Agency Runtime v12 — graph-aware retrieval

Use this skill when a coding task needs repository context before an agent edits code.

Contract

workspace_root enables deterministic indexing. Retrieval starts with a small lexical/path/symbol seed set, then expands up to context_graph_hops (0–2) through typed graph edges:

• imports: direct module/file dependency
• calls: symbol call relationship; reverse traversal supplies callers and forward traversal supplies callees
• implements: inheritance/interface relationship
• tests: conservative test-to-source association; this is context, not proof of line coverage
• configures: configuration-to-source association

Every edge has a confidence and reason. The context carries graph_trace, including seed paths, expanded paths, selected edges and explicit stop reasons.

Token discipline

The graph is used for navigation, not dumped into the model. Only bounded source windows become ContextChunks. context_token_budget remains the hard context ceiling and context_max_files limits the final file set. Files excluded by budget are reported in unknowns.

Orchestration flow

CodingTask -> adaptive recommendation -> graph-aware context -> TaskProfile.context -> specialist selection -> execution -> evidence/verification -> artifact regression -> release gate -> benchmark observation.

The host remains responsible for model calls, tools, permissions, concurrency and side effects. The portable runtime only supplies deterministic evidence and orchestration state.

Unknown policy

Do not silently infer missing code. Surface no seed match, unreadable files, budget omissions and graph stop conditions as explicit unknowns and record them in provenance. A graph edge is evidence of a relationship, not proof of runtime execution.

Evaluation

v11 deterministic retrieval/answer metrics remain compatible. Graph expansion improves the evidence set but does not bypass faithfulness, regression, release or provenance gates.