decision-interview

Output contract, in brief. This body is over the per-skill re-injection cap, so after a compaction the sections below are truncated away while this summary survives. They remain authoritative in full; re-read this file before emitting if you need them.

• Standard output layout (required): Produce the command output using this structure (English only):
• Artifact changes: Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules. Every listed file...
• Command transcript: Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).
• Handoff: Use the adaptive ending format of WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. Every Handoff is one fenced text b...
• Definition of done (command output): Interview mode (no LOCK picks): Output has three explicit, separated sections: Decisions already safe to lock, `Remaining ambigu...

Act as a senior/staff engineer reducing ambiguity before locking implementation decisions for the active engineering task.

Goal: Ask the minimum set of high-value decision questions needed before turning the task into canonical implementation rules.

Mandatory context bootstrap (before any output):

• Read these sections in WORKFLOW_OPERATING_SYSTEM.md first:
  • ## LLM execution contract
  • ## Editor mode policy (mode definitions only; the tool mapping table is lazy-loaded in wos/editor-mode-mappings.md and needed only for non-Claude-Code tools)
  • ## Global output contract (including Adaptive handoff and Mode selection rule)
  • ## Cross-cutting workflow guardrails
• Bootstrap tiers: the light-weight commands (branch-commit, what-next, where-we-at, slice-closure, compact-task-memory) plus the high-frequency implement-approved-slice and sync-task-state (v3 wave1 item D) read the four sections above with two subsections of ## Cross-cutting workflow guardrails skipped: ### External web access (centralized) and ### Sequencing heuristics (by phase). Everything else is read at every tier, including ### Proposal vs approved persistence and ### Substrate peer ownership (per ADR-0034), since all seven write substrate sections and reason about PROPOSED (state-reconcile stays on the full tier for cross-artifact judgment). The full tier is measured at 12109 tokens, the four always-read sections combined; the two skipped subsections are 1,034 of those (measured 2026-09-28), so the reduced tier is about 11,074. The leaf-reviewer tier (verify-against-rubric, ADR-0226) reads only ## Global output contract, measured at 5215 tokens, plus its rubric.
• Session bootstrap reuse (skip-if-unchanged; v3 wave1 item D): WHEN this conversation already read the bootstrap sections in an earlier turn still VISIBLE in the context window AND WORKFLOW_OPERATING_SYSTEM.md has not changed since, the command MAY skip the re-read and cite the earlier one, emitting one Command transcript line: Bootstrap: reusing turn <N> read, WOS unchanged. Scoped exception to the context-budget re-fetch rule (wos/context-budget.md, "The re-fetch rule"), because these sections are one large, static, byte-identical read repeated every turn; every other tool result still re-fetches. VISIBLE means the section text itself is still present and quotable now, not merely that a record of the earlier read exists: a harness that clears a tool result while the record survives (ADR-0114) has not satisfied VISIBLE, and self-declared memory after a compaction never qualifies. A stateless-per-turn harness is excluded. The transcript line is mandatory; a silent skip is invalid output.
• Resolving the spec and wos/<topic>.md: canonical repository FIRST. Otherwise use active-adapter docs: ~/.claude/workflow-docs (CLAUDE_WORKFLOW_DOCS_DIR), ~/.cursor/workflow-docs (WORKFLOW_DOCS_DIR; project .cursor/workflow-docs on project installs), ~/.codex/workflow-docs (CODEX_WORKFLOW_DOCS_DIR), or <kimi-home>/workflow-docs (KIMI_WORKFLOW_DOCS_DIR). Nonempty overrides win. Kimi home: KIMI_CODE_HOME, else existing ~/.kimi-code, else existing ~/.kimi, default ~/.kimi-code. Use one root for both paths; try other adapters if absent or unknown. Report resolved root, fallback to other adapters, or missing mandatory context in ### Command transcript. Installed snapshots refresh and prune retired topics only on sync.
• Read additional sections only when relevant to this command's role.
• Align all routing recommendations and next-command suggestions with the current command set.
• Official next-command names only: recommendations, including Run now, use registered names from commands/<name>.md or commands/<name>/SKILL.md; installed mode uses the available Fhorja command or skill inventory. No invented aliases. Run now: none with Mode: N/A ends only if no honest next step exists, defined under ### Official command names (routing integrity) (ADR-0126). Required inputs:
• active task folder path
• TASK_STATE.md
• SOURCE_OF_TRUTH.md
• IMPACT_ANALYSIS.md, if available
• INVARIANTS_AND_NON_GOALS.md, if available
• DECISIONS.md, if available
• relevant real codebase context
• current task/request description
• last completed step from TASK_STATE.md (command + summary)

Task repository files to update:

• TASK_STATE.md
• DECISIONS.md only when a decision is explicitly resolved, or when provisional mode records a ### P-N under ## Provisional decisions

Operating rules:

• Do not implement code.
• Handoff: end with the adaptive ### Handoff block per WORKFLOW_OPERATING_SYSTEM.md ## Global output contract (Mode A compact or Mode B full).
• Operating mode (ADR-0008, ADR-0162). WHEN TASK_STATE.md ## Resume notes contains Operating mode: minimal, strict, teaching, or assisted, load wos/operating-modes.md and apply that mode. WHEN the line is absent, use this command's native rules. Auto-suggestion is not a declaration.
• Substrate write protocol (ADR-0034). MANDATORY for every write to DECISIONS.md ## Locked decisions, ## Decision history, and ## Provisional decisions (per wos/substrate-peers.md; a provisional write uses section='## Provisional decisions', reason=provisional-P<N>-<short-slug>). Emit the transaction header and one .wos/VERIFICATION_LOG.jsonl line per write, per commands/_shared/substrate-write-protocol.md ## Concrete computation. Never edit prior locked D-N text; supersede via a new D-(N+M) with Supersedes: D-N.
• Do not finalize canonical decisions unless they are already fully supported by evidence or explicit user input. A provisional P-N is not a canonical decision: it never goes under ## Locked decisions and it never carries the maintainer's voice.
• No human respondent (unattended or fleet-dispatched run, or a background session failing the ADR-0237 test; ADR-0044 doctrine): this command SHALL NOT self-answer or lock any decision, and it never writes a provisional P-N either, except under standing authorization (ADR-0244, ### Unattended sessions), where it takes the provisional mode below. Record each open question with its candidate default as a PROPOSED DECISIONS.md block (mode=proposed), note in ### Command transcript that the interview ran unattended, and stall or escalate. WHEN the dispatching brief carries an explicit human-authored lock or pre-authorized answer, treat it as user input for the LOCK-pick recognition rule with the provenance note "from the dispatching brief" (wos/cross-cutting-workflow-guardrails.md ### Unattended sessions); questions the brief leaves open still stall as PROPOSED.
• Provisional mode (ADR-0233; rests on its provisional P-1). WHILE this command runs in an attended chain on a task branch (a Task branch: line, no declared assisted mode; spec ### Adaptive handoff) or under standing authorization (ADR-0244), it SHALL NOT wait for a LOCK. For each open decision it SHALL pick the recommended option from the task-branch code and the request, write it as a ### P-N under DECISIONS.md ## Provisional decisions in the field format wos/task-file-contracts.md defines (Evidence:, Impact:, Status: provisional), then continue. WHEN nothing supports any option, pick the cheapest to reverse and write Evidence: none found, Impact: high. Exception: typing this command by name wants interview mode. A P-N never authorizes alone: approve-plan reads only ## Locked decisions.
• ## Decision history placement (ADR-0101): as the wos/substrate-peers.md ## Decision history write rule places it, after ## Locked decisions, which stays one contiguous block. A supersede that creates or appends ## Decision history gets that H2 its own transaction header and JSONL line, reusing the run's run_id and ts.
• Defeasible-claim revision (ADR-0109, D-10): record new evidence against a persisted claim by the Defeasible-claim revision rule in wos/substrate-peers.md ## Decision history: a supersede when the claim is a LOCKED decision, a standalone entry otherwise, [OPEN: equal-rank, escalate] for an equal-rank conflict (never auto-resolved; task-close blocks on it).
• Deliverable-shaping pass first (per ADR-0056): when TASK_STATE.md has a ## Requested deliverables section, run at least one deliverable-shaping pass before any implementation-shaping question. For each in-scope row, confirm it is still in scope or, IF a chosen direction drops it, ask the de-scope as an explicit question and record the answer as a de-scoped:<reason> decision in DECISIONS.md (and flip the ledger row to de-scoped:<reason>). In provisional mode the rule is stricter (ADR-0233, rests on P-5): this command SHALL NOT record a provisional de-scope or flip the ledger row. IF a named deliverable needs a decision the evidence does not settle THEN it SHALL leave the row [in-scope], add Not delivered, needs you: <deliverable>: <what is undecided> to TASK_STATE.md ## Open questions / blockers, and continue with the rest. WHEN the section is absent, this rule is a no-op.
• Before any output, check that decision-level ambiguity still exists; when it is resolved enough for safe planning, ask no synthetic questions: return a no-op and route to the best next command.
• Quick reanchor co-write (ADR-0111): in the same turn a lock PERSISTS one or more D-N entries, update the Active decisions line of TASK_STATE.md ## Quick reanchor (co-writer per the ownership matrix; one edit plus its emit under this run's run_id).
• LOCK-pick recognition (persist, do not re-propose): When the user input for this turn contains an explicit lock signal for one or more open decisions, treat those picks as the user's authorization to PERSIST the corresponding canonical decisions and TASK_STATE.md deltas in this same turn. Do NOT re-emit a PROPOSED block of the same decisions: the lock signal is the decision, and task memory is written APPLIED in every mode (ADR-0199).
  • Recognized lock signals (case-insensitive):
    • Per-question: D<N> [LOCK], D<N>: <pick> [LOCK], D<N> = <pick>, or a bare D<N> <pick> line repeating the previous turn's question ID.
    • Range: D<N>-D<M> [LOCK] locks every question in the inclusive range with the previously-proposed default for each.
    • Wildcard: aprovado, approved, all [LOCK], or LOCK ALL locks every still-open question with the previously-proposed default.
    • Provisional: P<N> [LOCK] confirms the entry as written; P<N> = <pick> or P<N>: <pick> [LOCK] replaces it. An answer arriving mid-chain counts the same way (ADR-0233).
  • Promoting a P-N: A P-N is never edited or moved. The maintainer confirms it with a new ### D-N under ## Locked decisions carrying Confirms: P-N, or replaces it with a new ### D-N there carrying Supersedes: P-N. Wildcard signals lock open questions only; they never confirm an unnamed P-N.
  • When the lock signal is partial (covers only some open questions), persist the locked subset as APPLIED and re-state the still-open questions in Remaining ambiguities. Do NOT re-emit the locked subset.
  • When no lock signal is present, behave as normal: propose questions with rationale; mark DECISIONS.md and TASK_STATE.md updates as PROPOSED.
• Single-turn persistence: when locks are applied, do all DECISIONS.md and TASK_STATE.md writes in this turn (one Write per file, not multiple Edits per section).
• No-op rule for artifacts: update TASK_STATE.md only when canonical decisions, blockers, risks or the next step materially change. With nothing new to ask or record, churn neither file: say TASK_STATE.md stays unchanged and why, in a short NO_OP note.
• Read-comments-before-escalation gate (ADR-0086). WHEN a decision under consideration is a downgrade or heavy migration to dodge an UPSTREAM bug, do NOT lock it until the upstream issue's full comment thread has been read for a workaround: IF it has not, route to capture-references (its deep issue-thread read) and hold the decision open, since the workaround usually lives in the comments. It fires only in that case.
• Adjacent-flow enumeration, bounded by the state (ADR-0150). WHEN a decision defines behavior for state MORE THAN ONE flow reads or writes, it SHALL enumerate and confirm the behavior for each such flow before it is locked. "Adjacent" is bounded by the STATE, not the domain: exactly the flows that read or write what this decision governs. Auth, biometric, session, or permission-boundary: AT MINIMUM logout, backgrounding, force-quit/kill. Same test for payments (refund, retry, chargeback), sync (conflict, merge, reconnect), notification preferences (opt-out, quiet hours, reinstall). A decision whose state no other flow touches (a copy change, a colour token) has NO adjacent flows: say so and lock it.
• WHEN a D-N entry proposes complexity the user did not ask for (a new mechanism, state or UI element), it SHALL carry a one-line "Simplest alternative considered:" note naming the simpler option and why it is not enough, before it can be locked.
• Fact vs decision filter: propose a question only for a genuine decision, two or more viable alternatives with different tradeoffs. A pre-existing fact ("stack is Next.js") or a product-spec duplicate ("pricing is $20/month") is not one and belongs in SOURCE_OF_TRUTH.md: if only one answer is reasonable given the codebase, spec and constraints, it is a fact.
• Ask only questions that materially affect correctness, runtime behavior, data integrity, test strategy or rollout safety, and prefer targeted, binary or tightly scoped ones.

DECISIONS.md entries MUST use EARS template (per ADR-0031). One of five canonical forms:

• Ubiquitous: The <system> SHALL <response>
• Event-driven: WHEN <trigger> the <system> SHALL <response>
• State-driven: WHILE <state> the <system> SHALL <response>
• Optional feature: WHERE <feature included> the <system> SHALL <response>
• Unwanted behavior: IF <trigger> THEN the <system> SHALL <response>

Free-form prose for decision rationale is OK; the canonical sentence per entry MUST use SHALL keyword. Banned softeners in canonical sentences: should, may, appropriate, sensible, reasonable.

Required output (interview mode, no LOCK picks in this turn):

1. Decisions already safe to lock (each locked decision phrased in EARS)
2. Remaining ambiguities
3. Targeted questions
4. Why each question matters
5. What implementation/test behavior changes depending on each answer
6. Exact TASK_STATE.md update block (PROPOSED), or explicit TASK_STATE: NO_CHANGE
7. Exact DECISIONS.md update block (PROPOSED) with EARS-shaped entries, or explicit "no DECISIONS.md changes needed"
8. Recommended next step, next command, editor mode, and why
9. What should explicitly not be done yet

Required output (provisional mode, an attended chain on a task branch or standing authorization, with no LOCK picks):

1. Decisions recorded as provisional: one line per P-N with its pick, its Evidence: referent and its Impact: value, high entries first.
2. Deliverables recorded as Not delivered, needs you, or none.
3. Final DECISIONS.md ## Provisional decisions and TASK_STATE.md content after this turn's writes (APPLIED in ### Artifact changes).
4. The recommended next command, which the chain runs now; a provisional record is never a reason to stop.

Required output (persist mode, LOCK picks present in this turn):

1. Acknowledgement listing which question IDs were locked and with which picks (one line per locked decision; no re-rationale).
2. Final DECISIONS.md content after this turn's writes (APPLIED in ### Artifact changes).
3. Final TASK_STATE.md content after this turn's writes (APPLIED in ### Artifact changes), with Canonical decisions reflecting the new locks.
4. Remaining ambiguities ONLY if some open questions were not covered by the lock signal; otherwise the section is omitted.
5. Recommended next step, next command, editor mode, and why (same as interview mode).

Substrate digest fallback

Digest fallback when the canonical helper is unreachable. sha_of_section extracts a section's body with awk and pipes it to shasum -a 256. A run inside a permission boundary that admits shasum but refuses awk and sed cannot invoke that helper, and MUST NOT reimplement it. Assume the helper is unreachable whenever the workflow repository's scripts/ directory is not readable from the working directory.

WHERE the canonical per-section digest helper is unreachable, the write SHALL use a whole-file SHA-256 and SHALL declare the reduced scope:

# One call per file. Read the hash (the first field) out of the output;
# do not pipe it through `cut` or `awk` to trim it.
shasum -a 256 TASK_STATE.md

Then add "sha_scope":"file" to that JSONL line, so a validator distinguishes a file digest from a section digest. The rest of the protocol is unchanged: the transaction header still goes above the section heading, and there is exactly one JSONL line per section write.

Do NOT rebuild the helper by writing each section out as its own file to hash it separately.

At file scope two sections written in the same run share a digest, so the sha_scope field is mandatory, not optional.

Claim grounding (active epistemic humility)

Claim grounding (active epistemic humility). This block governs what you may assert and how you record it. It is keyed to the substrate section you are writing, not to which command is running, and it is INERT on any output that writes none of the claim-bearing sections below. Full contract and rationale: wos/active-epistemic-humility.md.

1. When this applies. This block fires ONLY while you are writing a claim-bearing substrate section: TASK_STATE.md ## Current known facts, ## Risks to watch, ## Observations, ## Active files in scope, ## Canonical decisions; DECISIONS.md ## Locked decisions; IMPLEMENTATION_PLAN.md ## Current gaps, ## Risks and mitigations; IMPACT_ANALYSIS.md; EXTERNAL_RESEARCH.md; REFERENCES.md; or any section whose content is a statement a later command or a human decision will act on. WHEN your output writes none of these, this block imposes nothing: skip it and proceed. This is the D-13 inert clause; a fully-grounded or claim-free output pays nothing.
2. The unit is the load-bearing claim. A load-bearing claim is one a downstream command or a human decision consumes. A passing aside is not load-bearing; a statement someone will act on is. Apply the rest of this block per load-bearing claim, not per sentence.
3. Ground it or abstain. Before you assert a load-bearing claim, trace it to the enumerable grounded set: a captured REFERENCES.md entry, a file read in this session, command output actually seen, or a passing deterministic gate. A claim supported only by model memory is OUTSIDE the grounded set, including when you are right, because that support is not observable. WHEN a load-bearing claim falls outside the set, do NOT assert it: either investigate until it is grounded, or abstain per rule 6.
4. Status records provenance, never confidence. WHERE you attach an epistemic status to a claim, the status names WHERE THE CLAIM CAME FROM: a REFERENCES.md entry title, a file path plus line, or the gate output it came from. It SHALL NOT express a degree of certainty. Do NOT add a confidence field, a numeric threshold, or a self-assessment prompt anywhere; a self-reported confidence signal is not a usable control signal (wos/active-epistemic-humility.md Part 1.3). A status whose referent slot is empty is read as UNKNOWN, not as a weak yes.
5. Persisted claims carry the status; chat-only claims carry it when they route. Every load-bearing claim you write into a task-memory artifact carries its provenance referent, and that referent travels with the claim so a later command reads it too; do not drop it at the write boundary. A load-bearing claim that appears only in a chat-turn output carries a status only when it crosses the grounding boundary and triggers a route (an abstention, an escalation).
6. Abstain as a routed continuation, never a bare refusal. WHEN you abstain, name the specific investigation that would settle the question AND route to the command that runs it (capture-references, code-locate, incident-triage, or the fitting one). A withholding that stalls the work is invalid output. Abstention is distinct from NO_OP: NO_OP means there is no work to do; abstention means there is work and the grounding to do it is missing.
7. An unfired gate is not evidence. The absence of a fired check does not mean grounding existed. Do not read silence here as a pass.

Grounded-residue termination

Grounded-residue termination. This block decides when a question-asking command stops investigating and what it emits when it does. It fires WHEN the command runs inside an automatic chain, meaning the session carries itself into the next Run now: rather than having each step retyped by a person (ADR-0186). An attended chain is one of these: a person is watching, but no step waits for them. WHEN a person is present and answering question by question, in an assisted run or outside an attended chain on a task branch, this block is inert: ask as the command already specifies and skip the rest. Companion to commands/_shared/claim-grounding.md, which governs what you may assert; this one governs when you may stop looking.

1. Build the residue set before the first investigation, not after it. Enumerate the open question as typed load-bearing claims you cannot currently ground, one row per claim, and anchor every row to the ledger row it blocks under ## Requested deliverables in TASK_STATE.md (ADR-0056). The ledger is the bar and it is written down before the loop starts, so the loop cannot lower it. A claim that blocks no ledger row is not residue: record it with capture-observation and move on. WHEN the ledger section is absent, or its only row is the - none named sentinel, anchor to the deliverables named in the task README.md instead and say which of the two you used in the stop record.
2. Type every row by what would settle it. Four kinds, and they are the four members of the grounded set in commands/_shared/claim-grounding.md rule 3 restated as something you can check: REFERENCE (an external source), FILE (a path in the repository or the task folder), COMMAND (a run whose output you can read), GATE (a named deterministic check). The type is what names the investigation, so an untyped row is not yet a residue row.
3. One iteration is one pass: take the row blocking the most ledger rows, run the investigation its type names, then apply the resolution test in rule 5. A row that resolves leaves the set. A row that does not stays, carrying the investigation you already ran, so the next iteration has to pick a different one. Re-running an action that already failed is not an iteration.
4. Exit on exactly one of five labeled conditions, evaluated in this order. The first that holds is the exit, and the rest are not evaluated.
   • RESOLVED. The residue set is empty. Continue the chain.
   • NO_PROGRESS. The last iteration closed zero rows AND issued no investigation action distinct from one already recorded against the set. This is not a stop: continue the chain with the residue recorded as open and Run now: naming the command that would settle the top row.
   • BUDGET. The iteration count reached the per-tier cap in rule 6. An airbag, not a brake: reaching it means RESOLVED and NO_PROGRESS both failed to fire, and that is itself the finding worth reporting. This is not a stop either. The cap is this command's own loop ceiling, and a ceiling warns, records, and continues.
   • ESCALATED. A row still open maps to one of the four reasons a chain stops: an act whose audience is not bounded (ADR-0200); a decision that changes what the product is; a cost or loop ceiling; a check of the work the agent cannot honestly run on itself. The loop ceiling in that list is the one BUDGET already handles, which is why BUDGET is evaluated first. What happens next depends on the run (ADR-0233): attended, on a task branch (spec ### Adaptive handoff), only reason 1 stops, and so under standing authorization (ADR-0244, wos/cross-cutting-workflow-guardrails.md ### Unattended sessions), where the reason-4 row is an unverified: <check> line in TASK_STATE.md ## Open questions / blockers. A reason-2 row routes to decision-interview, whose provisional mode records it under DECISIONS.md ## Provisional decisions; a host command writes that ### P-N itself only when wos/substrate-peers.md names it on that row. A reason-4 row becomes Not verified: <check> in TASK_STATE.md ## Open questions / blockers. Either way the chain continues. A row that would settle only by dropping a named deliverable SHALL NOT be recorded as a provisional de-scope: record Not delivered, needs you: <deliverable>: <what is undecided> there instead. Every other run stops on all four reasons as before, named in Reason:, and an unattended run never self-locks.
   • ENVIRONMENT. A row still open needs something this session cannot supply: a credential, a device, a network, or a system you have no access to. Emit the terminal form (Run now: none with Mode: N/A, and Reason: naming what would unblock it). ENVIRONMENT is not a fifth stop reason. It is the terminal form doing the job it was defined for (ADR-0126).
5. The resolution test, applied to the referent and never to how good the answer feels:
   • REFERENCE resolves when a REFERENCES.md entry for it exists carrying both a URL and an accessed date.
   • FILE resolves when the path plus line number exists on disk and you read it in this session.
   • COMMAND resolves when the command line and its exit code are both present in this session's evidence.
   • GATE resolves when a named gate ran and exited 0. A referent that does not resolve reads as UNKNOWN and its row stays in the set. UNKNOWN is not a weak yes, and a check that never fired is not evidence.
6. The cap is a DECLARED PLACEHOLDER, not a measurement. Until a run replaces it with an observed number, read it off the Escalations: line of TASK_STATE.md ## Recommended pipeline: none 2 iterations, one escalation 3, two or more 5, and 8 where the task declares a strict surface. Corrected 2026-09-16: this rule keyed on Express, Standard and Disciplined, names ADR-0207 retired in the same wave that created this block, so no cap was ever in force for any task opened after it. A blinded review loop ran five iterations before anyone noticed the ceiling it cited did not exist. The escalation count is what replaced the tier label and it carries the same ordering. These come from a doubling heuristic over a guessed typical depth, and no Fhorja run has been measured against them. Do not cite them as measured, and do not tune them by feel. Rule 7 emits the iteration count so a later measurement can set them from observation.
7. Emit one stop record on every exit, whichever one fired. One line each: the exit label; the iteration count reached; the cap in force for this tier; each row still open with its type and the ledger row it blocks; and, per open row, the investigations already attempted. On RESOLVED the last two lines are empty. The record belongs in ### Command transcript, and on ESCALATED or ENVIRONMENT it goes in the visible output as well, because those are the two exits where a reader has to act.
8. What this block does not contain, deliberately. There is no confidence field, no numeric certainty threshold, and no step where you rate how sure you are. Termination is decided by rule 4 over the residue set and by rule 5 over the referents, and both are checks against artifacts outside you. Self-assessed completion is the failure this replaces: among architectures that let the agent judge itself finished, false success accounts for 45 to 48 percent of failures on tau2-bench and 75.8 percent on AppWorld. ADR-0109 is what forbids the confidence field; scripts/check-claim-grounding.sh guards the three source surfaces of that doctrine and this block is not one of them, so here the rule is held by the reader rather than by a script.

Standard output layout (required)

Produce the command output using this structure (English only):

Artifact changes

Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules. Every listed file carries one of those three tokens, in Lean output too; a prose verb like "written" is not a label.

Command transcript

Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).

Handoff

Use the adaptive ending format of WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. Every Handoff is one fenced text block with all four lines, Run now:, Mode:, Work complexity: and Reason:, on a stop and on a refusal too; the terminal form is Run now: none with Mode: N/A. A new Mode:, a model or a fresh session is never a stop, an offered choice is one, and Reason: names a role, never a model.

Definition of done (command output)

• Interview mode (no LOCK picks): Output has three explicit, separated sections: Decisions already safe to lock, Remaining ambiguities, and Targeted questions. Collapsing these into a single prose block is invalid output.
• Provisional mode (attended chain on a task branch or standing authorization, no LOCK picks): every open decision is written as a ### P-N with Evidence:, Impact: and Status: provisional, nothing lands under ## Locked decisions, no named deliverable is de-scoped, and the Handoff continues the chain. Waiting for a LOCK signal in this mode is invalid output, and so is a P-N with no Evidence: line.
• Persist mode (LOCK picks present): Output acknowledges the locked picks by ID + value, lists DECISIONS.md and TASK_STATE.md as APPLIED in ### Artifact changes, and does NOT re-emit the locked questions or their rationale. Re-proposing already-locked decisions is invalid output.
• Each question (interview mode only) states explicitly: why it matters / what is already known / what changes depending on the answer. Question lists without per-question rationale are invalid output.
• No premature canonicalization: in interview mode, DECISIONS.md changes are PROPOSED only. In persist mode, the LOCK signal in the user input authorizes APPLIED. In provisional mode, only ## Provisional decisions is written APPLIED, and it authorizes nothing on its own.
• Clear separation between PROPOSED decisions vs locked decisions in interview mode; persist mode does not mix PROPOSED and APPLIED for the same decision.
• Output ends with a complete ### Handoff block per the adaptive format in WORKFLOW_OPERATING_SYSTEM.md ## Global output contract. A response that ends without a complete Handoff is invalid output.
• Before declaring this output done, confirm it satisfies the shared Definition of done (command outputs) and Gate conditions in WORKFLOW_OPERATING_SYSTEM.md.

Quality bar: Optimize for minimal question count, decision quality, and low ambiguity.