Baelfyre/Kaos

What is Orchestra?

Orchestra coordinates AI-assisted development across specialists for architecture, implementation, UI/UX, data, security, testing, documentation, and governance. It routes the task, preserves evidence, validates results, and respects human authority.

Orchestra is not an AI model. It works around compatible coding hosts and does not replace engineering judgment.

Quick start

Install through the host installation guide. Then describe the result you want and what Orchestra may do:

@Orchestra

Review this repository, identify the failing tests,
fix the issue, and run relevant checks.
You may edit files and prepare a PR. Do not merge or deploy.

Conductor chooses the smallest useful specialist route. Compatibility and getting started explain supported hosts.

Capabilities

Area	What it provides	Important limit
Specialist coordination	Focused roles, bounded execution, and handoffs	Tool access never creates permission
Assurance and governance	Evidence-based validation and human-controlled boundaries	Passing tests never authorizes a merge
Security and adversarial review	Cipher provides scoped, evidence-based source review; Dagger supports authorized adversarial validation planning	Cipher runtime and live penetration testing remain unqualified; Dagger execution requires exact target-owner authorization, Conductor routing, a non-production target, and a passing guardrail
Optional visual refinement	Cloak offers optional 8-point spacing and visual refinement	Advisory under project tokens, accessibility, and responsive behavior; rendered effectiveness needs rendered evidence
Artifact authoring and review	Scribe artifact guidance records requirements and backend evidence; Cloak artifact guidance covers visual hierarchy and accessibility	Guides do not establish backend capability, output correctness, or cloud-operation authorization
Adaptive memory	Local observations, opt-in read-only context, and shadow learning	No automatic authority, preference promotion, or private cloud sync
Adaptive discovery	Scoped capability-memory contracts and offline planning	Live host discovery and trusted private-memory writes are not activated
JEV advisory	Optional TypeSafe System One advisory components	Live TypeSafe/BYOK onboarding, R3-A host trust, and production wiring remain unqualified
Scribe documentation	Requirements, as-built reconstruction, drift reconciliation, and an optional six-dimension blueprint	Scribe does not approve technical decisions or impose six mandatory files

These are source and maturity claims, not a guarantee that each capability is enabled in every host or in the latest published release.

JEV and local API keys

JEV supports non-binding engineering advice when independently reviewed alternatives exist. The existing offline pilot provides a repeatable 14-scenario pytest replay with explicitly labeled synthetic responses and evidence-gap recording. It makes no TypeSafe calls or actual JEV inference. The original V6 packet and source artifacts are absent, so replayed V6 findings remain UNVERIFIED. The pilot remains PREPARED_NOT_QUALIFIED; further qualification and integration are deferred from v1.12. Live TypeSafe integration, R3-A host trust, and production wiring remain unqualified. The TypeSafe adapter and offline data-minimization components exist, but there is no verified general-purpose end-user BYOK activation flow. Ordinary Orchestra use needs no JEV key. Never paste credentials into prompts or tracked project files.

Read the offline pilot instructions, JEV local API setup and limitations, and the decision hierarchy.

Adaptive memory

A1 maintains scoped machine-local observations. A2 can provide opt-in, read-only context to an already-selected specialist. A3 compares inferred tendencies in shadow mode; they cannot silently change execution. Portable memory supports backend contracts, not built-in synchronization. D3B capability memory/discovery remains offline and does not authenticate a live host.

Read adaptive memory setup and limitations and portable memory.

Scribe and the pre-development blueprint

Scribe operates in SPEC_TO_SYSTEM, SYSTEM_TO_DOCS, and RECONCILE modes. It can organize sources and identify documentation gaps; relevant specialists retain ownership of architecture, UI, persistence, security, validation, and governance decisions.

The six coverage areas are product requirements, technical requirements, application flow, design brief, backend/data schema, and implementation plan. They are not six mandatory documents. Reuse existing artifacts, label missing evidence, and preserve unresolved decisions.

This blueprint guidance is merged to main, but post-v1.11.0 work is not necessarily in the latest published release. Read the blueprint guide, Scribe skill, and documentation applicability rules.

Permissions and privacy

TOOL_ACCESS != USER_AUTHORIZATION
TESTS_PASS != MERGE_APPROVAL
JEV_ADVICE != EXECUTION_AUTHORITY
MEMORY_CONFIDENCE != PERMISSION

No model output, learned preference, documentation matrix, or installed plugin may bypass protected decisions, transmit sensitive data without proper authority, or silently change policy.

Release governance uses a proportional evidence-first relevance, validation, and compliance review: deterministic issues may be corrected and retested within granted scope, while unresolved legal, privacy, IP, authority, and final publication decisions remain governed. A machine evidence result never authorizes a release.

More documentation

Installation · MCP · Architecture · Governance · Developer portal · Detailed documentation map · Machine index

Published release

The latest public release is v1.11.0. The development branch contains later work, which remains unreleased until the complete version-specific qualification and publication process is completed. See releases and changelog.

The v1.12.0 integration candidate was Squash-merged through PR #990; the separately approved evidence-first governance amendment followed through PR #996. These integrations are validated source states, not a published release. Release-specific rights, privacy, retention, and governance gates remain open; the owner has requested publication once those gates pass. See the candidate record, current release evidence and applicability screening, targeted source and naming rights assessment, third-party provenance, and historical local reconciliation.

Support

Support development · MIT License