You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[coordination][deferred] Typed task/handoff model (owner/status/result) #157
DEFERRED. Promote the minimal task handle from the near-term work into a first-class coordination record — task → owner → status → result — on the existing board + coding_timeline, so ALL delegation (not just Coder's) is observable + resumable. Part of #154.
Scope addition from docs/supervision.md: the typed task model also needs to carry the escalation ladder.
Under multi-layer supervision (#183), when a subordinate hits needs_human it should surface to its supervisor first — which may resolve it without waking anyone — and reach the actual human only after a bounded number of hops.
Without this, multi-layer supervision multiplies human interrupts, which is exactly backwards: the point of a hierarchy is to absorb them.
Related: goals also need a completion predicate and a deadline. #154's complaint about the old one-shot was fire-and-forget; "waits forever" is the opposite failure and just as real — supervisors block and the board fills with permanent running.
Escalation ladder done in f5f521a (lib/escalation.ts). An escalation climbs to the supervisor first and reaches the human only at the root or once hops are spent — so three subordinates under one Lead stop meaning three pings to the same person.
Deliberately tells, does not trigger: the supervisor gets a board card and a note naming the subordinate, its objective and what it can do next. Auto-resolution starts spending unprompted and belongs behind its own gate.
The typed task/handoff half of this ticket (owner/status/result as a first-class record) is not done — delegated runs are tracked via agent_loop_runs + board cards, which covers observability but is not the unified model this ticket described. Reopening as a follow-up if that shape is still wanted.
Activity
docs: coordination-primitives design (#154) — delegate goals, not one…
serge-ivo commentedon Aug 3, 2026
Scope addition from
docs/supervision.md: the typed task model also needs to carry the escalation ladder.Under multi-layer supervision (#183), when a subordinate hits
needs_humanit should surface to its supervisor first — which may resolve it without waking anyone — and reach the actual human only after a bounded number of hops.Without this, multi-layer supervision multiplies human interrupts, which is exactly backwards: the point of a hierarchy is to absorb them.
Related: goals also need a completion predicate and a deadline. #154's complaint about the old one-shot was fire-and-forget; "waits forever" is the opposite failure and just as real — supervisors block and the board fills with permanent
running.feat(coordination): supervision graph as configured platform data (#183)
feat: escalation ladder, repo setting attaches a repo, routing-filter…
serge-ivo commentedon Aug 4, 2026
Escalation ladder done in
f5f521a(lib/escalation.ts). An escalation climbs to the supervisor first and reaches the human only at the root or once hops are spent — so three subordinates under one Lead stop meaning three pings to the same person.Deliberately tells, does not trigger: the supervisor gets a board card and a note naming the subordinate, its objective and what it can do next. Auto-resolution starts spending unprompted and belongs behind its own gate.
The typed task/handoff half of this ticket (owner/status/result as a first-class record) is not done — delegated runs are tracked via
agent_loop_runs+ board cards, which covers observability but is not the unified model this ticket described. Reopening as a follow-up if that shape is still wanted.fix(workflows): a run row nobody closes is worse than a failed one (#291
fix(coding): a repo's address is stored once, and it can be corrected (…