SISO Agent Base — Repository Decomposition

Measured reality

23 GiBcurrent checkout

9.9 GiBGit database

≈12.47 GiBeligible for the current whole-root installer copy

≈5.4 MiBcandidate coherent runtime source before dependencies and owned tests

The public installer clones the repository and scripts/install-local.sh uses a root-level rsync excluding only .git, node_modules, and .DS_Store. It therefore copies unrelated repository material into an application installation. The working tree also had 586 existing dirty status entries during this review, so no source movement was attempted.

The actual application boundary

Current path	Observed responsibility	Target ownership
`install.sh`, `scripts/install-local.sh`	Clone, dependency setup, profile installation, Bifrost credential configuration, and launcher projection.	Runtime app
`bin/siso*`	Stable user-facing CLI: launch, doctor, update, inspect, and TUI entrypoints.	Runtime app
`extensions/siso-lifecycle`	Agent-session lifecycle capture.	Runtime app package
`extensions/siso-context-manager`	Context filtering, retrieval, memory, and librarian behavior.	Runtime app package
`extensions/siso-status`	Status and terminal HUD surfaces.	Runtime app package
`extensions/siso-agent-router`	Routing, child execution, tasks, goals, workflow, mailbox, tools, skills, worktrees, and scorecards across 80 files.	Runtime app; internally decompose by domain
`templates/profile`	Installed Pi profile, model configuration, system instructions, and currently bundled skills.	Runtime profile; skills become versioned dependency
`apps/siso-tui` + `packages/siso-tui`	The active SISO-owned OpenTUI application and renderer/runtime contracts.	Runtime app monorepo

Working name: SISO Agent Runtime or SISO Agent App. “SISO Agent Base” may remain a compatibility/product name, but it must not imply that the current warehouse repository is one architectural component.

Repository extraction ledger

Current area	Content evidence	Disposition	Probable repository
`siso-project-os/`	Clean nested Git repository with its own package, CLI, schemas, tests, MIT license, and public remote.	Already independent	`sisodias/siso-project-os`
`siso-agent-playbook/`	Nested Git repository with installer, skills, tools, prompts, telemetry, evidence, and public remote. Four local modifications exist.	Already independent; preserve dirt	`sisodias/siso-agent-playbook`
`foundry/` + owned pipeline code	Independent source-mining and reuse-knowledge product with a database pipeline and read-only web interface.	Extract product; separate code from datasets	`siso-foundry`
`apps/siso-os/`	2.9 GiB nested upstream cmux repository on a SISO WIP branch, with four dirty entries; its own commit says superseded by Herdr.	Preserve fork/WIP outside runtime	Dedicated fork or external-source ledger
`templates/profile/skills/`	Large cross-agent capability catalog installed by symlink projection.	Extract capability registry; promote independently useful skills	`siso-skills` plus individual skill repos
`brain/`	Composes shared human/agent instructions into Claude- and Codex-specific generated profiles.	Coherent package; decide app package vs own repo	`siso-agent-profile` candidate
`atlas/`	Read-only persistent architecture auditor with its own scripts, rubric, tickets, and operating loop.	Independent tool candidate	`siso-atlas` candidate
`siso-agentlabs/`	Former Pi Harness Lab with its own workspaces, experiments, research, and verification commands; current duplication against runtime has not yet been reconciled.	Compare before promote/archive	`siso-agent-labs` or vault
`apps/siso-web-chat/`	Independent Vite application described as a read-only web chat command shell.	Validate active product state	Own app repo if active; otherwise lab
`packages/design-tokens/`	Cross-product SISO design token package with non-agent consumers.	Move out of agent domain	SISO shared UI/design repository
`integrations/headroom`, `integrations/context-firewall`	Bounded, opt-in context optimization experiments with their own safety and promotion contracts.	Adapter/lab repos after promotion decision	`siso-agent-integrations` or individual adapters
`extensions/knowledge-engine`, `codex-engine`, `idea-engine`, `session-mirror`, `braind`, voice/Agent Zero extensions	Direct module review found coherent evidence engines, shared agent-state and session-intelligence services, and staged Agent Zero interface packages.	Mapped candidate Works	Open the module extraction map
`office/`, runtime databases, telemetry, generated eval output	Machine/private operating state, including database snapshots.	Never public runtime source	Private state/data store
`archive/`, `research/`, imports, upstream clones	Archives, evidence corpora, external source, or staged material with different lifecycles.	Never application install payload	Vault, Foundry data store, or external-source manifests
`scripts/`	238 root commands, including 208 `smoke:` package scripts spanning many unrelated domains.	Distribute tests to owning repos	Each extracted Work owns its checks

Target repository topology

great-library-of-siso          registry, relationships, releases, evidence
├── siso-agent-runtime        the installable SISO coding-agent app
├── siso-project-os           project-local operating system (already separate)
├── siso-agent-zero           coordination, boot, and staged interface packages
├── siso-agent-playbook       fleet operating package (already separate)
├── siso-skills               capability index/distribution; mature skills may graduate
├── siso-foundry              source-mining and reuse-knowledge product
├── siso-agent-brain          shared agent-state service
├── siso-session-intelligence evidence-backed learning from agent sessions
├── siso-evidence-engines     knowledge, coding-principles, and idea packages
├── siso-agent-integrations   reusable experiments with explicit promotion gates
└── siso-atlas                later, if active review confirms it

Private incubation and non-code ownership:
├── private siso-agent-lab                → unfinished agent R&D and prototypes
├── SISO_Internal_Lab                    → internal products; not agent-infra overflow
├── datasets / large databases / telemetry  → data store
├── external upstream clones                → clone manifest + upstream locator
├── private office/runtime state            → private system store
└── dead or superseded material             → reversible vault

How the sauce enters the Library

mixed source / warehouse
        ↓ read and inventory
Source Inventory unit
        ↓ coherent outcome + contract + preservation proof
stable Library Work
        ↓ independent source locator and release evidence
GitHub repository / package / dataset manifest
        ↓ contextual composition
Stack Assembly, playbook, or capability index

The Great Library does not need to physically contain every byte. It provides the stable identity, inventory, relationships, provenance, release manifests, and human documentation. GitHub repositories hold independently evolvable source; large datasets live in appropriate data stores; private state remains private.

The machine-readable starting point for this audit is registry/source-inventories/siso-agent-base-2026-07-30.json. Candidate units remain provisional until their direct comparison is complete.

When something earns its own repository

It has one coherent outcome and a clear public contract.
It can be installed, run, forked, released, or maintained independently.
Its dependency direction can be stated without reaching into another repository's internals.
It owns its tests, release evidence, documentation, and migration path.
Moving it does not strand sole-copy dirty work; a bridge or pinned dependency exists first.

A folder does not earn a repo because it is large. A skill does not earn a repo merely because it is a skill. Independent adoption and evolution earn the boundary. Stable Great Library Work IDs remain independent of whatever GitHub repository currently locates the source.

Safe migration order

Freeze an evidence inventory of current dirty and nested Git state. Do not move anything yet.
Define the minimal runtime file manifest and change installation from whole-root rsync to an allowlisted release payload.
Remove already-independent nested repositories from the runtime payload and replace them with Library links or pinned dependencies.
Extract one proven Work at a time using describe → bridge → move, preserving history and local changes.
Move each domain's checks with it; shrink the root's 238-script cross-domain test surface.
Only then rename or retire the warehouse repository.

The first code change should be the installer allowlist, not a giant filesystem reorganization. That immediately stops distributing gigabytes of unrelated material while preserving every current path for later controlled extraction.

Evidence and limitations

Read directly: root README, manifest, package configuration, installer, launcher, runtime extension entrypoint, architecture documents, product/package READMEs, nested Git remotes and status, top-level file/size counts, and existing working-tree state. Semantic code search was attempted through the registered project service but did not return in a bounded time, so targeted source reads were used as the documented fallback.

No files were changed in SISO Agent Base. Candidate labels are intentionally unverified where a direct content comparison is still required. The working tree's 586 pre-existing dirty entries and dirty nested repositories are preservation constraints, not cleanup targets.

Open the Agent Base module extraction map · Return to the SISO Agent Stack model · Open Agents