Repository layout

The source repository is DarbotLM/raybot. It holds 206 files: 163 PNG, 22 JSON, 7 Markdown, 4 PDF, 4 PowerShell, 2 Python, and one each of CSV, HTML, and SVG.

raybot/
├── README.md                     entry point and honesty notes
├── RUNBOOK.md                    main runbook, Phases 1-8, 570 lines
├── RUNBOOK.legacy.md             preserved narrative, Steps 0-20, 315 lines
├── screenshot-map.md             41 curated plus 35 superseded crosswalk
├── raybot-evalset.csv            6 Ray question-and-answer pairs
├── raybot-logo.svg               the only non-PNG brand asset in the root
├── NN_<phase>_....png            66 numbered originals, at the root
├── screenshots/                  41 curated phase-ordered renames
├── captures-2026-06-17/
│   ├── raw/                      26 unannotated captures
│   ├── annotated/                26 annotated captures
│   └── showcase/                 34 files, Incident Triage build
├── captures-2026-06-18/          70 files, showcase workflow catalog
│   ├── wfNN/                     stub-catalog stills (discarded set)
│   ├── wfNN-real/                connector-backed build stills
│   ├── annotated/                stub-catalog annotations
│   ├── annotated-real/           connector-backed annotations
│   └── node-rects.json           measured node geometry, per workflow
└── coral-schema/                 audit engine and ground truth
    ├── coral-ui-audit.html       self-contained interactive report
    ├── naming-schema.md          the two identifier layers
    ├── automation-candidates.md  25 ranked candidates, 2026-06-18 harvest
    ├── schema-validation.md      coverage narrative
    ├── schema-validation.json    coverage numbers
    ├── api-schema-map.json       57 endpoints, 11 domains, 7 hosts
    ├── gallery.json              95 curated captures
    ├── tab-audit-log.json        live designer-tab state, 2026-06-07
    ├── ui-elements-dump.json     raw per-surface element dump
    ├── elements-slim.json        1,627 rows, abbreviated keys
    ├── annotate_*.py             three Pillow annotators
    ├── surfaces/                 one JSON per audited surface
    └── darbot-validation/        independent validation pass

The screenshot layout has changed twice. Commit 7698a0d moved all 66 numbered originals into screenshots/; a later commit moved them back to the root and left screenshots/ holding only the 41 curated renames. The tree above is the current arrangement. Any document describing screenshots/ as holding everything predates the reversion.

The two runbooks

RUNBOOK.md is the current document: eight phases, a companion-audit index, and a process review. RUNBOOK.legacy.md is the original narrative written as twenty sequential steps. It was preserved rather than deleted because it contains detail the rewrite dropped:

• The original working directory, ...\Clawpilot\raybot-runbook, reached through filesystem-create_directory. That directory no longer exists on any local drive.
• The full Evaluate field catalogue, including the CSV template download link (fwlink 2335991), the six-pair conversation ceiling, and the 1000-character limit on the Reference field.
• A far more verbose account of the Teams failure, running from line 172 to line 237.

Where the two disagree, RUNBOOK.md is authoritative on what the runtime does and RUNBOOK.legacy.md is authoritative on what the original session did.

The capture folders

There are four, and they are not interchangeable.

Folder	Count	Naming	Source
repository root	66	NN_<phase>_<surface>_<detail>.png	Phases 1 through 6 originals
screenshots/	41	NN-<phase>-<detail>.png	Phases 1 through 6, curated renames
captures-2026-06-17/	86	NN-<topic>-<detail>.png	Phase 7 re-audit plus the Incident Triage build
captures-2026-06-18/	70	wfNN[-real]/NN-<state>.png	Phase 8 showcase catalog

The underscore versus hyphen distinction is load-bearing. The screenshot-move commands in session d106b6f4 used exactly that difference to route each filename to the right folder with two separate regular expressions. See Command reference.

captures-2026-06-18/ carries both the discarded stub catalog (wfNN/, annotated/) and the connector-backed rebuild that replaced it (wfNN-real/, annotated-real/). Neither folder name signals which is current. Read the -real variants unless specifically looking for the discarded set - see Phase 8.

The 95 captures surfaced in this wiki's gallery are the curated subset drawn from all four locations.

The audit folder

coral-schema/ is the ground-truth half of the project, and it is self-describing. Start with coral-ui-audit.html, which is a single self-contained file holding a searchable element table, the API map, and a screenshot gallery. Everything else in the folder is the data that report renders.

Read the JSON directly if you are building tooling:

File	Shape
elements-slim.json	Flat array of 1,627 rows with abbreviated keys
schema-validation.json	Totals, per-surface counts, label gaps, shared test ids
api-schema-map.json	domains array of 11, hosts object map of 7
tab-audit-log.json	Per-tab route, semantic route, and live state
surfaces/*.json	Per-surface harness output

The abbreviated keys in elements-slim.json are: s surface, t test id, tag, r role, l label, ls label source, i is-interactive, ir interactive reason, d disabled.

The scripts

Four PowerShell scripts and five Python files sit in coral-schema/.

rename-screenshots.ps1 maps the 66 numbered originals at the repository root to the 41 curated names in screenshots/. It originally hardcoded an absolute path into a OneDrive-synced authoring workspace, making it unrunnable outside one machine; it now derives paths from $PSScriptRoot and treats that workspace as an optional fallback. Eight of its 41 mapped sources - originals 69 through 76 - were never committed, so a default run resolves 33 of 41 and reports the rest by name. -RequireAllSources makes that gap fatal. Re-runs are byte-identical. See Errata.

Three of the Python files are Pillow annotators that turn raw canvas captures into labelled figures. See Annotation pipeline.

Commit lineage

The repository has an unusually short history.

Commit	Author	Change
16e4d1a	darbotlabs	Initial seed: 206 files, 71,796 insertions, no incremental history
7698a0d	dayour	Move all screenshots to screenshots/ and update references: 71 files changed
b9f9595	darbotlabs	Showcase workflow: Incident Triage and Routing, 17 annotated captures, Step 24
9206f4b	darbotlabs	11-workflow Coral showcase catalog plus session-harvest report
56562d8	darbotlabs	Rebuild the 11 showcase workflows with real connectors, agent, and human review
d18c852	dayour	Fix the four documented errata

The seed commit means the build itself produced no version history. Everything that happened before 16e4d1a exists only as prose, captures, and audit JSON.