esceptico/ntrp/wiki-automation

Wiki automation

Create a normal automation whose results live under automations/. The automation gets its own dedicated channel automatically. Do not attach the current chat; loop_create is the separate current-chat workflow.

Choose the page layout

• Default: automations/<stable-slug>.md.
• Repeated dated output, only when history is needed: automations/<stable-slug>/<YYYY-MM-DD>.md.
• Use another descendant of automations/ only when the user chooses it.

These are common wiki pages. Do not invent a page type or producer metadata.

Create directories deliberately

Every directory has a semantic README.md, not a generated file list. The wiki creates any missing ancestor README atomically with the first page. Automations must read a directory README before processing files in that directory. A single file directly in the existing automations/ directory needs no new subdirectory README.

• Do not create boilerplate README pages manually. Creating automations/<slug>.md also creates automations/README.md when needed.
• Creating dated output also creates automations/<slug>/README.md when needed.
• When a create or move reports a created or restored README, read it immediately. If it contains bootstrap text, edit it in the same run with the directory's exact purpose, producer and owned paths, consumers and read order, source/privacy/trust boundaries, and retention policy.
• On later runs, read the directory README before its pages and follow that contract.
• Any automation that consumes another automation's output must explicitly list/read its named input directory, or read its named file, in its prompt. Raw output is readable/searchable by any agent, but it never becomes a fact or canonical daily page without an explicit curation step.

Build a standalone prompt

The automation prompt must:

1. Gather its source data.
2. Call wiki_list_pages with the root directory (directory=""). A not_found

result while probing a missing target directory is normal before its first page.

1. For each existing named directory, read its README and then list/read the exact input pages before using its output.
2. Read the exact target path when it exists.
3. Call wiki_create_page when absent with its exact managed path, title,

aliases, and body; otherwise call wiki_edit_page with the exact path and body. The backend handles write safety.

1. Read every created or restored README reported by that write and specialize any bootstrap contract before completing the run.
2. Retry a write conflict only after another fresh list/read.
3. Archive old pages only when the user requested a retention policy; never hard-delete.
4. Report the page path and whether it created, changed, or archived anything.

Do not use filesystem tools for managed wiki pages.

Create the automation

Call automation_create once. Set:

• auto_approve=true;
• all_tools=true — the prompt must write wiki pages, and runs are read-only without it;
• the requested trigger and source tools;
• idempotency_key to wiki-automation:<stable-slug> and idempotency_scope="global". The slug names this producer's enduring job, not one run or page version;
• no result-channel argument—the dedicated channel is automatic.

The normal read-only wiki tools are available in every run; all_tools only unlocks writes.

If creation has an ambiguous result, first call automation_list. Retry with the exact same key only when that producer is absent. Never make a -2, date, or versioned key to work around a retry. When the producer already exists, call automation_update instead of creating another one. An explicit deletion releases the global key, so recreating the same producer uses the same key.