⚔ The party, in marching order

A run is a fixed relay of 48 turns. Each turn has a contract: what it consumes, what it must produce, and what evidence proves it. The cards below are generated from that contract registry — the same one the pipeline validates against — so this page cannot drift from reality.

In the tavern they are a raid party; on the clock they are a delivery team.

1.Intake Bucketherintake.bucket
The Oracle · Diviner (scout)
Takes in
The player's raw ask + the world's existing milestones, epics and stories
Produces
A bucket ruling: which tier and parent this ask belongs to, and how big it is (trivial/standard/major) — placement only, never the shape of the work
Runs as: a real model call (transcript role "questGiver").
2.UX Researcherintake.ux-research
The Scholar · Sage (scout)
Takes in
A campaign ask naming a real product (before any PRD exists)
Produces
The UX standards research brief the whole design phase is held to: Nielsen's 10 heuristics applied to THIS product, WCAG 2.2 AA floors with the right numbers, a numeric layout/type system, a visual language (palette, fonts, radii), SPA interaction conventions (optimistic mutations, keyboard map, designed error boundaries), and dated comparable-product conventions — machine-checked by arithmetic and truth tables, never vibes
Runs as: a real model call (transcript role "productManager").
3.Product Researcherintake.prd
The Cartographer · Chronicler (scout)
Takes in
A campaign ask naming a real product (e.g. 'a true clone of Todoist, web only')
Produces
A researched product requirements document: the product in one paragraph, its personas, its feature surface with must/should/later priorities, explicit non-goals, and the full PRD as readable markdown — drawn from the model's knowledge of the real product, never invented to look thorough
Runs as: a real model call (transcript role "productManager").
4.PRD Criticintake.critique.prd
The Examiner · Adjudicator (scout)
Takes in
The researched PRD (and the UX standards brief it must honor)
Produces
A citation-bearing ruling BEFORE the breakdown consumes the document: approve (quoting the document verbatim where it satisfies each judged dimension — quotes are machine-verified against the text) or revise (concrete must-fix findings with a place to aim); a rubber stamp is schema-invalid by construction
Runs as: a real model call (transcript role "productManager").
5.Work Breakdownintake.breakdown
The Storyweaver · Loremaster (support)
Takes in
The ask, the Oracle's settled bucket ruling, and the current subtree at that point
Produces
The whole tree below the bucket — epics/stories/tasks sized to the ruling, every story carrying Gherkin, every story's scenarios covered by its tasks, sibling order declared
Runs as: a real model call (transcript role "productManager").
6.Work Amendmentintake.amend
The Mender · Restorer (support)
Takes in
The Council's unresolved findings + an ID/title container index + full content only for entities the findings name
Produces
Explicit ID-addressed update/add/remove/move/setBlockers operations; the engine applies them to the persisted standing tree and revalidates the complete result through BreakdownTree before any write
Runs as: a real model call (transcript role "productManager").
7.Product Reviewcommittee.product
The Patron · Patron (support)
Takes in
Round 1: the tree + the framed bet. Round 2: the mockups + the stories they serve
Produces
A verdict on whether the tree delivers the framed bet and nothing the ask didn't buy — and, in round 2, whether the mockups deliver the stories' intent
Runs as: a real model call (transcript role "committeeProduct").
8.Business Analysis Reviewcommittee.businessAnalyst
The Auditor · Auditor (support)
Takes in
Round 1: every story's Gherkin + every task's acceptance and scenario allocation. Round 2: the mockups' states and copy
Produces
A verdict on whether the stories are INVEST, every scenario is Given/When/Then and observable, every acceptance is testable, and the scenario allocation is sensible
Runs as: a real model call (transcript role "committeeBusinessAnalyst").
9.UX Reviewcommittee.ux
The Pathfinder · Pathfinder (scout)
Takes in
Round 1: the tree's user-visible stories. Round 2: the UX mockups (wireframe/flow)
Produces
A verdict on whether every user-visible story names the surface it changes, and whether the flow is reachable — loading/empty/error states declared, keyboard-reachable, consistent with existing surfaces
Runs as: a real model call (transcript role "committeeUx").
10.Visual Design Reviewcommittee.visualDesign
The Illuminator · Illuminator (support)
Takes in
Round 2 only: the product mockups (visual/layout/copy) + the sibling surfaces they must match
Produces
A verdict on information hierarchy (does the plain answer lead), reuse of the .gw skin and ds-web primitives, contrast/AA and shape-not-color-only, plain-first copy, and jargon left unexplained
Runs as: a real model call (transcript role "committeeVisualDesign").
11.Legal Reviewcommittee.legal
The Advocate · Advocate (support)
Takes in
Round 1: the tree. Round 2: the mockups + the stories it flagged in round 1
Produces
Round 1: the stories that touch user data, third-party content, or claims needing a notice. Round 2: a constraint on each — attributions, notices, ToS-bound API use, disclaimers
Runs as: a real model call (transcript role "committeeLegal").
12.Compliance Reviewcommittee.compliance
The Registrar · Registrar (support)
Takes in
Round 1: the tree. Round 2: the mockups + the stories it flagged in round 1
Produces
Round 1: the stories that retain, export, or log personal data. Round 2: a constraint on each — retention, redaction, audit trail, data egress
Runs as: a real model call (transcript role "committeeCompliance").
13.Security Reviewcommittee.security
The Warden · Warden (tank)
Takes in
Round 1: the tree. Round 2: the mockups + the stories it flagged in round 1
Produces
Round 1: the stories that introduce a trust boundary, a credential, or an external call. Round 2: a constraint on each — secret handling, authz on new surfaces, input validation, injection surface
Runs as: a real model call (transcript role "committeeSecurity").
14.Architecture Reviewcommittee.techArchitect
The Wright · Wright (support)
Takes in
Round 1: the tree + this codebase. Round 2: the mockups + the existing architecture
Produces
A verdict on whether the tree is buildable here, whether any task hides a system-shaped decision belonging in a slice's ArchPlan, and whether the design's split fights the existing architecture
Runs as: a real model call (transcript role "committeeTechArchitect").
15.Bet Framerdecide
The Oracle · Diviner (support)
Takes in
The player's raw ask
Produces
A framed bet: problem, falsifiable hypothesis, success + guardrail metrics
Runs as: a real model call (transcript role "betFramer").
16.Product Managerdecide.stories
The Storyweaver · Loremaster (support)
Takes in
The framed bet
Produces
2–4 Gherkin user stories (Given/When/Then) the slices are designed against
Runs as: a real model call (transcript role "productManager").
17.Slice Plannerdesign
The Cartographer · Strategist (support)
Takes in
The framed bet + the product manager's user stories
Produces
A slice decomposition with acceptance criteria; the first slice is taken
Runs as: a real model call (transcript role "decomposer").
18.UX Designerdesign.ux
The Draughtsman · Illusionist (support)
Takes in
The ask + designed slices (runs only when the work has a user-visible surface)
Produces
A markdown/ASCII wireframe mockup of the surface; skipped honestly otherwise
Runs as: a real model call (transcript role "uxDesigner").
19.Design Reviewerdesign.review
The Curator · Arbiter (scout)
Takes in
The UX mockup
Produces
approve/revise verdict with a note; a revise sends the UX Designer back once
Runs as: a real model call (transcript role "designReviewer").
20.UX Designerdesign.mockup.ux
The Draughtsman · Illusionist (support)
Takes in
One story of the locked-to-be tree, plus the whole subtree for context
Produces
A MockupSpec (V50) — screens, queryable elements, declared loading/empty/error states, and a wireframe block — OR, when the story changes no user-visible surface, a SurfaceDetermination (V69.2) quoting the story verbatim and saying why there is nothing to draw
Runs as: a real model call (transcript role "mockupUx").
21.Product Designerdesign.mockup.product
The Illuminator · Illusionist (support)
Takes in
One user-visible story + the UX designer's wireframe for it
Produces
A MockupSpec (V50) of the product/visual design — layout, states and the verbatim copy each state shows
Runs as: a real model call (transcript role "mockupProduct").
22.Design Criticdesign.critique.wireframe
The Critic · Oracle (support)
Takes in
One rendered wireframe spec + the story it serves
Produces
A WireframeCritique verdict — approve (with ≥3 verified wireframe quotes across ≥3 judged dimensions) or redraw (with must-fix findings naming what and where). Scope: what ASCII can show — states, elements/roles, keyboard, heuristics-in-structure, flow. Pixel floors are the hi-res tier's machine checks.
Runs as: a real model call (transcript role "wireframeCritic").
23.Product Designerdesign.hires
The Illuminator · Illusionist (support)
Takes in
One Critic-approved wireframe + the UX brief's numeric design system
Produces
A HiresMockup — one self-contained HTML file (tokens verified against the brief by arithmetic, no raw colors, a labeled section per declared state, per-region 375px behavior) that then faces the rendered gate and the identity judge
Runs as: a real model call (transcript role "hiresDesigner").
24.Design Directordesign.critique.hires
The Beholder · Oracle (support)
Takes in
The desktop + mobile captures of one hi-res mockup (as images) + the PRD's visual identity
Produces
A HiresCritique — approve (≥3 verified identity citations across ≥3 aspects, each naming what the capture shows) or redraw (must-fix findings naming what reads wrong and where)
Runs as: a real model call (transcript role "identityJudge").
25.Slice Committerslice
The Quartermaster · Logistician (support)
Takes in
The chosen slice
Produces
A committed slice bound to the run
Runs as: the pipeline itself — real commands and state transitions, no model call.
26.Test Writerbuild.red
The Trapwright · Shadowblade (dps)
Takes in
Slice title + acceptance criteria
Produces
One failing test (slice.test.js) proven RED by a real `node --test`
Runs as: a real coding agent (the local codex CLI) editing a real git workspace.
27.Implementerbuild.green
The Forgemaster · Berserker (dps)
Takes in
The failing test
Produces
The minimal implementation (slice.js) proven GREEN by `node --test`
Runs as: a real coding agent (the local codex CLI) editing a real git workspace.
28.Refactorerbuild.refactor
The Polisher · Cleric (healer)
Takes in
The green implementation
Produces
A refactor with the tests still GREEN
Runs as: a real coding agent (the local codex CLI) editing a real git workspace.
29.Integration Gatebuild.integrate
The Gatekeeper · Guardian (tank)
Takes in
The completed TDD loop
Produces
The slice committed for real in the sandbox repo; fan-out to the assurance gates
Runs as: the pipeline itself — real commands and state transitions, no model call.
30.Security Reviewerassure.security
The Warden · Paladin (tank)
Takes in
The real diff
Produces
A security attestation naming the top risk and a verdict
Runs as: a real model call (transcript role "security").
31.Code Reviewerassure.code
The Inspector · Sentinel (tank)
Takes in
The real diff
Produces
A code-review verdict: pass/risk with the single most important finding
Runs as: a real model call (transcript role "codeReviewer").
32.Architecture Reviewerassure.architecture
The Architect · Runesmith (support)
Takes in
The real diff
Produces
An architecture verdict: pass/risk on boundaries, dependencies, and growth
Runs as: a real model call (transcript role "archReviewer").
33.Performance Reviewerassure.performance
The Loadmaster · Siege Engineer (dps)
Takes in
The real diff
Produces
A performance verdict with a budget to hold
Runs as: a real model call (transcript role "performance").
34.Constraint Attesterconstraint.attest
The Oathkeeper · Arbiter (scout)
Takes in
The constraint this seat itself raised in council round 2, verbatim, + the delivered artifacts
Produces
A ruling on its own requirement: satisfied or not, with at least one verbatim citation (file:line, transcript seq, or artifact ref) — an attestation with no citation fails the gate
Runs as: a real model call (transcript role "constraintAttester").
35.Coverage Exemption Reviewercoverage.exemption
The Assessor · Arbiter (support)
Takes in
The QA engineer's exemption request + the file's REAL content (never a description of it)
Produces
granted/denied with a specific reason — a denial leaves the 90% coverage gate standing and the repair loop must write the test
Runs as: a real model call (transcript role "coverageExemptionReviewer").
36.Acceptance Verifierverify.dev
The Adjudicator · Justicar (support)
Takes in
Ask + acceptance criteria + final diff
Produces
met / not_met with cited evidence
Runs as: a real model call (transcript role "verifier").
37.Staging Checkerverify.staging
The Proving Grounds · Ranger (scout)
Takes in
The dev-verified, committed slice
Produces
Tests re-run for real in a clean clone (STG_GREEN, or a FAIL wipe)
Runs as: the pipeline itself — real commands and state transitions, no model call.
38.Canary Monitorverify.canary
The Canary · Pathfinder (scout)
Takes in
The staged slice
Produces
A real stability bake: consecutive green test runs in the clean copy (a flake = a REAL wobble/wipe)
Runs as: the pipeline itself — real commands and state transitions, no model call.
39.Fix & Resumeverify.fixing
The Field Medic · Field Medic (healer)
Takes in
The REAL failure output of the wiped stage (remembered by the machine)
Produces
A real committed repair; PATCH_READY resumes exactly the stage that wiped
Runs as: a real coding agent (the local codex CLI) editing a real git workspace.
40.Browser-Spec Reviewerverify.e2e.review
The Journey Judge · Pathfinder (scout)
Takes in
The exact proposed Playwright spec + every slice title and acceptance promise it must prove
Produces
approve | revise, with one grounded ruling per slice and literal citations from the spec. Missing promised UI, unrelated journeys, and absence-as-proof are not accepted as browser evidence
Runs as: a real model call (transcript role "browserSpecReviewer").
41.Rollout Operatorverify.rollout
The Herald · Bard (support)
Takes in
A healthy canary
Produces
A real smoke load of the delivered module (nothing deploys from a local run — none is claimed)
Runs as: the pipeline itself — real commands and state transitions, no model call.
42.Release Managership
The Shipwright · Captain (support)
Takes in
The rolled-out slice
Produces
RELEASED — handed to the human QA lane; the code lives in the sandbox workspace
Runs as: the pipeline itself — real commands and state transitions, no model call.
43.Retrospective Scribelearn
The Chronicler · Sage (support)
Takes in
Everything the cycle shipped and measured
Produces
Signals re-emitted for the next bet; the cycle completes
Runs as: the pipeline itself — real commands and state transitions, no model call.
44.User-Docs Writerdocs.user
The Scribe · Lorekeeper (support)
Takes in
The stories this campaign SHIPPED + the ship evidence for each (what was really delivered, verbatim)
Produces
User-facing documentation: one section per shipped story, plain language first. A section for a story that did not ship is an invented feature and a machine refuses it before the reviewer reads a word
Runs as: a real model call (transcript role "docsWriter").
45.User-Docs Reviewerdocs.user.review
The Proofreader · Lorekeeper (healer)
Takes in
The drafted user docs + the SAME ship evidence the writer was handed
Produces
approve | revise (one bounded revise, then the phase FAILS). An approval must QUOTE the ship evidence it checked the docs against, verbatim — a machine verifies the quote really occurs in that evidence, so an approval from a reviewer that never opened it cannot be written
Runs as: a real model call (transcript role "docsReviewer").
46.Internal-Docs Writerdocs.internal
The Archivist · Chronologist (support)
Takes in
The delivered record: the locked tree, its constraints, and every slice's outcome
Produces
Architecture notes — each distilling ONE decision and quoting the record it came from verbatim (V36: 'verbatim, not re-imagined') — plus runbook entries for anything operable
Runs as: a real model call (transcript role "internalScribe").
47.Internal-Docs Reviewerdocs.internal.review
The Auditor · Chronologist (healer)
Takes in
The drafted internal docs + the SAME delivered record the archivist was handed
Produces
approve | revise (one bounded revise, then the phase FAILS). Same citation rule as the user-docs gate: an approval quotes the record, and the quote is checked against it
Runs as: a real model call (transcript role "internalDocsReviewer").
48.Retrospective Scribelearn.retro
The Chronicler · Sage (support)
Takes in
Every escape this campaign really took, swept from the record: each constraint discharged by attestation rather than a test, each task that shipped delivering no scenario, each committee note nobody had to act on
Produces
Real follow-up entities, FILED into the hierarchy — each naming the recorded escape it answers. An escape with no follow-up fails the phase; an empty list requires the explicit ruling 'none found' with reasons, because silence and diligence look identical in a record
Runs as: a real model call (transcript role "retrospectiveScribe").

How to read a member

The job title (bold) is what the member actually does — it is the primary label everywhere in the game.
The italic line is pure flavor: the character name, class, and raid role. It never replaces the plain label.
Each member’s input is the previous member’s output — the party marches in the order listed.