@getappz/agentflare-darwin-x64 1.2.0 → 1.3.1

package/README.md

@@ -2,29 +2,64 @@
 
 # agentflare
 
-**Optimize AI CLI agents for cost and performance. A single Rust binary, no
-Node, no runtime dependencies — across Claude Code, Codex, Cursor, Windsurf,
-VS Code, Cline, and Continue.**
+**Run AI coding agents efficiently, and coordinate more than one of them.**
+**A single Rust binary, no Node, no runtime dependencies — across Claude Code,
+Codex, Cursor, Windsurf, VS Code, Cline, and Continue.**
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
+[![Status: Beta](https://img.shields.io/badge/status-beta-yellow.svg)](STATUS.md)
 
 </div>
 
 ---
 
+## Status: Beta
+
+agentflare is under active development. The optimization layer (lean-ctx
+integration, memory, Caveman/Ponytail wiring) is the most mature part — CI-gated,
+tested, in daily use. The multi-agent coordination layer (tasks, review, coaching,
+artifacts, handoffs) is newer and still finding its shape: CLI flags, MCP tool
+names, and on-disk formats there can change without a major version bump.
+
+See **[STATUS.md](STATUS.md)** for what's stable vs. still moving, before you
+build automation on top of a specific flag or MCP tool signature.
+
 ## What this is
 
-Two non-overlapping layers, plus two Claude-Code-only companions:
+agentflare is two things bundled into one binary:
+
+**1. Optimization** — cut the token/cost overhead of running a single AI coding
+agent session.
 
 | Layer | What it compresses | Tool |
 |---|---|---|
 | **lean-ctx** | tool I/O *within* a session — reads, shell output, search, up to 99% | [yvgude/lean-ctx](https://github.com/yvgude/lean-ctx) |
-| **engram** | knowledge *across* sessions — decisions, facts, preferences that survive a session ending | [Gentleman-Programming/engram](https://github.com/Gentleman-Programming/engram) |
+| **memory** (built-in) | knowledge *across* sessions — decisions, facts, preferences that survive a session ending | ships in the binary, SQLite + FTS5, no separate install |
 | **Caveman** (Claude Code only) | conversation verbosity, ~65% | companion plugin |
 | **Ponytail** (Claude Code only) | code-writing over-engineering | companion plugin |
 
-lean-ctx and engram aren't substitutes for each other — one saves tokens inside a
-session, the other saves the re-explaining tax across sessions.
+**2. Coordination** — a lightweight, local-first backend for running *more than
+one* agent (or agent session) against the same body of work, exposed as MCP
+tools any MCP-capable agent can call:
+
+| Capability | What it's for |
+|---|---|
+| **Work items** (`item`, `claim`) | A shared backlog — create/list/search items, claim one before working it so two agents don't collide, mark done. |
+| **Review** (`review`) | Submit findings against a diff/PR, run consensus across reviewers, track scores over time. |
+| **Artifacts** (`artifact`) | Publish specs, plans, and docs as versioned, shareable pages — a durable handoff surface between agents (and to you) instead of scratch files that vanish with the session. |
+| **Handoff** (`handoff`) | Pass context to a specific agent/runtime, addressed and threaded, when work moves from one agent to another. |
+| **Coaching** (`coaching` CLI + hook) | Small persistent nudges surfaced to an agent — at session start today, moving toward contextual triggers (tool-name and prompt-relevance) so rules only show up when actually relevant. |
+| **Comments, labels, projects, webhooks, channel_send** | Threaded discussion on items, categorization, cross-project views, and outbound notifications. |
+
+Everything above is local-first (SQLite-backed), no daemon, reachable over the
+same stdio MCP transport agentflare already exposes for the optimization layer.
+
+lean-ctx and the built-in memory aren't substitutes for each other — one saves
+tokens inside a session, the other saves the re-explaining tax across sessions.
+The coordination layer is a different axis entirely: it's not about a single
+session's token bill, it's about multiple agents (or multiple sessions of the
+same agent, over time) staying out of each other's way and handing off work
+cleanly.
 
 **Why Rust, not Node:** Claude Code doesn't bundle or require Node.js — it's a
 standalone compiled binary. A plugin whose hooks shell out to `node` breaks on
@@ -42,14 +77,15 @@ that wiring ships as a small `.codex-plugin/` manifest instead.
 
 Numbers below are each project's own published, reproducible benchmarks — attributed,
 not blended into a fake combined total, and not accepted on faith. Where a claim had no
-supporting evidence in its own repo, it's flagged instead of repeated.
+supporting evidence in its own repo, it's flagged instead of repeated. These cover the
+**optimization layer** specifically — the coordination layer is too new for a
+comparable benchmark suite yet (see [STATUS.md](STATUS.md)).
 
 | Tool | Published claim | Methodology | Confidence |
 |---|---|---|---|
 | lean-ctx | 98.1% compression (`map` mode), 96.7% (`signatures`), ~99.99% cached re-read | CI-gated, reproducible via `lean-ctx benchmark report .`, measured on a 50-file repo with the GPT-4o tokenizer | High — real, reproducible, methodology named |
 | Caveman | 65% avg output-token reduction (range 22–87%, 10 prompts) | Committed in `benchmarks/`/`evals/` — and its own docs flag the failure mode: ~1–1.5k input-token overhead per turn can make it net-negative on already-terse workloads (`docs/HONEST-NUMBERS.md`) | High — reproducible, unusually transparent about limits |
 | Ponytail | ~54% less code (94% ceiling on best task), ~20% cheaper, ~27% faster, 100% safe | 12 real feature tasks on a FastAPI+React repo, Haiku 4.5, n=4 — self-corrected an earlier overgeneralized single-shot figure | High — reproducible, self-corrected once already |
-| engram | "95.8% accuracy on LongMemEval-S — #1 globally" | **Not citing this.** No benchmark directory, results file, or harness exists anywhere in the repo, despite being listed as a shipped feature. | Unverifiable — flagged, not repeated |
 
 ### Real usage, one live project
 
@@ -60,11 +96,11 @@ for a sense of scale. Not a controlled benchmark; one data point, your mileage v
 lean-ctx   34.2M tokens saved   92% compression   $88.45 saved   (lifetime; lean-ctx gain)
 caveman    1.16M tokens saved (~65%)                              (single session; caveman-stats hook)
 ponytail   23 `ponytail:` shortcut markers logged, no token figure (ponytail doesn't measure per-repo savings)
-engram     2 sessions, 11 observations tracked, across 2 projects  (engram stats)
+memory     2 sessions, 11 observations tracked, across 2 projects  (agentflare memory sessions/search)
 ```
 
 Check your own: `lean-ctx gain` · `/caveman-stats` (Claude Code) · `ponytail-debt` skill ·
-`engram stats`. Don't trust this table blindly either — re-run those commands yourself.
+`agentflare memory context`. Don't trust this table blindly either — re-run those commands yourself.
 
 ---
 
@@ -127,20 +163,14 @@ plugin loader:
 codex plugin marketplace add getappz/agentflare
 codex plugin install agentflare
 ```
-then `agentflare init --agent codex` for the rules/lean-ctx/engram setup (Codex's
+then `agentflare init --agent codex` for the rules/lean-ctx setup (Codex's
 hook wiring itself comes from the plugin manifest, not `init`).
 
-Each run: writes rule files (if absent), installs lean-ctx (`npm install -g
-lean-ctx-bin && lean-ctx onboard`) and engram (`go install`/`brew`, never the
-prebuilt Windows binary — see below) if missing, wires hooks/MCP where the host
-supports it. Detection-first — already-satisfied components are skipped, nothing
-gets clobbered.
-
-**engram's install safety**: the project's own docs say prebuilt Windows binaries
-get flagged as AV false positives, and explicitly recommend `go install` or
-Homebrew instead. `agentflare init` only auto-installs engram through one of those
-two paths; if neither is available, it prints the exact command instead of
-downloading something that might trip your AV.
+Each run: writes rule files (if absent), installs lean-ctx (native `curl | sh`
+or Homebrew installer) if missing, wires hooks/MCP where the host supports
+it. Detection-first — already-satisfied components are skipped, nothing gets
+clobbered. Persistent memory ships in the binary itself — nothing to install
+for it.
 
 ## Docs-only fallback (Aider, other AGENTS.md readers)
 
@@ -158,13 +188,19 @@ src/
 │                         # dirs::home_dir() ignores HOME/USERPROFILE overrides on
 │                         # Windows, learned the hard way)
 ├── state.rs              # ~/.agentflare/state.json — on/off flag for the hooks
-├── rule_text.rs           # shared rule copy (Exa, git, lean-ctx, engram usage)
-├── engram_install.rs      # engram's safe-install logic (go install/brew only)
+├── rule_text.rs           # shared rule copy (Exa, git, lean-ctx usage)
+├── memory/                # built-in persistent memory (SQLite + FTS5)
+├── compact.rs             # ephemeral FTS5/BM25 relevance scorer (PreCompact hook)
+├── coaching/               # session nudges: rule storage, CRUD, CLI presentation
+├── claims.rs, review.rs    # work-item claiming, review/consensus (coordination layer)
+├── artifacts.rs, channels.rs   # artifact publishing, outbound notifications
+├── mcp_server.rs, mcp_prompts.rs  # MCP stdio server exposing both layers as tools
+├── auth.rs, auth_crypt.rs, auth_db.rs, auth_runner.rs  # auth profile vault
 ├── components.rs          # registry: each entry checks + fixes itself, host-aware
 ├── init.rs                # `agentflare init --agent X` — runs every component,
 │                           # wires hooks directly for claude-code/cursor
-├── hook.rs                # `agentflare hook session-start|prompt-submit --agent X`
-└── main.rs                 # clap CLI, dispatch
+├── hook.rs                # `agentflare hook session-start|prompt-submit|... --agent X`
+└── main.rs                 # clap CLI, dispatch across 45 modules
 
 .codex-plugin/              # Codex only — its hooks require the plugin loader
 install.sh, install.ps1      # installers (checksum-verified download / local build)
@@ -178,15 +214,15 @@ Adding a new managed component means adding one entry to `components.rs` — nei
 
 ## What Gets Created
 
-**Claude Code**: `~/.claude/rules/{exa,git,lean-ctx,engram}.md`, `~/.claude/settings.json` hooks section, `~/.config/{caveman,ponytail}/config.json`, `~/.agentflare/`.
+**Claude Code**: `~/.claude/rules/{exa,git,lean-ctx}.md`, `~/.claude/settings.json` hooks section, `~/.config/{caveman,ponytail}/config.json`, `~/.agentflare/` (includes the built-in memory database).
 
 **Codex**: project-local `AGENTS.md` (only if absent), `~/.agentflare/`.
 
 **Cursor**: project-local `.cursor/rules/agentflare.mdc`, `.cursor/hooks.json`, `~/.cursor/mcp.json`, `~/.agentflare/`.
 
-**Windsurf/VS Code/Cline**: project-local rules file (see table above), MCP config for lean-ctx/engram.
+**Windsurf/VS Code/Cline**: project-local rules file (see table above), MCP config for lean-ctx.
 
-**Continue**: `.continue/mcpServers/{agentflare,engram}.json`.
+**Continue**: `.continue/mcpServers/agentflare.json`.
 
 Nothing is created if it already exists.
 
@@ -196,8 +232,7 @@ Nothing is created if it already exists.
 
 Remove the binary (see Install section above), then remove whatever `init`
 wrote for the hosts you set up — see "What Gets Created" above. Ponytail/
-Caveman/engram plugins themselves stay installed (uninstall separately if
-wanted).
+Caveman plugins themselves stay installed (uninstall separately if wanted).
 
 ---
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@getappz/agentflare-darwin-x64",
-  "version": "1.2.0",
+  "version": "1.3.1",
   "description": "Optimize AI CLI agents for cost and performance",
   "bin": {
     "agentflare": "bin/agentflare"