[FScholPer]

Summary

First step of the Agent Context Attention Layer. Turns this repo into a uv monorepo and lands Phase 0 only: the schema / data model. No adapters, graph composition, ranking, apm.yml generation, or MCP server logic yet — those are later phases; placeholder dirs mark where they go.

The layer's job is to turn "trace everything" across S-CORE's 80+ repos into "surface what matters for my current context". Everything becomes a typed, provenance-bearing graph node/edge; later phases rank them per (task, repo, role) and project the top-N into a generated APM manifest that compiles to AGENTS.md.

Layout

mcp-servers/
  pyproject.toml            # uv workspace root
  apm.yml / apm-policy.yml  # Microsoft APM manifest + starter (tighten-only) policy
  packages/score-context/   # engine LIBRARY — MCP-free, importable, CI-runnable
    src/score_context/schema/{nodes,edges,provenance}.py, context.jsonld
    src/score_context/delta.py
  servers/context/          # placeholder: Phase 6 get_context MCP server
  apm-packages/             # placeholder: authored vs generated APM packages

Schema (the substance)

Design principle: sphinx-needs is one signal source, not the foundation. Node vocabulary is split into two families:

• Needs-model nodes — IDs reused verbatim from score_metamodel.yaml (stkh_req, feat_req, comp_req, feat, comp, mod, logical interfaces, dec_rec, safety/security analysis types, testcase, …). Not forked.
• Trace/event nodes — the part deliberately outside the needs model, where most context change actually happens: issue, discussion, pull_request, commit, code_change, contract, repo, module, team, person, artifact, ci_result.

Edge relations are a superset: all needs_extra_links from the metamodel (implements, satisfied_by, derived_from, covers, consists_of, mitigated_by, fully_verifies, …) plus the ADR trace relations (affects, depends_on, discussed_in, blocks, supersedes, conflicts_with, authored_by, owns). ADR_RELATION_MAPPINGS documents how ADR terms map onto existing metamodel links where they overlap.

Invariants baked in now so adapters can't violate them later:

• Natural keys only — need id / dec_rec id / org/repo#123 / commit sha / file path / bazel module label. No synthetic UUIDs.
• Provenance mandatory on every node AND edge — {repo, sha?, adapter, confidence∈[0,1], observed_at}. confidence=1.0 for deterministic extraction, lower for LLM-assisted; required for the multi-company / safety audit story.
• All models are pydantic v2 with extra="forbid".

context-delta — the one unit every future adapter emits

class ContextDelta:            # delta.py
    signal: Signal             # {type, source: SourceRef, provenance, weight, decays}
    graph_ops: list[GraphOp]   # discriminated union: UpsertNode | UpsertEdge

context.jsonld maps every node/edge type to a stable IRI under https://eclipse-score.github.io/context-attention/ns# for git-friendly, vendor-neutral serialization.

Verification

uv sync, uv run ruff check ., ruff format --check, uv run pyright (0 errors), uv run pytest (4 passed) all green.