SYS_ONLINE[ALPHA ACCESS GRANTED]
AI Geeks & Freakscommunity · academy · factory
ai-geeks&freakssoftware-factory
04.1/ software-factory

A real factory,
not a metaphor.

An orchestrator runs a swarm of specialist agents through a strict phase pipeline — hard gates, full traces, automated GitHub, and a visualizer that shows every phase, cost, and decision.

SF.1Three Pillarswhat it does
The Orchestrator (Kaia)

A dedicated agent that launches factory runs, watches every trace, and steers via chat — it never writes code itself. You talk to it, it dispatches the work.

  • ▸Launches runs from chat or CLI
  • ▸Watches live event streams
  • ▸Steers mid-run with guidance
  • ▸Reports status, never intervenes
Specialist Agents

Each agent owns one role in the chain — scout, planner, builder, reviewer, documenter. Every agent call is traced, metered, and gated.

  • ▸Scout: read-only recon
  • ▸Planner: turns the ask into phases
  • ▸Builder: makes the code change
  • ▸Reviewer + Documenter: confirm and write up
The Visualizer

A read-only observability UI that shows every session, phase, event, cost, gate, and decision. Sessions, waterfall traces, stats, memory — all in one place.

  • ▸Session cards with per-agent timelines
  • ▸Waterfall trace with phase blocks
  • ▸Phase inspector: prompts, thinking, costs
  • ▸Stats dashboard: tokens, vendors, cache hits
SF.2How a Run Worksask → launch → gate → ship

1. You ask

Type a task in the orchestrator chat, or run `sssf run --roster <team> "your task"`. The orchestrator receives it and prepares the run.

2. Orchestrator launches

Kaia converges the environment (LM Studio, visualizer, tunnel, watchdog), spawns the ADW chain, and begins watching the event stream.

3. Agents run in sequence

Each agent takes its turn — scout reconnoiters, planner phases the work, builder codes, reviewer confirms, documenter writes. Every action is traced to SQLite.

4. Gates must clear

Phases have hard gates: `tests_pass`, review approval, quality checks. A phase defaults to fail until its gates pass. No gate = no ship.

5. GitHub is automated

Git branch, commit, test suite, PR creation, and diff capture are deterministic code phases — the pipeline runs them, not an agent improvising git.

6. Everything is recorded

Events stream to the visualizer live. Costs, tokens, prompts, thinking, decisions, and gate results are all stored. Every run is replayable.

agent chainspecialists
  • [01]scoutidle
  • [02]plannerrunning
  • [03]builderready
  • [04]reviewerready
  • [05]documenteridle
live trace · event stream
✓ converge lm-studio · visualizer · tunnel · watchdog
✓ [01] scout · recon · ok · 12 files
✓ [02] planner · plan · ok · 6 phases
● [03] builder · code · running
○ [04] reviewer · review · queued
○ [05] documenter · docs · waiting
gates: tests … · review pending · deploy locked
The one rule

You launch, converge & observe — the factory's agents own the work. If a run fails, the failing phase is visible in the trace; the fix is a deliberate config, prompt, or chain change.

SF.3The Visualizersee everything

A read-only observability UI backed by SQLite. Sessions, phases, events, costs, gates, prompts, thinking, memory, decisions — every run becomes a case study you can step through.

#/

Sessions Grid

Every run as a card — status chip, per-agent timeline dots, cost, runtime, tokens. Polls every 500ms. Filterable and searchable.

#/<run>

Waterfall Trace

Per-lane timeline: engineer, code workspace, and each agent. Phase blocks with colored gradients, tool-call tick marks, and queued phases shown dashed.

#/<run>/<phase>

Phase Inspector

Click any phase block. Compiled prompts (system.md + user.md), model thinking/reasoning, gate results with per-item evidence, token costs, output envelopes.

#/stats

Stats Dashboard

Total runs, success rate, token optimization, cache-hit ratio. Local vs online split. Per-model cost breakdown. Commercial vendor comparison.

#/memory · #/decisions

Memory & Decisions

Kaia's engram memory: semantic search, timeline, recent episodes. Decisions view groups failed runs by diagnosis type with fixes.

#/chat

Orchestrator Chat

Talk to Kaia. Launch runs, steer live sessions, switch models, view session side-panel with live progress. Tool calls and session launches inline.

500ms polling

sessions and live events refresh automatically

cursor-based events

pagination never re-fetches seen events

read-only server

Bun reads SQLite — no writes except archive

two themes

neutral dark and classic deep-space aurora

SF.4Infinite Ways to Build a Teamone config file
roster.yaml · the ONE file
stacks:
  local-fast:            # your roster name
    tier: local          # pulls every model from tiers:local
    agents: [planner, builder, reviewer]

  frontier:              # pin models per team
    agents:
      planner: { model: claude-opus-5, thinking: high }
      builder: { model: gemini-3.6-flash }

  you: any mix, any names → sssf run --roster you "ask"
$sssf run --roster local-fast "polish the homepage"

Named rosters

Any team under `stacks:` — run it with `--roster <name>`.

Model tiers

One preset pulls every agent's model. Or pin a model per agent.

Compose chains

Chain agents in any order — scout, planner, builder, reviewer, documenter.

Generate

`make_adw.py` scaffolds a new chain from the agents you name.

The chain says what runs; the config says who runs it. Build a team per project, per task, or per model budget — never recompile a thing.

SF.5Automated GitHub Actionsdeterministic code phases
kind="code" · never hidden in an agent
  • git branch feature/xopen the branch
  • git commit -m "feat: …"write the commit
  • quality.run_tests(run)run the suite (CI)
  • gh pr createopen the pull request
  • changes.capture(run, base=main)diff against main

Branch · commit · PR

Each git step is its own traced phase — the pipeline opens the branch and raises the PR, not an agent improvising git.

Hard gates

Success must be earned. Every phase defaults to fail until it clears review & the suite is green. `quality.py` is the CI — `tests_pass` is a gate.

Watch it live

The visualizer shows every phase block, gate result, and tool call tick. Click any block to see the full inspector.

SF.6Run Lifecycle & Controlpause · steer · stop
run controls · live
stop
pause
resume
steer mid-run
focus on error handling, skip the docs phase
diagnose & recover:
$ sssf audit <run-id>
$ sssf diagnose <run-id>
$ sssf recover <run-id>
$ sssf emergency-stop <run-id>

Read without blocking

The visualizer polls sessions every 500ms. Cursor-based event pagination means you never re-fetch seen events. Watch live without stalling anything.

Replay any run

Every session is saved — events, prompts, thinking, gates, costs. Click into any run and step through every phase in the waterfall trace.

Diagnose & recover

The Decisions view groups failed runs by diagnosis type and model. `sssf audit`, `sssf diagnose`, and `sssf recover` turn a failure into a deliberate fix.

Emergency stop

`sssf emergency-stop` halts a runaway run instantly — SIGTERM to children, finalize the DB. `sssf learn` records the lesson into knowledge.

SF.7Connected to Way of Teamsthe coordination backbone

The factory plugs straight into the team OS

Every plan item becomes a Way of Teams ticket before work starts. Decisions get recorded in Way of Teams memory & knowledge. The WOT MCP server exposes tickets, memory, knowledge, standups, teams, and GitHub review tools — so the factory's runs are coordinated like any other team's.

ticketsmemoryknowledgestandupsteamsgithub review
factory → productized

Factory runs are becoming a Way of Teams feature — per-tenant pipelines, visible runs, review gates, and results surfaced right in the platform. Paying teams run their own build & release loops.

the loop
ticket→factory run→review gate→memory & knowledge→shipped

Try it today: install the harness, then run your own factory on a hosted workspace.

Boot a workspace

Run your own software factory

Install the harness, boot a hosted workspace, and dispatch your first roster. The visualizer shows every phase, cost, and decision. The Discord is where operators trade runbooks.

05The Tools // What We Build Onstack: live

Custom-built extensions. Serious tools. No dead links.

We build on the real, open-source AI toolchain. These are the engines we extend, configure, and run — with links that actually go somewhere.

01
THE ENGINE

ai-harness

Our AI engineering harness — 51+ battle-tested skills, 14 agents, and workflows wired into 7 AI coding tools. Install once, use everywhere. MIT / Apache-2.0.

  • ▸one installer wires skills + agents + commands into every AI tool
  • ▸runs on OpenCode, Claude Code, Pi, Codex, Gemini, Wo Coder, Antigravity
02
THE SHELL

OpenCode

The open source AI coding agent we drive from. Terminal, desktop, or IDE — 75+ LLM providers through models.dev, plan mode, LSP, and your existing Copilot/ChatGPT subs.

  • ▸bring your own model or reuse existing subscriptions
  • ▸Plan mode, /undo, /share, multi-session — free and open source
03
THE MINIMALIST

Pi

The minimal, extensible coding agent by Mario Zechner. Adapt Pi to your workflow — not the other way around — via extensions, skills, prompt templates, and packages.

  • ▸15+ LLM providers · tree-saved sessions · MIT open source
  • ▸install: curl -fsSL https://pi.dev/install.sh | sh