abcd plans unattended and asks only what a person owns

Press Release

A person hands abcd an issue or an idea and says build it. abcd takes it from there without a terminal of its own: it promotes the issue, conducts the planning interview, plans the intent and starts the build, launching the person's own claude where no host session is driving. Along the way it puts to the person only the decisions that are theirs to make, each in their role's words (outcomes and choices for the product thinker, mechanism and trade-offs for the technical facilitator), and settles everything the record already answers itself, writing down what it settled and why. A decision only a person can make parks the work that depends on it, and nothing else; abcd never adopts a product decision on a person's behalf, and a question left unanswered stays asked.

Why This Matters

On 2026-10-24 a product thinker asked for one issue to be built (/abcd:build iss-2610140652230444). Getting from that request to a running build took six verbs, three refusals and two questions about abcd's own record-keeping, and it then stopped at a planning interview the person could not run: the session sent them to the terminal verb, the verb needs a runner on a paid API key, their account is a subscription, and the verb's refusal never said that the session itself may conduct the interview (commands/intent.md, "Planning interview (host-run, with the human present)"). Each refusal was correct by its own rule. Together they make a person do the orchestration abcd exists to do, and they ask questions (which category, which link, which route) that are about abcd rather than about the product.

The person's own words: "too complicated. Not working … the user (TF / PT) is only presented with the options it can answer."

Mechanism

We expect a person to reach a running build from one request, answering only their own decisions, because three things change together: (1) the unattended path gets a model it can actually start: the person's own claude, under their own login (iss-2610240932215918), or the host session already driving (iss-2610240932444000); (2) deciding whether to ask is taken away from the agent that writes the plan and given to a separate detection pass with a structured output, each open decision carrying an owner (product thinker, technical facilitator, or the record) and the evidence that settles it, because models asked to plan almost never ask unprompted, even on badly underspecified input (Ambig-SWE, ICLR 2026, https://arxiv.org/abs/2502.13069v3; Su and Cardie 2026, https://arxiv.org/abs/2605.25284), and separating detection from execution is what measurably helped (https://arxiv.org/abs/2603.26233); and (3) a person-owned decision becomes a durable parked question that blocks only the work depending on it, resumed by session id when answered (the deferred-tool round trip: https://code.claude.com/docs/en/hooks; LangGraph's interrupt model: https://docs.langchain.com/oss/python/langgraph/interrupts). Shown wrong if, over the next ten requests run this way, the person is asked a question whose answer the record already held, or a product decision is found settled without a person's recorded answer.

Scope Conditions

• One person's own machine and own model account; abcd starts the person's unmodified claude binary and never reads, mints or stores a credential.
• Unattended means the person answers within a working day; work longer than that between answers is outside what this claim is about.
• The repository is managed by abcd, with its decision store, its issue ledger and its intent store, so the record a pass reads to settle a decision exists.
• Claude Code alone: the deferred-tool round trip is its feature, and another runner is outside this claim until it offers a structured output naming the open decisions and a resume by session id.

Acceptance Criteria

• Given an issue build <iss-N> hands back on its category or its severity, when a person runs it, then the refusal also names the route that reaches a build from there: capture promote <iss-N> when no intent delivers it, or build <itd-N> naming the intent that does (the promoted refusal already names its intent, internal/core/capture/eligible.go:238-245).
• Given a draft intent and a host session driving, when the planning interview runs, then the session conducts it through the same answers record the terminal interview writes, and no runner route is required.
• Given a draft intent and no host session, when the planning interview runs on a machine whose model account is a subscription, then abcd starts the person's own claude under that login and the interview proceeds; an interview whose run did not load abcd's question hook is refused before its first turn, naming the hook.
• Given a planning pass over a draft, run separately from the agent that writes the plan, when it finishes, then every open decision it found is recorded with its owner (product thinker, technical facilitator, or the record) and its evidence; each whose owner is the record cites the record id that settles it and is written into the intent's Decisions with that id; and only the rest are put to a person.
• Given a seeded draft whose one open decision a cited decision record already answers, when the pass runs, then no question is put to a person and the Decisions line names that record (the GRILL rule that a decision with one defensible answer is recorded, not asked, made checkable by the cited id).
• Given a decision whose owner is the product thinker, when it is put, then it is asked in product terms with no record ids or internals, and one whose owner is the technical facilitator is asked with the mechanism and ids, as the GRILL register rule already says; the new part is that the owner is recorded before the question is drafted.
• Given a parked question, when the person has not answered it, then the work that depends on it stays parked, other work proceeds, and nothing adopts a default; an answer resumes the parked session by its id.
• Given a request carried through to a running build this way, when the run's record is read, then it lists every question put to a person, its owner, and every decision settled from the record with the id that settled it.

Decomposition

Hand-run four-piece table (DECOMPOSITION, itd-84), 2026-10-24. Routing awaits a person's confirmation.

Open Questions

• Does this reverse dcn-2607180000000002 (plan sign-off is implicit in a human running intent plan; agents are barred from unattended planning)? The intent proposes that an agent may run intent plan when every product-owned decision in the plan carries a person's recorded answer. Reversal flag, advisory: the product thinker rules it.
• Does this reverse ruling RN2 (dcn-2610020000000016, the claude runner on a paid key only)? iss-2610240932215918 carries the evidence and the probe it waits on. Reversal flag, advisory: RN2 was the technical facilitator's ruling, so the technical facilitator rules it.

Typed Links

• reverses dcn-2607180000000002 (advisory, unconfirmed): agents barred from unattended planning.
• reverses dcn-2610020000000016 (RN2, advisory, unconfirmed), and with it the --bare premise of dcn-2610161624025493.
• refines iss-2610222000271964 (a drain run from a plain terminal launches claude itself): the same launch, for planning.
• builds on itd-2610030810370060 (plain-Terminal interviews): the written interview this intent makes reachable without a paid route.
• builds on itd-2609201916056194 (the command-line runner).

Audit Notes

Empty. Populated by intent-auditor when intent moves to shipped/.