix fuzzy anchor validation and tighten error handling
• Fix prepend into empty file correctly
• Handle empty range deletes and empty edit results
• Avoid emitting sentinel anchors and correctly track changed span
• Compute updated anchors in final document coordinates
• Use final-document offsets for append tracking
• Preserve non-string legacy key values instead of silently dropping
• Return advisory for empty files, hide terminal newline sentinel
• Remove noisy warning heuristic for line-shift noise
• Enable chained anchors for legacy top-level replace

Docs

• Document read offset/limit, chained edits, and diff preview
• Replace pseudo-code payload with full JSON examples for edit
• Require edits to be a JSON array, not a JSON-encoded string

Other

• Sync with pi 0.63+ types and tighten peer dep minimums

New Contributors

• @yuiti-ara made their first contribution

Full Changelog: v0.4.1...v0.5.3

v0.4.1

Fixed

• Fixes #7: restore Pi tool registration by publishing the edit tool schema as a top-level JSON object, while preserving legacy oldText/newText compatibility.
• Fix EOF append behavior for files ending with a terminal newline so append inserts at the real end of file instead of creating an unintended blank line.

Verification

• Added regression coverage for schema publication and terminal-newline append handling.
• bun test

v0.4.0

What's Changed

Added

• Compact diff preview in edit results. Each successful edit now returns a condensed Diff preview: block showing changed lines with +/- markers and their new LINE#HASH anchors, making quick follow-up edits possible without a full re-read.
• Legacy compatibility mode. When a caller sends a top-level oldText/newText or old_text/new_text payload (the built-in edit format), the tool attempts an exact unique match and applies it. Usage is surfaced to the interactive UI as a warning — not to the model — so operators can see when hashline mode is not being used.
• Compatibility notifications. A turn-end notification is emitted to the UI when one or more edits in a turn fell back to legacy mode, with a count of affected edits.
• Input autocorrections. The tool now automatically strips accidental LINE#HASH: display prefixes or diff +/- markers copied into replacement lines, and corrects \t-escaped tab indentation when PI_HASHLINE_AUTOCORRECT_ESCAPED_TABS=1 is set.
• Binary and image file detection. Both read and edit now classify files before processing: images (JPEG, PNG, GIF, WebP) are handled as attachments; binary files are rejected with a descriptive error; only UTF-8 text files proceed to hashline processing.
• Out-of-range read offset reporting. Requesting an offset beyond the end of file now returns a clear advisory with the file's actual line count and valid offset range.

Fixed

• Stale anchor error snippets include retry anchors. On hash mismatch the error snippet marks stale lines with >>> and includes their current LINE#HASH for immediate retry, with surrounding context for range edits.
• Atomic writes preserve symlink targets, file mode, and detect symlink loops.
• Caller-owned edit arrays are not mutated — applyHashlineEdits clones its input before deduplication.
• Schema validation accepts legacy payloads and rejects mixed camelCase/snake_case keys.
• Hash semantics tightened — symbol-only lines use their line number as hash seed to reduce collisions on } / { etc.
• Unsafe truncated previews are rejected — read returns an advisory instead of a partial hashline when the first line exceeds the byte budget.

Changed

• Edit tool is strict hashline-only. Free-text relocation removed; legacy oldText/newText is a hidden compatibility fallback only.
• grep tool override removed — use the built-in grep tool.
• Migrated from npm to Bun (bun.lock replaces package-lock.json).
• Test suite reorganized into layered directories (core/, tools/, extension/, integration/).
• tsconfig.json added for explicit TypeScript strict-mode checking.

Full Changelog: https://github.com/RimuruW/pi-hashline-edit/blob/master/CHANGELOG.md