Release notes
Conduit v0.5.4: reliable agent runs
Conduit adds strict agent contracts, safer isolated worktrees, clearer live activity, and bounded local runtime retention.
Conduit v0.5.4 makes multi-agent runs safer to execute and much easier to understand while they are active.
Structured agent contracts
Assignments and responses now use strict, versioned JSON contracts. Conduit validates response structure, ownership, verification, blockers, questions, findings, and artifact claims before accepting a role as complete. Valid partial responses, assignment-policy failures, provider errors, and malformed responses are reported as distinct outcomes.
Custom roles can select an explicit semantic role kind while retaining sensible defaults for built-in roles. Active feature packets are captured into read-only assignment context so agents receive the approved specification even when it is newer than the worktree’s checked-out commit.
Safer isolated worktrees
Worktrees now live under the configured state directory by default, have recorded lifecycle and retention metadata, and are protected by internal ignore rules. Automated checkout avoids project checkout hooks, while existing Node and Composer dependency trees are reused inside each isolated workspace.
Conduit-owned assignment files are excluded from change attribution. Agent claims are checked against independently observed worktree changes before results are accepted.
Clear live activity
The worker monitor keeps its compact header limited to phases such as reading, thinking, writing, and running. The event list retains bounded reasoning summaries, commands, outputs, errors, and file changes, follows new events by default, and preserves manual history scrolling.
Architect refinement uses structured runner events for live progress, keeps detailed output out of the loading header, and presents packet review controls on stable rows at supported terminal sizes.
Bounded local state
Custom state directories now ignore every runtime artifact class. Raw diagnostics and completed worktrees have configurable retention, imported legacy data is archived after verification, duplicated transcript storage is reduced, and credential fallback protection is described accurately in persisted metadata and documentation.
Read the documentation to configure roles and run an approved feature packet.