Agent runtime
The runtime is Vekrun’s own Rust loop; it does not depend on any other agent product. The same loop runs in the desktop app and from the CLI (vekrun agent run; exact flags unverified on this site, use --help from your build).
The loop
Section titled “The loop”Model call, tool proposals, policy decisions, tool execution, evidence back into the timeline. Every tool action goes through the policy engine and leaves an audit event. Cancellation stops the loop and the subprocesses it started. Structured outputs are validated against schemas, which is how Plan mode produces a typed PlanRevision.
Modes as policy profiles
Section titled “Modes as policy profiles”Ask, Plan, Build, Debug and Review are policy profiles, not personalities: the model is the same, the authority differs. Ask cannot write (tested). Build shows every write as a diff and applies it only after approval, with checkpoints and rollback. Debug adds approved command execution. Review reads a change set and reports. A Custom profile with its own role, tools and permissions is specified but not yet exposed.
Context
Section titled “Context”On open, project discovery reads marker files, a language census, instruction files and Git status into the system prompt, all as data. The context inspector shows exactly what each turn sent. The read_terminal tool reads a Rust-side terminal model rather than scraping the renderer. The egress preview shows what is about to leave the machine before the first contact with a hosted provider in a session, including the framed ticket content when work starts from the Board, and requires an explicit Send.
Memory
Section titled “Memory”Six scopes: user, project, team, agent, task and council session. Access is enforced per scope through a memory-access capability; every memory records provenance; retention is configurable; memories enter the prompt as data. Recall and remember are tools; the Memory panel edits and deletes.
Plans, councils and teams
Section titled “Plans, councils and teams”Approved plan revisions are materialised into Factory tasks with intrinsic traceability. Council orchestration runs an isolated first round, critique rounds on positions only, moderator synthesis with evidence cited only from read-only tool ledgers, enforced member, round, budget and timeout limits, and ends with the user’s decision (defaults: three members, two rounds, visible estimate). Delegation never widens authority; see Rust internals.
Not yet built
Section titled “Not yet built”Durable background execution, remote workers, scheduling, user-defined agent services and their APIs, and the extension host. See Progress.