# Changelog All notable changes will be documented in this file. The project follows [Semantic Versioning](https://semver.org/) and uses Conventional Commits during development. ## [Unreleased] ## [0.7.0] - 2026-09-03 ### Added - A Node-compatible Streamable HTTP MCP server exposes the complete `signal_grep` search, pagination, source inspection, and structural-analysis contract through `/mcp`. - MCP sessions have a documented global count bound and proactive idle expiry while retaining independent cursor state for active clients. ### Security - Streamable HTTP rejects unconfigured browser origins to prevent DNS rebinding. Standard MCP clients without an `Origin` header are unaffected, and explicitly allowed browser origins receive the CORS preflight and exposed session headers required for direct access. ### Correctness - Malformed request targets and invalid UTF-8 fail as bounded HTTP errors instead of terminating the server or silently changing search input. - Initialization requests rejected before session registration release their service immediately and continue to count against the session bound until cleanup completes. - MCP shutdown reports protocol or service cleanup failures, and the published Node artifact has an executable startup, real-search, and shutdown smoke test. - Pi host packages are optional peers, so installing the standalone MCP server does not pull in a Pi runtime that it never executes. ## [0.6.6] - 2026-09-03 ### Added - A zero-result search scoped to a subpath automatically retries from the project root. Project-wide matches are returned as ordinary evidence, while a project-wide zero result states the effective scope and retained filters. - Analysis responses expose independent coverage for exact occurrences, syntax classification, related tests, navigation, and retained storage. Structural operations also report enumerated, parsed, skipped, cached, timed, and budget-exhaustion statistics. - `redact=true` optionally masks credential-like values and private-key bodies in text and structured output. It is disabled by default, never changes search admission or counts, and survives ordinary and analysis cursor pagination. - Cursor failures expose stable `E_CURSOR_*` categories for malformed, missing, expired, wrong-kind, conflicting-option, and invalid-offset requests. - Structural analysis accepts a strictly validated `maxFilesToParse` from 1 through 2,000; the default remains 200 and never limits the preceding full candidate search. ### Changed - Analysis modes now set top-level `totalMatches`, `storedMatches`, and `returnedMatches` from their actual analysis items instead of returning a misleading zero. - Related-test discovery uses a fast full-scope ripgrep prefilter and parses only likely test entries. Parsed syntax facts are cached by language and content hash within bounded service memory, with automatic invalidation after content changes. - `anyOf` accepts 2–64 literals. Requests above eight terms run as bounded parallel eight-term chunks and merge into one ordered, version-checked, pageable result. - Code-only analysis on an entirely unsupported source set fails explicitly. Mixed repositories retain supported evidence and collapse repeated unsupported-language diagnostics into one bounded summary. - Numeric input outside the declared `limit`, `context`, and structural parse bounds is rejected rather than silently clamped. - Git capability detection is cached per executable environment. Git 2.45+ uses `--no-lazy-fetch`; older Git works for full repositories and gives a focused capability error for partial/promisor clones. ### Fixed - Exact occurrence counts no longer become untrustworthy merely because syntax classification, related-test augmentation, display paging, or storage retention has a different coverage status. - Invalid cursors are validated before continuation-only parameter checks, so forged cursors no longer receive a misleading mode or option error. - High-cardinality unsupported-language diagnostics no longer consume the response with one repeated reason per file. ## [0.6.2] - 2026-09-03 ### Changed - Explicit absolute paths and `..` traversal can search and inspect outside the current working directory. Requests without a path remain cwd-scoped, Git change comparison remains cwd-scoped, and canonical path checks reject `.git` internals and known external credential and special-system areas. ## [0.6.0] - 2026-09-02 ### Added - `anyOf` accepts 2–8 distinct, case-sensitive, single-line literals and returns every retained exact occurrence in one version-bound analysis snapshot. Different input terms may overlap; ordered per-term counts, changed-line containment, pagination, and executable inspection remain explicit. - `mode="impact"` selects one reliably parsed JS/TS/TSX symbol by direct source target or ordinary search match, returns it first, inventories every exact same-spelling workspace candidate by syntax role, and appends existing related-test evidence. ### Changed - Analysis storage now derives `anyOf` and impact counts after applying the shared 50,000-item/32 MiB retention bound, so partial counts describe only pageable evidence. - Analysis call rendering identifies multi-term and impact requests while result rendering keeps the existing fail-open original-text boundary. - Tool registration is fixed to the independent `signal_grep` name. The extension no longer changes another search tool or exposes user commands. - The session status now reports returned queries, complete results, partial results, and automatic file organization in plain language. Cursor pages and source continuations do not inflate the query count. - Existing configuration files keep their interface locale; retired settings are ignored. ### Removed - Removed token-comparison accounting and its normal-format baseline text. - Removed the health, snapshot-clear, tool-override, and metrics command surfaces. Snapshot and process cleanup remain automatic and session-bound. ### Correctness - The syntax parser now runs from a published JavaScript worker, so Node can execute parser-backed operations when the npm package is installed under `node_modules`. The release gate rebuilds the worker and fails if the published artifact differs from its TypeScript source. - Impact rejects ambiguous, unsupported, stale, anonymous, and anonymous-default targets before its workspace content scan. Same-spelling candidates explicitly do not prove binding; related tests remain `not-run` with assertion coverage `not-evaluated`. - Unsupported occurrence files remain visible as unclassified evidence. Supported parser failures retain exact occurrences, mark the result partial, and never become an empty success. - Impact storage retains exact target/occurrence evidence before derived test candidates, merged tests use stable case identity instead of stale page indices, and high-cardinality diagnostics remain within an explicit 64-reason/4 KiB bound. - Workspace search targets are passed to ripgrep relative to the request cwd, so root-relative include/exclude globs apply identically during candidate revision enumeration and content matching instead of being defeated by an absolute search-root prefix. ## [0.5.8] - 2026-08-31 ### Added - Reliable JS/TS/TSX source parsing now identifies concrete implementations, including methods, private/static/accessor members, arrows, generics and JSX-adjacent syntax. Go syntax roles are also supported without requiring a Go toolchain at query time. - `mode="inspect"` returns complete UTF-8 source whenever the selected implementation fits. Larger source is represented as raw byte fragments with an executable `sourceCursor` continuation that remains bound to the same worktree revision or Git commit/blob. - Batched inspection merges same-file overlapping ranges before assigning the shared 16 KiB output budget, reads/parses a file once per request, and gives unused budget to later source blocks. - Explicit `allOf` accepts two or three case-sensitive literal terms. It can require the terms in one file or, for JS/TS/TSX, in one implementation's own code while excluding nested functions, comments, static strings, regular expressions and type areas. - A single-pattern `roles` filter distinguishes declaration, call, import, export, comment, string, JSX text, code and unknown occurrences in JS/TS/TSX/Go. Go call-versus-conversion and short-declaration ambiguity are labelled as candidates. - `changes` searches a fixed Git comparison across changed files or only one side's changed lines. Worktree comparisons use final working-tree content and include unignored untracked files; historical inspection stays pinned to commit/blob evidence. - `mode="outline"` pages version-bound JS/TS/TSX symbols. `mode="imports"` follows bounded static named/default ESM import and named re-export links. `mode="tests"` reports direct, indirect and weak related-test candidates without claiming coverage or test success. ### Changed - Broad summaries now reserve file navigation first, then show bounded source preview windows: at most 30 file rows, five previewed files, two non-overlapping windows per file and seven lines per window. - Advanced analysis output uses its own units (`files`, `functions`, `symbols`, relationships or evidence items). - Runtime parsing uses installed, pinned ast-grep native packages in an owned short-lived process. Runtime configuration cannot preload code, load environment files, or install missing parsers. ### Fixed - Empty and whitespace-only cursors fail before scanning. Every returned follow-up is complete JSON with its required cursor and retained path selection. - Source continuation rejects changed, expired, malformed and forged offsets; replaying the same valid token returns the same page. - Historical Git reads do not invoke repository hooks, external diff/text conversion, filters, lazy fetches or user ripgrep configuration. Current ignore rules continue to protect historical content. ## [0.5.6] - 2026-08-31 ### Added - Broad summaries now show bounded real first-retained-match samples across displayed files, with stable match numbers and explicit sample-omission counts. Samples are not relevance ranking. - `mode="inspect"` accepts either cursor-bound `matchIndices` or cursorless `targets`, each containing 1–5 entries. The complete batch shares 16 KiB, reports every target's outcome, deduplicates same-revision source lines, and provides explicit single-target retry requests for budget-limited evidence. ### Fixed - Summary cursors and follow-up instructions are included in model-facing text rather than only structured details. - Context windows no longer emit future matching lines early and duplicate them on later cursor pages. - Snapshot source revisions are bound only when metadata collected before and after the content scan agrees; changed, new, unreadable, and uncached files retain matching evidence but are explicitly unverified for context and inspection. - Git exclusions now take precedence over positive user globs, and explicit Git-internal search paths are rejected after resolving symbolic links and filesystem case aliases. - Human-facing summary rendering recognizes actual service responses, including their blank-line layout, samples, and cursor instructions. - Ctags and ripgrep share owned-process cleanup that terminates and awaits children after cancellation or protocol failures, preserving startup errors and failing explicitly when cleanup cannot finish. - Cursor inspection keeps a late occurrence visible in a long source line and reports source-range and clipped-line omissions in text and details; direct source inspection also rechecks revisions. - Dense same-line matches retain their path, line and stable match marker within bounded output; displayed occurrence ranges are capped at 20 per line with explicit omitted-range counts, without deleting snapshot occurrences. - Search and inspection excerpts preserve distant matches in non-UTF-8 files. Inspection and context keep ripgrep's LF-based line numbers when source contains bare carriage returns. - Implicit auto mode returns a navigable summary when one matching line exceeds its estimated-token detail target; hard byte limits and execution failures remain explicit errors. ### Changed - Candidate source metadata is bounded to 50,000 files with 16 concurrent reads. The extra names-only traversal and metadata reads do not limit the matching set or imply a search-speed improvement. - Summary formatting has a dedicated module; detail formatting, process ownership, source verification, and batch composition each have explicit responsibilities. - Existing single-target request forms remain valid. - Tool guidance distinguishes ordinary auto searches from deliberate detail paging, explains when exact matching evidence is sufficient, and gives inspection-only selectors and corrective errors instead of leaving models to guess which search parameters must be omitted. Publisher: **宝儿**. ## [0.5.2] - 2026-08-30 ### Added - Responsive Pi TUI renderers for search calls, ranked summaries, grouped matches, bounded inspection, partial-retention warnings, and errors in English and Simplified Chinese. - Collapsed result views adapt across narrow, medium, and wide terminals while expanded views preserve the complete original tool text. ### Changed - Human-facing rendering now fails open to the original result when details or text are not safely recognized, or when custom layout rendering fails; search text, structured details, cursors, and non-TUI behavior remain unchanged. ## [0.5.1] - 2026-08-29 ### Added - Count-ranked, cursor-paged file summaries expose every file distribution page without rerunning ripgrep. - One summary cursor can repeatedly select exact retained files, including bounded `paths` batches, while selection-bound match cursors fail closed if a continuation changes its file set. - Every rendered match carries a stable 1-based `{match #N}` marker; cursor-scoped `mode=inspect` accepts `matchIndex` instead of manually repeating `path` and `line`. ### Changed - Long matching lines use occurrence-centered excerpts so the reported match remains visible while absolute UTF-16 or byte ranges stay unchanged. - Oversized enclosing symbols return a byte-bounded source window centered on the requested line and report omitted line counts before and after it. - Surrounding context is revision-safe: changed files keep their retained matching lines but omit current-file context with explicit text and structured `contextChangedFiles`. ### Fixed - Universal Ctags invocation no longer passes an unsupported `--` separator, and health checks now validate the exact JSON/field options used by inspection instead of checking only `ctags --version`. - File summaries rank by descending match count before applying the per-page file limit, preventing a dominant late-alphabet file from being hidden. - Completing one filtered selection no longer destroys the underlying summary snapshot; the original summary cursor remains reusable until TTL eviction, bounds eviction, explicit clear, or session shutdown. - Inspection output now reserves response metadata space so the complete result remains within the documented 16 KiB boundary. ## [0.5.0] - 2026-08-29 ### Added - Context-aware implicit `auto` detail budgets use Pi's reported context remainder: full (2,000 estimated tokens), tight (1,000), or critical (500). Compact complete results still return directly, while explicit limits, `matches`, inspection, and cursor pages retain the default budget. - Search details expose `budgetTier`, `contextRemainderPercent`, and `resultTokenBudget` when a context decision is available; tight and critical responses also attribute the adjustment in model-facing text. - Human-facing interface text supports the persisted `locale` values `en` and `zh-CN`; English remains the default for existing configs. ### Fixed - Tool-name conflicts fail visibly instead of changing the requested search behavior silently. ## [0.4.0] - 2026-08-28 ### Added - Exact per-occurrence byte and UTF-16 match ranges derived from ripgrep JSON submatches. - `mode=inspect` for bounded source inspection with optional enclosing-symbol details from Universal Ctags. - Cursor continuation can select one retained file without rerunning the original search. - Source revision checks reject stale cursor-scoped inspections. ### Changed - Merged overlapping context windows so adjacent matches do not repeat source lines. - Replaced unbounded JSON-line buffering with an explicit bounded LF reader that preserves valid Unicode separators. - Search and inspection paths are confined to the working directory. ## [0.3.0] - 2026-08-27 ### Added - Opt-in persistent mode that overrides Pi's built-in `grep` with Signal Grep while registering exactly one public search tool; enabling Metrics performs the override/reload handoff automatically. - Independent differential tests that verify match parity with raw `rg` across case, literal, glob, exclusion, hidden-file, ignore, and Unicode semantics. ### Changed - Replaced the fixed 20-match default with an adaptive result budget: compact complete searches return directly, while oversized searches summarize before detail. - Metrics now derive an exact Pi-style normal baseline from the same stable search snapshot, eliminating duplicate scans and unsupported-comparison warnings. - Explicit `limit` requests return detail immediately in auto mode, while completed or cursorless snapshots are released as soon as they become inaccessible. - Empty and whitespace-sensitive patterns, paths, globs, and exclusions are preserved instead of being trimmed, matching ripgrep semantics. ## [0.2.0] - 2026-08-27 ### Added - Opt-in, session-local Status Line metrics comparing cumulative Signal Grep result-text token estimates with Pi's normal grep baseline. - Honest cumulative accounting for cursor pages, negative savings, and searches that normal grep cannot represent equivalently. ## [0.1.1] - 2026-08-27 ### Fixed - Included the contributor documentation referenced by the packaged README so npm and Pi Gallery links resolve. ## [0.1.0] - 2026-08-27 ### Changed - Relicensed the project from MIT to GNU AGPL v3.0 only. ### Added - Adaptive `signal_grep` tool with summary-first broad search. - Stable in-memory cursor snapshots with explicit partial-state reporting. - Bun 1.4 and Node.js 22 compatibility validation. - Type-aware Oxlint and deterministic Oxfmt quality gates. - Synthetic context-shape contract benchmark. - English and Simplified Chinese documentation. - AI-specific pull request and review protocol.