The sauce inside Agent Base.

The answer in one view

Recommended estate: the Great Library plus ten intentional public agent repositories and one private agent-lab incubator—not twenty-two repositories copied from twenty-two folders. The Library can still expose all twenty-two modules independently and later move any Work to its own repo without changing its identity.

How the current Library projection fits together

Great Library
├── Agents
│   ├── Operating stack
│   │   ├── SISO Project OS              project-local agent operating contract
│   │   ├── SISO Agent Runtime           installed coding-agent application
│   │   │   ├── Agent Control Core       routing, execution, goals, tasks, safety
│   │   │   ├── Session Lifecycle        corrections, lessons, checkpoints, restore
│   │   │   ├── Context Manager          budgets, memory, retained evidence, compaction
│   │   │   ├── Status & Evidence        operator visibility and commands
│   │   │   ├── Agent TUI Kit            terminal components, adapters, runtime UI
│   │   │   └── Profile Composer         shared instructions projected per tool
│   │   ├── Herdr                        workspace in which the stack is operated
│   │   └── Agent Zero                   user-facing coordinator
│   │       ├── Boot & Rehydration       wakes from shared state
│   │       └── Voice Gateway            staged voice-to-agent interface
│   ├── Capabilities
│   │   ├── Skills Catalog               atomic repeatable capabilities
│   │   └── Agent Playbook               compositions of skills for scenarios
│   ├── Shared intelligence
│   │   ├── Agent Brain Service          shared state, memory, tasks, fleet, timeline
│   │   └── Session Intelligence         learns from explicit agent sessions
│   └── Assurance & integrations
│       ├── Atlas                        read-only architecture auditor
│       ├── Context Firewall             shadow measurement and promotion rail
│       └── Headroom Adapter             opt-in lossless context sidecar
│
├── Research
│   ├── Foundry                          source discovery, evaluation and preservation
│   │   └── Research Topic Registry      stable research buckets and artifact links
│   └── Evidence Engines
│       ├── Knowledge Engine             sources → grounded knowledge and taxonomy
│       ├── Principles Engine            evidence → canonical engineering guidance
│       └── Idea Engine                  evidence → ranked improvement backlog
│
└── Source records
    └── Agent Base Source Warehouse      retired extraction source, not architecture

This is an ownership and dependency map, not a permanent folder taxonomy. Relationships live in registry data, so a module can move between repositories or graduate from package to repo without breaking its Library address.

Modules that should become Library Works

The repository plan

Later, only if proven: siso-atlas, a standalone voice gateway, independently forkable skill repositories, and independently released evidence engines. Stable Library IDs mean those later splits do not break links.

Where the current Agent Base material goes

Promotion flow: private siso-agent-lab → proven public package or Work → its owning public repo → indexed by the Great Library. The private lab is temporary incubation, never the canonical home of a released capability.

What must not become public module repos

Public extraction gates

Security gate: direct review found an embedded provider credential in the warehouse Session Intelligence configuration. The public repository was clean-room implemented and contains neither that configuration nor its Git history. The original provider credential still requires external revocation and rotation.

1. Preserve: record dirty files, nested repository heads, uncommitted patches, and sole-copy material.
2. Describe: give the module a stable Library Work ID, outcome, owner, public contract, dependency list, data boundary, and lifecycle status.
3. Sanitize: remove credentials, personal paths, machine names, private hosts, private datasets, session content, and generated runtime state; rotate anything exposed.
4. Bridge: make the current Agent Base import or install the new package before moving the source of truth.
5. Prove: move the module's fixtures, tests, smoke checks, license, and release evidence with it.
6. Move: publish one coherent module family at a time; update the Library source locator without changing the Work identity.

Extraction status after V14

1. Metadata first: register the module identities and relationships from this map; no risky source moves are required.
2. Stop the leak: rotate the exposed credential and replace the Agent Base whole-root installer copy with an allowlisted runtime release payload.
3. Connect what already exists: register Project OS and Agent Playbook by source locator, preserving Playbook's local modifications.
4. Separate incubation: preserve Agent Labs, Web Chat, and unfinished spikes in a private siso-agent-lab; do not move them into SISO_Internal_Lab.
5. Completed public boundaries: Agent Zero, Skills, Foundry, Agent Brain, Session Intelligence, Evidence Engines, Agent Integrations, and Agent Playbook now have verified source releases.
6. Runtime foundation complete: lifecycle, context, status, control, TUI event, and profile contracts now ship inside Agent Runtime; split repositories only when independent adoption appears.
7. Skill boundary pass complete: all 28 current entries now have provisional evidence-backed destinations; none is promoted to a repository by folder reflex.

Evidence and limits

The verdicts above come from direct reads of module entrypoints, type contracts, READMEs, architecture documents, migrations, pipeline stages, safety gates, package manifests, and nested Git state. Semantic search through the registered project service was attempted but did not complete in a bounded time, so targeted source reads were used as the required fallback.

The source inventory remains partial: Runtime parity beyond the clean foundation, Atlas current-use verification, Agent Labs and Web Chat supersession comparisons, and individual Skill promotion evidence still require direct follow-up. No files were moved or edited inside the dirty Agent Base warehouse.

Repository decomposition · Skill repository map · SISO Agent Stack model · Open Agents