Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Group Human Decisions: Tie-Breaks and Tasks

A group can put a person where it would otherwise guess. This page covers two places:

  • tiePolicy: HUMAN_DECIDES — a tied (or quorum-failed) vote pauses the discussion until a named person picks the winner (Tie-breaks);

  • humans as task-force executors — a HUMAN member works the PLAN/EXECUTE/VERIFY tasks assigned to them (below).

Both build on humans as group members — read that first.

Tie-breaks

Configuration

{
  "name": "Board vote",
  "type": "VOTE",
  "turnOrder": "PARALLEL",
  "contextScope": "NONE",
  "voteConfig": {
    "method": "MAJORITY",
    "optionsSource": "EXPLICIT",
    "options": ["Double down on SEO", "Launch the paid pilot"],
    "tiePolicy": "HUMAN_DECIDES",
    "humanTieBreak": { "decider": "founder@acme.com", "timeout": "PT24H", "onTimeout": "MODERATOR_DECIDES" }
  }
}
Field
Default
Meaning

decider

the group's single HUMAN member

The principal id that decides. It need not be a member — a board can defer to someone who does not sit on it, like a facilitator's escalateTo. A group with no HUMAN member, or several, must name one: saving without a resolvable decider is a 400

timeout

humanMemberConfig.turnTimeout, else none

ISO-8601 duration the decider has. Must be positive. With neither set, the pause waits indefinitely

onTimeout

NO_DECISION

NO_DECISION keeps the honest NONE record; MODERATOR_DECIDES lets the moderator choose (a 400 when the group has no moderatorAgentId); ABORT cancels the discussion

The Manager's phase editor shows these fields when On a tie is set to A human decides.

What happens

  1. The VOTE phase tallies as usual. On a tie or a quorum failure the discussion records the NONE decision with its tally, persists it and fires decision_reached for it, so the tally is visible while it waits.

  2. The discussion pauses in AWAITING_HUMAN_INPUT, as a non-speaker human turn (pauseType: HUMAN_TURN, the same machinery as a member's turn and a facilitator escalation). The pending input carries detail.kind: "TIE_BREAK" and the options; the status and MCP summaries show pendingKind: "TIE_BREAK" and, to a caller who may read the pending content (the owner, the decider or an admin), pendingOptions — an approver who is neither gets the summary without it, so clients must handle its absence; the inbox shows a Tie-break badge, the human_input_requested event carries the detail, and Slack posts a notice naming the decider and the options. The decider's prompt is built from the option texts and the weighted totals only, never from member statements.

  3. The decider answers on the human-input endpoint with a choice:

    choice must name exactly one of the options (an exact match, or the same whole-word rule a moderator's tiebreak reply gets); anything else is a 400 that lists them, and nothing changes. content is an optional rationale. Over MCP, submit_group_human_input takes the same choice argument — optional in the tool's schema, because other human turns have none, but required for a tie-break.

  4. The decision becomes a VOTE record — winner the chosen option, outcome "… chosen by <decider> (tie break)", method vote+human-tiebreak, the original tally, and the losing ballots' statements as dissents (re-read from the transcript, not stored twice). A peer-visible HUMAN_INPUT entry records who broke the tie and why.

  5. The discussion moves on from the phase after the vote. A VOTE phase with requiresApproval pauses for that approval next — the decision first, then the gate — and approving resumes after the vote as usual. A VOTE that is the last phase completes the discussion.

Who may answer

Two checks, the access rule of every human turn. The caller needs a role the endpoint admits (eddi-user, eddi-editor, eddi-approver or eddi-admin — an eddi-viewer cannot answer, even as the decider), and must be the decider's own principal unless they are an eddi-admin. So eddi-approver lets a decider answer but never lets anyone answer for the decider: setting the outcome of someone else's vote is not reviewing it. The principal check needs authentication — with it disabled, no caller is compared with the decider, so any client that passes the request guards can answer. The second of two racing answers (or an answer racing the timeout) gets 409.

Timeouts

onTimeout

Result

NO_DECISION

A SKIPPED entry and a NONE record, "No decision from within "; the discussion moves on

MODERATOR_DECIDES

One moderator turn chooses among the options, gated on maxTurns and the cost ceiling exactly as the policy is when a vote ties; with the budget spent, or an unreadable reply, the NONE record stands

ABORT

The discussion is cancelled, as for a human turn's ABORT

Timeouts are scheduled, re-armed after a restart and bound to the pause they were armed for, like every other human turn.

Spend and turns

A human choice makes no model call and is not charged as a turn. The MODERATOR_DECIDES timeout fallback is one turn, like the policy it borrows.

Audit and metrics

With the audit ledger enabled, each resolution writes one hitl.approval audit entry with verdict: "TIE_BROKEN", approvalCategory: "group.vote_tiebreak", reasonCode (HUMAN_CHOICE, TIMEOUT_NO_DECISION or TIMEOUT_MODERATOR) and the options; the choice is recorded only for a human submission (HUMAN_CHOICE) — a timeout entry carries none, and TIMEOUT_MODERATOR does not record the moderator's eventual selection. It is a human turn, not an approval, so it carries no decisionOutcome — approval analytics skip it. When a meter registry is present (the default with metrics enabled), counter eddi.group.vote.tiebreaks{policy, resolution} counts resolutions (human, timeout_none, timeout_moderator).

Interactions

  • A facilitator escalation deferred to the boundary where the vote tied is recorded as a FACILITATION note and dropped: the decider is about to be asked anyway, and a boundary pauses once. A facilitator's CALL_VOTE move is still applied.

  • The group conversation schema is v5 from this release, and every new discussion is stamped v5, whether or not it uses HUMAN_DECIDES. A pod that predates it refuses to resume any v5 document — a tie-break, but also an ordinary approval or human turn — rather than re-running a vote. During a rolling upgrade, finish the rollout before relying on resumes.

  • Slack stays notification-only for tie-breaks: a Slack user is not the decider's EDDI principal, so the choice is made in EDDI.

Humans as task-force executors

A HUMAN member can sit in a TASK_FORCE (or any group with PLAN/EXECUTE/VERIFY phases) and execute tasks: a legal sign-off, a customer call, a judgement nobody wants an agent to make.

Assignment is explicit

A task reaches a person only when it names them: assignToRole: "ROLE:<role>" or their principal id, a task an agent files for them, or the LLM planner naming them (the planner is told which members are human and that their work is slow). Round-robin ("ALL") never picks a HUMAN member, and neither do bids — humans do not bid, and a BID task nobody bid on falls back to ROLE assignment, which may land on a person.

Agents first, then one task per pause

The EXECUTE wave loop leaves HUMAN-assigned tasks out of its waves, so a person never holds up a wave and the wave budget is unchanged. Once no agent task is left to run but a human task is executable, the discussion pauses in AWAITING_HUMAN_INPUT for the highest-priority one (priority 0 first): the prompt is the same task input an agent would get, the pending input carries detail.kind: "TASK" and the task id (pendingKind: TASK, pendingTaskId in the summaries), the inbox shows a Task badge and Slack says a task is ready. The member answers on the usual human-input endpoint with content — their result.

  • The result completes the task (recorded as a TASK_RESULT entry) — or, when the EXECUTE phase gates tasks for approval (hitlConfig.granularity: TASK and requiresApproval), submits it for approval exactly as an agent's result would be. The resumed wave loop then runs whatever depended on it, and pauses for the next human task.

  • A human task counts as a turn (the human-turn rule) and costs $0.

  • A SKIP_TURN timeout fails the task ("No response from within "); ABORT cancels the discussion. Reassigning an unanswered task to an agent is not supported.

Verification

Agents verify human work like any other. maxReworks skips HUMAN-assigned tasks: a failed human task stays FAILED with the verifier's feedback and a "Not reworked: … is assigned to a human" note, because reworking it would pause from inside VERIFY, and resuming a VERIFY phase replays the whole verification pass. For human review of agent work, use TASK-granularity approval (HITL).

Still refused at save time

  • A HUMAN planner or verifier. In a group with PLAN or VERIFY phases, the moderator — or, without one, the first member by speaking order, who stands in — must be an agent: planning and verification parse a structured reply inside one turn that has no resume point, and prose from a person would silently auto-verify every task.

  • targetEachPeer phases with humans. A person would owe one critique per peer and be a target; the resume bookmark has no (speaker, target) coordinate, and adding one changes the resume path of every pause kind.

  • Nested groups containing humans. A sub-group that pauses is cancelled and its member turn recorded as SKIPPED; supporting it needs a parent pause that follows the child's, across two documents.

A standing team's cadence reclaims a run whose claim outlives eddi.groups.cadence.claim-ttl (default PT24H), so give human members a turnTimeout well inside it.

Last updated

Was this helpful?