---

type: adr id: adr-0089 status: accepted created: 2026-06-20 updated: 2026-06-20

ADR-0089 — The decision handoff: an ## Open decisions section that frames an open decision for the human (convention, fact-not-verdict, evidence-grounded)

Context

A developer adopting corpus rarely runs one agent — they orchestrate several, each deep in its own domain (a parser change here, a docs sweep there, a benchmark in a third worktree). At the moment a unit of work closes and a decision is owed, that developer is context-poor: pulled across domains, holding none of any single agent's depth. The agent that just did the work is the only party still holding the full context. A closeout that ends in three terse bullets ("next: do X / consider Y / maybe Z") forces the developer to re-derive context the agent already has — and re-contextualizing the decision is the agent's job, not the developer's. (This is the owner's stated working preference, raised 2026-06-20.)

A /deep-research run (2026-06-20) and a follow-up 2026 recency check grounded the question in primary sources. The findings, in short:

• A lone recommendation has documented failure modes in both directions: over-trust of the recommendation [OVERRELIANCE-REVIEW] [AICODE-INSECURE], and, once the agent errs even once, algorithm aversion — the human over-discounts it thereafter [ALGOAVERSION].
• More justification does not fix this and can worsen it: explanations often persuade rather than help a person evaluate [OVERRELIANCE-REVIEW]; a light cognitive forcing function reduces over-reliance better than explanation alone [OVERTRUST-CFF]. The job of the "why" is to make checking cheap, not to convince.
• Choice overload is not a law of option count — count alone has a near-zero average effect; harm is conditional on decision difficulty and set complexity [CHOICEOVERLOAD]. So the lever is comparable, well-structured options, not a hard cap on how many.
• The "present evidence for and against the options, not a single recommendation" stance comes from a position paper [EVALAI] and is contested — a reasoned design stance, not a proven result. The 2026 recency check [PERSUASIONPARADOX] is mixed and task-dependent — explanations hurt on visual reasoning but helped on language-based logical reasoning (the task class closer to a developer's decision) — so it corroborates only weakly and is cited as preliminary, never load-bearing.

The review packet already routes exceptions to human attention (ADR-0060): unverified rows, scope drift, risky files. But an open decision is not an exception in the work — it is a fork the human must choose. The packet has no home for it: ## Suggested decision is the reviewer's merge call only. This ADR adds that home.

Decision

1. Adopt the decision handoff as a convention. When a unit of work closes with a genuinely open decision the human must make, the closeout frames that decision rather than dumping bullets. The concrete home is an optional ## Open decisions section in the review packet (extending ADR-0060's section list — the ADR-0076 precedent, which added the optional Task status section to that same list); the same convention applies to a task's ## Run summary and to an agent's chat handoff. Level: convention (ADR-0063) — nothing enforces it; review may inspect it. It routes a decision to the human — a fact, never a verdict (ADR-0077 Decision 8): the section presents and recommends; the human decides.

2. The fields are fixed; the content is present only when a decision is open. Each open decision carries: the decision (one line) · 2–4 comparable options, each with its tradeoff (the case for and against) · a recommendation + a brief why · the context/impact the agent holds and the human may not · what it blocks. The for-and-against framing guards both failure modes — over-trust [OVERRELIANCE-REVIEW] and algorithm aversion [ALGOAVERSION] — by giving the human something to re-engage with rather than reflexively accept or reject.

3. Present only when something is open; no counts ceremony. When nothing is open, the section is absent — that is the default and there is no ritual to perform. Options are listed, never tallied; the 2–4 range is a soft heuristic, not a hard cap, because the lever is decision difficulty, not option count [CHOICEOVERLOAD]. This keeps the convention low-ceremony — consistent with the no-counts-ceremony rule.

4. Short calibration cues over persuasion; the recommendation is skippable. The "why" is brief and verification-oriented — how sure · what would change it · what it blocks — not a persuasive essay, because explanation alone does not cure over-trust and longer explanation can worsen it [OVERTRUST-CFF] [OVERRELIANCE-REVIEW] [PERSUASIONPARADOX]. "What it blocks" is a lightweight, forcing-function-inspired cue — but honestly, a static markdown section cannot implement an interaction-level cognitive forcing function (the [OVERTRUST-CFF] mechanisms — decide-before-seeing, withhold-until-requested, a brief wait — need a live interaction). So the over-trust mitigation here rests on the for-and-against framing (itself contested, Decision 6), not on a true CFF. The recommendation is present-but-ignorable — the human may skip it — never a forced second opinion, which would trade over-trust for under-trust; a genuine decide-before-seeing interaction is available only in the chat-handoff channel, not the persisted packet.

5. A convention this cycle — no check, no contract mint. ## Open decisions is prose; it surfaces a decision for a human, with no closed-value form a checker could reconcile (unlike C012–C015). So this ships as a convention only — no checks.yaml row, no contract-version bump — matching ADR-0088's convention-first posture. Any future check (e.g. "a closed task left an undocumented owner-decision") is deferred to a measured ADR, and would risk the over-reliance an always-on prompt induces.

6. Recorded honestly as a well-reasoned convention, not a proven win. The options-over-recommendation prescription rests substantially on a contested position paper [EVALAI] and on adjacent human-AI-decision evidence, not on a measured study of agent-orchestration handoffs; interventions of this family reduce over-reliance but do not reliably produce appropriate reliance [RELYORNOT]. The 2026 recency check strengthens the underlying over-reliance findings and adds developer-domain support [AICODE-INSECURE] — but the convention is adopted as a reasoned design stance, revisitable if measured evidence in the orchestration setting lands.

Alternatives considered

Alternative	Why weaker
End with a terse single recommendation (the status quo)	A lone recommendation invites over-trust [OVERRELIANCE-REVIEW] and is brittle to one error → algorithm aversion [ALGOAVERSION]; it also forces the context-poor orchestrator to re-derive the fork the agent already understood.
A required, always-present ## Open decisions section	An always-on second opinion induces over-reliance and trades over-trust for under-trust; present-only-when-open keeps it a signal, not boilerplate.
A hard option-count cap (e.g. "exactly 3")	Unsupported — option count alone has a near-zero average effect; the lever is decision difficulty and option comparability [CHOICEOVERLOAD]. Hence a soft 2–4, not a rule.
Long, persuasive justification prose	Explanation alone does not cure over-trust and detailed explanation can worsen it [OVERTRUST-CFF] [OVERRELIANCE-REVIEW]; the "why" should make checking cheap, so short calibration cues win.
Mint a check / contract rule now	No closed-value form to reconcile; premature for a prose convention, and an always-on prompt risks the very over-reliance the evidence warns of. Convention-first (ADR-0088 precedent); a check defers to a measured ADR.

Consequences

Accepted: the review packet gains an optional ## Open decisions section; docs/reference/artifact-formats.md and the kit's templates/review.md carry it; this extends ADR-0060's frozen review-packet section list (the ADR-0076 precedent — which added the optional Task status section the same way; the original set is not rewritten). It is convention-level: nothing is enforced, no check fires, no contract version moves. Honors ADR-0077 Decision 8 (a fact, never a verdict) and ADR-0063 (convention, never enforcement-claimed). Because the prescription rests on a contested position paper and adjacent evidence rather than a measured orchestration study, it is recorded as a reasoned convention — to be revisited if the benchmark or field evidence ever measures it. Generalizes the owner's "frame the decision" preference into a framework affordance, so every adopter's agents close a unit of work the same way.

Propagation

docs/reference/artifact-formats.md (the ## Open decisions entry in the review section list + a load-bearing bullet) · ../corpus-starter-kit/templates/review.md (the optional section with authoring guidance) · docs/research/sources.md (the eight evidence entries, tiered) · docs/adrs/README.md (the index row) · the corpus-works trail (SPEC/TASK/REVIEW + the persisted research). Extends 0060; honors 0077 Decision 8 and 0063; convention-first like 0088. No contract change, no check.