# Canonical ShiftPlanner — measurement record

Date: 2026-11-09
Scope: five divergent copies of the same application found under `spaces/`.
No source file was edited, moved, or deleted. All copies were read only, then
copied into this product tree.

## 1. Measured feature table

Produced by one pass over each file with `wc`, `md5sum` and `grep -o | wc -l`.
Pasted verbatim:

```text
FILE                                            LINES    BYTES                                MD5 OLLAMA AIWIRE  DND EXPORT   CDN  VERSION
spaces/mindtech/shiftplanner.html                2163   110972   5dc466b446ae3c19fffd9cd80f91461a      0      0    0      4     0      2.1
spaces/mindtech/downloads/shiftplanner-v12-ai.html   2789   136212   37207b2227899603f90c42e59ff1875a      4    126    0      4     0      2.1
spaces/mindtech/downloads/shiftplanner.html      1910    98435   a59dedc7608243985aad1d59b880bddd      0      0    0      4     0      2.0
spaces/mindtech/shiftplanner_v1.3.html           1434    70415   daa2a71c94c2d0e5cb585857be5e4721     23     26    4      2     0      1.2
spaces/agando/shiftplanner_v1.3.html             1627    81158   710875062b293ed78698d774696058ec     45     27    4      2     0      1.3
```

Column definitions (all counts are occurrences of the regex, not lines):

| Column | Meaning |
|---|---|
| `LINES` / `BYTES` | `wc -l` / `wc -c` |
| `MD5` | `md5sum` of the source file |
| `OLLAMA` | `grep -o -i 'ollama'` — case-insensitive occurrences |
| `AIWIRE` | `grep -o -i -E '\bai\b\|assistant'` — AI/assistant wiring tokens |
| `DND` | `grep -o -i -E 'draggable\|dragstart'` — real drag-and-drop handlers |
| `EXPORT` | `grep -o -E 'exportCSV\|toBlob\|download'` — export code paths |
| `CDN` | `grep -o -E 'src="http'` — external `<script src="http…>` dependencies |
| `VERSION` | the version literal the file itself declares (`version:'x.y'` in its save/export payload, or `version:'2.x'` in state) |

Notes on measurement honesty:

- `DND` deliberately excludes the bare word `drop`, because `drop` also matches
  `dropdown` and `backdrop-filter`. Under the looser
  `draggable|dragstart|drop` regex the three v2.x files score 1–2, but those
  hits are CSS/UI strings only: `dragstart` is 0 and `draggable` is 0 in all
  three v2.x files, so they have **no** drag-and-drop.
- `CDN` is 0 for every file. A wider sweep for `src="http`, `href="http`,
  `@import` and `fonts.googleapis` also returned nothing, so all five copies
  are self-contained single files with no network dependency at load time.
- `AIWIRE` 126 vs. the "~96" figure in the task context: 96 is reproducible as
  the count of *lines* containing an AI token. The table reports occurrence
  count, which is the more meaningful measure of wiring density.

### Reconciling the "ollama" numbers

The task context states the canonical copy has 0 `ollama` refs and the two
v1.3 copies have 12 and 21. Measured result:

| File | `grep -o 'ollama'` (case-sensitive occurrences) | `grep -c 'ollama'` (matching lines) | `grep -o -i 'ollama'` (any case) |
|---|---|---|---|
| canonical (v12-ai) | **0** | 0 | 4 |
| mindtech v1.3 | 13 | **12** | 23 |
| agando v1.3 | 25 | **21** | 45 |

So the context figures are case-sensitive *line* counts. They are correct as
stated, but they understate the canonical file's Ollama usage: the canonical
contains four capitalised `Ollama` strings and calls the Ollama HTTP API
directly. The claim "the canonical has no Ollama path" would be false. What is
true is that the canonical's AI path is **Ollama-native and endpoint-driven**,
while the v1.3 copies carry a **multi-provider** abstraction (Ollama + OpenAI +
Anthropic) — a different architecture, not a strictly larger one. See §4.

## 2. Decision

**Canonical: `spaces/mindtech/downloads/shiftplanner-v12-ai.html`**
→ published here as `src/index.html` (md5 `37207b2227899603f90c42e59ff1875a`,
136,212 bytes, 2,789 lines).

One-line reason: it is a strict functional superset — it contains every
function present in both other v2.x copies (verified by set difference of
`function` names: **zero** functions unique to either sibling) and adds the
entire AI advisor panel on top, while remaining a single self-contained file
with no CDN dependency.

Evidence for the superset claim — the function-name set of each sibling minus
the canonical's set is empty:

```text
=== spaces/mindtech/shiftplanner.html  (functions NOT in canonical):

=== spaces/mindtech/downloads/shiftplanner.html  (functions NOT in canonical):

```

The two v1.3 copies, by contrast, are a **different, older codebase**
(1,434 / 1,627 lines vs 2,789) with roughly 50 functions absent from the
canonical. They have **no** shift templates (`saveTemplate`/`applyTemplate`
count 0; their only "template" matches are `grid-template-columns` CSS), **no**
day notes (`dayNote` count 0) and **no** CSV export (`exportCSV` count 0).
They do keep automatic generation (`autoGeneratePlan`), fairness options
(`Wochenende fair verteilen` / "distribute weekends fairly") and conflict
handling (`detectAllConflicts`, `clearAllConflicts`), under a different
codebase and with a **German-language UI**. They also add real drag-and-drop
and a multi-provider AI chat. They are therefore not candidates for canonical,
but they are worth archiving as feature donors.

## 3. Archived copies

| Archive file | Origin | Verdict |
|---|---|---|
| `archive/mindtech__shiftplanner.html` | `spaces/mindtech/shiftplanner.html` | **nothing found** — identical feature set to the canonical minus the AI panel; 0 functions and 0 features unique to it. |
| `archive/mindtech__downloads__shiftplanner.html` | `spaces/mindtech/downloads/shiftplanner.html` | **nothing found** — the v2.0 predecessor; a strict subset of the canonical in both functions and UI strings. |
| `archive/mindtech__shiftplanner_v1.3.html` | `spaces/mindtech/shiftplanner_v1.3.html` | **adds real drag-and-drop** and a 3-provider AI layer (see below). |
| `archive/agando__shiftplanner_v1.3.html` | `spaces/agando/shiftplanner_v1.3.html` | **adds real drag-and-drop** plus the most complete Ollama integration of the five (proxy path, model refresh, timeouts, connection test). |

What the two v1.3 copies have that the canonical does not, verified by grep:

- **Drag-and-drop scheduling** — `draggable="true"`, `ondragstart`,
  `ondragover`, `ondrop`, and handlers `onDragStart` / `onDrop` /
  `onDragOver` / `onDragLeave` / `onDragEnd`. The canonical has **0** of these
  (`draggable` 0, `dragstart` 0); it assigns workers via modals instead.
- **Multi-provider AI** — both define `callLLM`, `sendChat`, `processChatMessage`,
  `applyChatAction`, `chatAction`, `mockAgent`, `onProviderChange`; the
  canonical's AI is single-provider (Ollama) and conversational only.
  Endpoints observed:
  - `http://localhost:11434` (Ollama)
  - `https://api.openai.com/v1/chat/completions` + `https://api.openai.com/v1/models`
  - `https://api.anthropic.com/v1/messages` (with `anthropic-version` and
    `anthropic-dangerous-direct-browser-access` headers in the agando copy)
- **Different localStorage schema** — key `schichtplan_v12` with
  `loadState` / `saveState`, versus the canonical's key `shiftplanner_v2` with
  `loadFromStorage` / `saveToStorage`. Any migration between the two needs an
  explicit field mapping: the v1.3 payloads carry
  `employees` / `plan` / `vacations` / `absences` / `customShifts` / `settings`,
  the canonical carries `workers` / `shifts` / `assignments` / `customShifts` /
  `templates` / `settings`.
- **Extra functions unique to the agando copy** — `refreshOllamaModels`,
  `testConnection`, `withTimeout`, `callProvider`, `providerName`,
  `dateKeyOf`, `esc`.
- **Export formats differ in the other direction** — both v1.3 copies export
  **JSON only** (`exportCSV` count 0); the canonical exports **both JSON and
  CSV**. This is a genuine canonical advantage.

## 4. What the canonical's AI path actually calls

The AI panel in `src/index.html` is wired to a locally running Ollama server.
Exact strings, from grep:

```text
921:    let aiEndpoint = 'http://localhost:11434';
922:    let aiModel = localStorage.getItem('sp_ai_model') || 'llama3.2';
2478:        fetch(aiEndpoint + '/api/tags', {
2479:            method: 'GET',
2556:        fetch(aiEndpoint + '/api/generate', {
2486 (testAIConnection)  models = data.models || []
869:  <input ... id="aiEndpointInput" value="http://localhost:11434" placeholder="http://localhost:11434">
870:  <div class="help-text">Ollama server URL. Leave default for local Ollama.</div>
```

- **Base URL:** `http://localhost:11434` (editable in Settings, persisted to
  `localStorage['sp_ai_endpoint']`).
- **Endpoints called:**
  - `GET  {endpoint}/api/tags` — connectivity check / model discovery.
  - `POST {endpoint}/api/generate` — chat completion, body:
    `{ model, prompt, stream: false, options: { temperature: 0.4, num_predict: 1024 } }`,
    reply read from `data.response`.
- **Model identifiers offered in the UI select:** `llama3.2` (default),
  `mistral`, `llama3.1`, `phi3`, `gemma2`, `minimax-m3:cloud`.
  The model choice persists to `localStorage['sp_ai_model']`.
- **No cloud inference endpoint is contacted.** There is no `api.openai.com`,
  `api.anthropic.com` or `generativelanguage` reference anywhere in the
  canonical; the only network calls are to the user-configured Ollama base URL.
- **Graceful degradation:** on failure the panel reports
  `Not connected (start Ollama)` and the rest of the app is unaffected, so the
  scheduler works fully with no server running.

## 5. Provenance

`src/index.html` is a byte-for-byte copy of the canonical source:

```text
37207b2227899603f90c42e59ff1875a  products/shiftplanner/src/index.html
37207b2227899603f90c42e59ff1875a  spaces/mindtech/downloads/shiftplanner-v12-ai.html
```

All four archived files were verified md5-identical to their sources after
copying. The originals under `spaces/` are unchanged.