v0.2.1

v0.2.1

Added

• universal_find_references: with more than one searched root the report now lists the searched roots, marks the root each file path is relative to (file.rs (relative to <root>)) and states how many duplicate files were skipped — so a path like project/src/lib.rs can no longer be mistaken for a second copy of src/lib.rs.
• universal_find_references: the audit log entry now records searchedRoots and duplicatesDropped.
• 10 new tests (symbols.test.ts): nested-root pruning, duplicate detection via a junction, sibling roots with an identical relative path, single-root parity with universalFindReferences, and the report format (total 348).
• CI (.github/workflows/ci.yml): npm ci + npm run build + npm test on windows-latest and ubuntu-latest, plus an ubuntu-latest job that runs the same suite with MCP_PROJECT_ROOT / MCP_EXTRA_ROOTS / MCP_PROJECT_NAMES set (the root registry and project aliases are read from the environment at import time, so both cases have to keep passing). CI status badge added to README.

Changed

• run_command_grep: a zero-match response no longer returns a bare (žiadna zhoda) — it now states what was scanned: (žiadna zhoda — 0 z N riadkov výstupu, exit code 0) on the normal path, and (žiadna zhoda — výstup prázdny, exit code M) / exit code neznámy (signal/timeout) in the catch block. Nájdených N zhôd: … is unchanged. Motivation: the same text was returned for "output scanned, nothing matched" and "command produced nothing", which was misread as a lost command (feedback 2026-09-16 — the only actionable part of a report that was otherwise not reproducible). New helper formatZeroMatch() in src/format.ts + 3 tests (total 354).

Fixed

• universal_find_references: every match was reported twice when cwd was omitted — all registered roots were searched, and because the registry contains nested pairs (D:\W\TS + the projects inside it, …\Rust + …\Rust\pixel-blaster-engine) the same file was reachable through two roots. The dedup key was the root-relative path (different per root) and totalMatches summed every root, so ColAlign reported 56 matches for 28 real ones, the listing was truncated mid-output, and the two path prefixes made one file look like two. Nested roots are now pruned (pruneNestedRoots), files are deduplicated by resolved real path, and totalMatches is the sum over the files actually listed. Distinct files that exist in two sibling roots under the same relative path are no longer silently dropped.
• universal_find_references: cwd was documented as "default: primary project root" while the implementation searched every registered root — the description now matches the behaviour (tool description, cwd parameter, help_tool, README).
• The test suite is no longer Windows-only. Fixtures that were hardcoded D:\... / Z:\... literals are relative paths on POSIX, so the affected tests could only pass on Windows (resolveCwdRequested, resolveFilePath, resolve_cwd, buildRegistrations / findMatchingRegistration, findOutOfRootWriteTargets, findSuspiciousCrossRootReads, findAliasByName / findAliasByPath, findEscapeReason — git -C bypass). They now build a real tree under os.tmpdir() with path.join/path.sep — the pattern the multi-root fallback block already used — and separator-agnostic path assertions. Coverage is unchanged (348 tests, still green on Windows); POSIX is enforced by the new CI job.
• POSIX: the command-target safety checks were silently Windows-only. isFlag() tested ^\/[a-z] unanchored, and on POSIX every absolute path starts with /<letter> — so an absolute POSIX path was read as a cmd flag and findOutOfRootWriteTargets never saw the target of mkdir /tmp/x, mkdir -p /tmp/x or copy a.txt /tmp/x (worst case the copy/move branch fell back to a.txt). QUOTED_PATH (cross-root reads, active with MCP_BLOCK_CROSS_ROOT_READS=1) matched C:\, ../ and ~/ only, so quoted absolute POSIX reads were invisible. Both are fixed: isFlag() now accepts a /x flag only as a 1–2 letter token, and QUOTED_PATH gained a POSIX absolute branch (a quoted token is a path, never a flag). The 5 previously skipped assertions run again, plus 3 new platform-neutral regression tests (mkdir/copy at the filesystem root, quoted absolute read) that fail on the old code even on Windows.

Known Limitations (documented)

• Residual: a 1–2 letter directory directly under the filesystem root (/a) is still read as a cmd flag by isFlag() — use a longer name or the unquoted full path form for such targets.

v0.2.0

v0.2.0

Added

• close_feedback: closing an entry now also moves every closed entry out of the active log into .mcp/FEEDBACK_ARCHIVE.md, so .mcp/FEEDBACK.md keeps holding open items only (it had grown to 39 entries / 65 KB). The move is append-first (block written to the archive before it is removed from the log, so a crash can at most duplicate, never lose an entry), idempotent (entries already archived are skipped) and self-healing (entries closed before this feature are migrated with the next close). close_feedback returns archived (IDs moved) and archivePath.
• list_feedback: new archived parameter — true lists .mcp/FEEDBACK_ARCHIVE.md (closed entries) instead of the active log; all existing filters (type, tool, status) apply to the archive too.
• archiveClosedEntries() (src/feedback.ts) — bulk-migrates closed entries from the active log to the archive; also reused by close_feedback.
• batch_apply_edits: explicit machine-readable outcome — every response now carries message (human summary), appliedEdits (edits whose changes are on disk), written (files changed on disk) and applied: true|false on each preview entry. Failure responses additionally carry reason (validation_failed | write_failed), reverted (files rolled back) and nothingWritten. A validation failure now says plainly VALIDATION FAILED on edit #N of M — NO edits were written to disk instead of returning a preview that looks like a partial success; when writes did happen but were all rolled back it says NO net changes were left on disk: N edit(s) had been written and M file(s) were rolled back […]. The MCP tool result is the message followed by the JSON payload.
• batch_apply_edits: preview is padded to map 1:1 to the edits array — edits that were never reached get an action: "error" entry with not evaluated — batch stopped at edit #N.
• batch_apply_edits: excludePatterns parameter — when replaceAll is used with excludePatterns, occurrences inside excluded regions (e.g. #[cfg(test)] blocks) are skipped. Handles annotation-on-separate-line with 5-line lookahead for opening brace.
• batch_apply_edits: sequential validation in Phase 2 — each edit is re-validated against current file state (after previous edits) with rollback on failure. Phase 1 defers validation for chained edits on same file.
• findEscapeReason(): optional cwd parameter — when a command contains both git -C/--git-dir/--work-tree and a cd/pushd escape pattern, the function checks if the git path resolves inside cwd before flagging as escape. safetyCheck() now passes cwd to findEscapeReason().
• 9 new tests (total 317, up from 308)
• 6 new tests for explicit batch_apply_edits outcome reporting (total 323)
• report_tool_feedback: the reported tool name is now validated against this server's tool registry — unknown names are rejected without writing anything, with a "did you mean …?" suggestion for typos (Unknown tool 'batch_apply_edit' … Did you mean 'batch_apply_edits'?); when nothing is close, the error lists the tools this server exposes. New optional parameter allowUnknownTool: true bypasses the check for missing-capability reports about the server as a whole. Validation runs only when the handler supplies the registry, so the pure function stays usable without one.
• 8 new tests for the feedback archive (self-healing migration, idempotency, verbatim blocks, archived filters, log hygiene) (total 331)
• 7 new tests for tool-name validation + legacy log normalization (total 338)

Changed

• resolveFilePath() multi-root fallback now triggers even when cwd is provided but the file doesn't exist at the resolved location. Scans all allowed roots for the relative path — 1 match auto-resolves, 2+ shows disambiguation error.
• ALLOWED_ROOTS now includes paths from MCP_PROJECT_NAMES (friendly project names). Projects registered only via MCP_PROJECT_NAMES and not via MCP_EXTRA_ROOTS are now recognised as allowed roots — fixes file-based tools failing to resolve paths for alias-registered projects.
• run_command_grep: catch block now extracts stdout/stderr from exec error object. Non-zero exit code with output returns matches instead of error.

Fixed

• batch_apply_edits: sequential application — edit 2 can now find text created by edit 1 on the same file (was failing with "search string not found" because Phase 1 validated all edits against original file content).
• batch_apply_edits: replaceAll no longer breaks test code — use excludePatterns: ["#[cfg(test)]"] to skip test modules.
• batch_apply_edits: relative/absolute paths now resolve correctly when the target project is registered only via MCP_PROJECT_NAMES (was resolving against devkit root instead of project root).
• batch_apply_edits: partial rollback — when a chained edit fails, only files modified by the failed edit and subsequent edits are reverted. Files successfully modified by earlier edits are preserved (was: full atomic rollback lost all valid edits).
• extract_code_block: relative and absolute paths now resolve correctly for alias-registered projects (was returning "Cannot read file" for files within the project).
• run_command_grep: zero matches with non-zero exit code now returns empty result instead of error.
• batch_apply_edits: validation-failure response was ambiguous — the returned preview showed action: "edit", matchCount: 1 for edits that had validated while no edit was written at all, so callers could not tell that the batch was a complete no-op. The response now states VALIDATION FAILED on edit #N of M — NO edits were written to disk plus machine-readable appliedEdits, written, reverted, nothingWritten and per-entry applied.
• extract_code_block: relative paths now resolve across all allowed roots even when cwd is provided.
• report_tool_feedback: appending an entry no longer stacks --- separators — a trailing separator left by the previous write is removed first, so the log keeps exactly one separator between entries (.mcp/FEEDBACK.md had accumulated doubled separators over time).
• close_feedback / archiveClosedEntries(): rewriting the active log after an archive no longer leaves a doubled separator and extra blank lines between the header and the first entry.

Known Limitations (documented)

• batch_apply_edits: same-file chain failures revert the entire file (can't partially undo changes to a single file).
• run_safe_command: 60s client-side timeout limit (MCP client, not server) — use run_commands tool for long-running commands.
• Windows cmd.exe: ; is not a command separator — use & or separate tool calls.
• universal_find_references: symbols with special characters (colons, spaces) may fail when sent from certain MCP clients due to client-side JSON-RPC serialization issues.

v0.1.1

v0.1.1

Added

• Supported Languages section in README.md (Rust, TypeScript, Python, C++)
• Multi-root fallback for resolveFilePath(): relative paths without cwd now scan all allowed roots — unique match auto-resolves, multiple matches produce disambiguation error (6 new tests)
• 6 new tests for resolveFilePath multi-root fallback (total 300, up from 294)
• Tool Annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) for all 17 tools per MCP spec
• src/__tests__/tool-integration.test.ts — 23 new tests covering list_tools, help_tool, read_log_slice, list_allowed_roots, resolve_cwd
• M8ven Trust Index badge in README.md
• Release automation script (scripts/release.cjs)
• Pre-commit hook: blocks sensitive files (.env, .key, .pem, etc.)
• Pre-push hook: gitleaks scan + path scanner for tracked files
• 7 refactoring tools (Phase 1–3):
  • universal_find_references — language-agnostic symbol search with optional role detection (declaration/import/usage)
  • extract_code_block — code block extraction with annotation-aware, string/comment-safe bracket matching
  • split_file_by_declarations — split large files into modules with index generation (mod.rs/index.ts/init.py)
  • batch_apply_edits — atomic multi-file edits with validation-first and rollback on failure
  • generate_module_skeleton — generate module files by extracting symbols from source
  • verify_refactor_safety — semantic diff with 5 checks (function count, signatures, exports, imports, comment ratio)
• report_tool_feedback — agent feedback tool for bugs, improvements, feature requests (writes to .mcp/FEEDBACK.md, idempotent, project-protected)
• list_feedback — list feedback entries with optional filters (type, tool, status)
• close_feedback — close feedback entries by ID with optional resolution text
• list_tools — lists all available MCP tools with markdown-formatted descriptions, filterable by category
• help_tool — detailed help for any tool including markdown parameter table
• textResult() / jsonResult() helpers in src/format.ts for consistent MCP response formatting
• Test fixtures (src/__tests__/fixtures/) with Rust, TypeScript, Python test data
• Bypass test suite: 14 tests documenting security heuristic coverage and known limitations (base64, variable indirection, encoded cd)
• Integration test suite (src/__tests__/handlers.test.ts): 7 tests verifying tool output format
• Shared resolveFilePath() helper in src/safety.ts — unified file path resolution with alias lookup + allowed roots validation for all file-based tools
• cwd parameter for split_file_by_declarations, generate_module_skeleton, and batch_apply_edits (relative path resolution)
• 6 new tests for resolveFilePath() (total 269, up from 263)
• 116 new tests (total 263, up from 147)
• Shared parsing helpers exported from src/symbols.ts (tokenizer, bracket matching, annotation detection)
• tsconfig.json: exclude test fixtures from compilation

Changed

• resolveFilePath() now checks file existence on disk before falling back to scan other allowed roots — prevents silent misresolution when multiple projects share common filenames (e.g. index.ts, README.md)
• README.md: added cwd parameter to extract_code_block documentation, updated test badge to 300, described multi-root path resolution in batch_apply_edits and Typical workflow section
• PROJECT_CONTEXT.md: added language support entry, updated test count to 300
• index.ts refactored — extracted src/commands.ts (safetyCheck, executeCommand, executeGrep, resolveToolCwd) and src/helpers.ts (temp paths, audit log, line parsing); index.ts reduced from 1349 to 950 lines (-30%)
• Audit logging removed from list_tools and help_tool (discovery tools are now truly read-only)
• Upgraded @modelcontextprotocol/sdk from 1.29.0 to 1.30.0
• Refactored extract_code_block to use shared resolveFilePath() (replaced inline ad-hoc path resolution with safety validation)

Fixed

• README.md: test count badge updated from 294 to 300
• Closed feedback: batch_apply_edits relative paths now resolve across allowed roots
• Closed feedback: list_tools raw JSON envelope — identified as Cline client display issue, not server bug
• batch_apply_edits: added cwd parameter + resolveFilePath() — now resolves relative paths (was returning "Cannot read file" for files in newly created subdirectories)
• split_file_by_declarations & generate_module_skeleton: multi-line use/import blocks now collected completely (was truncated at first line, breaking generated imports)
• extract_code_block & split_file_by_declarations: removed # → line comment handling from tokenizer — Rust #[derive(...)]/#[allow(...)] attributes no longer cause missing closing braces in extracted functions
• split_file_by_declarations: now includes top-level use/import statements in generated modules
• split_file_by_declarations: detects and appends impl blocks associated with extracted types (Rust)
• split_file_by_declarations & generate_module_skeleton: relative file paths now resolve correctly via shared resolveFilePath() helper (was returning "Cannot read file")
• split_file_by_declarations: finds private declarations (fn, const, struct) and pub(crate)/pub(super) items
• generate_module_skeleton: filters extractCodeBlock matches to declaration-only — no more garbled output from usage sites
• generate_module_skeleton: imports now only collected from top-level (no indented use statements)
• split_file_by_declarations: cross-module references now generate use super::TypeName; imports
• split_file_by_declarations: cfg-aware cross-module imports — wraps with #[cfg(...)] when symbol is only used in cfg-gated code, falls back to #[allow(unused_imports)]
• split_file_by_declarations: mod blocks (e.g. #[cfg(test)] mod tests) now extractable
• batch_apply_edits: multiple edits to same file now accumulate correctly (was overwriting previous edits)
• batch_apply_edits: CRLF line endings now handled — normalizes to LF for comparison, preserves original endings in output
• extract_code_block: added cwd parameter — relative paths now resolved against cwd or primary root
• universal_find_references: without cwd, now searches ALL allowed roots (was defaulting to primary root only)
• LICENSE restructured — standard GPL text first, project copyright as appendix

v0.1.0

v0.1.0

Added

• 6 MCP tools: run_safe_command, run_destructive_command, read_log_slice,
  run_command_grep, list_allowed_roots, resolve_cwd
• Safety layers: dangerous-pattern blacklist, directory-escape detection,
  write-target checks, destructive confirmation, missing target guard
• Output normalization: Windows UTF-8, ANSI stripping, redirect reporting
• GPL-3.0-or-later license (full text in LICENSE)
• 147 passing tests (vitest)
• Branch protection: linear history, force push blocked
• README badge shields (Version, License, Tests, Node.js)
• GitHub About: description, homepage, topics