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
Configurable acceptance for coordinated business deliverables #187
Problem observed in synthetic multi-agent acceptance
Execution completeness and business acceptance are separate. A five-work-item task can finish with complete model replies and a correct primary deliverable, yet the final report says a cohort passes a support-ticket rule per buyer using only order counts. A QA employee's caveat can be lost during synthesis.
The existing Reviewed task path and criterion checks should be reused where possible, not duplicated. Coordinated tasks need an explicit final-deliverable acceptance boundary.
Proposed scope (next product priority, not shipped)
Opt-in, user-configured checks per task/template. Simple tasks keep the current minimal workflow.
Bind checks to the selected deliverable(s), with immutable check/version/config snapshots.
Typed facts/calculations with explicit numerator, denominator and units. Missing inputs produce an unknown/needs-review result, never a fabricated pass.
Start with required fields and deterministic rate/threshold checks; do not hardcode financial rules into the core.
Distinguish execution status from acceptance status in the API and Console. Preserve the original output and failed-check evidence.
Gate downstream delivery/external actions where configured. No irreversible action on failed/unknown acceptance.
Allow explicit human acceptance with audit. Keep model review advisory rather than a substitute for deterministic checks.
Complete correct inputs => deterministic acceptance with source evidence.
Conflicting prose such as unknown support compliance + all metrics passed => remains unaccepted pending correction/review.
No acceptance policy => simple task behavior unchanged.
Retry/replan/replay cannot alter pinned checks or silently accept another deliverable.
Separate follow-ups: measured tokens vs pricing estimates, and observable true worker parallelism. Production external-memory credential/deletion reconciliation remains tracked in #183.
Activity
0YHR0 commentedon Sep 30, 2026
Implemented and merged in #188, including bilingual Console, pinned typed/unit-aware checks, audited decisions, gated primary export, memory/notification guards and live-model qualification. Semantic review, automatic revision and external publishing remain separate follow-ups.