core/docs/handoff/WORKBENCH-UI-WAVE-BRIEF-PACK.md

Workbench UI Wave — Brief Pack (W1..W4)

Goal: Stand up the workbench frontend so the math teaching corridor is operable by one operator at human-realistic throughput. Implements the wave shape committed in docs/handoff/WORKBENCH-UI-WAVE-SCOPING.md under the trust boundary pinned by ADR-0173.

Blocker. Every brief here depends on ADR-0173 W0 landing first (PR #394). Do not dispatch W1..W4 until that PR merges.

Bundling rule. feedback-batch-during-research. This is implementation mode, not research — one PR per wave (W4 splits into two). No spec churn in the implementation PRs; if doctrine needs to move, it moves in its own ADR PR first.

Production-line discipline. feedback-production-line-pattern. Dispatch lines are copy-paste-ready; each brief opens with a worktree add line; brief pack is the single source of truth.

---

Dependency DAG

ADR-0173 (PR #394)
  │
  ▼
W1 — Scaffold (one frontend PR; blocks W2/W3/W4)
  │
  ├──────────┬──────────┐
  ▼          ▼          ▼
 W2         W4a        W4b
(read)    (replay)   (eval)
  │
  ▼
W3 (ratification corridor — needs W2's queue + detail)

Parallel-safe pairs: {W2 ‖ W4a ‖ W4b} once W1 lands. W3 is sequential after W2 because the RatificationCommandPanel is wired to the ProposalDetailPanel that W2 ships.

Recommended sequence (single-operator): W1 → W2 → W3 → W4a → W4b. Recommended sequence (parallel, 2+ operators): W1 → (W2 ‖ W4a ‖ W4b) → W3.

---

Operator profiles

Per feedback-parallel-dispatch-pattern:

Wave	Profile	Why
W1	Codex (mechanical)	Vite/TS/Tailwind/shadcn scaffold from ADR-0162 spec; deterministic toolchain bootstrap; minimal judgment.
W2	Sonnet (tight-scope)	Frontend work over a fully-specified API contract; TanStack Query + Zustand; semantic color mapping from ADR-0162 ratified enums.
W3	Opus (load-bearing wrong=0 surface)	Same profile as CC-2; ratification dispatch with hazard-pin discipline, partition tests, case 0050 pin.
W4a	Sonnet (tight-scope)	Replay surfaces over existing API routes.
W4b	Sonnet (tight-scope)	Eval surfaces over existing API routes.

---

Brief W1 — Frontend Scaffold

Operator profile: Codex (mechanical; Vite/TS/Tailwind bootstrap) Branch: feat/workbench-ui-w1-scaffold Base: origin/main (post-ADR-0173 merge) Style: Frontend bootstrap. No app routes yet. Tokens + shell + palette + status footer only.

Dispatch

git fetch origin main && \
  git worktree add /tmp/wt-wb-ui-w1 origin/main && \
  cd /tmp/wt-wb-ui-w1 && \
  git checkout -b feat/workbench-ui-w1-scaffold

Reads required FIRST

• docs/decisions/ADR-0160-core-workbench-v1.md §"Architecture choice" + §"Trust boundary"
• docs/decisions/ADR-0162-workbench-design-system.md §"Token namespace", §"Typography", §"Color semantics", §"Motion", §"Keyboard contract", §"Layout shell", §"Component map — v1 must-ship", §"The no-go list", §"Implementation plan — Branch 1"
• docs/decisions/ADR-0173-workbench-ratification-trust-boundary.md §"Keyboard contract", §"What the workbench MUST NOT do"
• docs/handoff/WORKBENCH-UI-WAVE-SCOPING.md §"Guardrails"

Outcome

workbench-ui/ directory at repo root, fully bootstrapped per ADR-0162 §"Implementation plan — Branch 1":

1. workbench-ui/package.json — pinned deps:
   • React 18.x, ReactDOM 18.x
   • Vite 5.x
   • TypeScript 5.x
   • Tailwind 3.x
   • shadcn-style primitives (Radix UI under shadcn)
   • TanStack Query 5.x
   • Zustand 4.x (state)
   • lucide-react for icons (bundled, no CDN)
2. workbench-ui/vite.config.ts — dev server on 127.0.0.1, no network proxies to remote hosts, build output to dist/.
3. workbench-ui/tsconfig.json — strict TS.
4. workbench-ui/tailwind.config.js + postcss.config.js.
5. workbench-ui/src/design/tokens/tokens.css — every token from ADR-0162 §1–§4 (color, typography, motion, spacing).
6. workbench-ui/src/design/badges/ — four badge primitives:
   • EpistemicStateBadge (15 values from core/epistemic_state.py)
   • NormativeClearanceBadge (4 values)
   • ReviewStateBadge (4 values from teaching/proposals.py)
   • GroundingSourceBadge (6 values) Each binds color + glyph + text label per ADR-0162 §3. No color-only encoding.
7. workbench-ui/src/shell/ —
   • WorkbenchShell.tsx (the outer layout)
   • TopBar.tsx
   • LeftNav.tsx
   • StatusFooter.tsx
   • CommandPalette.tsx (focusable via : per ADR-0173)
8. workbench-ui/src/keyboard/ —
   • KeyboardHelpOverlay.tsx populated from ADR-0173 §"Keyboard contract"
   • global useKeyboard() hook
   • ? opens the help overlay
   • Esc dismisses
   • No Cmd/Ctrl chords (ADR-0173 constraint)
9. workbench-ui/src/api/client.ts — fetch wrapper for the existing backend (http://127.0.0.1:<port>); typed against workbench/schemas.py shapes via hand-written TS types in src/api/types.ts.
10. workbench-ui/dist/ — added to root .gitignore.
11. CI job — .github/workflows/workbench-ui-build.yml:
    • triggers on PR touching workbench-ui/**
    • runs npm ci && npm run build
    • does not commit dist/
    • fails the PR if build fails

Hard requirements

• Tokens snapshot test. A snapshot test pins the computed token values from tokens.css so future PRs cannot drift them silently.
• Fonts bundled locally per ADR-0173 Q4 — no CDN font fetch, no @import url(...google...). License compliance is W1's responsibility.
• Icons bundled locally — lucide-react is installed as a dependency; no remote icon fetches.
• No app routes yet. W1 ships the shell; /proposals, /evals, /trace routes are W2/W3/W4's job.
• Backend binds 127.0.0.1. UI verified to refuse cross-origin requests; CORS configured restrictively in workbench/server.py if not already (read current state first; do not relax).
• Keyboard help overlay populated from ADR-0173. Verbatim bindings table.
• Accessibility: no icon-only buttons without aria-label; all badges carry text + shape, not color alone (ADR-0162 no-go list).
• No-go list cross-check. Reviewer must verify: no chat-clone bubbles, no "AI thinking" affordances, no glassmorphism, no purple-neon-cyberpunk, no graph-builder canvas, no dashboard splash metrics.

Tests

• workbench-ui/src/**/__tests__/*.test.tsx using Vitest + React Testing Library
• shell.test.tsx — WorkbenchShell renders with TopBar, LeftNav, StatusFooter
• command_palette.test.tsx — : focuses palette, Esc dismisses
• keyboard_help.test.tsx — ? opens overlay with ADR-0173 bindings
• badges.test.tsx — all four badge primitives render correct color + glyph + label for every enum value
• tokens.snapshot.test.ts — pins computed token bytes

Deliverables

• workbench-ui/ directory complete per outcome list
• root .gitignore updated (one line: workbench-ui/dist/)
• .github/workflows/workbench-ui-build.yml
• Tests green; CI workbench-ui-build job green
• No backend changes
• No changes to core/, chat/, teaching/, field/, generate/, algebra/, language_packs/

Forbidden

• Adding app routes (W2/W3/W4 territory)
• Adding ratification UI (W3 territory)
• Adding any ProposalQueue/Card/Detail components (W2 territory)
• Adding any TraceDrawer/Replay/Eval components (W4 territory)
• Touching backend code (workbench/api.py, workbench/readers.py, etc.)
• Committing dist/
• Remote CDN dependencies (fonts, icons, runtime)
• Cmd/Ctrl keyboard chords
• Color-only state encoding
• Auto-dismiss toasts for audit events

---

Brief W2 — Read Surfaces (Proposal Queue)

Operator profile: Sonnet (tight-scope frontend over fully-specified API) Branch: feat/workbench-ui-w2-proposal-queue Base: origin/main (post-W1 merge) Style: Frontend only. Read-only against existing routes.

Dispatch

git fetch origin main && \
  git worktree add /tmp/wt-wb-ui-w2 origin/main && \
  cd /tmp/wt-wb-ui-w2 && \
  git checkout -b feat/workbench-ui-w2-proposal-queue

Reads required FIRST

• workbench/api.py — routes /proposals, /proposals/{id}, /math-proposals, /math-proposals/{id} (exact payload shapes)
• workbench/schemas.py — ProposalSummary, ProposalDetail, MathProposalSummary, MathProposalDetail, MathReasoningStep
• docs/decisions/ADR-0162-workbench-design-system.md §"Component map", §"StableJsonViewer", §"Empty/error/loading state contract"
• docs/decisions/ADR-0173-workbench-ratification-trust-boundary.md §"Keyboard contract" (queue + detail bindings)
• docs/decisions/ADR-0167-audit-as-teaching-evidence.md §"Partition guarantees"

Outcome

Six components + two routes + two stores.

1. src/proposals/ProposalQueue.tsx — paginated list of pending proposals. Renders math + cognition in separate tabs (partition discipline). Default sort: oldest pending first. Filter chips: state, source kind, replay status.
2. src/proposals/ProposalCard.tsx — one-row representation: proposal_id (truncated, CopyableHash), EpistemicStateBadge, ReviewStateBadge, age_proposals, replay_equivalent icon, handler_name (if math).
3. src/proposals/ProposalDetailPanel.tsx — full detail view with StableJsonViewer over evidence, replay_evidence, proposed_chain / proposed_change_payload, and reasoning_trace_steps rendered as a step list (not a graph — ADR-0162 no graph-builder).
4. src/json/StableJsonViewer.tsx — canonical JSON renderer; sorted keys; no re-ordering; no syntax-highlight that mutates bytes; copy-to-clipboard per top-level node.
5. src/json/CopyableHash.tsx — truncated SHA with full-on-hover
   • copy button.
6. src/json/ArtifactLink.tsx — typed link to /artifacts/{id} route (lazy-loads the artifact).
7. src/routes/proposals.tsx — /proposals (cognition) + /math-proposals (math) routes; URL state binds to selected proposal id.
8. src/state/proposals.ts — Zustand store: queue cache, selected id, filter chips. TanStack Query owns the fetch.

Keyboard wires (per ADR-0173)

• g p → navigate to /proposals (or /math-proposals based on last-viewed tab)
• j / ↓ → focus next proposal in queue
• k / ↑ → focus previous
• Enter → open detail panel for focused proposal
• Esc → close detail, return to queue
• / → focus filter input

Hard requirements

• Partition discipline. Cognition and math queues are separate routes, separate Zustand slices, separate component trees. Math ProposalCard cannot render a cognition payload and vice versa. Verified by component test.
• StableJsonViewer does not alter bytes. A round-trip (json.stringify(parsed)) of the rendered payload equals the fetched payload byte-for-byte (modulo whitespace which is rendered but not stored).
• No ratification UI. No r keybind. No "ratify" button. No "accept" or "reject" affordance. That is W3's surface.
• Empty/error/loading per ADR-0162 §6. Empty state names what's missing; error state names the failure; loading state is a skeleton, not a spinner-soup splash.
• Replay-state visualization. replay_equivalent: true|false|null renders three distinct states (passed / failed / not-yet-replayed) — not collapsed to a two-state truthy/falsy icon. Per scoping brief pain point #6.
• suggested_ratify_cli rendered (read-only). Operator can copy via y keybind even though r (ratify) is not yet wired. This unblocks audit-fallback workflows during W2 before W3 ships.

Tests

• src/proposals/__tests__/queue.test.tsx — renders 0, 1, many proposals; sort/filter behavior
• src/proposals/__tests__/card.test.tsx — one card per state permutation
• src/proposals/__tests__/detail.test.tsx — full detail render; StableJsonViewer byte-stability
• src/proposals/__tests__/partition.test.tsx — math card refuses cognition payload; vice versa
• src/proposals/__tests__/keyboard.test.tsx — j/k/Enter/Esc// behavior per ADR-0173 contract
• src/proposals/__tests__/replay_state.test.tsx — three distinct states render distinctly (passed / failed / not-yet-replayed)
• src/json/__tests__/stable_json.test.tsx — round-trip byte-equality on a sample MathProposalDetail payload
• e2e — e2e/proposal_queue.spec.ts — Playwright; start backend with fixture proposals, open UI, verify the queue populates and the detail panel renders for one math + one cognition proposal

Deliverables

• Components 1–8 above
• Routes + store wired
• Tests + e2e green
• Workbench backend untouched (read-only against existing routes)

Forbidden

• Adding ratification UI (W3)
• Adding trace / replay / eval surfaces (W4)
• Backend route changes
• Mutating any existing component from W1
• Mixing cognition and math proposals in one queue/component

---

Brief W3 — Ratification Corridor (the throughput multiplier)

Operator profile: Opus (load-bearing wrong=0 surface; ratification handler discipline; mirror CC-2 rigor) Branch: feat/workbench-ui-w3-ratification-corridor Base: origin/main (post-W2 merge, and post PR #393 / CC-2 merge for CompositionClaim handler availability) Style: Frontend + thin backend dispatch test. Load-bearing.

Dispatch

git fetch origin main && \
  git worktree add /tmp/wt-wb-ui-w3 origin/main && \
  cd /tmp/wt-wb-ui-w3 && \
  git checkout -b feat/workbench-ui-w3-ratification-corridor

Reads required FIRST

• docs/decisions/ADR-0173-workbench-ratification-trust-boundary.md (entire ADR)
• docs/decisions/ADR-0167-audit-as-teaching-evidence.md §"Partition guarantees"
• docs/decisions/ADR-0168-frameclaim-ratification.md §"Acceptance gates"
• docs/decisions/ADR-0169-compositionclaim-ratification.md §"Acceptance gates"
• teaching/math_lexical_ratification.py — apply_lexical_claim(), exception types
• teaching/math_frame_ratification.py — apply_frame_claim(), exception types
• teaching/math_composition_ratification.py — apply_composition_claim(), exception types
• workbench/api.py — POST /math-proposals/{id}/ratify route + MathRatifyResult shape
• tests/test_math_frame_ratification.py::test_case_0050_hazard_pin (the template)
• tests/test_math_composition_ratification.py::test_case_0050_hazard_pin (same template)

Outcome

One component, one keyboard surface, one e2e per handler.

1. src/proposals/RatificationCommandPanel.tsx — bottom-sheet panel attached to the focused proposal in ProposalDetailPanel.
   • Visible only when a proposal is focused AND in pending state AND replay_equivalent == true AND handler_name is in the admitted set (Lexical / Frame / Composition).
   • Disabled with a named status message when any precondition fails (per ADR-0173 §"Keyboard contract" — no silent failures).
   • Renders: handler name, claim signature digest, target JSONL artifact path, "ratify (r)" / "reject with note (x)" / "defer (d)" / "copy CLI (y)" affordances.
2. src/api/ratify.ts — typed wrapper around POST /math-proposals/{id}/ratify. On 200, returns MathRatifyResult. On 4xx, surfaces the structured error verbatim into the status footer (do not translate exception messages — operator audit fallback depends on raw text).
3. src/state/ratify.ts — Zustand slice for post-ratify auto-advance (focus next pending after a successful ratify).
4. Operator-action telemetry wire in workbench/api.py — emit operator_ratify / operator_reject / operator_defer events to the existing chat/telemetry.py sink per ADR-0173 Q3. Event schema:

   {"event": "operator_ratify", "proposal_id": "...", "handler": "...", "outcome": "applied|rejected_precondition", "ratifier_kind": "workbench", ...}
   

5. Ratification record extension — ratifier_kind: "workbench" is written to the JSONL artifact's provenance record. Audit forensic field only; not a permission gate. Verified by test that strips ratifier_kind and confirms replay still works (per ADR-0161 §5 "Replay invariants").

Keyboard wires (per ADR-0173)

• r → ratify focused proposal (only when preconditions hold; otherwise no-op + status footer message)
• x → reject with note (opens a single-line text input; Enter commits, Esc cancels)
• d → defer (proposal stays pending; operator-defer telemetry emits)
• y → copy suggested_ratify_cli to clipboard
• Post-ratify: focus auto-advances to next pending + replay_equivalent==true proposal

Hard requirements

These mirror ADR-0173 §"Acceptance gates for W1..W4 → W3" verbatim.

• Same Python entrypoint. Every r keypress executes the same apply_*_claim() function as the corresponding CLI invocation. Verified by parametrized test (one per handler) that:
  1. constructs a fixture proposal
  2. ratifies via the UI (HTTP POST through the API)
  3. ratifies an equivalent fixture via the CLI
  4. diffs the resulting JSONL artifact rows
  5. asserts byte-equal except for ratifier_kind
• Case 0050 hazard pin holds end-to-end. Ratifying any admitted handler via the UI must not cause case 0050 to admit. Verified by e2e: ratify a synthetic CompositionClaim under each safe category, then run core eval gsm8k_math (via the existing /evals/run route) and assert case 0050 stays refused.
• Exception surface verbatim. AlreadyRatified, WrongClaimSubType, WrongCompositionCategory, EvidenceTampered, replay-regression errors — all surface verbatim into the status footer. No translation, no smoothing.
• r is a no-op on any proposal not in pending AND replay_equivalent == true AND admitted handler. The no-op emits a status-footer message naming the failing precondition.
• Partition. Math ratifications cannot mutate cognition artifacts and vice versa. Verified by test that attempts a cross-domain ratification and asserts the API returns 400 with a partition-violation reason.
• No auto-ratify. Replay-passed proposals remain pending until an explicit r keypress on a focused, expanded detail. Verified by test: poll the queue with replay-passed fixtures; assert no state transitions until a keypress arrives.
• No batch ratification. Each r press ratifies exactly one focused proposal. Multi-select UI is not introduced.
• No CORS relaxation. Backend continues to bind 127.0.0.1 only; cross-origin requests refused. Verified by test.
• Operator-action telemetry redact-by-default. Events do not include surface text or token content. Verified by test: emit, parse, assert no proposal evidence text leaks.

Tests

Mirror tests/test_math_composition_ratification.py rigor.

• tests/test_workbench_ratify_lexical.py — end-to-end UI ratify → JSONL artifact byte-diff vs CLI ratify
• tests/test_workbench_ratify_frame.py — same shape for FrameClaim
• tests/test_workbench_ratify_composition.py — same shape for CompositionClaim
• tests/test_workbench_ratify_case_0050_hazard_pin.py — mandatory
• tests/test_workbench_ratify_partition.py — math/cognition isolation
• tests/test_workbench_ratify_no_auto.py — auto-ratify forbidden
• tests/test_workbench_ratify_exception_surface.py — exceptions surface verbatim
• tests/test_workbench_operator_telemetry.py — events emit; no content leak
• tests/test_workbench_ratify_idempotent.py — duplicate ratify via UI returns AlreadyRatified
• src/proposals/__tests__/ratification_panel.test.tsx — frontend unit tests for visibility, keyboard, no-op behavior
• e2e/ratification_corridor.spec.ts — Playwright end-to-end for all three handlers

Deliverables

• RatificationCommandPanel.tsx, src/api/ratify.ts, src/state/ratify.ts
• Backend: operator-action telemetry wire in workbench/api.py
• workbench/readers.py updates if needed for ratifier_kind surfacing (read current state first; minimal diff)
• All tests above, green
• core test --suite teaching -q green
• core test --suite runtime -q green
• core eval gsm8k_math green
• Case 0050 remains refused

Forbidden

• Admitting a handler not in {Lexical, Frame, Composition}
• Adding auto-ratify of any flavor
• Adding batch / multi-select ratification
• Bypassing handler preconditions in any code path
• Adding remote operator auth, login, or token-bearer
• Relaxing CORS on the backend
• Adding Cmd/Ctrl keyboard chords
• Mutating engine_state/* outside the existing checkpoint path
• Treating ratifier_kind as a permission gate
• Adding a parallel telemetry log

---

Brief W4a — Replay Surfaces

Operator profile: Sonnet (tight-scope frontend over existing routes) Branch: feat/workbench-ui-w4a-replay-surfaces Base: origin/main (post-W1; parallel-safe with W2/W3/W4b) Style: Frontend only. Read-only.

Dispatch

git fetch origin main && \
  git worktree add /tmp/wt-wb-ui-w4a origin/main && \
  cd /tmp/wt-wb-ui-w4a && \
  git checkout -b feat/workbench-ui-w4a-replay-surfaces

Reads required FIRST

• workbench/api.py — /trace/..., /replay/... routes
• workbench/schemas.py — ReplayComparison, ReplayDivergence, ReplayDivergenceSeverity, ReplayStatus
• docs/decisions/ADR-0160-core-workbench-v1.md §"Trace Drawer", §"Replay Theater"
• docs/decisions/ADR-0162-workbench-design-system.md §"Component map"

Outcome

Four components + one route.

1. src/trace/TraceDrawer.tsx — slide-in drawer showing the trace for a focused turn / proposal / ratification. Renders trace hash + step list (not graph). Opens via g t keybind.
2. src/replay/ReplayTheater.tsx — full-width panel hosting replay comparison + diff.
3. src/replay/ReplayComparisonPanel.tsx — side-by-side original-vs-replay hash + equivalent boolean + divergence count.
4. src/replay/ReplayDiffViewer.tsx — divergence list with severity coloring (info / warning / failure per ADR-0162); uses StableJsonDiffViewer from new component below.
5. src/json/StableJsonDiffViewer.tsx — canonical JSON diff; structural-aware (key/value/array-index), not line-based.
6. src/routes/trace.tsx — /trace/:id route.
7. src/routes/replay.tsx — /replay/:id route.

Hard requirements

• Read-only. No replay-mutation UI. No "re-run replay" button in v1 (the backend route is read; replay is run server-side by separate mechanisms).
• Diff is structural. StableJsonDiffViewer operates on parsed JSON, not text lines. Byte-equal payloads produce zero divergences regardless of formatting.
• Severity rendering. info / warning / failure render distinctly without color-only encoding (icon + text + color).

Tests

• src/trace/__tests__/trace_drawer.test.tsx
• src/replay/__tests__/replay_theater.test.tsx
• src/replay/__tests__/replay_diff.test.tsx
• src/json/__tests__/stable_json_diff.test.tsx — round-trip; byte-equal payloads → zero divergences
• e2e/replay_theater.spec.ts

Forbidden

• Adding replay-mutation routes or buttons
• Line-based diff (must be structural)
• Color-only severity encoding

---

Brief W4b — Eval Surfaces

Operator profile: Sonnet (tight-scope frontend over existing routes) Branch: feat/workbench-ui-w4b-eval-surfaces Base: origin/main (post-W1; parallel-safe with W2/W3/W4a) Style: Frontend only. Drives existing read-only eval routes.

Dispatch

git fetch origin main && \
  git worktree add /tmp/wt-wb-ui-w4b origin/main && \
  cd /tmp/wt-wb-ui-w4b && \
  git checkout -b feat/workbench-ui-w4b-eval-surfaces

Reads required FIRST

• workbench/api.py — /evals, /evals/{id}, POST /evals/run routes
• workbench/schemas.py — EvalLaneSummary, EvalRunResult
• docs/decisions/ADR-0160-core-workbench-v1.md §"Eval Center"
• docs/decisions/ADR-0166-measurement-capability-sequencing.md (do not introduce new eval lanes)

Outcome

Three components + one route.

1. src/evals/EvalCenter.tsx — top-level eval page.
2. src/evals/EvalLaneList.tsx — list of available lanes from GET /evals. Each row: lane name, latest run status, version, read-only badge, description. Clicking a row opens detail.
3. src/evals/EvalFailureViewer.tsx — for a failed eval run, list failing cases with their MathReaderRefusalEvidence (or cognition-equivalent) + ArtifactLink to the run JSON.
4. src/routes/evals.tsx — /evals route.

Hard requirements

• No new eval lanes. ADR-0166 holds. The UI runs existing lanes only.
• POST /evals/run is the only mutation surface. It is read-only in effect (runs an eval; does not mutate corpus/packs). Other than that, every interaction is a GET.
• Failure surfacing tied to evidence. A failing case must link to the audit row that produced it (via ArtifactLink).
• No vanity dashboard splash. ADR-0162 no-go list: no "total turns" / "active sessions" / pulsing-orb metrics on the eval landing page.

Tests

• src/evals/__tests__/eval_center.test.tsx
• src/evals/__tests__/lane_list.test.tsx
• src/evals/__tests__/failure_viewer.test.tsx
• e2e/eval_center.spec.ts — list lanes, open one, run via the read-only route, view results

Forbidden

• Introducing new eval lanes (ADR-0166)
• Adding splash metrics / "AI cognition theater" panels
• Mutating corpus/packs through the eval surface

---

Anti-regression invariants (all four briefs)

• wrong == 0 on core eval gsm8k_math — preserved
• ADR-0166 — no new eval lanes
• ADR-0057 replay-equivalence — inherited
• ADR-0167 partition — math/cognition isolation in UI as on server
• Pinned-lane SHAs — should not require updates
• engine_state/* — never committed
• Case 0050 hazard — pinned in W3's test suite (mandatory)
• Backend bound to 127.0.0.1 — never relaxed
• No remote runtime dependencies — fonts, icons, telemetry all local

---

Memory pointers

• feedback-batch-during-research — implementation mode; one PR per wave
• feedback-no-self-dispatch-of-subagents — Shay dispatches operators
• feedback-production-line-pattern — this brief pack pattern
• feedback-parallel-dispatch-pattern — operator profile mapping
• feedback-parallel-agent-worktrees — fresh worktree per brief
• feedback-wrong-zero-hazard-case-0050 — W3 mandatory pin
• feedback-cleanup-as-you-find — apply if W1 finds dead workbench backend stubs
• user-circumstances — operators on library wifi; local-first matters
• milestone-adr-0172-tier1-2026-05-27 — context for throughput-vs-capability framing
• adr-0167-audit-as-evidence-wave — parent corridor

---

What ships when all PRs land

• W1. Operator can run core workbench serve and open a bootstrapped UI shell with command palette, keyboard help, and state badges. Nothing actionable yet.
• W2. Operator can read every pending math + cognition proposal, inspect evidence/replay/reasoning trace, and copy the suggested ratify CLI for shell fallback.
• W3. Operator can ratify a Lexical / Frame / Composition proposal via r keypress in under 10 seconds per row for a row whose evidence they understand. Case 0050 remains refused. The compounding loop runs at human-realistic throughput.
• W4a. Operator can inspect any trace or replay artifact inline; replay divergences are structurally diffed.
• W4b. Operator can run and inspect eval lanes inline; failing cases link back to their audit-row evidence.

Together: the math teaching corridor becomes operable by one operator without context-switching to the shell. The five operator pain points from the scoping brief are individually retired:

Pain point	Retired by
HITL burden high	W3 keyboard corridor
No keyboard corridor for ratification	W3 r/x/d/y bindings
Evidence is text-block JSON	W2 StableJsonViewer + structured detail
Reasoning traces are flat	W2 step-list rendering of MathReasoningStep
Replay equivalence is opaque hex	W2 three-state replay visualization; W4a ReplayComparisonPanel
No "stuck queue" diagnosis	W2 filter chips + replay-state visualization

---

Copy-paste dispatch lines (when ready)

After ADR-0173 (PR #394) lands, dispatch in order:

# W1 — Codex (mechanical scaffold)
Read docs/handoff/WORKBENCH-UI-WAVE-BRIEF-PACK.md §"Brief W1".
git fetch origin main && git worktree add /tmp/wt-wb-ui-w1 origin/main && cd /tmp/wt-wb-ui-w1 && git checkout -b feat/workbench-ui-w1-scaffold

# (wait for W1 merge)

# W2 — Sonnet (read surfaces)
Read docs/handoff/WORKBENCH-UI-WAVE-BRIEF-PACK.md §"Brief W2".
git fetch origin main && git worktree add /tmp/wt-wb-ui-w2 origin/main && cd /tmp/wt-wb-ui-w2 && git checkout -b feat/workbench-ui-w2-proposal-queue

# W4a — Sonnet (parallel with W2)
Read docs/handoff/WORKBENCH-UI-WAVE-BRIEF-PACK.md §"Brief W4a".
git fetch origin main && git worktree add /tmp/wt-wb-ui-w4a origin/main && cd /tmp/wt-wb-ui-w4a && git checkout -b feat/workbench-ui-w4a-replay-surfaces

# W4b — Sonnet (parallel with W2/W4a)
Read docs/handoff/WORKBENCH-UI-WAVE-BRIEF-PACK.md §"Brief W4b".
git fetch origin main && git worktree add /tmp/wt-wb-ui-w4b origin/main && cd /tmp/wt-wb-ui-w4b && git checkout -b feat/workbench-ui-w4b-eval-surfaces

# (wait for W2 merge AND #393 CompositionClaim merge)

# W3 — Opus (ratification corridor; load-bearing wrong=0)
Read docs/handoff/WORKBENCH-UI-WAVE-BRIEF-PACK.md §"Brief W3".
git fetch origin main && git worktree add /tmp/wt-wb-ui-w3 origin/main && cd /tmp/wt-wb-ui-w3 && git checkout -b feat/workbench-ui-w3-ratification-corridor

---

Status zsh trap (operator convenience)

For local dispatch monitoring (mirror the pattern used during the ADR-0172 Tier 1 wave):

status() {
  for pr in 394 393; do
    gh pr view "$pr" --json number,title,state,mergeStateStatus,statusCheckRollup \
      | python3 -c "
import json,sys
d=json.load(sys.stdin)
checks=d.get('statusCheckRollup') or []
from collections import Counter
states=Counter(c.get('conclusion') or c.get('status') for c in checks)
print(f\"#{d['number']} {d['state']:<6} {d['mergeStateStatus']:<16} {dict(states)}  — {d['title'][:60]}\")
"
  done
}

Run status to get a single-line summary per gating PR.