nrupala/TransformationSpine

Provider-neutral transformation platform. Models are workers; the spine owns
context, memory, decisions, governance, and provider selection — so outcome
convergence holds no matter which engine runs beneath (llama.cpp, Codex,
openenry, Claude Code, any OpenAI-compatible endpoint).
Core Idea
**Context is contextual to its usage.** Like variables in a compiled language,
the spine assigns every piece of context a lifetime:
| Scope | Analogy | Durability | Persisted |
|-------|---------|-----------|-----------|
| TRANSIENT | temporaries | one call | no |
| SESSION | stack frame | one task/session | while open |
| PROJECT | module state | project lifecycle | yes |
| PERSISTENT | globals in code | forever | yes |
The spine assembles the *minimum sufficient* in-scope context for each prompt,
promotes proven facts up a scope, demotes superseded decisions down, and
retires the ephemeral — without the model ever owning memory.
Outcome Convergence
From the OCS projects: every transformation is
a gated cycle. The spine computes an error signal per provider result; a
result converges when the gap to the committed artifact is zero. This is what
makes a 8B llama.cpp model and a 70B Codex model converge to the same outcome —
swapping an engine is a configuration change, not a data loss event.
Package
Architecture
See docs/spine-architecture.mmd for the full architecture diagram.
Build Plan
See BUILD_PLAN.md for the phased v0.2.0 upgrade plan covering:
- Multi-backend adapters (HuggingFace, skill-level)
- Expanded MCP/ACP connectors (ServiceNow, Databricks, Confluence)
- Audit logs & telemetry
- Local/cloud/hybrid profiles
- Graphical workflow authoring
- Prebuilt agent libraries
- Enterprise integrations roadmap
- Tool-call abstraction
- OS-level safeguarding
All phases maintain ASF gate compliance (G0-G8) with verified evidence.
Install
Test / verify
Status
v0.1.0 — Core spine with provider abstraction, context lifecycle, and a
CTST ledger.
v0.2.0 — Released 2026-10-09 after an independent audit and a full
remediation program (see CHANGELOG.md and ASFQC/GATES.md): 178 tests,
coverage held to a minimum of 80% in CI, CI green on Python 3.11 and
3.12. Serving surfaces: REST API, MCP (HTTP and stdio), an A2A agent
card, ACP runs, a browser UI, and the spine CLI. Connector
auto-discovery via entry points and plugin directories; a CTST ledger
with a verified hash chain; durable context persistence.
v0.3.0 — Released 2026-10-10. Closes the current development plan:
estimator calibration (the token engine now learns per-route
correction factors from estimate-vs-actual telemetry), a