The phase negotiator — a Socratic agent that proposes phases and grounds every trade-off

Terminology note. The how layer is named the spec. This ADR's prose was updated by the spec-terminology-rename ADR (adr-11).

Context

adr-9 made the phase abcd's sequencing layer: an ordered stretch of work that bundles intents and plumbing and ends in a milestone, with a product-authored ## Expectation and ## Phase Acceptance.

adr-9 gave sequencing a home. It did not give it a negotiator. Deciding which intents go in which phase, and in what order the phases run, is currently the product thinker's unchallenged judgement. That has a known failure mode, and the user named it directly: the product thinker "wants everything." They hold a vision of the whole product; they have no natural adversarial reader at the phase-planning grain who will tell them, honestly, what they give up when they pull an intent forward, defer one, or split a phase.

abcd already has adversarial readers at the other grains — intent-fidelity- reviewer (does reality match the intent's acceptance?), the /abcd:intent grill skill (Socratic interrogation of an intent before it is planned), and dual-backend review of specs (adr-8). The phase-planning grain has none.

The hard part is not "add an agent that lists trade-offs." An LLM asked "what are the trade-offs of this phasing?" will fluently invent plausible ones. A hallucinated trade-off is worse than no trade-off: it spends the product thinker's trust on a fiction. The real design problem is grounding — every trade-off the agent raises must trace to a fact that already exists in the project, not to the model's generative fluency.

Decision

abcd adds a phase negotiator: an agent that proposes a phase (or a re-sequencing of phases) and then discusses the trade-offs with the product thinker. It is captured as a user-facing capability in itd-41.

The negotiator is built from two patterns abcd already has, deliberately — it is the same family as intent-fidelity-reviewer and the grill skill, applied one grain up, at the phase-planning grain:

1. Socratic, not assertive (from the grill skill). The negotiator does not declare "deferring itd-3 costs you X." It interrogates the proposed phasing with questions, each tagged with a named Socratic move, the same way /abcd:intent grill interrogates an intent. "You have placed itd-3 in Phase 3 — what in Phase 2 depends on the rules loader and now waits?" The product thinker's own answer surfaces the trade-off. An agent that only asks cannot hallucinate an answer it does not give.

1. Grounded comparison (from intent-fidelity-reviewer). Where the negotiator does assert, every assertion must cite a fact that already exists in the project — exactly as intent-fidelity-reviewer cites a specific ## Acceptance Criteria bullet rather than a generated standard. The grounding sources, in priority order:
   • the dependency DAG in brief/06-delivery/01-build-sequence.md and intent/spec dependency declarations — "deferring itd-3 delays every item whose dependency chain runs through the rules loader" is a checkable graph fact;
   • each phase's ## Phase Acceptance (per the adr-9 amendment) — moving an intent out of a phase means naming which ## Phase Acceptance bullet can no longer be met;
   • the ## Expectation prose — what the phase's user-truth would lose.

The grounding rule (load-bearing): the negotiator MUST NOT present a trade-off it cannot tie to a named DAG edge, a named ## Phase Acceptance bullet, or a named ## Expectation clause. If a concern has no such anchor, the negotiator raises it as a question (Socratic mode), never as an asserted trade-off. "Honest trade-offs only" is enforced by construction: assertions are grounded; everything else is a question.

The negotiator's output is a proposal plus a discussion, not a decision. The product thinker still owns the phase plan — they own ## Expectation and ## Phase Acceptance, per adr-9. The negotiator is the adversarial reader that makes "wants everything" meet "here, concretely and truthfully, is the cost."

Alternatives Considered

1. No negotiator — phase planning stays the product thinker's unchallenged call. Rejected: it leaves "wants everything" with no counter-force. adr-9 gave sequencing an artefact; without a negotiator the artefact is just an unreviewed wish-list, and the phase audit only catches the mismatch after the work is built.
2. A trade-off agent that asserts trade-offs directly (no Socratic layer). Rejected: this is the hallucination trap. An agent licensed to assert trade-offs will assert ungrounded ones fluently. Restricting assertions to grounded facts and routing everything else through questions is what makes the output trustworthy.
3. A pure DAG/graph tool — compute the dependency consequences mechanically, no LLM, no discussion. Rejected: the mechanical consequences are part of the grounding, but a graph diff is not a negotiation. The product thinker needs the discussion — the Socratic surfacing of what an Expectation would lose, which is a judgement exchange, not a graph query. The negotiator uses the graph as grounding and the Socratic loop as the interface.
4. Fold it into the existing grill skill as another mode. Rejected: grill operates on a single intent before it is planned; the negotiator operates on a phase (a set of intents plus plumbing) and on the ordering of phases — different input, different grain. Shared lineage (Socratic moves) does not mean shared surface. It is its own capability.

Consequences

Gains:

• The phase-planning grain gains the adversarial reader the other grains already have. abcd's review family becomes complete: grill (intent, pre-plan) → phase negotiator (phase, pre-commit) → dual-backend review (spec) → intent-fidelity-reviewer / phase audit (delivered reality).
• "Honest trade-offs" is a structural guarantee, not a hope: assertions are DAG/acceptance-grounded, the rest are questions.
• The ## Phase Acceptance blocks (adr-9 amendment) gain a second consumer — they are both the phase-audit target and the negotiator's grounding input.

Costs / obligations:

• The negotiator needs the dependency DAG to be accurate and machine-readable enough to ground claims against. Where the DAG is prose-only, the negotiator is limited to Socratic questioning for that area — an honest degradation, but a real limit.
• One more agent to maintain, with the grill skill's Socratic-move taxonomy as a shared dependency.
• The negotiator must not drift into a decision-maker. Lint or prompt-level discipline keeps its output advisory (proposal + discussion), preserving the product thinker's ownership of the phase plan per adr-9.

Downstream decisions enabled:

• itd-41 is the user-facing intent that specifies the negotiator's surface, flow, and acceptance criteria.
• A future shared "Socratic move taxonomy" module, if grill and the negotiator converge enough to warrant extracting one.