Local Coding Agent v4.4.3 - ChatGPT Web Compact & Resume

English

Long ChatGPT Web sessions can become difficult to continue. Opening a new chat often means explaining the repository, decisions, and unfinished work again. Local Coding Agent v4.4.3 adds a practical Compact & Resume workflow for that handoff.

What's new

• compact_context saves the current goal, summary, decisions, constraints, completed work, open tasks, next action, and touched files.
• resume_context restores the latest structured checkpoint in a new conversation.
• context_status shows the latest checkpoint and an MCP-traffic-based context pressure estimate. It does not claim to read ChatGPT Web's actual context window.
• The local dashboard can copy ready-to-use Compact and Resume prompts.
• Checkpoints are workspace-specific, immutable, retention-limited, and redact common credentials before they are written.
• Legacy checkpoint and resume tools remain available for compatibility.

Maintenance included

• Quiet tray health probes no longer flood the server request log.
• Manual health checks, MCP traffic, and failed requests remain visible for diagnosis.
• Network Doctor provides clearer guidance for tunnel HTTP 503 and reset-before-headers events.
• Public release checks now inspect tracked and untracked non-ignored files before publishing.

Update

Windows:

scripts\lca.cmd update

macOS/Linux:

bash scripts/lca update

Restart the tray app or server after updating. Windows users can also download the self-contained x64 executable attached below.

Verification

• Context memory: 3/3
• Pro regression: 28/28
• Agent integration: 30/30
• Security: 7/7
• Hardening: 31/31
• Evaluation suite: 20/20 (100%)
• Public release boundary: 72 files checked, 0 private-tree/channel leaks

Tiếng Việt

Khi cuộc trò chuyện ChatGPT Web kéo dài, việc chuyển sang chat mới thường buộc người dùng phải giải thích lại repo, các quyết định và phần việc còn dang dở. Local Coding Agent v4.4.3 bổ sung quy trình Compact & Resume để việc chuyển tiếp này gọn hơn.

Có gì mới

• compact_context lưu mục tiêu, tóm tắt, quyết định, ràng buộc, việc đã hoàn thành, việc còn lại, bước tiếp theo và các file đã chạm tới.
• resume_context khôi phục checkpoint có cấu trúc gần nhất trong cuộc trò chuyện mới.
• context_status hiển thị checkpoint gần nhất và ước tính áp lực ngữ cảnh dựa trên lưu lượng MCP. Công cụ không tuyên bố đọc được context window thật của ChatGPT Web.
• Dashboard local có nút copy sẵn prompt Compact và Resume.
• Checkpoint được tách theo workspace, không ghi đè, có giới hạn lưu giữ và tự che các thông tin đăng nhập phổ biến trước khi ghi xuống ổ đĩa.
• Hai tool cũ checkpoint và resume vẫn được giữ để tương thích.

Các bản vá đi kèm

• Health probe nội bộ của tray không còn làm đầy log request mỗi ba giây.
• Health check thủ công, lưu lượng MCP và request lỗi vẫn được giữ lại để chẩn đoán.
• Network Doctor giải thích rõ hơn lỗi tunnel HTTP 503 và reset-before-headers.
• Cổng kiểm tra public release giờ quét cả file đã track và file chưa track nhưng không bị ignore.

Cập nhật

Windows:

scripts\lca.cmd update

macOS/Linux:

bash scripts/lca update

Khởi động lại tray app hoặc server sau khi cập nhật. Người dùng Windows cũng có thể tải file EXE x64 self-contained đính kèm bên dưới.

Kết quả kiểm tra

• Context memory: 3/3
• Pro regression: 28/28
• Agent integration: 30/30
• Security: 7/7
• Hardening: 31/31
• Evaluation suite: 20/20 (100%)
• Public release boundary: 72 file, 0 dấu vết private/preview

Project status / Trạng thái dự án

Local Coding Agent is an independent community project. It is not an official OpenAI product and is not affiliated with or endorsed by OpenAI.

Local Coding Agent là dự án cộng đồng độc lập, không phải sản phẩm chính thức của OpenAI và không được OpenAI bảo trợ hoặc xác nhận.

Local Coding Agent v4.4.1 Prodev — Anti-Lag Customer Setup

Local Coding Agent v4.4.1 Prodev

Customer-ready Prodev release focused on reducing ChatGPT Web lag and making customer setup/support easier.

Highlights

• Reduced default tool payloads so long ChatGPT Web threads stay responsive:
  • read_file default: 12k chars
  • run_command default: 8k chars
  • read_many batch cap: 120k chars
• Added customer AI prompt generator:
  • node scripts/local-coding-agent.mjs prompt setup
  • node scripts/local-coding-agent.mjs prompt update
  • node scripts/local-coding-agent.mjs prompt diagnose
• Added one-command setup wizard report:
  • node scripts/local-coding-agent.mjs setup-wizard --workspace "C:\path\repo"
• Added skills doctor:
  • node scripts/local-coding-agent.mjs skills doctor
• Added dashboard copy buttons for setup/update/diagnose prompts.
• Added redacted support report helper for customer diagnostics.
• Updated README with AI Agent Quick Setup at the top in English and Vietnamese.
• Updated Windows tray app to 4.4.1-prodev.

Public release cleanup

• Removed the unpublished experimental v5 standalone folder from this release branch/tag.
• Release notes and public docs only describe stable Prodev features.

Fixes

• Fixed Windows CLI npm spawning by routing npm commands through cmd.exe /c.
• Lower default payload sizes reduce chat-frame lag caused by giant logs, diffs, base64, and repeated single-file reads.

Verification

• node --check server/server.mjs
• node --check scripts/local-coding-agent.mjs
• node --check scripts/support-report.mjs
• node scripts/validate-skills.mjs
• npm --prefix server run test:agent
• npm --prefix server run test:pro
• npm --prefix server run test:security
• npm --prefix server run test:hardening
• dotnet build tray-app/LocalCodingAgentTray.csproj -c Release --no-restore
• dotnet publish tray-app/LocalCodingAgentTray.csproj -c Release -r win-x64 --self-contained true -p:PublishSingleFile=true

Local Coding Agent v4.4.0-pro — Skill System and Customer Update Flow

What's new in v4.4.0-pro

This release focuses on the Skill System and Customer Update Flow so customers can ask their own AI coding agent to install, update, validate, and diagnose Local Coding Agent more safely.

Added

• Official shipped skill pack:
  • setup-local-coding-agent
  • update-local-coding-agent
  • debug-tunnel-network
  • customer-support
  • release-manager
  • security-hardening-review
  • skill-creator
  • code-review
• scripts/validate-skills.mjs for validating skill frontmatter, duplicate names, secret-like content, mojibake, and formatting issues.
• Universal CLI commands:
  • update
  • skills list
  • skills validate
• docs/CUSTOMER_UPDATE_PROMPT.md, a copy-paste prompt customers can give to Codex, Claude Code, Cursor, or another AI coding agent to update their local clone safely.
• CI coverage for the new CLI, network doctor, and skill validator entry points.

Changed

• Version bumped to 4.4.0-pro across server, npm package, tray app, and tests.
• Shipped skill docs and code-review skill were cleaned up and normalized.
• README now points users to the customer update prompt and documents update/skill commands in both English and Vietnamese.

Verified

• test:agent: 30/30
• test:pro: 24/24
• test:security: 7/7
• test:hardening: 26/26
• Eval suite: 20/20
• Skill validation: 8/8
• Windows tray build: 0 warnings, 0 errors

Local Coding Agent v4.3.0-pro — faster tool batches and safer approvals

What's new in v4.3.0-pro

This release focuses on faster local-agent workflows, safer approvals, and more reliable launch behavior.

Added

• run_commands: run up to 12 bounded commands in one MCP round-trip, sequentially or with limited parallelism.
• Enhanced read_many: supports structured per-file ranges, bounded concurrency, and a combined response cap.
• request_approval_batch: approve 2-20 exact actions at once without wildcard permissions.
• Approval TTL configuration via AGENT_APPROVAL_TTL_MINUTES.
• Batch read response cap via AGENT_MAX_BATCH_READ_CHARS.
• Health metadata for launcher/dashboard checks: pid, policy, allow_dangerous, config_id, and dashboard_port.
• CI syntax checks for PowerShell and Bash launchers.

Fixed and hardened

• Approval records are now one-time, exact-action grants with replay protection.
• Concurrent calls can no longer reuse the same one-time approval.
• Audit redaction now covers approval_token, mcp_auth_token, and control_plane_api_key.
• Path resolution now canonicalizes paths before root validation, improving symlink/junction escape protection.
• Launchers restart the managed server when effective config changes, and stop only the intended server process.

Verified

• test:agent: 25/25
• test:pro: 24/24
• test:security: 7/7
• test:hardening: 26/26
• Eval suite: 20/20
• Windows tray build: 0 warnings, 0 errors

Local Coding Agent v4.2.0-pro — Pro workflow, hardening & dashboard health

Local Coding Agent v4.2.0-pro

Bản 4.2.0-pro là big update tiếp theo sau v4.0.0, tập trung vào 3 thứ quan trọng nhất khi dùng ChatGPT/Codex qua MCP local: nhanh hơn khi thao tác qua tunnel, an toàn hơn khi cho agent sửa/chạy lệnh, và có bằng chứng kiểm thử rõ ràng trước khi bàn giao.

Release này gộp các nâng cấp nền của 4.1.0, lớp Pro snapshot của 4.1.1-pro, và bộ workflow Pro hoàn chỉnh trong 4.2.0-pro.

Điểm nổi bật

• Nâng version server, npm package và tray app lên 4.2.0-pro.
• Thêm bộ công cụ Pro mới cho agent:
  • workspace_snapshot
  • workspace_doctor
  • quality_gate
  • session_report
• Dashboard local có Health Score, Pro speed/safety tips, pending approvals, metrics latency chi tiết và mini file/diff viewer.
• Chính sách an toàn AGENT_POLICY được enforce thật: strict, balanced, full.
• Approval flow một-lần qua dashboard cho thao tác rủi ro.
• MCP hardening mạnh hơn: chặn browser origin mặc định, bearer token chỉ qua header, giới hạn body dạng streamed/chunked.
• State/backup/index/approval được tách theo workspace để tránh lẫn dữ liệu giữa nhiều project.
• Undo/backup mạnh hơn cho file mới, thư mục rename/move, destination rename và patch delete.
• CI/test suite mở rộng cho Pro + hardening.

Công cụ Pro mới

workspace_snapshot

Tool “mở màn” cho mỗi phiên làm việc. Một call trả về nhiều thông tin để giảm round-trip qua tunnel:

• roots/workspace đang active
• mode/policy/auth/safety model
• project profile, language/framework/package manager
• important files
• compact tree
• git status
• test/build/lint/typecheck commands
• optional symbol sample
• metrics summary
• health score
• next-best actions

Khuyến nghị: agent nên gọi workspace_snapshot trước khi đọc repo sâu.

workspace_doctor

Chẩn đoán readiness của workspace:

• cấu hình roots
• mode/policy
• MCP auth/origin posture
• ripgrep availability
• git state
• project profile
• quality commands
• README/security docs
• score và recommendations

Dùng tốt khi muốn hỏi: “repo này đã sẵn sàng để agent làm việc an toàn/chất lượng chưa?”

quality_gate

Runner kiểm thử có cấu trúc cho lint/typecheck/test/build:

• detect command tự động từ repo/profile
• hỗ trợ dry-run để xem plan trước
• chạy theo thứ tự gate
• có compact pass/fail summary
• hỗ trợ stop-on-failure

Dùng sau khi sửa code, trước khi báo hoàn thành.

session_report

Báo cáo cuối phiên:

• health score
• bottlenecks
• metrics/top tools
• git state
• doctor summary
• recommendations
• recent errors

Dùng trước handoff để người dùng biết agent đã làm gì, còn rủi ro gì, và nên làm tiếp bước nào.

Nâng cấp bảo mật & safety

AGENT_POLICY thật sự được enforce

• strict: chặn write/execute/mutation tools, phù hợp audit/read-only.
• balanced: dùng hằng ngày; thao tác bình thường vẫn mượt, thao tác rủi ro cần approval một lần.
• full: toàn quyền cho trusted automation/local regression test.

Các thao tác rủi ro trong balanced gồm delete, install/network commands, mutating git, risky background processes và destructive patch operations.

Approval queue trên dashboard

Dashboard local có hàng chờ approval:

• approve-once
• deny
• approval bị consume sau khi dùng
• không replay lại được
• dashboard action có kiểm tra origin

MCP boundary cứng hơn

• /mcp không còn nhận token qua query string ?token=....
• Nếu dùng token, phải dùng Authorization: Bearer <token>.
• Browser-origin MCP calls bị chặn mặc định.
• Chỉ mở bằng MCP_ALLOWED_ORIGINS khi thật sự tin tưởng origin đó.
• Request body limit tính cả streamed/chunked payload, tránh bypass giới hạn size.

Dashboard & telemetry

Dashboard local được nâng cấp thành “control panel” hữu dụng hơn:

• version/tier hiển thị 4.2.0-pro / pro
• Health Score
• Pro speed & safety recommendations
• latency avg/p50/p95/p99
• success rate
• calls/minute
• per-tool average/p95 latency
• recent call duration
• pending approvals
• file tree refresh
• read-only file viewer
• git diff viewer

Mục tiêu là nhìn vào dashboard biết ngay agent đang nhanh/chậm ở đâu, workspace có rủi ro gì, và có approval nào đang chờ không.

Tray app / launcher

• Tray app hiển thị version 4.2.0-pro.
• Thêm Policy dropdown để chọn strict, balanced, full.
• Tray truyền AGENT_POLICY xuống server process.
• .env.example, PowerShell launcher và shell launcher có thêm biến cấu hình policy/origin/approval.

Workspace-scoped state

Dữ liệu runtime được chuyển về dạng workspace-scoped:

server/data/workspaces/<workspace-id>/

Bao gồm:

• notes
• checkpoint
• repo index
• patch history
• backups
• approvals

Điều này giúp nhiều workspace không dùng nhầm patch history/backup/approval của nhau.

Undo/backup tốt hơn

Bản này cải thiện độ tin cậy của undo:

• backup cả file trước/sau trong patch diff
• backup created files để undo có thể xoá lại file mới tạo
• backup directory move/rename
• backup destination khi rename đè hoặc tạo mới
• patch history tách theo workspace
• lịch sử backup cũ được prune an toàn

CI & regression tests

Release đã thêm/chạy các bộ test:

• npm run test:pro
• npm run test:hardening
• npm run test:agent
• npm run test:security
• npm run test:security:baseline
• npm run eval
• npm audit --audit-level=moderate
• tray app Release build gate

Kiểm thử trước khi publish commit này:

• test:pro: 24/24 pass
• test:hardening: 15/15 pass
• git diff --check: pass
• server syntax checks: pass
• tray build trước đó: pass

Upgrade notes

1. Update source

git pull origin main
cd server
npm ci

2. Chạy server

npm start

Hoặc dùng launcher/tray app như trước.

3. Policy khuyến nghị

Cho dùng hằng ngày:

AGENT_POLICY=balanced

Audit/read-only:

AGENT_POLICY=strict

Automation/test tin cậy:

AGENT_POLICY=full

4. Nếu dùng MCP auth token

Không dùng URL kiểu:

/mcp?token=...

Hãy dùng header:

Authorization: Bearer <token>

5. Nếu dashboard vẫn hiện version cũ

Kiểm tra tray/client đang trỏ đúng thư mục server của repo mới nhất. Với Windows tray app, trường MCP app folder phải trỏ tới:

<repo>/server

Sau đó Stop/Start lại server từ tray và refresh dashboard.

Breaking / behavior changes đáng chú ý

• Query-string bearer token đã bị tắt vì lý do bảo mật.
• Browser-origin MCP request bị chặn mặc định.
• AGENT_POLICY=balanced có thể yêu cầu approval cho một số thao tác trước đây chạy ngay.
• Runtime data chuyển sang workspace-scoped path, nên metrics/history cũ có thể không còn nằm cùng vị trí như v4.0.0.

Commit

• Commit: 739bd8aa6b8eb269eaa374e7f3f8ef9a53264fe4
• Compare: v4.0.0...v4.2.0-pro

---

Nói ngắn gọn: v4.2.0-pro biến Local Coding Agent từ một MCP server chạy được thành một local coding cockpit có policy, approval, telemetry, quality gates và session report đủ nghiêm túc để dùng lâu dài.

Local Coding Agent v4.0.0

Local Coding Agent v4.0.0

V4 is the consolidated cross-platform release. It combines the complete v3
tool expansion with the v4 runtime, security, tunnel-authentication, tray-app,
and CI improvements.

Highlights

• 58 MCP tools for repository discovery, file editing, patching, command
  execution, process management, Git, skills, testing, review, planning,
  policies, approvals, and workspace profiles.
• Cross-platform runtime for Windows, macOS, and Linux, with native support
  for cmd, PowerShell, bash, sh, and zsh.
• Reliable process-tree cleanup using taskkill on Windows and POSIX
  process groups on macOS/Linux.
• Dashboard mini-IDE with workspace tree, read-only file viewer, Git diff,
  tool metrics, recent calls, and metrics reset.
• Windows tray app v4.0.0, targeting .NET 10 and provided below as a
  self-contained Windows x64 executable.

Repository intelligence and editing

• Repository profile, important-file discovery, repo map, symbol scan, and
  index freshness reporting.
• Patch preview and validation before writes.
• Automatic backup batches and undo_last_patch.
• Structured git_status and git_diff tools.
• Skill creation and deletion in addition to skill discovery/loading.

Test, review, and agent workflow

• Automatic test/build/lint command detection.
• Full and changed-test runners.
• Diff review, secret scanning, TODO scanning, and change summaries.
• Persistent task plans, task state, decision logs, policy controls, approval
  requests, and workspace profiles.
• A 20-scenario eval suite covering edits, undo, tests, path confinement,
  audit redaction, Git safety, repository mapping, checkpoints, planning,
  policies, and patch previews.

Security and tunnel hardening

• Canonical path confinement blocks symlink and junction escapes.
• Windows path checks are case-insensitive; POSIX checks remain case-sensitive.
• Unsafe raw-Git flags such as --output and -c are blocked in safe mode.
• Recursive audit redaction covers commands, patches, file contents, tokens,
  passwords, authorization headers, and nested secret-like fields.
• MCP_AUTH_TOKEN is forwarded through the tunnel using environment variables,
  without placing secrets in command-line arguments.
• The tray app now defaults to safe mode for new or empty configurations.
• AGENT_EXTRA_ROOTS_JSON provides unambiguous multi-root configuration.

CI and verification

• GitHub Actions runs on Windows, macOS, and Ubuntu.
• Windows tray publish gate using .NET 10.
• Local release verification completed:
  • MCP integration: 23/23 passed
  • Runtime security: 7/7 passed
  • Baseline security: 6/6 passed
  • Eval suite: 20/20 passed (100%)
  • npm audit: 0 vulnerabilities
  • Tray build: 0 warnings, 0 errors

Upgrade notes

• Building the tray app requires the .NET 10 SDK. The attached Windows x64
  executable is self-contained and does not require a separate .NET runtime.
• Keep AGENT_MODE=safe unless unrestricted command execution inside the
  configured roots is explicitly required.
• Restart both the MCP server and tunnel after upgrading, especially when
  MCP_AUTH_TOKEN is enabled.
• Port 8788 is reserved by the tunnel client; the dashboard defaults to
  8790.

---

Tóm tắt tiếng Việt

V4.0.0 hợp nhất toàn bộ nhóm tool v3 với runtime và bảo mật v4: tổng cộng 58
MCP tool, chạy trên Windows/macOS/Linux, có patch preview + undo, test/build/
lint runner, review và secret scan, planner/policy/profile, dashboard mini-IDE,
chặn symlink/junction escape, audit redaction đệ quy, truyền bearer token qua
tunnel an toàn và CI đa hệ điều hành.

File tray Windows x64 đính kèm là bản self-contained v4.0.0. Người dùng chỉ cần
.NET 10 SDK khi tự build từ source.

v3.0.0 — Codex-style coding agent (repo intelligence, safe patching, tests, review, planner, policy)

The biggest release yet — Local Coding Agent grows from a file/command tool into a Codex-style local coding agent: it can understand a repo fast, edit safely with undo, run the right tests, review its own diff, remember work across chats, and enforce a permission policy. 58 tools total.

Not a replacement for Codex/CLIs — a free, open-source MCP server you drive from ChatGPT Web (or any MCP client). Not affiliated with OpenAI/Anthropic. AGPL-3.0.

✨ What's new in 3.0

🧠 Repo intelligence (understand a repo in 1–2 calls)

• repo_map — tree + manifests + package scripts + project profile in one call (cached 5 min)
• project_profile — detects languages/frameworks/package managers (Node/React/Next/Nest, Flutter, Python, Go, Rust, .NET, Java…)
• important_files — README, tsconfig, CI workflows, Dockerfile, env examples…
• repo_symbols — functions/classes/routes via fast regex scan
• index_status — cache freshness

🩹 Safe patch engine + undo

• preview_patch (dry-run), validate_patch (conflict check), undo_last_patch
• Every write (write_file / replace_in_file / apply_patch / move / delete) now backs up first → mistakes are reversible

🧪 Smart test/build runner

• detect_test_commands, run_tests, run_build, run_lint, run_changed_tests (maps changed files → nearby tests) — with structured failure summaries

🔍 Review mode

• review_diff (P1/P2/P3 findings + PASS/WARN/BLOCK verdict), security_scan (secret/dangerous-call detection), todo_scan, change_summary

🗺️ Planner / thread memory

• task_plan, task_state, decision_log saved under .agent/state/; checkpoints snapshot the plan → open a fresh chat, resume, keep going

🛡️ Permission / approval layer

• AGENT_POLICY = strict | balanced | full (default balanced) on top of safe/full mode
• policy_status, explain_risk, request_approval, approve_request, deny_request — dangerous actions are classified and never run silently

🧩 Workspace profiles

• profile_status, reload_profile — auto-loads <workspace>/.agent/profile.json (mode, test commands, conventions…)

✅ Quality

• New eval suite (npm run eval, 20 scenarios). CI runs functional + security + eval.
• Tests this release: test:agent 23/23 · test:security 6/6 · eval 20/20 (all re-verified).

Notes

• The Windows tray app is unchanged in this release (server-side only). It still launches/manages v3.0.0.
• Carries forward all prior hardening (realpath root confinement, raw-git lockdown, recursive audit redaction, catastrophic-command blocklist, loopback-only bind).

v2.0.2 - security hardening v2

Raw git lockdown (blocks --output/--no-index/--ext-diff/--git-dir/--work-tree/-c/-C and remote/exec flags that could write files, run programs, or escape the repo), recursive audit redaction (nested apply_patch content/edits never hit data/audit.log), and a new security regression suite (npm run test:security) wired into CI. 6/6 security checks + 23/23 functional pass. No breaking changes.

v2.0.1 - security hardening

Hardening from an external code review: realpath root confinement (blocks symlink/junction escape), safe-mode git is read-only (mutations need AGENT_MODE=full), git_status/git_diff handle non-repos cleanly, and the audit log redacts secret/content fields. No breaking changes; full test suite 23/23.

v2.0.0 - mini-IDE dashboard, git diff, skill authoring, CI

The mini-IDE release.

New MCP tools

• git_status — parsed git status --porcelain (branch, per-file index/worktree codes, rename from -> to, staged/untracked), confined to roots.
• git_diff — git diff with optional path and staged, confined to roots.
• create_skill — author a skill: writes <skillsdir>/<name>/SKILL.md with YAML frontmatter + body (default <PRIMARY_ROOT>/.claude/skills); list_skills sees it immediately.
• delete_skill — remove a skill folder (only within a skills directory).

Dashboard mini-IDE (local-only, never tunneled)

• GET /api/tree, GET /api/file, GET /api/diff JSON endpoints on the dashboard server (port 8790).
• New Files section in /ui: read-only file tree + viewer + syntax-colored git diff toggle.
• Clear metrics button (POST /api/clear-metrics).

Tooling

• .github/workflows/ci.yml — Ubuntu, Node 20: install deps, NUL-byte guard, syntax check, start server on a temp workspace, wait for /healthz, run npm run test:agent.
• CHANGELOG.md summarizing v1.3.0 → v2.0.0.

Compatibility

All 27 existing tools unchanged; the full test:agent suite (23 checks) plus the new-tool checks pass. Safe defaults and the catastrophic-command blocklist are intact.