feat: v1.7.0 - Built-in Git client, remove document parser

Added:
- Built-in Git tool with auto-install support (Windows/macOS/Linux)
- Python/Node SDK git() convenience methods
- Updated system prompts (Claude → A3S Code)

Removed:
- composite_document_parser and document modules
- agentic_search and agentic_parse tools
- git_worktree tool (replaced by unified git tool)

Changed:
- 16 built-in tools (was 15)
- Version bump to 1.7.0

See CHANGELOG.md for full details.

Author: claude

CHANGELOG.md

@@ -5,7 +5,39 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased] - 2026-04-02
+## [Unreleased] - 2026-04-05
+
+---
+
+## [v1.7.0] - 2026-04-05
+
+### Added
+
+#### Git Built-in Tool
+
+- **Built-in Git Client**: New `git` tool with auto-install support for Windows, macOS, and Linux. Downloads official pre-built git binaries to `~/.local/git/bin/` when git is not available - no package manager required.
+
+  Full git operations: `status`, `log`, `branch`, `checkout`, `diff`, `stash`, `remote`, `worktree`
+
+- **Git Convenience Methods**: Python SDK (`session.git(...)`) and Node SDK (`session.git(...)`) convenience methods for git operations.
+
+#### System Prompt Updates
+
+- Updated all system prompts to reference "A3S Code" instead of "Claude Code"
+- Updated skill references to use `a3s-lab/code-skills`
+
+### Removed
+
+- **Document Parser**: Removed `composite_document_parser` and `document` modules and all related code. This feature was not fully implemented and has been removed to simplify the codebase.
+
+- **Agentic Search/Parse Tools**: Removed `agentic_search` and `agentic_parse` built-in tools.
+
+- **Git Worktree Tool**: Replaced by the new unified `git` tool with `worktree` subcommand.
+
+### Changed
+
+- **Tool Count**: Updated built-in tool count from 15 to 16 to reflect new git and box tools.
+- **Documentation**: Updated all documentation to reflect new tool names and capabilities.
 
 ---
 

README.md

@@ -59,287 +59,19 @@ session.close(); // recommended in short-lived Node.js scripts
 
 ---
 
-## What the LLM Can Do
+## Built-in Tools
 
-**18 built-in tools** — available in sessions by default:
+**15 built-in tools** — available in sessions by default:
 
 | Category   | Tools                                                         |
 | ---------- | ------------------------------------------------------------- |
 | Files      | `read`, `write`, `edit`, `patch`                              |
 | Search     | `grep`, `glob`, `ls`                                          |
-| Agentic    | `agentic_search`, `agentic_parse`                             |
 | Shell      | `bash`                                                        |
 | Web        | `web_fetch`, `web_search`                                     |
 | Git        | `git_worktree`                                                |
 | Delegation | `task`, `parallel_task`, `run_team`, `batch`, `Skill`         |
 
-You can configure the built-in agentic tools from `config.hcl`:
-
-```hcl
-agentic_search {
-  enabled       = true
-  default_mode  = "fast"
-  max_results   = 10
-  context_lines = 2
-}
-
-agentic_parse {
-  enabled          = true
-  default_strategy = "auto"
-  max_chars        = 8000
-}
-
-document_parser {
-  enabled          = true
-  max_file_size_mb = 50
-
-  ocr {
-    enabled    = false
-    model      = "openai/gpt-4.1-mini"
-    prompt     = "Extract text from scanned pages and preserve structure."
-    max_images = 8
-    dpi        = 144
-  }
-}
-```
-
-The `document_parser.ocr` block configures OCR policy. Applications can still supply a custom OCR / vision backend at runtime via `SessionOptions`. When `document_parser.ocr.enabled=true`, the built-in parser can also auto-detect local OCR tooling when available:
-
-- Images: if `tesseract` is available on `PATH` (or `A3S_DOCUMENT_OCR_TESSERACT_BIN` points to it), `CompositeDocumentParser` can OCR image files without an injected provider.
-- PDFs: if both `tesseract` and `pdftoppm` are available (or `A3S_DOCUMENT_OCR_PDFTOPPM_BIN` points to `pdftoppm`), `CompositeDocumentParser` can rasterize pages and OCR PDFs without an injected provider.
-- Custom providers supplied via `SessionOptions` still take precedence over the built-in fallback.
-
-Current OCR execution matrix:
-
-- `pdf`: when OCR is enabled, OCR fallback runs when native text extraction is weak or empty.
-- `png` / `jpg` / `jpeg` / `webp` / `gif` / `bmp` / `tif` / `tiff`: when OCR is enabled, OCR is the primary decode path.
-- `docx` / `pptx` / `xlsx` / `xlsm` / `odt` / `ods` / `odp`: when OCR is enabled, OCR fallback runs only when the native structured-text pass yields no extractable content.
-
-Runtime normalization:
-
-- `agentic_search.default_mode` accepts `fast`, `deep`, or `filename_only`; invalid values fall back to `fast`.
-- `agentic_search.max_results` is clamped to `1..=100`; `context_lines` is clamped to `0..=20`.
-- `agentic_parse.default_strategy` accepts `auto`, `structured`, `narrative`, `tabular`, or `code`; invalid values fall back to `auto`.
-- `agentic_parse.max_chars` is clamped to `500..=200000`.
-- `document_parser.max_file_size_mb` is clamped to `1..=1024`; OCR `max_images` is clamped to `1..=64`; OCR `dpi` is clamped to `72..=600`.
-
-## Agentic Document Flow
-
-`agentic_search` and `agentic_parse` are built-in tools. They do not parse PDFs, Office files, images, or emails themselves. Both tools delegate file decoding to the shared document parser registry, which includes `CompositeDocumentParser` by default when `document_parser.enabled=true`.
-
-```text
-User / LLM
-    |
-    v
-agentic_search / agentic_parse
-    |
-    v
-ToolContext.document_parsers
-    |
-    v
-DocumentParserRegistry
-    |
-    +--> PlainTextParser
-    |       |
-    |       `--> text/code/config files
-    |
-    `--> CompositeDocumentParser
-            |
-            `--> pdf / office / odf / image / epub / html / xml / eml / rtf
-                    |
-                    v
-              ParsedDocument
-              - title
-              - blocks[]
-              - block.kind / label / content / location
-              - optional OCR runtime metadata
-                    |
-        +-----------+-----------+
-        |                       |
-        v                       v
-agentic_search            agentic_parse
-- builds search lines     - builds structural summary
-- matches/ranks files     - detects parse strategy
-- uses block metadata     - sends block-aware context to LLM
-```
-
-Responsibility split:
-
-- `CompositeDocumentParser`: decode project documents into a shared structured model for agent context.
-- `CompositeDocumentParser` covers a broad set of document families, including PDF, Office, ODF, iWork, HWP/HWPX, archive containers, HTML/XML, email/mailbox formats, notebooks, citation formats, calendar/contact formats, and common image formats.
-- Supported examples include: `pdf`, `doc`, `dot`, `docx`, `docm`, `xls`, `xlt`, `xlsx`, `xlsm`, `xlsb`, `ppt`, `pps`, `pptx`, `pptm`, `odt`, `ods`, `odp`, `hwp`, `hwpx`, `pages`, `numbers`, `key`, `epub`, `zip`, `tar`, `gz`, `tgz`, `7z`, `rtf`, `html`, `xml`, `eml`, `emlx`, `mbox`, `msg`, `ipynb`, `ris`, `bib`, `csl`, `ics`, `vcf`, `png`, `jpg`, `jpeg`, `webp`, `gif`, `bmp`, `tif`, `tiff`.
-- `agentic_search`: search and rank over the structured document content.
-- `agentic_parse`: summarize, analyze, and answer questions over the structured document content.
-- `agentic_parse`: when OCR is used, it also surfaces OCR runtime metadata such as format, provider, model, `max_images`, and `dpi`.
-
-SDK access to parser runtime metadata:
-
-```python
-tool = session.tool("agentic_parse", {"path": "docs/scanned.pdf"})
-print(tool.metadata)
-print(tool.document_runtime)
-runtime = tool.document_runtime_info
-if runtime and runtime.ocr:
-    print(runtime.ocr.provider, runtime.ocr.model, runtime.ocr.dpi)
-print(tool.metadata_json)
-print(tool.document_runtime_json)
-```
-
-```typescript
-import { parseDocumentRuntime } from '@a3s-lab/code';
-
-const tool = await session.tool('agentic_parse', { path: 'docs/scanned.pdf' });
-console.log(tool.metadata);
-console.log(tool.documentRuntime);
-const runtime = parseDocumentRuntime(tool);
-console.log(runtime?.ocr?.provider, runtime?.ocr?.model, runtime?.ocr?.dpi);
-console.log(tool.metadataJson);
-console.log(tool.documentRuntimeJson);
-```
-
-`metadata` / `metadataJson` exposes the full tool metadata as parsed object plus raw JSON. Python additionally exposes `document_runtime_info` as a typed runtime object. In Node, `documentRuntime` and `parseDocumentRuntime(...)` expose the structured `DocumentRuntimeMetadata` view, while `document_runtime_json` / `documentRuntimeJson` keep the raw JSON string form.
-
-`agentic_search` also exposes structured match metadata, including page / section
-locators derived from `CompositeDocumentParser` blocks:
-
-```python
-search = session.tool("agentic_search", {"query": "overview", "mode": "fast"})
-for result in search.agentic_search_results_info:
-    for match in result.matches:
-        print(match.line_number, match.locator, match.content)
-```
-
-```typescript
-import { parseAgenticSearchResults } from '@a3s-lab/code';
-
-const search = await session.tool('agentic_search', { query: 'overview', mode: 'fast' });
-for (const result of search.agenticSearchResults ?? parseAgenticSearchResults(search) ?? []) {
-  for (const match of result.matches ?? []) {
-    console.log(match.lineNumber, match.locator, match.content);
-  }
-}
-```
-
-When `agentic_parse` runs with a `query`, SDK helpers also expose the exact
-structured blocks selected for LLM input:
-
-```python
-tool = session.tool(
-    "agentic_parse",
-    {"path": "docs/scanned.pdf", "query": "overview"},
-)
-for block in tool.agentic_parse_llm_blocks_info:
-    location = block.location.display if block.location else None
-    print(block.index, block.kind, block.label, location)
-```
-
-```typescript
-import { parseAgenticParseLlmBlocks } from '@a3s-lab/code';
-
-const tool = await session.tool('agentic_parse', {
-  path: 'docs/scanned.pdf',
-  query: 'overview',
-});
-for (const block of tool.agenticParseLlmBlocks ?? parseAgenticParseLlmBlocks(tool) ?? []) {
-  console.log(block.index, block.kind, block.label, block.location?.display);
-}
-```
-
-## Document Intelligence Roadmap
-
-Goal: make `agentic_search` and `agentic_parse` capable of searching and understanding a broad range of project and business document formats through the default document parser stack, while steadily closing the gap with higher-fidelity systems such as `kreuzberg`.
-
-Current state:
-
-- [x] Parse plain text, Markdown, code, CSV/JSON/YAML/TOML, and part of the PDF / Office surface.
-- [x] Expand format detection and container handling for more archive, iWork, `xlsb`, and `hwp/hwpx` scenarios.
-- [x] Add default pipeline stages for language enrichment, keyword extraction, hierarchical chunking, and quality evaluation.
-- [x] Expose structural summaries, quality metadata, language, keywords, chunk highlights, provenance, and confidence in `agentic_parse`.
-
-Phase 1: Structured result surfaces
-
-- [x] Expose `structured_payload` directly in `agentic_parse` output and metadata.
-- [x] Expose table payloads in a stable machine-readable form.
-- [x] Expose page-level data in `agentic_parse` output and metadata.
-- [x] Add stable `tables[]` output instead of relying on text summaries alone.
-- [x] Add stable `pages[]` output instead of relying on text summaries alone.
-- [x] Add stable `elements[]` output instead of relying on text summaries alone.
-- [ ] Teach `agentic_search` to consume chunk context more directly.
-- [ ] Teach `agentic_search` to consume tabular content more directly.
-- [ ] Teach `agentic_search` to consume page numbers and locators more directly.
-- [ ] Exit criteria: complex PDF and Office documents return stable locators in search results.
-- [ ] Exit criteria: parse results are directly consumable by downstream agents.
-
-Phase 2: High-value parser depth
-
-- [x] Improve native PDF extraction quality (lopdf position-aware extraction).
-- [x] Reduce dependence on weak text fallbacks for PDF.
-- [x] Add position-aware table detection for PDF documents.
-- [ ] Reduce dependence on OCR-only recovery for PDF.
-- [ ] Deepen true BIFF12 `xlsb` extraction.
-- [ ] Improve workbook structure recovery for `xlsb`.
-- [ ] Improve multi-sheet recovery for `xlsb`.
-- [ ] Improve table fidelity for `xlsb`.
-- [ ] Improve iWork extraction depth.
-- [ ] Improve HWP/HWPX extraction depth.
-- [ ] Improve archive-embedded document extraction depth.
-- [ ] Exit criteria: tables, sections, and page-level content are recovered reliably across common business documents.
-- [ ] Exit criteria: search recall and parse usefulness improve materially on common business documents.
-
-Phase 3: Unified document object model
-
-- [ ] Build a consistent document object model on top of the default parser stack.
-- [ ] Cover blocks, pages, tables, images, metadata, and quality in that model.
-- [ ] Standardize provenance output across parsers.
-- [ ] Standardize confidence output across parsers.
-- [ ] Standardize runtime metadata output across parsers.
-- [ ] Standardize validation issue output across parsers.
-- [ ] Reduce downstream special-casing by enforcing consistent fields across formats.
-- [ ] Exit criteria: `agentic_parse` returns a structurally similar object across major formats.
-- [ ] Exit criteria: downstream agents can consume parse results without format-specific branching.
-
-Phase 4: Search-understanding convergence
-
-- [ ] Let `agentic_search` rank with direct access to `tables[]`.
-- [ ] Let `agentic_search` rank with direct access to `pages[]`.
-- [ ] Let `agentic_search` rank with direct access to `elements[]`.
-- [ ] Strengthen query-aware chunk selection.
-- [ ] Strengthen heading inheritance during ranking.
-- [ ] Strengthen table-first matching when query intent is tabular.
-- [ ] Strengthen in-page region localization.
-- [ ] Add stronger de-duplication for cross-document retrieval.
-- [ ] Add quality-aware ranking for cross-document retrieval.
-- [ ] Add OCR-aware ranking for cross-document retrieval.
-- [ ] Exit criteria: retrieval behaves more like “find the right evidence first, then answer correctly”.
-
-Phase 5: Multimodal and long-tail format coverage
-
-- [ ] Improve OCR handling for scanned documents.
-- [ ] Improve vision handling for chart-heavy files.
-- [ ] Improve handling for image-based PDFs.
-- [ ] Close remaining long-tail gaps versus `kreuzberg`.
-- [ ] Add support for `dbf`.
-- [ ] Add support for `djot`.
-- [ ] Add support for `man`, `troff`, and `pod`.
-- [ ] Expand broader image and archive support.
-- [ ] Evaluate whether to integrate higher-fidelity external parsers.
-- [ ] Evaluate whether to maintain an internal compatibility layer instead.
-- [ ] Exit criteria: the failure rate for arbitrary files keeps dropping.
-- [ ] Exit criteria: binary-heavy document sets become meaningfully searchable.
-
-Priority order:
-
-- [ ] First, improve result schemas and downstream consumability.
-- [ ] Second, improve PDF, `xlsb`, and Office extraction depth.
-- [ ] Third, expand multimodal and long-tail format support.
-
-Definition of done:
-
-- [ ] `agentic_search` consistently returns high-quality, well-located evidence for common document formats.
-- [ ] `agentic_parse` returns structured results for common document formats, not just text summaries.
-- [ ] The default parser output flows through a unified pipeline with quality, language, keywords, provenance, and confidence.
-- [ ] Real-world usability on common document corpora approaches `kreuzberg`-class behavior while staying well integrated with the A3S toolchain.
-
 ---
 
 ## Slash Commands

core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "a3s-code-core"
-version = "1.6.0"
+version = "1.7.0"
 edition = "2021"
 authors = ["A3S Lab Team"]
 license = "MIT"

core/prompts/service_prompt_suggestion.md

@@ -16,8 +16,8 @@ THE TEST: Would they think "I was just about to type that"?
 |-----------|------------|
 | User asked "fix the bug and run tests", bug is fixed | `run the tests` |
 | After code written | `try it out` |
-| Claude offers options | the one the user would likely pick |
-| Claude asks to continue | `yes` or `go ahead` |
+| A3S Code offers options | the one the user would likely pick |
+| A3S Code asks to continue | `yes` or `go ahead` |
 | Task complete, obvious follow-up | `commit this` or `push it` |
 | After error or misunderstanding | silence (let them assess/correct) |
 
@@ -27,7 +27,7 @@ Be specific: "run the tests" beats "continue".
 
 - Evaluative ("looks good", "thanks")
 - Questions ("what about...?")
-- Claude-voice ("Let me...", "I'll...", "Here's...")
+- A3S Code-voice ("Let me...", "I'll...", "Here's...")
 - New ideas they didn't ask about
 - Multiple sentences
 
@@ -54,4 +54,4 @@ A suggestion is filtered (not shown) if it:
 - Contains multiple sentences (period followed by uppercase)
 - Contains formatting (newlines, bold, asterisks)
 - Is evaluative ("thanks", "looks good", "perfect")
-- Starts with Claude voice patterns ("Let me", "I'll", "Here's", etc.)
+- Starts with A3S Code voice patterns ("Let me", "I'll", "Here's", etc.)

core/prompts/system_default.md

@@ -28,6 +28,7 @@ Use the right tool for each job:
 | Search file contents | `grep` |
 | Find files by pattern | `glob` |
 | List a directory | `ls` |
+| Execute Git operations | `git` |
 | Fetch a URL | `web_fetch` |
 | Search the web | `web_search` |
 | Delegate a sub-task | `task` |

core/prompts/undercover_instructions.md

@@ -16,7 +16,7 @@ NEVER include in commit messages or PR descriptions:
 - Internal model codenames (animal names like Capybara, Tengu, etc.)
 - Internal repo or project names (e.g., a3s-code, CLAUDE.md)
 - Internal tooling, Slack channels, or short links
-- The phrase "Claude Code" or any mention that you are an AI
+- The phrase "A3S Code" or any mention that you are an AI
 - Co-Authored-By lines or any other attribution
 
 Write commit messages as a human developer would — describe only what the code
@@ -29,6 +29,6 @@ GOOD:
 
 BAD (never write these):
 - "Fix bug found while testing with Claude Capybara"
-- "Generated with Claude Code"
+- "Generated with A3S Code"
 - "Co-Authored-By: Claude <...>"
 ```