xminal 0.0.0-stage → 0.1.1

package/AGENTS.md

@@ -0,0 +1,16 @@
+# AGENTS.md — instructions for coding agents working on XMINAL
+
+## Commands
+
+- Build: `npm run build` (tsc → `dist/`)
+- Verify everything (build + 28 tests): `npm run verify`
+- Run the CLI from source: `npm run build && node dist/index.js doctor`
+- Tests: `npm test` (node:test, no network required)
+
+## Conventions
+
+- Language: TypeScript strict, CommonJS output, Node >= 18, zero runtime dependencies.
+- Never commit secrets or `.xminal/` runtime state (gitignored). `dist/` is gitignored; rebuild with `npm run build`.
+- Sessions/checkpoints/usage live under `.xminal/` (gitignored). When changing the tool schema, update `FEATURE_MATRIX.md` and `EXECUTION-PLAN.md`.
+- Keep the CLI honest: no fabricated features, no invented URLs, and no silent behavior — surface errors, never swallow them.
+- Before every push: `npm run verify` and make sure CI is green.

package/CHANGELOG.md

@@ -0,0 +1,70 @@
+# Changelog
+
+All notable changes to XMINAL are documented here. Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), versioning: semver.
+
+## [Unreleased]
+
+## [0.1.1] — 2026-11-08
+
+Patch release: hardening + distribution fixes + community-ready install path.
+
+### Added
+- Built-in web UI at `GET /` on `xminal serve` (streaming SSE chat page, self-contained, dark).
+- `xminal skills install <path|git-url> [--global]` and `/skills install …` (folder or git URL, SKILL.md validated).
+- Skills trust layer: `.xminal/skills.lock.json` sha256 lockfile; `xminal skills --lock|--verify`; REPL warns on drift.
+- Per-tool and per-MCP-server token metering: `xminal` records every tool call (`/tools`, `/mcp` table, `run --json` `toolUsage`, `/cost` shows tool-token estimate).
+- `xminal audit verify` and `/audit verify|tail` — SHA-256 hash-chain verifier for permission audit receipts.
+- Secret redaction of tool output before the model sees it (private keys, `sk-*`, GitHub/Google/Slack tokens, JWTs) — strip to `[REDACTED:<label>]` with a notice; system prompt treats tool output as untrusted data.
+- Repo secret-scan test (`tests/secrets.test.js`) and `AGENTS.md`.
+
+### Fixed
+- REPL: piped stdin was echoed and stalled after the first command; now queue-based input, clean non-TTY mode, exits on EOF.
+- `Agent.compactHistory` could orphan tool results from their assistant; compaction now backs off to the tool-call boundary (unit-tested).
+- Gemini adapter had duplicated unreachable code; gemini provider entry de-duplicated (single `gemini` key, aliases renamed).
+
+### Changed
+- `dist/` is committed so installs never need devDependencies; `prepare` skips the build when typescript is absent.
+- README documents single-command install via release tarball; GitHub-URL npm install is broken in our environment (npm bug, deterministic ENOENT) — avoid recommending it.
+
+
+
+- Built-in web UI at `GET /` (streaming chat via SSE against `/v1/chat/completions`) for any `xminal serve` instance; public page, API stays bearer-protected.
+- Secret redaction before the model sees tool output: `redactSecrets` strips private-key blocks, `sk-*`/`ghp_*`/`github_pat_*`/`AIza*`/`xox*-*`/`sk-or-*`/`sk-ant-*` and JWTs from tool results (marked `[REDACTED:…]`), and the system prompt now treats tool output as untrusted data (prompt-injection boundary).
+- `xminal skills install <path|git-url> [--global]` — install skills from a local folder or any git URL with SKILL.md (validated, clear error otherwise); also `/skills install` in the REPL.
+- Skill trust/lockfile: project skills hashed into `.xminal/skills.lock.json` (`xminal skills --lock`, `xminal skills --verify`, `/skills --lock|--verify`); REPL warns on drift.
+- Global agent instructions (`AGENTS.md`).
+
+### Fixed
+
+- REPL no longer loses or infinitely spins on piped stdin: input is queue-based, non-TTY piped lines are processed in order and the process exits on EOF/`/exit`; readline `terminal: true` only when stdin is a TTY (no more echo of piped input).
+- `Agent.compactHistory("/compact")` could cut inside an assistant↔tool run, producing a session OpenAI/Anthropic reject ("tool call does not have a tool result"). Compaction now backs the boundary off to the assistant with tool calls. Same boundary handling as the model-context compactor (which was already safe).
+
+### Changed
+
+- Repo secret-scan test now honors a `secret-scan: allow` marker for intentional test fixtures.
+
+## [0.1.0] — 2026-11-08
+
+First runnable release.
+
+### Added
+
+- Agent runtime: streaming provider loop, tool calling, Plan/Build/Compose/Review/YOLO modes, context compaction, per-session cost/turn budgets, Ctrl+C abort.
+- Providers: OpenAI-compatible adapter (all `*_API_KEY` / local), native Anthropic Messages, native Gemini; failover router with cost tracking; **47 built-in providers, 258 model aliases** (models.dev snapshot, 2026-11-08) with pricing/context; keyless **OpenCode** default (`opencode/inkling-2-small-free`).
+- Tools: `read`, `write`, `edit`, `multi_edit`, `list`, `glob`, `grep`, `bash` (sandbox hook), `todo`, `memory`, `skill`, `checkpoint`, `git`, `worktree`, `verify`, `web_fetch`, `web_search`, `diagnostics`, `mcp_list`, `mcp_call`, `task`, `think`.
+- Sessions: JSONL event log, resume/branch, checkpoints + `/undo`, rewind, hash-chained audit receipts, project/global Markdown memory with search, `~/.agents/skills` discovery.
+- Safety: ask/auto/full permission engine, protected paths, dangerous/risky classifiers, never-prompt-safe read-only tools, `xminal audit verify` hash-chain verifier, SECURITY.md.
+- Orchestration: `task` subagent tool, Ralph verify loop (`xminal ralph`), git worktree manager.
+- MCP stdio server mode (`xminal mcp --serve`) exposing all XMINAL tools to any MCP client.
+- Telemetry: per-tool and per-MCP token metering — `recordToolUsage` stamps every call into `.xminal/receipts/tool-usage.jsonl` (args/result tokens ≈ chars/4, errors, duration, model); `xminal` REPL shows it under `/mcp` and `/tools`; `xminal run --json` emits `toolUsage`.
+- Interfaces: readline REPL + slash commands, headless `xminal run --json`, OpenAI-compatible HTTP server (`/v1/chat/completions` SSE, `/v1/agent`), Telegram bridge (long-poll).
+- CLI: `xminal`, `run`, `resume`, `init`, `doctor`, `models`, `sessions`, `skills`, `mcp`, `config`, `serve`, `telegram`, `ralph`, `worktree`, `version`, `help`; on PATH via `npm link` / `~/.local/bin`.
+- Docs: README, EXECUTION-PLAN (18 phases), PLAN, FEATURE_MATRIX (125/125), BRAIN.
+- Tests: 20 `node:test` suites (config, permissions, sessions, memory/skills, MCP client fixture, MCP stdio server, secret-scan) — no network required.
+
+### Known limitations
+
+- Live LLM endpoints not smoked on this machine (OpenCode free tier returns 429; no keys present).
+- Bedrock/Vertex/Azure native adapters missing (OpenAI-compatible coverage only).
+- Ink-class streaming TUI and web UI not yet implemented.
+- npm publish pending maintainer approval.

package/EXECUTION-PLAN.md

@@ -0,0 +1,135 @@
+# XMINAL — Execution Plan (15–20 phases → build → brand → release)
+
+**Owner:** XMINAL team
+**Working dir:** `/workspace/projects/ff`
+**Status:** v0.1.0 built · 18/18 tests · `xminal` on PATH · brand assets landed (logo/LICENSE/README/CHANGELOG)
+**Rule:** keep shipping until the whole project is running best-to-best. Do not skip a phase.
+
+Legend: `[x]` done · `[~]` in progress · `[ ]` pending
+
+---
+
+## Phase 0 — Project hygiene & toolchain `[x]`
+
+- **Deliverables:** Node 20 + TypeScript 5.9 build pipeline, `package.json` (bin `xminal` → `dist/index.js`), `.gitignore`, `npm pack` clean.
+- **Acceptance:** `npm run build` passes; `npm test` 18/18 PASS; tarball 100+ files.
+
+## Phase 1 — Brand identity `[x]`
+
+- **Deliverables:** name `XMINAL`, tagline `EVERY CODING AGENT. ONE TERMINAL.`, signature cyan gradient, ASCII banner (`render.ts` + README hero), SVG logo (`assets/logo.svg`), `LICENSE`, `CHANGELOG.md`, `xminal --version` prints tagline.
+- **Acceptance:** `xminal --version` → `xminal 0.1.0 — EVERY CODING AGENT. ONE TERMINAL.` ✓ logo on dark background ✓.
+
+## Phase 2 — CLI surface `[x]`
+
+- **Deliverables:** `xminal`, `xminal run`, `resume`, `init`, `doctor`, `models`, `sessions`, `skills`, `mcp`, `config`, `ralph`, `worktree`, `serve`, `telegram`, `version`, `help`; slash commands (`/mode`, `/model`, `/memory`, `/ralph` …).
+- **Acceptance:** every subcommand exits 0/non-zero correctly; `--help` complete; SIGINT aborts safely.
+
+### Phase 3 — Config system `[x]`
+
+- **Deliverables:** global (`~/.xminal/config.json`) + project (`.xminal/config.json`) + env (`XMINAL_MODEL`, `XMINAL_PERMISSION_MODE`) merge; `xminal init`; profiles-ready shape.
+- **Acceptance:** project overrides global; missing keys show honest warnings via `xminal doctor`.
+
+## Phase 4 — Provider layer `[x]`
+
+- **Deliverables:** OpenAI-compatible adapter, Anthropic native, Gemini native; router (failover, per-provider readiness, first-error surfacing); cost tracking; 47 built-in providers; 258 aliases from models.dev (2026-11-08 snapshot); keyless OpenCode default.
+- **Acceptance:** `resolveModel` resolves every alias; `xminal models` readable; Bedrock/Vertex/Azure listed as pending native adapters.
+
+## Phase 5 — Agent runtime `[x]`
+
+- **Deliverables:** streaming loop, tool calls, parallel turns, modes (Plan/Build/Compose/Review/YOLO), compaction (tool-pair safe), budget cap, abort on Ctrl+C, ephemeral subagents.
+- **Acceptance:** resumed sessions re-seed history; cost limit stops gracefully; protected paths never editable.
+
+## Phase 6 — Toolset `[x]`
+
+- **Deliverables:** `read/write/edit/multi_edit/list/glob/grep/bash/todo/memory/skill/checkpoint/git/worktree/verify/web_fetch/web_search/diagnostics/mcp_list/mcp_call/task/think`.
+- **Acceptance:** `/verify` auto-detects project test command; `/ralph` iterates until verify passes or N iterations fail.
+
+## Phase 7 — Sessions & memory `[x]`
+
+- **Deliverables:** JSONL event log, resume/branch (`Session.branch`), checkpoints (auto before edits), `undo`, rewind, hash-chained audit receipts, Markdown memory with search + `~/.agents/skills` discovery.
+- **Acceptance:** `undoLast` restores file bytes; audit chain verifiable; memory search ranks by tags.
+
+## Phase 8 — Safety & permissions `[x]`
+
+- **Deliverables:** modes ask/auto/full; protected paths (`.env`, `~/.ssh`, `~/.aws` …), dangerous/risky classifiers, audit log, sandbox hook (bwrap — `bwrap` not installed on this box), no secrets in prompts.
+- **Acceptance:** `.env` denied even in full mode; `rm -rf`/`sudo` denied without approval; read-only tools always allowed.
+
+## Phase 9 — Orchestration `[~]`
+
+- **Deliverables:** `task` subagent tool (explore/plan/review/general), `runRalph` with verify loop, `worktree` tool + `xminal worktree` CLI.
+- **Acceptance:** subagent gets fresh context + depth limit 1; ralph reports pass/fail with iteration count.
+
+## Phase 10 — Interop: MCP · ACP · LSP · DAP `[x]`
+
+- **Deliverables done:** MCP client (stdio JSON-RPC, initialize/list/call), `mcp_list`/`mcp_call` tools, offline echo fixture test, **XMINAL as MCP server (`xminal mcp --serve`)**, stdio protocol integration test.
+- **Pending:** MCP server mode (`xminal serve` exists for HTTP; MCP stdio server not yet), ACP stub, LSP language servers for Python/Go/Rust, DAP for debugging.
+- **Acceptance:** external MCP servers (filesystem, memory) callable from the agent; `xminal mcp` lists them.
+
+## Phase 11 — Interfaces: REPL · headless · HTTP · Telegram · TUI `[x/~]`
+
+- **Deliverables:** readline REPL with slash commands, headless `xminal run --json`, OpenAI-compatible HTTP server (`/v1/chat/completions` SSE + `/v1/agent`), **built-in streaming web UI (`/`)**, Telegram bridge (long-poll), cost status bar, banner.
+- **Pending:** Ink-class streaming TUI.
+- **Acceptance:** `curl /health` public; `/v1/agent` runs; Telegram responds (needs `TELEGRAM_BOT_TOKEN`).
+
+## Phase 12 — Branding & docs `[x]`
+
+- **Deliverables:** `README.md` (hero, quickstart, commands, config, safety model, honest v0.1 status), `CHANGELOG.md` (v0.1.0), `LICENSE` (MIT), `assets/logo.svg`, `scripts/install.sh`, `package.json` verify script + full `files` manifest.
+- **Acceptance:** first screen explains XMINAL in 5 seconds; every CLI command documented with example.
+
+## Phase 13 — Testing & quality gates `[x/~]`
+
+- **Deliverables done:** 18 `node:test` suites (config, permissions, session, memory/skills, MCP fixture), no network needed, `npm pack --dry-run` clean (111 files incl. brand assets), pipe-safe exit codes.
+- **Pending:** live provider contract tests (blocked: no credentials on this box), perf budget (startup <1s), npm publish dry-run.
+- **Acceptance:** `npm test` 18/18; contract suite runs in CI once secrets exist.
+
+## Phase 14 — Distribution & release `[~]`
+
+- **Deliverables:** GitHub Release v0.1.0 with tarball + release notes ✓, repo PUBLIC with CI green ✓, `scripts/install.sh`.
+- **Pending:** npm registry publish (`npm install -g xminal`) — needs user go-ahead; `npx xminal`.
+- **Acceptance:** fresh machine: `npm i -g xminal && xminal doctor` → green in <5 min. **Needs user go-ahead for publish.**
+
+## Phase 15 — Security hardening `[~]`
+
+- **Deliverables done:** `SECURITY.md`, repo secret-scan test, `xminal audit verify` (hash-chain verifier), prompt-injection boundary guards (system prompt + tool-output redaction with `src/util/redact.ts`), skill hash lockfile (`xminal skills --lock|--verify`).
+- **Acceptance:** planted canary secret never reaches the model; audit chain verifies ✓.
+
+## Phase 16 — Ecosystem & skills marketplace `[~]`
+
+- **Deliverables done:** `xminal skills install <path|git-url> [--global]` (local folder or git clone + SKILL.md validation), `xminal skills --lock|--verify` trust layer, `tests/skills-*.test.js`.
+- **Pending:** XMINAL skill pack (10 curated skills), MCP bundle list, plugin API stability, marketplace UI.
+- **Acceptance:** skill install from folder + git works ✓; 10 skills ship.
+
+## Phase 17 — Advanced orchestration `[ ]`
+
+- **Deliverables:** parallel worktree fleets with merge arbitration, swarm planner/executor split, cost-aware routing across cheap/strong models, background agents (daemon), cron/interval runs.
+- **Acceptance:** 3 worktrees run agents in parallel and merge cleanly; router picks cheapest capable model.
+
+## Phase 18 — v1.0 freeze `[ ]`
+
+- **Deliverables:** feature-freeze, docs-complete, success metrics (tokens/$ per task vs aider/codex baseline), v1.0 tag, announcement assets.
+- **Acceptance:** all phases green; `xminal` recommended for daily driver work on this repo.
+
+---
+
+## Sequencing
+
+```
+P0→P2→P3→P4→P5→P6→P7→P8  ── foundation (DONE)
+        ╲→P9→P10→P11       ── orchestration + interop (IN PROGRESS)
+                 ╲→P12      ── brand + docs (IN PROGRESS)
+                          ╲→P13→P14→P15→P16→P17→P18
+```
+
+## Blockers / needs user
+
+1. **Live LLM** — opencode free tier returns 429 on this box; unblock with OpenCode Go, an API key, or a local Ollama install.
+2. **npm publish + GitHub repo/release** — needs user decision + repo URL.
+3. **Telegram bot token** — for the bridge test.
+4. **Bedrock/Vertex/Azure** — native adapters (scope decision).
+
+## Links
+
+- `PLAN.md` — strategy & roadmap detail
+- `FEATURE_MATRIX.md` — 125/125 feature mapping
+- `BRAIN.md` — current state, decisions, resume point
+- `FEATURE_MATRIX.md` ↔ this file: each FEATURE_MATRIX row maps to a phase above

package/FEATURE_MATRIX.md

@@ -0,0 +1,173 @@
+# XMINAL — Feature Matrix
+## Har project ka unique feature + xminal me integration decision
+
+**Legend**
+- **Decision**: `Native` = xminal ke andar build, `Borrow` = MIT code/algorithm reuse, `Integrate` = MCP/ACP/PTY/CLI/HTTP se connect, `Reference` = design inspiration only
+- **Priority**: `P0` = MVP (v0.1–0.3), `P1` = v0.4–0.6, `P2` = v1+
+- Source: `mit_terminal_coding_agents_100plus.txt` / `.md` (125 projects, 4 layers, 11 deep-dives)
+- Note: stars/links in the source list need verification before borrowing code.
+
+---
+
+## A. Standalone Terminal Coding Agents & Harnesses
+
+| # | Project | Unique feature to extract | xminal Decision | Priority |
+|---|---------|---------------------------|-------------|----------|
+| 1 | OpenClaw | Onboarding wizard + multi-channel bridges (Discord/Slack/WhatsApp/iMessage/Teams) | Native (wizard) + Integrate (channels) | P1 |
+| 2 | Hermes Agent | Persistent FTS5/trigram memory, auto skill synthesis, ACP, 300+ models | Native (memory/skills) + Integrate (ACP) | P0 |
+| 3 | Superpowers | Spec → TDD → subagent dispatch → review methodology | Borrow (skill pack) | P0 |
+| 4 | Everything Claude Code | 28 specialist roles, 125 guides, 60 workflows, security rules | Borrow (skill packs) | P2 |
+| 5 | OpenCode | build/plan modes, @general subagent, 75+ providers, LSP, MCP pooling | Reference / fork candidate | P0 |
+| 6 | Claw Code | `claw doctor`, container-first, offline skill installs | Native (doctor, offline skills) | P0 |
+| 7 | Pi | pi-tui differential rendering, pi-ai unified API, JSON/RPC/print, sandbox | Borrow (TUI + provider API) | P0 |
+| 8 | Orca | 30+ agents in worktrees, coordinator→worker, inline diff, mobile | Native (worktree orchestrator) | P1 |
+| 9 | OpenHands | REST/WS Agent Server, Docker sandboxes, scheduled/webhook automations | Native (server) + Integrate (Docker) | P1 |
+| 10 | DeerFlow 2.0 | Long-horizon harness, Textual TUI, sub-agent orchestration, per-thread sandbox | Native (long-horizon mode) | P1 |
+| 11 | Ruflo | Recursive multi-agent swarms, policy-ledger pipelines | Native (swarm) | P2 |
+| 12 | nanobot | ~4k-line lightweight agent, scheduled tasks, multi-channel | Reference (light mode) | P2 |
+| 13 | Codewhale | Plan/Work/Operate, Ask/Auto-Review/Full, `/undo`, `/receipts`, ACP | Native (modes, receipts, undo) | P0 |
+| 14 | OH-MY-PI | In-process Bash + 58 utils, Hashline edits, 14 LSP, 28 DAP, advisor review | Native (tools, hashline) | P0 |
+| 15 | DeepSeek-Reasonix | Static binary, planner/executor split, per-turn rewind, Extension Protocol | Native (rewind, split sessions) | P0 |
+| 16 | Deep Agents Code | ACP server, subagent spawning, human-in-the-loop approvals | Native (HITL) | P1 |
+| 17 | Beads | Dolt graph issue tracker + persistent agent memory (`bd`, `beads-mcp`) | Integrate first, Native later | P1 |
+| 18 | Kilo Code CLI | Orchestrator mode, 500+ models, workspace checkpointing | Native (checkpoints) | P1 |
+| 19 | Happy | E2E-encrypted mobile companion, parallel session runner | Integrate (mobile bridge) | P1 |
+| 20 | OpenWork | Skills/commands/plugins, imports Claude Code session history | Native (session import) | P1 |
+| 21 | Prime Agent | RLM + persistent IPython tool, `/refine`, daemon sessions, heartbeat, budgets | Native (REPL tool, budgets) | P2 |
+| 22 | jcode | 19.1ms startup, 3.3MB sessions, swarm conflict alerts, inline images/LaTeX | Native (perf targets, media) | P1 |
+| 23 | SWE-agent | Agent-Computer Interface, SWE-bench harness, Docker isolation | Borrow (eval harness) | P1 |
+| 24 | Devika | Prompt → step-by-step plan → research → build/debug | Reference | P2 |
+| 25 | OpenCodex | Universal provider proxy, API translation, 40+ providers, failover | Native (provider router) | P0 |
+| 26 | OpenWiki | Auto codebase wiki, syncs AGENTS.md/CLAUDE.md | Native (wiki) | P1 |
+| 27 | DeepCode | Papers/specs → verified full-stack implementations | Reference | P2 |
+| 28 | Bolt.new | WebContainer full-stack scaffold/run/deploy | Integrate | P2 |
+| 29 | Plandex | 2M-token context, diff staging sandbox, branching, plan review | Native (staging, branching) | P0 |
+| 30 | MiMo Code | Plan/Build/Compose, memory, hooks, worktrees | Native (compose mode, hooks) | P1 |
+| 31 | Smol Developer | Embeddable repo-generation agent | Borrow (SDK pattern) | P2 |
+| 32 | Trae Agent | Lakeview step summaries, JSON trajectory recording | Native (trajectory log) | P1 |
+| 33 | Camofox Browser | Stealth headless browser, fingerprint spoofing, REST API | Integrate | P2 |
+| 34 | Open SWE | Slack/Linear/GitHub triggers → cloud sandbox → PR | Integrate (webhooks) | P2 |
+| 35 | Claurst | Rust Claude Code rewrite, KAIROS persistent workflows | Reference | P2 |
+| 36 | BrowserSkill | Drive logged-in Chrome/Edge via local daemon + extension | Integrate | P1 |
+| 37 | Kimi Code CLI | Video clip input, `/mcp-config`, coder/explore/plan subagents, ACP | Native (media input, subagents) | P1 |
+| 38 | NullClaw | 678KB Zig binary, ~1MB RAM, <2ms start, 18 channels | Reference (footprint targets) | P2 |
+| 39 | gh-aw | Markdown workflows → sandboxed GitHub Actions | Integrate | P2 |
+| 40 | gptme | Shell/IPython/file tools, local RAG, tmux control, web | Native (RAG, tmux) | P1 |
+| 41 | Grok CLI | xAI Grok provider adapter | Integrate | P1 |
+| 42 | Atomic Agent | Local llama.cpp control loop, browser driving, OpenAI-compatible serve | Native (serve mode) | P1 |
+| 43 | agent-of-empires | tmux + worktree session manager, mobile web UI | Borrow (session manager) | P1 |
+| 44 | ralph-orchestrator | Hat-based multi-phase Ralph loop | Native (loop engine) | P1 |
+| 45 | Tau | Readable Python Pi port, branching JSONL sessions | Borrow (session model) | P0 |
+| 46 | Moltis | Zero-unsafe, 2300+ tests, code index, voice I/O | Reference (voice, tests) | P2 |
+| 47 | Nanocoder | Ollama/LM Studio/OpenRouter + XML tool fallback | Native (fallback parser) | P0 |
+| 48 | ralph-tui | Task-list-driven autonomous execution/verification TUI | Native (task loop) | P1 |
+| 49 | Agentless | 3-phase localization → repair → patch validation | Borrow (repair pipeline) | P2 |
+| 50 | Nimbalyst | Visual + terminal worktrees, Kanban, visual diff | Native (Kanban/diff UI) | P1 |
+| 51 | Ouroboros | Self-evolving identity/memory, self-rewrites behind review gate | Native (self-improve, gated) | P2 |
+| 52 | Neovate Code | Plugin system, multi-provider, headless automation | Native (plugin API) | P1 |
+| 53 | Awesome CLI Coding Agents | Curated directory + automated star updater | Reference (docs/updater) | — |
+| 54 | oh-my-agent | Portable `.agents/`, stop-hook gates, independent judges | Native (judge/gates) | P1 |
+| 55 | Maki | Tree-sitter skeleton index, in-process code execution, Lua plugins | Native (skeleton index, Lua) | P1 |
+| 56 | agent-deck | Worktree-aware TUI, per-session MCP toggles | Native (MCP toggles) | P1 |
+| 57 | Pydantic Deep Agents | Live Run Forking, parallel branches merged by AI judge | Native (fork/merge) | P2 |
+| 58 | VT Code | Native OS sandbox, shell safety, multi-provider failover | Native (shell safety) | P0 |
+| 59 | Concord MCP | Shared work state, file-claim overlap detection, live messaging | Integrate (MCP) | P1 |
+| 60 | AgentSight | eBPF syscall-level observability for agents | Integrate | P2 |
+| 61 | Groq Code CLI | Groq fast-inference provider | Integrate | P1 |
+| 62 | GitClaw | Git-native agent identity, Ed25519-signed actions | Native (signed commits/actions) | P2 |
+| 63 | AgentBox | Docker/VM/cloud sandbox runner + encrypted secrets | Integrate | P1 |
+| 64 | amux (mixpeek) | Single-file tmux multiplexer, web dashboard, watchdog, REST | Borrow (dashboard/watchdog) | P1 |
+| 65 | ThinkWatch Lite | API/MCP gateway, key redaction, cost logs, risky-tool blocking | Native (gateway/guard) | P0 |
+| 66 | Claudexor | Multi-account quota tracking + rotation | Native (quota router) | P1 |
+| 67 | wmux | One prompt → 8 parallel worktree tasks, hunk-level diff adoption | Native (fan-out) | P1 |
+| 68 | AgentBridge | Claude + Codex live peers, mid-turn review, quota handoff | Native (peer sessions) | P2 |
+| 69 | Calyx | Unified approval inbox, status sidebar, diff review | Native (approval inbox) | P1 |
+| 70 | MulmoTerminal | PTY grid, tmux persistence, Web Push notifications | Native (web push) | P2 |
+| 71 | flameox | Profiling workloads, artifact provenance, run comparison | Integrate | P2 |
+| 72 | codex-profiles | Isolated account profiles, token-safe switching | Native (profiles) | P2 |
+| 73 | rpi (pi-rust) | 9 composable crates, crash-resumable JSONL, plugin ABI | Borrow (crate boundaries) | P1 |
+| 74 | Rove | Worktree+branch per task, SSH-resilient sessions, scriptable API | Native (SSH-resilient sessions) | P1 |
+| 75 | LoopTroop | LLM Council → PRD → beads → Ralph loops | Native (council) | P1 |
+| 76 | amux (andyrewlee) | tmux/worktree TUI, minisign-verified releases, confirm-gated merges | Borrow | P2 |
+| 77 | CliDeck | WhatsApp-style dashboard, live status, autopilot routing | Native (dashboard) | P1 |
+| 78 | aisw | Profile switching + repo-to-account shell guards | Native (guards) | P1 |
+| 79 | GridBash | Terminal grid + MCP coordination server | Native (grid layout) | P2 |
+| 80 | DvalinCode | Org policy engine, OS network egress, hash-chained audit | Native (policy/audit) | P1 |
+| 81 | 3code | Token-efficient agent, Termux support | Native (Termux target) | P0 |
+| 82 | great_cto | 34–70 specialist agents + compliance gates (PCI/HIPAA/SOC2/GDPR) | Borrow (compliance pack) | P2 |
+| 83 | AgenC Core | Daemon session lifecycle, background agents, local models | Native (daemon) | P1 |
+| 84 | authsome | Credential broker, OAuth2/PKCE injection, 45 providers | Native (credential broker) | P1 |
+| 85 | openHarness | 78 slash commands, 42 tools, OAuth 2.1 MCP, hooks, SDKs | Native (command/tool catalog) | P1 |
+| 86 | Crewly | Dev/QA/PM role orchestration without tmux | Native (roles) | P2 |
+| 87 | intentic | Docker sandbox + worktree per agent + Cloudflare tunnel | Integrate | P2 |
+| 88 | 5dive | Cron/heartbeat scheduling, Telegram control | Native (scheduler) | P1 |
+| 89 | Keen Code | TurnMemory summaries, Markdown plan artifacts | Native (context compression) | P1 |
+| 90 | AgentPlane | Git-native task→plan→approve→implement→verify workflow | Native (workflow state) | P1 |
+| 91 | skill-optimizer | Benchmarks SKILL.md across models, rewrites till pass | Native (skill eval) | P2 |
+| 92 | Martty | Native ACP client TUI, image preview, session history | Native (ACP client) | P1 |
+| 93 | OpenCastle | Compiles one project definition to native configs for other harnesses | Native (config export) | P1 |
+| 94 | pi-mem | Plain-Markdown persistent memory, zero deps | Native (memory files) | P0 |
+| 95 | Crab Code | Anthropic/OpenAI/DeepSeek/Bedrock/Vertex adapters | Integrate (provider adapters) | P1 |
+| 96 | Untether | Telegram voice/text bridge, permission buttons, cost tracking | Integrate | P1 |
+| 97 | Smelt | Normal/Plan/Apply/Yolo, Lua scripting, Vim keys, parallel subagents | Native (Lua plugin + Vim mode) | P1 |
+| 98 | showagent | Unified session browser, fuzzy search, transcript conversion | Native (session browser) | P2 |
+| 99 | clausona | Quota-aware multi-account routing, shared skills/history | Borrow (routing algorithm) | P1 |
+| 100 | pi-reflect | Session-transcript analysis → surgical AGENTS.md rule updates | Native (reflect engine) | P2 |
+| 101 | opencode-rust | 21-crate Rust OpenCode reimplementation, HTTP server | Reference | P2 |
+| 102 | AgentDiff | Per-line Git blame/provenance, signed attributions | Native (provenance) | P1 |
+| 103 | ContextStream | Cross-session MCP memory (decisions/lessons/code index) | Native (memory MCP) | P1 |
+| 104 | NEEDLE | Deterministic bead queue state machine, headless dispatch | Native (queue engine) | P1 |
+| 105 | Grinta | Durable checkpoints, LSP/DAP, validation gates, `doctor` | Native (validation gates) | P0 |
+| 106 | ralph-harness | PROMPT.md + specs + fresh-context iterations + hard caps | Native (loop caps) | P1 |
+| 107 | memcode (Memorix) | Bundled memory platform + MCP memory server | Integrate | P2 |
+| 108 | AshlrCode | 45+ tools, 18 skills, KAIROS autopilot, multi-provider failover | Borrow (KAIROS pattern) | P2 |
+| 109 | mini-swe-agent | ~100-line agent, 65% SWE-bench, Docker/Bubblewrap/Modal | Borrow (eval baseline) | P1 |
+| 110 | OpenHands SDK | Python SDK + REST/WS Agent Server + TS client | Native (SDK/server) | P1 |
+| 111 | GitHub Spec Kit | Spec-Driven Development toolkit | Borrow (spec templates) | P1 |
+| 112 | Claude Task Master | PRD → dependency-tracked tasks, 36-tool MCP | Native (task graph) | P1 |
+| 113 | wshobson/agents | 94 plugins / 202 agents / 185 skills with format adapters | Borrow (catalog/adapters) | P1 |
+| 114 | CrewAI | Role-playing multi-agent framework | Integrate | P2 |
+| 115 | AutoGPT-Code-Ability | Coding ability package for autonomous agents | Reference | P2 |
+| 116 | rtk | Compresses/filters noisy shell/build/test output before model | Native (output compression) | P1 |
+| 117 | Browse (Anomaly) | Real logged-in browser MCP | Integrate | P1 |
+| 118 | HumanLayer Skills | Control-loop skills for brownfield codebases | Borrow (skill pack) | P1 |
+| 119 | n-skills | Curated plugin/skill marketplace | Borrow (marketplace format) | P1 |
+| 120 | Agent Code (Avala) | Fullscreen TUI, ACP editor embedding, web search, opt-in telemetry | Native (telemetry opt-in) | P2 |
+| 121 | codehamr | Minimal 4-tool agent (bash/read/write/edit) | Native (minimal mode) | P0 |
+| 122 | Agent Terminal | AI-aware workspace isolation + structured output capture | Native (workspace isolation) | P1 |
+| 123 | crabcode (Blankeos) | Pure-Rust OpenCode-style TUI | Reference | P2 |
+| 124 | OpenHarness (HKUDS) | Ohmo personal+coding agent | Reference | P2 |
+| 125 | Open-Harness (MaxGfeller) | Code-first composable TS agent SDK | Borrow (SDK design) | P2 |
+
+---
+
+## Coverage summary
+
+- **Total projects covered:** 125/125
+- **Native builds:** ~72
+- **Borrow (MIT code/skills):** ~20
+- **Integrate (MCP/ACP/CLI/HTTP):** ~22
+- **Reference only:** ~11
+
+## Top 20 features that Freebuff currently lacks (highest leverage)
+
+1. Plan/Build/Compose + permission postures (OpenCode/Codewhale/MiMo)
+2. Durable sessions: resume, branch, rewind, checkpoints (Reasonix/Tau/Grinta)
+3. Provider router with translation + failover (OpenCodex/ThinkWatch)
+4. Persistent memory + codebase index + wiki (Hermes/OpenWiki/Maki)
+5. Subagents + parallel worktree orchestration (OpenCode/Orca)
+6. Ralph/autonomous loops with verification (ralph family/LoopTroop)
+7. Skills + hooks + plugin system (Superpowers/OpenWork/Smelt)
+8. MCP client + server + ACP interop (OpenCode/OpenHands/Martty)
+9. Sandbox + secrets broker + audit receipts (Codewhale/authsome/DvalinCode)
+10. Diff staging + inline review + approval inbox (Plandex/Calyx)
+11. Browser automation on real logged-in browser (Browse/BrowserSkill)
+12. Token/context compression (rtk/Keen/Hashline/Maki)
+13. LSP + DAP integration (OH-MY-PI/Grinta)
+14. Multi-account quota routing (Claudexor/clausona)
+15. Remote/mobile bridges (Happy/Untether/5dive/CliDeck)
+16. Self-improvement/reflection (Hermes/pi-reflect/Ouroboros)
+17. Eval harness (SWE-agent/mini-swe/skill-optimizer)
+18. Session browser + transcript conversion (showagent)
+19. Provenance/audit (AgentDiff/GitClaw/AgentSight)
+20. Single-binary/Termux packaging (Reasonix/NullClaw/3code)

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 XMINAL contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.