Skip to content

Agents, clients and model routing

These guides cover how coding agents and their clients are wired to a governed repository: lifecycle hooks, client and editor configuration, the context and transcript tools, and how work is routed to a model.

Read Agent lifecycle coverage first: it separates the shared repository policy from the client-specific adapters the other pages configure.

  • Agent lifecycle coverage: the shared Lefthook jobs and AGENTS.md context versus each client's native lifecycle adapter.
  • Context cache bands: the head, config and tail markers in AGENTS.md that keep the compiled files a stable prompt prefix, and --verify-stable.
  • Agent hooks: the single praetorctl hook <client> <event> entrypoint and the native hook coverage of each client.
  • Issue claims: praetorctl issue claim, status and release report which session works an issue and at which stage; the dispatch gate refuses an issue another live session holds.
  • Shared client configuration: praetorctl clients projects the MCP server registry into client settings.
  • Client schemas: the pinned upstream schemas every rendered client config and hook fixture is checked against, the types generated from them and the formats that have none.
  • Configurable backend connections: praetorctl clients connect applies one private deployment profile to that registry.
  • Repository-aware editor generation: praetorctl editors generate and verify, and the repository capabilities both resolve.
  • Local context review candidates: standardsctl context-optimize proposes smaller rules, skills and memory files for review.
  • Local transcript ingestion: praetorctl harvest transcript caches Antigravity and Claude Code transcript events privately.
  • Declared-task model routing: offline model selection with praetorctl models route from .config/models/routing.yaml.
  • Tribunus data sync: the data the Tribunus graph router needs before it can route work to a model.