Add user wiki feature

[sysread]

=head1 SYNOPSIS

A new flat encyclopedia about the user, peer to memory and journal.
Articles are titled, single-level (no nesting), and never auto-injected
into the chat - the assistant reaches them only through the always-on
wiki_search tool.

=head1 PURPOSE

Currently the app maintains three forms of long-running user knowledge:

• Memories - atomic labelled facts surfaced inline by opening recall.
• Journal - dated reflective prose; today's automatic entry is primed
  in the system prompt on the first turn.
• Chats - the raw conversations themselves.

There is no surface for topical, encyclopedic knowledge that spans
many conversations - "what is the recipe project", "who is Maya", "where
does the Lisbon trip stand". This is the gap the wiki fills.

=head1 DESCRIPTION

=head2 Layer 1: how the existing subsystems behave

The journal worker acquires a cross-tab Web Lock
(nak:journal-worker), claims a thread via
claim_next_thread_for_journal (gated on a same-day cooldown using
threads.updated_at), runs JournalAgent.run (single Venice call,
response_format=json_object, no tools), and lands the entry plus the
pointer-advance through the atomic upsert_journal_entry_and_mark_thread
RPC. The reflection worker uses the same scaffold but runs
runHeadlessToolLoop against memoryToolbox; tool calls are the side
effects, the model's final text is discarded. The generic embedding
worker round-robins polymorphic source adapters (memories, threads,
samskara substrate, journal entries), each adapter shaping a
(claim, save) pair against per-table RPCs. The chat-side toolbox
has every read tool registered always-on so the model never has to flip
a toolbox to search; writes still gate.

=head2 Layer 2: what this PR adds, parallel to layer 1

The wiki worker acquires nak:wiki-worker, claims via
claim_next_thread_for_wiki, runs WikiAgent.run (which uses
runHeadlessToolLoop against the new wikiToolbox), and lands the
pointer-advance through mark_thread_wiki_processed_if_claimed -
unconditionally on done. Two deliberate divergences from the
journal claim RPC:

• Newest-message timestamp, not threads.updated_at. The
  eligibility lateral reads messages.created_at directly via a
  second lateral so the gate stays stable against future bumps to
  threads.updated_at from unrelated writes.
• Strict-yesterday gate:
  (newest.created_at at time zone tz)::date < (now() at time zone tz)::date.
  Effect: chat Monday -> eligible Tuesday; resume Wednesday -> the
  newest msg lands on Wednesday and the inequality fails again until
  Thursday.

The wiki embedding adapter is a clone of the memories adapter (input
shape ${title}\n\n${content}, 16k char cap, same claim/save pair). The
generic worker picks it up alongside the existing four sources.

wiki_search is registered in the always-on toolbox so every chat
turn can call it without a toolbox toggle. wiki_create/wiki_update/
wiki_delete live only in the agent-only wikiToolbox -
the main LLM has no path to write to the wiki.

The per-article manual flow (WikiAgent.updateOne) runs synchronously
on the main thread: one Venice call, response_format=json_object, no
tool loop. UI mirrors Journal.svelte's regenerate flow with
Accept / Try again / Cancel preview. The prompt explicitly forbids
discarding existing facts unless the user asks for them removed, and
forbids fabrication.

unique(user_id, title) lets wiki_create raise on collision; the tool
rephrases the unique-violation as actionable text the autonomous agent
reads and falls through to wiki_search + wiki_update.

A new WIKI_BLOCK in src/lib/chat-prompt.ts tells the main LLM the
wiki exists and to call wiki_search whenever the user references
something topical or factual about themselves, a project, a person, or
a place. Articles are explicitly never auto-injected.

A new wikiAutomaticEnabled setting (defaulting on, mirroring
journalAutomaticEnabled) gates the autonomous worker. Manual edits and
the per-article "ask agent to update" button keep working when off. The
wiki shares the journal's tz preference - one user-tz setting covers
both background subsystems.

=head2 Layer 3: how this resolves PURPOSE

The user (and a background agent) now have a fourth long-running surface
sized for "what is X" rather than atomic facts or dated reflections. The
assistant can ground "what's the recipe project about" answers without
guessing. Articles accrete coherently because the autonomous agent
search-then-mutates and is told to preserve facts. The user can still
correct, rewrite, or delete by hand, or hand the agent explicit
instructions for a single article.

Notes for AI reviewers:

• The wiki agent is intentionally allowed wiki_delete (unlike
  memoryToolbox which uses memory_invalidate). The prompt
  constrains it to consolidation only - "an article you just updated
  now strictly subsumes another one". Never delete on the basis of
  "the user said something different today" alone.
• Pointer-advance is unconditional on done. Even a no-op cycle
  advances the pointer; otherwise every cycle would re-process the
  same "no topic warranted writing" conversation.
• embedding_claim_expires (no _at) on wiki_articles is intentional -
  matches the existing memories / journal_entries convention. The
  thread-side wiki_claim_expires_at DOES have the suffix, matching
  the journal claim columns.
• The autonomous prompt's "preserve every existing fact unless the
  conversation explicitly contradicts it" line is load-bearing. A
  reviewer suggesting "let the agent rewrite for clarity" would
  silently churn the wiki.

https://claude.ai/code/session_015XcR7xzLdij66ZbYERUdLH

---

Generated by Claude Code