3.8 KiB
ADR-0012 - Intent To Work Planning Foundation
Status: Accepted Date: 2026-08-17 Accepted under: DEV-000063 milestone decision
Decision
work remains the canonical owner for Goal, Task, Decision and the new V1 planning foundation. No new canonical owner is introduced.
Help is the central user-facing concept. Goal, Work and Plan are optional durable coordination primitives, not mandatory records for every conversation. A simple explanation, ordinary answer, short cognitive response, direct reminder capability, safe deterministic operation, or capability-gap response may bypass durable Work entirely.
Human chat must not reduce Domovoy to chat -> prompt -> model or to a Codex-specific task generator. The canonical pipeline is:
When durable coordination is actually needed, the canonical path is:
HumanUtterance -> OriginalHumanIntent -> Goal -> PlanCandidate -> accepted Work/Plan state -> CapabilityRequirement -> DelegationCandidate/Action/Investigation/Observation/Outcome.
Prompts, Codex instructions, mail payloads, reminders and provider requests are executor-adapter representations. They are not canonical human intent, Goal or Plan.
Ownership
interactioncontinues to own append-only human utterances, original intent provenance and typed utterance interpretations.workowns durable Goals, lifecycle, planning state, plan candidates, accepted plan summaries, capability requirements and general delegation candidates.capabilityowns available capability/binding registry and health. Work may reference capability requirements by typed IDs/kinds but does not select a concrete provider as canonical truth.epistemicsowns verified/unverified/unknown evidence semantics. Work planning must preserve epistemic state and may add verification requirements before risky work.authorityandcommitremain required for any real mutation or external side effect. Planning creates no permission and no execution claim.
Boundaries
Work PlanCandidates are structured candidates that must be validated before persistence or acceptance. Cognitive planning may later propose candidates only through the Cognitive Gateway/Router; model output never directly mutates canonical Work state. DEV-000063 may implement deterministic V1 candidate planning and keep cognitive candidate generation as a typed extension point.
CapabilityRequirement is executor-neutral. Examples include READ_EMAIL, SEND_EMAIL, READ_DATABASE, QUERY_LOGS, INSPECT_PROJECT, MODIFY_CODE, RUN_TESTS, WEB_RESEARCH, COGNITIVE_REASONING, and SCHEDULE_REMINDER. This is an extensible vocabulary, not a closed ontology and not an executor list.
DelegationCandidate is general. Codex is only a possible future implementation for MODIFY_CODE; mail, research, scheduler, device and database adapters can use the same canonical Work model without changing OriginalHumanIntent or Goal.
Consequences
/chat can choose a HelpPath first. Only the durable-work path creates a planning-only Goal and Plan. Other paths can answer, use cognition, report a missing capability, or describe a direct capability requirement without materializing a Goal/Plan. It must not pretend work was completed. Completion remains tied to outcome semantics, not to a model response.
Project is only one possible context reference. The Work foundation stores project context as optional opaque references and must later permit people, emails, documents, images, servers, organizations, research subjects and other context types without redesigning Goal/Plan.
Capability gaps are first-class planning information. "I cannot do this yet" is not necessarily terminal in future milestones: the system may later discover, connect, configure or build a missing capability and return to the original need. DEV-000063 does not implement self-extension; it preserves the extension path.