warded control plane coherence

New issue

Idealistic release thesis: warded should feel like one dependable control plane for agent work, not a set of adjacent launch scripts. A director should be able to choose a harness and model, dispatch at useful capacity, respect human intervention, survive routine broker/container failure, expose what happened, and read its defaults from one obvious config path.

Do not pull issues into this milestone automatically. This milestone starts empty. Candidate lanes below come only from open issue titles reviewed on 2026-07-15.

Candidate lanes:

Run correctness - human comments block active PR/issue work, read-only director surfaces still merge when authorized, PR pressure and global engineer limits behave consistently, director startup does not do surprising backlog work.
Dispatch resilience - broker availability, broker failure recovery, Docker name-conflict reproduction, stale reservations, stopped container cleanup, Docker kill priority, and safe ~/.ward state layout.
Observability and recovery - first-class broker logs, structured run summaries, friction events, director log drain, prominent hard resource warnings, and issue-scoped agent logs that explain missing output.
Config coherence - one authored smart-defaults source, fleet tunables single-sourced, selected bundle agent defaults honored or explicitly ignored, per-role/per-harness model overrides, Goose model selection, and KDL/spec artifact cleanup.
Release and adoption path - promote-only release branch behavior, release retag/draft artifact design, canonical first-time user flow, config-blocking onboarding, and a validation ladder after ward exec retirement.

Exit shape:

A normal director burndown is safe to run without babysitting broker state.
A failed or paused run leaves enough structured evidence for the next actor.
Model and harness configuration is one documented API across Claude, Codex, Goose.
Default values have a single visible owner.
The release path reflects one promoted, already-gated commit instead of re-testing by accident.

Issue selection happens in a later pass.

No due date

81% Completed