[loopx-agent]

Goal And Delivered Outcome

The core research-reader scenario previously exercised two Grid arms from one task. It now also reads complete original Ad placement and Schemathesis histories: 255 + 94 events, including a question with 39 saved versions, a rejected local variant, inherited records and a long recording gap. All display pages and exports retain the original events; the existing reader remains the consumer.

Selection combines recency and reading complexity. The inspected newer 26 October Ad/Schema captures contain only two/four events; these record-rich sources exercise the missing cases without inventing data. This follows the reader delivered in #8003, which is already merged. Intended base: main at 2087f540bb07dcc96e063911950ba32f3796f0ee.

Author Declaration

• Written by: model_agent — OpenAI Codex; signed off by LoopX Agent.
• Specification: maintainer request for real benchmark state as the core frontend reference, and task-first frontend delivery.

Criterion	Disposition	Implementation / evidence
Realistic complete populated references	implemented	Two original display-state wrappers, independently pinned JSONL byte sizes and digests; no event sampling or rewriting
Real consumer and long history	implemented	Existing production history HTTP owner; core research-real-benchmark scenario reads all display pages and exports all records
Original versions, negatives and unknowns	implemented	First/last of the 39 Ad versions, full negative text, original identities/measurements and absent relationship edges retained
Desktop, narrow and keyboard reading	implemented	Packaged 1440/390px journey, complete original text and Escape return to the selected record

Self-check: inspected the complete selected display-state files with typed validation and credential/private-path screening, compared their independent original bytes, and ran the checks below. Reused the existing backend, export and comparison owners; no new abstraction was needed.

Scope And Continuation

This is a complete, reversible reference-data/consumer-validation slice. The core now covers three different tasks (Grid, Ad and Schema), plus the existing historical saved Graph analysis. Ad remains runner_invalid; Schema remains cancelled; neither run is score-countable. Source-authored confirmation is not a claim of positive benefit. The Schema recording span includes inherited records and does not measure active solver time.

The broader research journey still needs installed adoption, complete corresponding trajectory publication and qualified contribution/efficiency evidence. This PR does not close that work or add a model/solver invocation. No product UI implementation changed in this follow-up.

Validation

• Tested revision: 5ea23c98bf67bc3e2c1ab709b34551a7e218c31c
• Run state: finished
• Input classes: public_fixture, authorized_private_read_only, synthetic (isolated backend/process boundary only)

Check kind	Result	Evidence / limitation
integration	passed	uv run --extra test python -m pytest tests/presentation/test_explore_history.py tests/presentation/test_research_history_fixture.py -q — 13 checks, including real isolated SQLite/loopback reads and immutable source digests
real_entrypoint	passed	LOOPX_PERSONAL_WORKSPACE_PACKAGED=1 LOOPX_PERSONAL_WORKSPACE_SCENARIO=research-real-benchmark node examples/personal-workspace-browser-smoke.mjs — 11.686s; all four history captures, existing saved Graph judgments/costs, pagination/export, 39-version comparison, keyboard return and source preservation
static	passed	npm run build:chat; touched Python Ruff check; JavaScript syntax; Git whitespace and public-boundary scan
manual	passed	Reviewed real Ad/Schema whole pages and expanded evidence in Chinese at desktop/390px; reviewed the current-revision packaged English captures below. Source readback remained unchanged with zero mutation attempts
static	not_run	New-head GitHub CI and independent review are pending; local passing checks do not establish merge readiness

Coverage: complete selected Explore logs only, with original Goal/run identities. No raw session transcripts, hidden evaluator, execution store or credentials are included. Installed App execution and scientific claims are outside this reference-data slice.

Frontend / Visual Evidence

UI impact: reference content changed; layout and production UI code unchanged. Before: the core reference covered two arms of one task. After: the same reader handles three tasks, real display pagination and 39 original question versions. Main history stays visible first; provenance, complete text and version comparison remain on demand. The Schema gap and missing relationships stay visible.

Current revision 5ea23c98 — packaged UI with production HTTP/SQLite in disposable display-only scaffolding; not an installed/live worker. Reproduce with uv run --extra test python examples/personal-workspace-browser/research-history-backend.py --fixture apps/presentation/dashboard/src/data/fixtures/edgebench-ad-20261023-history.json after building Chat. The remaining views use the corresponding Schema fixture or the core scenario.

Attention review: the first view exposes the recording span/density and complete loaded count; optional originals and 39-version detail do not fill the overview. Narrow expanded titles wrap and Escape returns to the chosen record. No new layout/craft claim is made.

Type of Change

• Documentation update
• Test update

LoopX Area

• Public docs or presentation surface (dashboard reference data)

Shared-authority RFC fixture impact

N/A — no authority, runtime-routing or protocol promotion change.

Boundary Checklist

• Selected display-state fixtures and screenshots screened; no credentials, private machine paths, session traces or hidden verifier output.
• Reuses existing benchmark state and frontend consumers; no duplicated benchmark execution or scoring owner.
• Scoped to real frontend reference data and its current consumer.
• No English/Chinese RFC pair changed.
• Actual first viewports and narrow detail appear directly above.
• The commit includes a DCO sign-off.