FESTIVE TRIALS · TILE 01

Generic Setup

Agent Bootstrap System — Deconstructed

Copy-paste a folder. Point any coding agent at it. Progressive disclosure loads only what's needed — and nothing else.

Hot Path 28 lines
Context Cut 96%
Files 124
Council Score 8.4 / 10
Agents 7

Three failures this kills.

Every agent project hits the same three walls. Generic Setup exists to break them — not with theory, with three mechanisms wired into the kit.

01

Agents invent APIs

The model hallucinates a function signature, ships code that calls it, the build fails, the human debugs for 20 minutes. Repeat.

fixed by library cards — pre-validated API references the agent loads on demand.
02

Orphaned files

Agent writes a new module, forgets to wire it into the barrel/index. File exists, nothing imports it. Dead weight accumulates.

fixed by wire-first protocol — index updates before the file is written.
03

Context bloat

Every invocation loads 500+ lines of context the agent doesn't need. Tokens burn, attention drifts, replies bloat.

fixed by progressive disclosure — 4 tiers, only HOT loads by default.

Four tiers, one rule.

Tier 0 loads on every invocation. Everything above it loads only when the task actually needs it. Cold/Frozen tiers require explicit invocation — never auto.

TIER 0
HOT
INDEX.md (28 lines) + TODO.md — the only files the agent sees on every run.
every invocation · 28 lines
TIER 1
WARM
Agent definition file + extensions — loaded per-task, scoped to the work being done.
per-task · ~670 tokens
TIER 2
COLD
Library cards — 21 pre-validated API references (Python, TS, Rust, Go). Loaded on demand.
on demand · 21 cards
TIER 3
FROZEN
Swarm runtime, research agents, Archon guardrails — explicit invocation only.
explicit only · never auto

The token math.

Before Generic Setup, the same agent invocation dragged 558 lines of hot context. After: 28. The reduction isn't cosmetic — it's a 95% cut on the path that runs on every single call.

Layer Before After Reduction
HOT (every invocation) 558 lines 28 lines 95%
CLAUDE.md 260 lines 68 lines 74%
Per-task context ~9,000 tokens ~2,500 tokens 72%
Reply size unbounded ≤ 20 lines 80%

Three modes. One default.

The agent picks a mode based on the task. Swarm is never auto-selected — it costs 5.6× more than the default and only pays off on parallelizable work.

Mode 01
SIMPLE
Ollama 8B local model. Free, offline, fast for trivial tasks.
~445tokens / call
DEFAULT
Mode 02
SINGLE SPECIALIST
One capable model, scoped to one task. The 90% case.
~670tokens / call
5.6× cost
Mode 03
SWARM
Parallel agents writing to isolated staging. Only on explicit ask.
~2,060tokens / call
!
Swarm is NEVER auto-triggered. Only fires when the user explicitly says "use swarm". Default mode handles 90%+ of real tasks.

Instructions, extensions, hooks.

Three layers stack — each more deterministic than the last. Instructions drift. Extensions drift less. Hooks can't drift.

L1
Instructions
CLAUDE.md, AGENTS.md — prose the model is asked to follow.
CAN DRIFT
L2
Extensions
Loaded per-task, scoped to the work. Less surface area = less drift.
LESS LIKELY
L3
Hooks
Shell scripts. Deterministic. The model cannot talk its way out of these.
CAN'T DRIFT

Index before file. Always.

Five steps, in order. The index update happens first — so even if the agent crashes mid-task, the barrel knows the file should exist. No orphans.

STEP 01
Update barrel/index
wire before write
STEP 02
Create file
empty shell
STEP 03
Fill
implement
STEP 04
Test
run + pass
STEP 05
Lint
clean exit

10-Point Checklist

01Barrel updated before file creation
02File path matches barrel entry
03Exports declared in barrel
04No orphaned imports
05Tests written & passing
06Lint clean (zero warnings)
07No dead exports
08Type signatures present
09No invented API calls
10Wire confirmed by hook

719 lines. Real, not docs.

A working shell runtime that spawns parallel agents, tracks them, integrates their work, and recovers from interruption. Not a sketch — production code.

swarm-run.sh — 719 lines · bash
$swarm-run.sh run tasks.yaml
spawns parallel agents from a YAML task spec
$swarm-run.sh status <id>
check progress of a running swarm
$swarm-run.sh integrate <id> --commit
merge staging branch into live tree + commit
$swarm-run.sh resume <id>
recover an interrupted swarm without losing work
Runtime
719 lines
Staging
Isolated
Hooks
Enforced
Recovery
Resume

High council review.

Five-axis scorecard from the council review. After fixes: 8.4 average. Before fixes: 3.1. The delta is the whole story.

Runtime Engine
9.0/10
Context Isolation
8.5/10
Structural Strategy
9.0/10
Sandbox Security
7.0/10
after fix
Vector Search
8.0/10
after fix
Overall · council aggregate
3.1 / 10 before  →  8.4 / 10 after fixes
8.4

Honest, both sides.

Not a sales pitch. The kit works — and it has real limits. Both columns are the truth.

Pros 7

  • 96% context reduction on hot path
  • Real swarm runtime (not just documentation)
  • Wire-first prevents orphaned files
  • Library cards prevent invented APIs
  • Swarm is explicit-only (token efficient)
  • Works with Claude Code, Grok CLI, Codex
  • Reply format enforced (no essays)

Cons 6

  • 120 files is still a lot (council recommends ~25 core)
  • Sandbox escape was real (fixed with CLI flags, not containers)
  • Extensions depend on LLM compliance (not fully deterministic)
  • Research swarm is a separate system bolted on
  • No cross-session state persistence (without Tasktron)
  • Library cards only cover Python, TS, Rust, Go (no Java/C# yet)

By the numbers.

The full inventory — what's actually in the kit, counted.

124
Files total
28
Lines hot path
21
Library cards
7
Agent definitions
7
Extensions
719
Lines swarm runtime
9
Research agents
3
Layer enforcement
4
Tier disclosure