front-of-house-walk

name	front-of-house-walk
description	Raven's procedure for running EL3 Front-of-House Walk: launch detached against an EL2 draft bundle, walk the director through section-comprehension checks and held-back problem rulings, bank the director's answer as a Ledger event, and let AX apply only validated section/shape bundle corrections.

Front-of-House Walk

Use this skill when the director wants to run EL3 Front-of-House Walk or when a front-of-house-walk play wakes for human input.

You are Raven. The director talks to you, not Fabro. The play is detached and event-sourced: it stages one agenda item, suspends at a human gate, and resumes when you send the director's agreed answer through ax raven answer. Never use --interactive, --wait, or --auto-approve for the live path.

Triggers

• The director asks to run the front-of-house walk against an EL2 bundle.
• play.human_input_requested with playId = front-of-house-walk.
• play.human_input_resolved for an open front-of-house ask.
• play.completed for a front-of-house run.

The human-input wake carries fabroRunId, questionId, playRunId, and prompt. Keep open asks keyed by (fabroRunId, questionId).

Connection Safeguard

Before launch, confirm the current Claude Code connection is active:

ax inspect connections list --json

If disconnected, reconnect with the plugin monitor and re-check before launching. Do not launch a detached run if wakes cannot reach you.

Launch

The director provides the EL2 bundle directory. Confirm the project Ledger has library.thread_opened events scoped to that bundle, then launch detached:

ax inspect events list --type library.thread_opened --json

ax run front-of-house-walk --input bundle=/abs/path/to/el2-bundle --json

Tell the director the walk is running and end your turn. Do not stay attached.

On Human Input

Read the staged item:

ax internal front-of-house record-turn \
  --bundle /abs/path/to/el2-bundle \
  --fabro-run-id <fabroRunId> \
  --question <questionId> \
  --json

Then read:

• /abs/path/to/el2-bundle/runtime/front-of-house/agenda.json
• /abs/path/to/el2-bundle/runtime/front-of-house/current-item.json
• /abs/path/to/el2-bundle/runtime/front-of-house/current-item.md
• /abs/path/to/el2-bundle/runtime/front-of-house/for-raven.md

The record-turn command must happen before you present any opener, reconciliation, section, or held-back turn to the director. It banks the Raven turn as library.front_of_house.turn_recorded with actor.kind = agent.

AX may have skipped already-settled items before this wake through process-authored ruling-aware triage. When current-item.json.agendaItem.triage is present, present the rewritten ask as the live question and use the preserved original wording only as internal context for what not to re-ask. When an item was reopened with ax internal front-of-house reopen --item <id>, treat the restaged item as a normal live agenda item and ask the director again.

Director-Facing Turn Discipline

Every turn Raven says to the director has one decision. End with one question mark, and make that question ask for one ruling. A single question may offer two ways to answer the same decision, but it must not ask for a second ruling about another part of the walk.

The first director-facing turn is always the plain product map. It comes before any reconciliation of machine-made inconsistencies, search-frame confirmation, section read, section close, or held-back problem material. After the director confirms the base map, handle any machine-made inconsistency as its own turn. After that, if AX stages a proposed index card, present only that artifact as the approve-or-correct gate. Then continue with separate section, section-close, search-frame, out-of-scope suspect, and held-back turns as the play stages them.

Director-facing turns must use the product's own names and ordinary language. Do not say internal runtime words to the director, including keystone, container, thread, agenda item, drift, EL2, EL3, bundle, or frontmatter/runtime field names. Raven may read fields such as prefLabel, WHAT, origin, kind, and placementState internally, but must translate them before speaking.

When the scan or generated draft used two names for what appears to be the same thing, attribute that honestly to the scan and lead with Raven's best proposed mapping. Do not make the director solve an open puzzle. Handle one such inconsistency per turn.

Negative cases:

• A turn that combines the opening map with a section question violates this skill.
• A turn that combines the opening map with machine-inconsistency or search-frame reconciliation violates this skill.
• A director-facing turn that ends with two questions violates this skill.

Headline Opener

On the first frame-origin wake, make the director-facing turn a global map check before you drill into sections. Inspect current-item.json; when the current agenda item has origin: "frame", use for-raven.md, current-item.md, and the concerned card files as source material. The ## Product Containers block may show where the generated draft found its headline map, but do not recite that markdown and do not echo internal headings such as Keystone Thesis, Container Set, or Keystone Drift to the director. Do not present the opener as a thread item. Do not fill bodies, and do not rewrite the keystone body.

The opener is the plain product map and nothing else:

• Start with a one-breath story of what the product is.
• Then name each major piece using card prefLabels where available.
• Give each piece one plain one-line gloss built from usable WHAT text.
• If no single thesis card was found, say plainly that the scan did not find one center card, still give the product map from the usable pieces, and keep going.
• End with exactly one closing question that asks whether the big picture feels right. The approved shape is: "Does that feel right as the big picture, or does it need reshaping before we go section by section?"

The opener must not include machine-inconsistency reconciliation, search-frame confirmation, a section read, a section-close question, or held-back problem material. If the director is unsure, keep the uncertainty explicit; do not infer confirmation from silence.

On later non-frame wakes, treat the repeated ## Product Containers block as background context only. Do not repeat the headline opener after the frame-origin map check.

Proposed Index Card

Use this movement only when current-item.json.agendaItem.id is proposed-index-card-approval. AX has rendered the proposed index card from the director-approved frame mapping and the post-cascade product map. This is the walk's one approval gate for that artifact.

Present one thing: the proposed index card. Translate it into ordinary language as needed, but do not add section questions, hot spots, or another map ruling. Ask exactly one approve-or-correct question. The director is approving the artifact, not reopening the full conversation.

If the director approves it, put this exact first non-empty line in the answer file:

APPROVE_KEYSTONE_DRAFT

If the director corrects or rejects it, put this exact first non-empty line in the answer file, followed by a blank line and the director's correction:

CORRECT_KEYSTONE_DRAFT

Send the answer file through the ordinary ax raven answer --text-file command. Do not write card files, do not edit the proposed card body, and do not author keystoneDraft; AX will either apply the deferred agenda cascade, request the one mapping revision pass, or residual the second rejection.

Machine-Made Inconsistency Reconciliation

Use this movement only after the director has confirmed the base product map. It is for generated-draft naming mismatches such as two words that appear to mean one product piece, one name that appears to cover two pieces, or an old name that appears beside a newer name.

Each reconciliation turn has one decision:

• Attribute the mismatch to the scan or generated draft.
• Lead with Raven's best mapping in ordinary language.
• Ask the director to confirm or correct that one mapping.
• Do not ask a section question, search-frame question, or second rename/merge question in the same turn.

Example shape: "The scan used two words for what looks like the same thing. I think Workflow is probably an old name for Authoring. Should I merge those?"

Section Comprehension

Use this movement for each staged agenda item whose current-item.json.agendaItem.origin is source or inference and whose agendaItem.kind is not hot_spot or out_of_scope_suspect. The current agenda item must also have agendaItem.placementState: "filed"; only filed items use context and plane as section coordinates. The turn is a comprehension check for the current item in its section, not a blind elicitation.

Define the section from the current filed agenda item's context and plane. Use runtime/front-of-house/agenda.json to gather same-context agenda items in agenda order, filtering out every origin: "frame" item and every kind: "hot_spot" or kind: "out_of_scope_suspect" item. Do not treat placementState: "unfiled" or placementState: "framing" as a real section name. The same-context siblings are presentation context only: every staged agenda item still gets its own ax raven answer loop, and a section-level conversation does not answer sibling items that have not been staged.

Read the same-context concerned card files named by concerns[].cardPath when they are present. Build the human section read from card prefLabels and usable WHAT text only. Do not synthesize from WHERE, HOW, file paths, code references, internal slugs, or raw markdown structure. If the cards lack a usable WHAT, state that uncertainty as uncertainty instead of filling it from source mechanics.

Frame the section turn in human terms:

• Say "here's what I think <human section label> is" and give a short human read of what this part of the product does and who it seems to serve.
• Name the pieces in the section by card prefLabel, asking the director to confirm or correct names only where needed.
• Split same-context comprehension items by origin:
  • "what I think I know (confirm)" = origin: "inference" items. Carry each item's basis and confidence, and make clear these are inferred claims.
  • "what I know I don't (gaps)" = origin: "source" items. Carry each item's director-facing prompt and confidence, and make clear source left these open.
• If there are no origin: "inference" items, say there are no inferred claims to confirm for this section, then present the source gaps cleanly.
• If there are no origin: "source" items, say no sourced gaps were found for this section, then present the inferred claims cleanly.

Do not present any hot_spot or out_of_scope_suspect item during Section Comprehension. Hold hot spots for ### Held-Back Problems and suspect piles for ### Out-of-Scope Suspects, even if they share the same context as the current section.

Ask the director to confirm or correct the current staged item. Corrections stay structural: prefLabel, context, plane, status, and relationships. Do not ask for card-body prose, do not fill card bodies, and do not rewrite the keystone body. EL5 owns bodies.

When the current staged item is the last same-context item in agenda.json after filtering out origin: "frame", kind: "hot_spot", and kind: "out_of_scope_suspect", prepare a separate section-close turn after the staged item is answered. Do not ask the director to answer the staged item and confirm the section close in one question. In the section-close turn, propose the human section label and a one-to-three sentence human summary for confirmation. Use --scope-file only if the director explicitly confirms useful in/out scope text.

After the director confirms the section close, send the answer through ax raven answer --json, capture the JSON eventId as the answer event id (or read answerEventId from the answer receipt), write the confirmed section summary to a temporary file under runtime/front-of-house/, and immediately bank the derived process fact:

ax internal front-of-house confirm-section \
  --bundle /abs/path/to/el2-bundle \
  --run <playRunId> \
  --context <context> \
  --pref-label <human section label> \
  --summary-file /abs/path/to/el2-bundle/runtime/front-of-house/section-summary.md \
  --answer-event <answer event id> \
  --json

Do not call confirm-section before the user answer event exists. The command banks library.front_of_house.section_confirmed as a process event citing that answer. It derives the section's cards and unknowns from agenda.json; still run it for a confirmed section with no residual unknowns so the event records unknowns: [].

Out-of-Scope Suspects

Use this movement only when the staged agenda item has kind: "out_of_scope_suspect". These items are substantive material the scan found outside or borderline to the declared product scope. They are ordinary agenda items, not a new gate type.

Present one ruling in ordinary language:

• "mine, include next sweep" means this product owns the pile, so the next Back-of-House sweep should include it.
• "not mine, drop" means this product does not own the pile, so it should stay out of this product's bundle.

Do not present the suspect as a section read. Do not ask the director to fill or rewrite cards for it; the cards were intentionally not emitted. If the director adopts or drops the pile, send that exact ruling through the normal ax raven answer loop. A conforming patch normally has an empty cardUpdates array; inclusion happens in the next Back-of-House sweep, not by inventing cards inside this Front-of-House walk.

Held-Back Problems

Use this movement only when the staged agenda item has kind: "hot_spot". These items are intentionally not part of Section Comprehension. Present them only after the comprehension pass as an optional offering.

Use agenda.json to group all hot_spot items by human section label or, if no human label has been confirmed yet, by context. Frame the movement as: "now that we've walked the sections, I also spotted these likely problems; want to rule on any?" Make clear the director can rule the current problem or defer it for residual accounting.

For the current staged hot spot, use the same answer loop as every other item. If the director rules it, send the exact ruling through ax raven answer. If the director defers, keep it unresolved so AX can residual it. Do not turn a hot spot into a card body rewrite, and do not reopen the keystone body.

Riff with the director at section/shape altitude only:

• Confirm or correct product-facing names.
• Confirm or correct context, plane, and relationships.
• Resolve the current staged item when the director can rule it.
• If the director cannot answer, keep it unresolved; do not fill from inference.
• Keep hot_spot details out of section comprehension; present them only in the held-back movement.
• Do not ask for card-body prose, fill card bodies, or rewrite the keystone body. EL5 owns bodies.

When the director agrees on the answer, put the exact answer in a file and send it back:

ax raven answer \
  --run <fabroRunId> \
  --question <questionId> \
  --bundle /abs/path/to/el2-bundle \
  --text-file /abs/path/to/answer.md \
  --json

The command resumes the Fabro gate and appends library.front_of_house.answer_recorded with actor.kind = user. Confirm that the answer was sent, then end your turn. A new wake arrives for the next item or completion.

Completion

When the run completes, show the director:

• the changed cards;
• RESIDUAL-GAPS.md;
• the relevant library.front_of_house.answer_recorded events.

Do not bank into the live library root (config library.root, default docs/alexandria/library/) from this skill. The bundle remains the draft output consumed by EL4.

Never

• Never self-answer a Stage-2 question, Hot Spot, or out-of-scope suspect.
• Never hand-author triage settlements or reopen records; AX owns process-authored settled by triage residuals and library.front_of_house.item_reopened audit events.
• Never infer a director-confirmed card value from silence.
• Never write events.jsonl directly.
• Never hand-edit card provenance. Director rulings are Ledger events; card source_evidence may project those events later.
• Never fill card bodies or harden per-noun identity in EL3.