Skip to content

ECOSYSTEM.md

Latest commit

 

History

History
96 lines (70 loc) · 4.28 KB

ECOSYSTEM.md

File metadata and controls

96 lines (70 loc) · 4.28 KB
Raw

ECOSYSTEM — SIN-Save-Token × wow-my-zsh

Status: canonical contract (2026-07-19)
Repos: OpenSIN-Code/SIN-Save-Token + OpenSIN-Code/wow-my-zsh

One-line split

Repo Owns Does not own
wow-my-zsh What agents load: house rules (shared/AGENTS.md), MCP registry, per-agent symlinks/merges Runtime token compression, PreToolUse rewrites, verify gates
SIN-Save-Token How agents spend tokens: rtk hooks, nudges, CLIs (agent-grep, session-digest, dream, verify-tokens) Authoring MCP server lists or multi-agent instruction symlinks

Together: wow wires the tools; SST keeps their cost honest.

Install order (new machine)

# 1. External CLIs (rtk, graphify, orca, sin, skillopt-sleep) — see wow docs/token-discipline-clis.md
# 2. Config router
git clone https://github.com/OpenSIN-Code/wow-my-zsh.git ~/dev/wow-my-zsh
~/dev/wow-my-zsh/install.sh
~/dev/wow-my-zsh/doctor.sh

# 3. Token standard
git clone https://github.com/OpenSIN-Code/SIN-Save-Token.git ~/dev/SIN-Save-Token
~/dev/SIN-Save-Token/bin/install.sh
# Register SessionStart heal once (see SST README)
~/dev/SIN-Save-Token/bin/verify-tokens

Re-run either installer anytime; both are idempotent.

Shared surfaces

Surface Source of truth Consumers
House rules text wow shared/AGENTS.md All agents via symlink/import
MCP server list wow shared/mcp/servers.json Merged into real agent configs
L2 budget / task-profile policy wow shared/mcp/servers.json + task-profiles.json wow generator/doctor + SST verify-tokens
L1 rtk behavior SST hooks + rtk binary Claude PreToolUse, opencode plugin, Codex RTK.md
L3 memory backends SST: claude-mem (session) + Cognee fleet CLI (domain graph) All runtimes
agent-grep binary SST bin/agent-grep wow doctor PATH check; house rules

Health commands

wow-my-zsh/doctor.sh          # symlinks, registry merge presence, env vars, CLI PATH
SIN-Save-Token/bin/verify-tokens   # L0–L4 compliance; fail-loud on regression

Both should be green after install. If they disagree, gates are wrong — open a P0.

L2 MCP budget (shared)

Source of truth: wow-my-zsh shared/mcp/servers.json fields:

  • tier: core | optional | experimental
  • always_on: bool — only true servers ship under default install profile
Profile Command Effect
minimal (default) ./install.sh or --profile minimal Zero managed MCP servers; lowest schema tax
core ./install.sh --profile core Only always_on=true servers; currently zero
task bin/sin-mcp-profile <task> <agent> Capability-based set with a hard cap of one or two servers
full SIN_ADMIN_CONFIRM=1 ./install.sh --profile full All registry servers for each agent; explicit administrative opt-in

Enforcement:

  • wow doctor.sh — any managed server under minimal, or disallowed server under core, is drift
  • SST verify-tokens — managed MCPs outside the selected budget fail unless an explicit diagnostic override is set

Core allowlist (current): empty. Tool access is selected per task, not globally preloaded.

Change rules

  1. Adding an MCP server → wow registry only; re-run install.sh; never hand-edit only one agent.
  2. Changing token policy / hooks → SST; re-run install.sh --heal / verify.
  3. Changing house-rule prose that names a CLI → update wow AGENTS.md and ensure SST or brew still ships that CLI.
  4. Secrets never in either repo — ${VAR} only.

Memory stack (current + proposed)

code structure         graphify (CLI, 0 LLM)
code symbols/LSP       Simone (primary for symbol navigation)
session observations   claude-mem (short-lived, pull-based)
domain graph memory    Cognee fleet (only canonical durable memory; CLI HTTP :8011)
                         embed: NVIDIA NIM / local fallback (:8012)
                         LLM cognify: OmniRoute
curated staging         gbrain → one-way curated export to Cognee
archive / plans         global-brain, on-demand; no default prompt injection
resume / lessons        session-digest, dream (SST CLIs)

Do not add a fourth overlapping “brain” without a measured ROI gate.